University of Southern California Dissertations and Theses

Robust and generalizable knowledge acquisition from text

(USC Thesis Other)

Robust and generalizable knowledge acquisition from text

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Robust and Generalizable Knowledge Acquisition from Text
by
Wenxuan Zhou
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(COMPUTER SCIENCE)
August 2023
Copyright 2023 Wenxuan Zhou
To my family and friends,
who support me during my Ph.D. journey.
ii
Acknowledgements
I express my heartfelt gratitude to my Ph.D. advisor, Prof. Muhao Chen, for his guidance, support,
and mentorship throughout my Ph.D. journey. His expertise in NLP and knowledge bases have been
invaluable in shaping my research agenda and in guiding my technical development. In addition,
I would like to extend my sincere appreciation to Prof. Yangqiu Song and Prof. Xiang Ren for
their supervision, valuable insights, and encouragement during my early research years. Their
mentorship and guidance were invaluable in helping me establish a strong foundation in the field.
I would like to thank my defense committee members, Prof. Laurent Itti, Prof. Jonathan May,
Prof. Tianshu Sun, and Prof. Robin Jia, for their critical feedback and constructive suggestions.
I would also like to thank Prof. Fred Morstatter for serving as a member of my qualification
committee.
I would like to extend my appreciation to all of my friends at the USC LUKA lab and USC
INK lab, including Bill Yuchen Lin, James Huang, Tenghao Huang, Woojeong Jin, Bangzheng
Li, Keming Lu, Ehsan Qasemi, Qin Liu, Fei Wang, and Nan Xu, for their support, friendship, and
fruitful collaborations. I am also grateful for the contributions of my collaborators from other labs
and universities, including Tianqing Fang, I-Hung Hsu, Fangyu Liu, Mingyu Derek Ma, Yiwei
Wang, Hongming Zhang, and Huan Zhang.
I am indebted to my mentors during internships, Dr. Heba Elfardy, Dr. Jing Huang, Dr. Hang
Li, Prof. Tengyu Ma, Dr. Tristan Naumann, Dr. Qiang Ning, Dr. Hoifung Poon, Dr. Kevin Small,
and Dr. Sheng Zhang, for their valuable guidance, support, and the opportunity to work on exciting
research problems.
Finally, I want to express my deep appreciation to my parents for their unwavering support,
encouragement, and understanding throughout my academic journey.
iii
Table of Contents
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Chapter 2: Document-level Relation Extraction . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Enhanced BERT Baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Encoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Binary Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Adaptive Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Localized Context Pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6.2 Experiment Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.4 Results on Biomedical Datasets . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.5 Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.6 Analysis of Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.7 Analysis of Context Pooling . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.7 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Chapter 3: Noisy Label Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
iv
3.2.1 Learning Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.2 Co-regularization Objective . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.3 Joint Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.2 Base Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.3 Model Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.4 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.5 Noise Filtering Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4.6 Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Chapter 4: Out-of-distribution Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.2 Framework Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3.3 Contrastive Representation Learning . . . . . . . . . . . . . . . . . . . . . 38
4.3.4 OOD Scoring Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4.2 Experimental Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4.4 Novel Class Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4.5 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Chapter 5: Resolving Knowledge Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3.1 Opinion-based Prompting . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3.2 Counterfactual Demonstration . . . . . . . . . . . . . . . . . . . . . . . . 51
5.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.4.2 Knowledge Conflict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.4.3 Prediction with Abstention . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.4.4 Additional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.4.5 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Chapter 6: Dataset Annotation by Neural Rule Grounding . . . . . . . . . . . . . . . . . . 60
v
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.3 Neural Rule Grounding (NERO) . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3.1 Framework Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3.2 Labeling Rule Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3.3 Relation Classifier ( RC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3.4 Soft Rule Matcher (SRM) . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.3.5 Joint Module Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.4 Model Learning and Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4.1 Parameter Learning of NERO . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4.2 Model Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.5.1 Data Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.5.2 Compared Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.5.3 Experiment Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.5.4 Performance Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.5.5 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5.6 Model Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.5.7 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Chapter 7: Continual Contrastive Finetuning . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.3.1 Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.3.2 Pretraining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7.3.3 Finetuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.3.4 Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.4.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.4.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.4.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.4.4 Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.4.5 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Chapter 8: Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
8.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
vi
List of Tables
2.1 Statistics of the datasets in experiments. . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Hyper-parameters in training. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Main results (%) on the development and test set of DocRED. We report the mean
and standard deviation of F
1
on the development set by conducting 5 runs of
training using different random seeds. We report the official test score of the best
checkpoint on the development set. . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Test F
1
score (%) on CDR and GDA dataset. Our ATLOP model with the SciBERT
encoder outperforms the current SOTA results. . . . . . . . . . . . . . . . . . . . 15
2.5 Ablation study of ATLOP on DocRED. We turn off different components of the
model one at a time. These ablation results show that both adaptive thresholding
and localized context pooling are effective. Logsumexp pooling and group bilinear
both bring noticeable gain to the baseline. . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Result of different thresholding strategies on DocRED. Our adaptive thresholding
consistently outperforms other strategies on the test set. . . . . . . . . . . . . . . . 17
3.1 Data statistics of TACRED and CoNLL03. . . . . . . . . . . . . . . . . . . . . . . 27
3.2 F
1
score (%) on the dev and test set of TACRED.♣ marks results obtained from
the originally released implementation. We report the median of F
1
on 5 runs of
training using different random seeds. For fair comparison, the CR results are
reported based on the predictions from model f
1
in our framework. . . . . . . . . 28
3.3 F
1
score (%) on the dev and test set of CoNLL03.♣ marks results obtained using
the originally released code. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 F
1
score (%) of using different number of models on the relabeled test set of
TACRED. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5 F
1
score (%) of alternative noise filtering strategies on the test set of TACRED. The
best results are achieved whenδ = 2% for both methods. . . . . . . . . . . . . . . 31
3.6 F
1
score (%) of different functions for q q q on the relabeled test set of TACRED. . . . 31
3.7 F
1
score (%) under different noise rates on the relabeled set of TACRED. . . . . . . 32
4.1 Statistics of the datasets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2 OOD detection performance (in %) of RoBERTa
LARGE
trained on the four ID
datasets. Due to space limits, for each of the four training ID dataset, we report
the macro average of AUROC and FAR95 on all OOD datasets (check Appendix
for full results). Results where the contrastive loss improves OOD detection on
both evaluation metrics are highlighted in green .“w/oL
cont
+MSP” thereof is the
method in Hendrycks, Liu, Wallace, Dziedzic, Krishnan & Song [1]. . . . . . . . 42
4.3 Novel class detection performance. . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Accuracy of the trained classifier. . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Average OOD detection performance of different distance metrics. . . . . . . . . . 45
vii
4.6 Average OOD detection performance of other pretrained Transformers. . . . . . . . 45
5.1 Results (in %) in the knowledge conflict setting. The overall best results are
highlighted in bold. The best and the second best results in each setting are
highlighted in green and orange , respectively. . . . . . . . . . . . . . . . . . . . 52
5.2 Results (in %) on RealTime QA. The overall best results are highlighted in bold.
The best and the second best results in each setting are highlighted in green and
orange , respectively. As all prompts achieve perfect accuracy (100%) on the
HasAns subset, it is not included in the table. . . . . . . . . . . . . . . . . . . . . 54
5.3 Results (in %) on the filtered evaluation set of natural questions with original
(factual) contexts and answers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4 Examples of prompts and LLMs’ corresponding predictions. In the “Prompt”
row, we show and highlight the added parts from different prompting templates
including attributed prompts, instruction-based prompts, and opinion-based prompts. 58
6.1 Statistics for TACRED and SemEval datasets. . . . . . . . . . . . . . . . . . . . . 72
6.2 Performance comparison (in %) of relation extraction on the TACRED and
SemEval datasets. We report the mean and standard deviation of the evaluation
metrics by conducting 5 runs of training and testing using different random seeds.
We use LSTM+ATT (S
matched
+P) as the base model for all semi-supervised
baselines and our models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.3 Performance on predicting unseen relations. NERO applies the learned soft rule
matcher on unseen relation rules to make predictions. . . . . . . . . . . . . . . . . 77
6.4 Ablation Study of Different Training Objectives on TACRED dataset. We remove
each loss term one at a time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.5 Ablation study of different soft-matching models for NERO on the TACRED dataset. 78
7.1 Probing results (in F
1
) on the test set of BioRED and Re-DocRED. . . . . . . . . . 87
7.2 Results on the test set of Re-DocRED. . . . . . . . . . . . . . . . . . . . . . . . . 90
7.3 F
1
on the test set of BioRED. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.4 F
1
on the test set of BioRED with different pretraining objectives. We use MCCL
in finetuning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
viii
List of Figures
2.1 An example of multi-entity and multi-label problems from the DocRED dataset.
Subject entity John Stanistreet (in orange) and object entity Bendigo (in green)
express relations place of birth and place of death. The related entity mentions are
connected by lines. Other entities in the document are highlighted in grey. . . . . . 8
2.2 An artificial illustration of our proposed adaptive-thresholding loss. A TH class
is introduced to separate positive classes and negative classes: positive classes
would have higher probabilities than TH, and negative classes would have lower
probabilities than TH. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Illustration of localized context pooling. Tokens are weighted averaged to form the
localized context c
(s,o)
of the entity pair(e
s
,e
o
). The weights of tokens are derived
by multiplying the attention weights of the subject entity e
s
and the object entity e
o
from the last transformer layer so that only the tokens that are important to both
entities (highlighted in light yellow) receive higher weights. . . . . . . . . . . . . 13
2.4 Dev F
1
score of documents with the different number of entities on DocRED. Our
localized context pooling achieves better results when the number of entities is
larger than 5. The improvement becomes more significant when the number of
entities increases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Context weights of an example from DocRED. We visualize the weight of context
tokens a
(s,o)
in localized context pooling. The model attends to the most relevant
context born and died for entity pair (John Stanistreet, Bendigo). . . . . . . . . . 19
3.1 Illustration of our co-regularization framework. The base models are jointly
optimized with the task-specific loss from label y and an agreement loss, which
regularizes the models to generate similar predictions to the aggregated soft target
probability q q q. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 F
1
score (%) on the clean set of TACRED. Classifiers trained with our framework
are more noise-robust compared to baselines (γ = 0). . . . . . . . . . . . . . . . . 30
4.1 Illustration of our proposed contrastive loss. The contrastive loss seeks to increase
the discrepancy of the representations for instances from different training classes,
such that OOD instances from unknown classes can be better differentiated. . . . . 38
4.2 Visualization of the representations for positive, negative instances in SST2 and
OOD ones. The discrepancy between ID and OOD representations is greater on
representations obtained withL
margin
. . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1 Examples of knowledge conflict and prediction with abstention. LLMs may ignore
the provided context and make unfaithful predictions based on their parametric
knowledge before Q4 2021. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Memorization ratios across different sizes of InstructGPTs, evaluated in the
zero-shot setting using natural questions. . . . . . . . . . . . . . . . . . . . . . . . 56
ix
5.3 Brier scores across different sizes of InstructGPTs, evaluated in the zero-shot
setting of RealTime QA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.1 Current rule-based methods mostly rely on exact/hard matching to raw corpus
and suffer from limited coverage. For example, the rule body of p
3
only matches
sentence s
1
but is also similar to s
2
and s
3
, which express the same relation as p
3
. A
“soft rule matching” mechanism is desirable to make better use of the corpus for
label generation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.2 Comparison between previous work and the proposed NERO framework. (A)
Bootstrapping. (B) Data Programming. (C) Self-Training. (D) NERO. The neural
rule grounding process enables the soft-matching between rules and unmatched
sentences using the soft rule matcher. . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.3 Overview of the NERO framework. Each unmatched sentence is first annotated by
the soft rule matcher (SRM) to generate pseudo labels, and then fed into the relation
classifier ( RC) to update the model parameters. The whole framework is trained
iteratively and jointly, based on multiple loss functions as introduced in Sec. 6.3.5. 65
6.4 Detailed architecture of the soft rule matcher (SRM). The cosine similarity between
two embeddings indicates the degree of matching between rules and sentences. . . 68
6.5 The contrastive loss in learning the SRM, which increases the matching score of
rules with the same relation type and decrease the matching score otherwise. . . . 69
6.6 Performance w.r.t. different number of rules and human-annotated labels on
TACRED. We show different models’ F1 scores and number of rules or labels used
for training the corresponding model. . . . . . . . . . . . . . . . . . . . . . . . . 75
6.7 Performance of different semi-supervised models trained using various amounts of
unlabeled sentences randomly sampled from the raw corpus on TACRED. . . . . . 77
6.8 Sensitivity analysis ofτ,α,β , andγ on TACRED. We report the mean and standard
deviation F1 score by conducting 5 runs of experiments using different random seeds. 79
6.9 Output visualization of SRM. Left: attention weights of words and the soft matching
scores between a rule and three sentences. Right: cosine similarity matrix between
word embeddings learned with the contrastive loss. . . . . . . . . . . . . . . . . . 80
6.10 Study on label efficiency. Average number of rules / sentences labeled by annotators
(dashed line) are shown on the x-axis over the left-hand side; and the performance
of models trained with these corresponding labeled rules / sentences (solid line) are
shown on the x-axis over the right-hand side. We use NERO and LSTM+ATT as the
base model for the labeling rules and sentences, respectively. . . . . . . . . . . . . 80
7.1 F
1
using different temperatures on 1% of BioRED. . . . . . . . . . . . . . . . . . 92
7.2 F
1
under different percentages of BioRED training data. . . . . . . . . . . . . . . . 92
7.3 Visualization of relation embedding finetuned with different objectives on BioRED.
NA instances are shown in grey. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
x
Abstract
Knowledge acquisition involves identifying and understanding concepts that are expressed in text,
and then inferring and consolidating their relationships. With the ever-increasing volumes of digital
text generated every day, the need for automated knowledge acquisition is becoming more crucial.
However, real-world knowledge acquisition systems face more complex scenarios where knowledge
is often expressed implicitly or within long contexts, and require models that are robust and able
to handle these complexities. Additionally, data annotation for knowledge acquisition can be
expensive, particularly in high-stakes domains, necessitating the development of efficient methods
that minimize human involvement. In this dissertation, we propose solutions to build robust and
generalizable knowledge acquisition systems, addressing challenges in: (i) extracting knowledge
from long contexts efficiently and effectively; (ii) improving system robustness in various scenarios
including noisy annotations, out-of-distribution instances, and knowledge conflicts; and (iii) data-
efficient knowledge acquisition, utilizing cheap resources such as labeling rules or unlabeled
corpora.
In the first part of this thesis, we study the extraction of knowledge from long contexts such as
documents. To handle the long-context problem, we introduce a novel pretrained-attention-guided
neural structure that finds context that both entities attend to in an entity pair. We show that our
model achieves state-of-the-art performance in both general and biomedical domains.
In the second part of this thesis, our focus is on enhancing the robustness of knowledge
acquisition systems. To start with, we investigate the mitigation of noisy labels during training.
Since the data used for knowledge acquisition is often sourced from crowdsourcing or distant
supervision, it is common for the data to be imperfect. We explore a co-regularization based
framework to address this issue effectively. Moreover, many knowledge acquisition systems are
typically developed for specialized domains or relations, making them susceptible to erroneous
predictions on instances from other domains or relations. To overcome this challenge, we propose
an out-of-distribution detection framework to identify such instances during inference and either
alert users or redirect them to other dedicated models for improved accuracy. Another challenge is
that the world is continually evolving, and the knowledge acquired by these systems may become
outdated. We explore techniques for updating the model’s predicted knowledge using context
retrieval and opinion-based prompts.
Finally, we study how to build knowledge acquisition systems with minimal human effort.
Our approach involves weak data creation, where we can generate a large amount of labeled data
by utilizing labeling a few rules developed by humans. Then, we investigate enhancing the data
efficiency of knowledge acquisition systems by self-supervised learning. To accomplish this, we
propose a framework that pretrains a model using a large unlabeled corpus, which can then be
fine-tuned with a small number of labeled instances. Additionally, we propose a novel continual
xi
contrastive finetuning framework that aims to minimize the discrepancy between the pretraining
and fine-tuning stages.
xii
Chapter 1
Introduction
1.1 Motivations
Knowledge acquisition involves recognizing concepts described in the text, inferring their rela-
tionships, and consolidating them. With large amounts of unstructured text generated every day,
automated knowledge acquisition is a critical tool for efficiently managing and processing this
information. The extracted knowledge plays a crucial role in various NLP tasks such as question
answering [2, 3], fact verification [4, 5], and dialogue [6, 7]. Moreover, it provides access to expert
knowledge for domain-specific research [8–10] and clinical applications [11, 12].
Pretrained language models (PLMs; [13]) have been successful in automatic knowledge acquisi-
tion in recent years. Previous works [14–16] mainly focus on knowledge expressed within a single
sentence. However, in real-world applications, large amounts of knowledge are expressed through
multiple sentences. For example, the knowledge triple (Bill Gates, founder, Microsoft) can only be
inferred from two sentences in the following Wikipedia example:
William Henry Gates III (born October 28, 1955) is an American business magnate,
philanthropist, and investor. He is best known for co-founding software giant Microsoft,
along with his late childhood friend Paul Allen.
According to a quantitative study [17], at least 40.7% of knowledge triples in Wikipedia can only be
inferred from multiple sentences. Therefore, extending automatic knowledge acquisition beyond
individual sentences is a crucial research problem, in order to achieve a higher recall of knowledge
contained in text.
Another research problem is the discrepancy between the datasets used in academic research and
real-world applications. Typically, in academic datasets [18–20], the data for training and testing are
usually sampled from similar distributions, from similar time scopes and corpora. However, in real-
world applications, there is less control over the distribution of data during inference. This can cause
knowledge conflicts between training and testing data, particularly in scenarios where knowledge
acquisition is used to extract knowledge from new publications, where finding novel knowledge
is more important than extracting known ones. Moreover, testing data may come from different
corpora or semantic classes from those used for training in real-world applications, making the
trained knowledge acquisition systems less applicable and requiring higher robustness to distribution
shifts. These discrepancies highlight the need for knowledge acquisition systems to be more robust
in real-world applications. However, most existing studies on knowledge acquisition do not take
into account these discrepancies, raising concerns about their effectiveness in real-world settings. In
1
addition, since the data in real-world scenarios are often obtained by crowdsourcing, they may have
noisy labels that degrade the system’s performance. Therefore, enhancing the system’s robustness
to noisy training labels is also a crucial problem.
Furthermore, building data-efficient knowledge acquisition systems is a challenge, especially
for high-stake domains such as biomedicine and finance. Annotators for these domain-specific
tasks require domain expertise, leading to the high cost of manual data annotation. As a result,
most existing efforts focus on distant supervision [21, 22], which consists of linking entities to an
existing knowledge base and using the relation in the knowledge base to annotate the text. However,
distant supervision is not applicable when relations of our interest do not exist in the knowledge
bases. Besides, it can suffer from the knowledge conflict problem mentioned earlier, since the
knowledge base may lack current knowledge. Another line of work focuses on few-shot knowledge
acquisition [23, 24]. This process consists of training a knowledge acquisition model on (abundant)
data from other relations first, and then adapting it to relations of our interest using a few labels.
However, such methods are not suitable for high-stakes domains where data for other relations is
also scarce. Therefore, constructing data-efficient knowledge acquisition systems remains a crucial
research problem.
In this dissertation, we aim to build robust and generalizable knowledge acquisition systems
that are: (i) able to extract knowledge from context beyond individual sentences; (ii) robust to the
shifts between training and testing data, and noise in the training data; and (iii) requiring minimal
human efforts.
1.2 Challenges
In this dissertation, we address several key challenges. First, to extract knowledge from long
contexts, the knowledge acquisition system must be able to identify salient contextual information
between pairs of entities. In addition, since a pair of entities may have multiple relations, multi-label
prediction also represents a challenge. Second, since current datasets for knowledge acquisition
do not address the issue of robustness, we need to develop experimental settings to measure the
robustness and dedicated frameworks to improve it. Third, to enhance data efficiency, we need to
address the problem from two angles: (1) obtaining annotated data with minimal human effort and
(2) utilizing labels more effectively through dedicated frameworks. The details of the challenges we
address in this dissertation are listed below:
• Knowledge acquisition from long contexts. To efficiently and effectively extract knowledge
from long contexts, the system needs to address two challenges:
(i) Identifying salient context. Current datasets for long-context knowledge acquisition can
contain up to thousands of words [25], whereas a single knowledge triple is usually expressed
within only a few sentences. Therefore, it is important for systems to identify relevant context
in order to avoid being overwhelmed by large amounts of irrelevant information. However, it
is usually hard to annotate salient context by humans, and it can be challenging to learn how
to locate relevant context from scratch. To address this issue, systems may need to leverage
additional syntactic and semantic information in order to accurately identify relevant context.
2
(ii) Multi-label prediction. Since an entity pair may appear multiple times in context, it often
has more than one relation types, making knowledge acquisition from long context a multi-
label prediction problem. For example, in the DocRED dataset [17], over 15% of entity pairs
have more than one relation type. Although the common practice is to convert this problem
into several binary classification problems, it involves adjusting various thresholds for each
relation type on development data, which may not generalize to the test data. Additionally,
this approach encounters a calibration issue as different entity pairs may have different optimal
thresholds.
• Robustness of knowledge acquisition systems. As mentioned earlier, robustness in this
context comprises several dimensions, such as distributional shifts, knowledge conflicts, and
noisy training labels. However, these aspects have been less explored in previous research,
which calls for setting up appropriate experimental settings and metrics to evaluate these
aspects. Another challenge in improving robustness lies in the fact that it is impossible to
know the distribution of testing data before deploying a system. As a result, it is essential to
develop strategies that can improve robustness in an unsupervised way.
• Data-efficient knowledge acquisition. In order to enhance data efficiency, there are two
primary challenges that must be addressed. First, data annotation is a necessary step, unless
distant supervision can be utilized. While for high-stake domains, data annotation is both
time-consuming and highly expensive. As a result, a major challenge is how to decrease the
cost of data annotation. Second, since unlabeled data is often abundant, another challenge is
how to effectively leverage this data to improve the system.
1.3 Thesis Contributions
• We propose a framework for document-level relation extraction that addresses both the salient
context and the multi-label prediction challenge. Specifically, we introduce localized context
pooling, which leverages attention in PLMs to identify relevant context for entity pairs.
Additionally, we propose the use of an adaptive-thresholding loss, which allows the model to
learn an adaptive threshold is adaptive to entity pairs.
• We present a co-regularization framework that can efficiently train supervised models using
noisy datasets, which consists of multiple models with different initialization and a regulariza-
tion loss that enforces the outputs of different models to be similar. We also discuss various
design strategies and balance efficiency with effectiveness.
• We propose to identify and reject instances that are from different distributions to training,
making the model abstain from making unconfident predictions. To achieve this, we propose a
framework for unsupervised out-of-distribution detection using contrastive learning to expand
the boundaries between different classes. We also explore various combinations of contrastive
learning losses and out-of-distribution scoring functions to achieve the best performance in
this challenging task.
• We study the knowledge conflict problem in knowledge acquisition, where the system needs
to extract the knowledge faithfully to the context and avoid memorizing knowledge in the
3
pretraining corpus. we focus on large language models and present two techniques, namely
opinion-based prompts and counterfactual demonstrations.
• We propose to reduce the annotation cost using labeling rules. To solve the low coverage prob-
lem of rules, we propose a framework that learns to generalize labeling rules to semantically
similar instances by neural rule grounding.
• To leverage unlabeled data in knowledge acquisition, we propose to first pretrain the PLMs
based on matching-the-blanks objectives, and then finetune PLMs on task-specific data. To
minimize the gap in training objectives between pretraining and finetuning, we propose to
use a consistent contrastive loss in pretraining and finetuning, and use kNN-based inference.
Furthermore, we design an MCCL finetuning objective, allowing one relation to form mul-
tiple different clusters, thus further reducing the distributional gap between pretraining and
finetuning.
1.4 Thesis Overview
Part I: Knowledge Acquisition in Documents
• Chapter 2:
This chapter introduces the ATLOP (Adaptive Thresholding and Localized cOntext Pooling)
framework for document-level relation extraction. We propose two techniques, namely
adaptive thresholding and localized context pooling, to address the challenges of multi-
label classification and identifying salient context. Adaptive thresholding overcomes the
limitations of a global threshold in prior work by introducing a learnable, entity-dependent
threshold. Localized context pooling directly transfers the attention from PLMs to locate
context that is salient for relation classification. Our experiments on three benchmark datasets
for document-level RE, including DocRED in the general domain, as well as CDR and GDA
in the biomedical domain, demonstrate the effectiveness of our proposed framework.
Part II: Robust Knowledge Acquisition
• Chapter 3:
In this chapter, we introduce a co-regularization framework that addresses the challenge of
noisy training labels. Our approach is motivated by studies showing that noisy labels are
more difficult to memorize during training and are more prone to being forgotten than clean
labels, making them identifiable during training. To mitigate this issue, we propose a simple
co-regularization framework for information extraction. This framework consisits of several
neural models with identical structures but different parameter initialization. These models are
jointly optimized with task-specific losses and are regularized to produce similar predictions
based on an agreement loss, which prevents overfitting on noisy labels. Our extensive
experiments on two information extraction benchmarks demonstrate the effectiveness of our
proposed framework.
4
• Chapter 4:
In this chapter, we present our framework for detecting out-of-distribution (OOD) instances,
which can lead to significant semantic shift problems during inference. In practice, it is crucial
for a reliable model to identify such instances and either reject them during inference or pass
them on to models that can handle them. To address this issue, we propose an unsupervised
OOD detection method that uses only in-distribution (ID) data during training. Our approach
involves fine-tuning PLMs with a contrastive loss, which improves the compactness of
representations, making it easier to distinguish OOD instances from ID ones. We use the
Mahalanobis distance in the model’s penultimate layer to accurately detect OOD instances.
Our comprehensive experiments demonstrate near-perfect OOD detection performance.
• Chapter 5:
This chapter focuses on techniques for mitigating knowledge conflicts in large language
models (LLMs). LLMs have demonstrated exceptional performance in knowledge-driven
NLP tasks due to their ability to encode parametric knowledge about world facts. However,
this reliance on parametric knowledge may cause them to miss contextual cues, leading to
inaccurate predictions in context-sensitive NLP tasks. To address this issue, we propose
techniques to assess and enhance LLMs’ contextual faithfulness in two aspects: knowledge
conflict and prediction with abstention. We show that LLMs’ contextual faithfulness can be
significantly improved using carefully designed prompting strategies, namely opinion-based
prompts and counterfactual demonstrations. Opinion-based prompts reframe the context
as a narrator’s statement and inquire about the narrator’s opinions, while counterfactual
demonstrations use instances containing false facts to improve faithfulness in knowledge-
conflict situations.
Part III: Data-efficient Knowledge Acquisition
• Chapter 6:
In this chapter, we introduce the NERO framework, which enables efficient data annotation by
creating labeling rules instead of seeking more instance-level labels from human annotators.
These rules can be automatically mined from large text corpora and generalized using a soft
rule-matching mechanism. Our framework consists of a relation extraction (RE) module and
a soft matching module. The RE module can be instantiated with any neural model, while the
soft matching module learns to match rules with semantically similar sentences. This process
allows raw corpora to be automatically labeled. We conduct extensive experiments and
analysis on two widely-used public datasets, demonstrating the effectiveness of the proposed
NERO framework compared to both rule-based and semi-supervised methods. Through user
studies, we find NERO yields a 9.5x speedup of annotation compared to human annotation.
• Chapter 7:
In this chapter, we introduce continual contrastive finetuning for data-efficient knowledge
acquisition. The approach involves self-supervised learning, where the relation embedding is
pretrained using a RE-based objective and then finetuned on labeled data using a classification-
based objective. To bridge the gap between pretraining and finetuning, we propose using
consistent objectives of contrastive learning for both stages of training. We also propose
5
a multi-center contrastive loss that allows one relation to form multiple clusters in the
representation space, which better aligns with pretraining. Our experiments on two document-
level RE datasets, demonstrate the effectiveness of our approach.
Part IV: Conclusion
• Chapter 8:
This chapter summarizes the contributions of the thesis and discusses several future directions.
6
Chapter 2
Document-level Relation Extraction
Document-level relation extraction (RE) poses new challenges compared to its sentence-level
counterpart. One document commonly contains multiple entity pairs, and one entity pair occurs
multiple times in the document associated with multiple possible relations. We propose two
novel techniques, adaptive thresholding and localized context pooling, to solve the multi-label
and multi-entity problems. The adaptive thresholding replaces the global threshold for multi-label
classification in the prior work with a learnable entities-dependent threshold. The localized context
pooling directly transfers attention from pre-trained language models to locate relevant context that
is useful to decide the relation. We experiment on three document-level RE benchmark datasets:
DocRED, a recently released large-scale RE dataset, and two datasets CDR and GDA in the
biomedical domain. Our ATLOP (Adaptive Thresholding and Localized cOntext Pooling) model
achieves an F1 score of 63.4, and also significantly outperforms existing models on both CDR and
GDA.
2.1 Introduction
Relation extraction (RE) aims to identify the relationship between two entities in a given text and
plays an important role in information extraction. Existing work mainly focuses on sentence-level
relation extraction, i.e., predicting the relationship between entities in a single sentence [26–28].
However, large amounts of relationships, such as relational facts from Wikipedia articles and
biomedical literature, are expressed by multiple sentences in real-world applications [17, 29]. This
problem, commonly referred to as document-level relation extraction, necessitates models that can
capture complex interactions among entities in the whole document.
Compared to sentence-level RE, document-level RE poses unique challenges. For sentence-level
RE datasets such as TACRED [19] and SemEval 2010 Task 8 [18], a sentence only contains one
entity pair to classify. On the other hand, for document-level RE, one document contains multiple
entity pairs, and we need to classify the relations of them all at once. It requires the RE model
to identify and focus on the part of the document with relevant context for a particular entity
pair. In addition, one entity pair can occur many times in the document associated with distinct
relations for document-level RE, in contrast to one relation per entity pair for sentence-level RE.
This multi-entity (multiple entity pairs to classify in a document) and multi-label (multiple relation
types for a particular entity pair) properties of document-level relation extraction make it harder
than its sentence-level counterpart. Figure 2.1 shows an example from the DocRED dataset [17].
7
John Stanistreet was an Australian politician. He was
born in Bendigo to legal manager John Jepson
Stanistreet and Maud McIlroy. (…4 sentences…) In 1955
John Stanistreet was elected to the Victorian Legislative
Assembly as the Liberal and Country Party member for
Bendigo. Stanistreet died in Bendigo in 1971.
Subject: John Stanistreet Object: Bendigo
Relation: place of birth; place of death
Figure 2.1: An example of multi-entity and multi-label problems from the DocRED dataset. Subject
entity John Stanistreet (in orange) and object entity Bendigo (in green) express relations place of
birth and place of death. The related entity mentions are connected by lines. Other entities in the
document are highlighted in grey.
The task is to classify the relation types of pairs of entities (highlighted in color). For a particular
entity pair (John Stanistreet, Bendigo), it expresses two relations place of birth and place of death
by the first two sentences and the last sentence. Other sentences contain irrelevant information to
this entity pair.
To tackle the multi-entity problem, most current approaches construct a document graph with
dependency structures, heuristics, or structured attention [30–33], and then perform inference
with graph neural models [34, 35]. The constructed graphs bridge entities that spread far apart
in the document and thus alleviate the deficiency of RNN-based encoders [36, 37] in capturing
long-distance information [38]. However, as transformer-based models [39] can implicitly model
long-distance dependencies [40, 41], it is unclear whether graph structures still help on top of
pre-trained language models such as BERT [13]. There have also been approaches to directly apply
pre-trained language models without introducing graph structures [42, 43]. They simply average
the embedding of entity tokens to obtain the entity embeddings and feed them into the classifier to
get relation labels. However, each entity has the same representation in different entity pairs, which
can bring noise from irrelevant context.
Instead of introducing graph structures, we propose a localized context pooling technique. This
technique solves the problem of using the same entity embedding for all entity pairs. It enhances
the entity embedding with additional context that is relevant to the current entity pair. Instead of
training a new context attention layer from scratch, we directly transfer the attention heads from
pre-trained language models to get entity-level attention. Then, for two entities in a pair, we merge
their attentions by multiplication to find the context that is important to both of them.
For the multi-label problem, existing approaches reduce it to a binary classification problem.
After training, a global threshold is applied to the class probabilities to get relation labels. This
8
method involves heuristic threshold tuning and introduces decision errors when the tuned threshold
from development data may not be optimal for all instances.
We propose the adaptive thresholding technique, which replaces the global threshold with a
learnable threshold class. The threshold class is learned with our adaptive-threshold loss, which
is a rank-based loss that pushes the logits of positive classes above the threshold and pulls the
logits of negative classes below in model training. At the test time, we return classes that have
higher logits than the threshold class as the predicted labels or return NA if such class does not exist.
This technique eliminates the need for threshold tuning, and also makes the threshold adjustable to
different entity pairs, which leads to much better results.
By combining the proposed two techniques, we propose a simple yet effective relation extraction
model, named ATLOP (Adaptive Thresholding and Localized cOntext Pooling), to fully utilize
the power of pre-trained language models [13, 44]. This model tackles the multi-label and multi-
entity problems in document-level RE. Experiments on three document-level relation extraction
datasets, DocRED [17], CDR [45], and GDA [46], demonstrate that our ATLOP model significantly
outperforms the state-of-the-art methods. The contributions of our work are summarized as follows:
• We propose adaptive-thresholding loss, which enables the learning of an adaptive threshold
that is dependent on entity pairs and reduces the decision errors caused by using a global
threshold.
• We propose localized context pooling, which transfers pre-trained attention to grab related
context for entity pairs to get better entity representations.
• We conduct experiments on three public document-level relation extraction datasets. Experi-
mental results demonstrate the effectiveness of our ATLOP model that achieves state-of-the-art
performance on three benchmark datasets.
2.2 Problem Formulation
Given a document d and a set of entities{e
i
}
n
i=1
, the task of document-level relation extraction is
to predict a subset of relations fromR∪{NA} between the entity pairs(e
s
,e
o
)
s,o=1...n;s̸=o
, where
R is a pre-defined set of relations of interest, e
s
,e
o
are identified as subject and object entities,
respectively. An entity e
i
can occur multiple times in the document by entity mentions{m
i
j
}
N
e
i
j=1
. A
relation exists between entities(e
s
,e
o
) if it is expressed by any pair of their mentions. The entity
pairs that do not express any relation are labeled NA. At the test time, the model needs to predict the
labels of all entity pairs(e
s
,e
o
)
s,o=1...n;s̸=o
in document d.
2.3 Enhanced BERT Baseline
In this section, we present our base model for document-level relation extraction. We build our
model based on existing BERT baselines [17, 42] and integrate other techniques to further improve
the performance.
9
2.3.1 Encoder
Given a document d=[x
t
]
l
t=1
, we mark the position of entity mentions by inserting a special symbol
“*” at the start and end of mentions. It is adapted from the entity marker technique [15, 19, 47]. We
then feed the document into a pre-trained language model to obtain the contextual embeddings:
H=[h
1
,h
2
,...,h
l
]= BERT([x
1
,x
2
,...,x
l
]). (2.1)
Following previous work [29, 48], the document is encoded once by the encoder, and the classifica-
tion of all entity pairs is based on the same contextual embedding. We take the embedding of “*” at
the start of mentions as the mention embeddings. For an entity e
i
with mentions{m
i
j
}
N
e
i
j=1
, we apply
logsumexp pooling [49], a smooth version of max pooling, to get the entity embedding h
e
i
.
h
e
i
= log
N
e
i
∑
j=1
exp

h
m
i
j

. (2.2)
This pooling accumulates signals from mentions in the document. It shows better performance
compared to mean pooling in experiments.
2.3.2 Binary Classifier
Given the embedding(h
e
s
,h
e
o
) of an entity pair e
s
,e
o
computed by equation 2.2, we map the entities
to hidden states z with a linear layer followed by non-linear activation, then calculate the probability
of relation r by bilinear function and sigmoid activation. This process is formulated as:
z
s
= tanh(W
s
h
e
s
), (2.3)
z
o
= tanh(W
o
h
e
o
), (2.4)
P(r|e
s
,e
o
)=σ(z
⊺
s
W
r
z
o
+ b
r
),
where W
s
∈R
d× d
, W
o
∈R
d× d
, W
r
∈R
d× d
,b
r
∈R are model parameters. The representation of
one entity is the same among different entity pairs. To reduce the number of parameters in the
bilinear classifier, we use the group bilinear [50, 51], which splits the embedding dimensions into k
equal-sized groups and applies bilinear within the groups:
h
z
1
s
;...;z
k
s
i
= z
s
,
h
z
1
o
;...;z
k
o
i
= z
o
,
P(r|e
s
,e
o
)=σ

k
∑
i=1
z
i⊺
s
W
i
r
z
i
o
+ b
r
!
, (2.5)
where W
i
r
∈R
d/k× d/k
for i= 1...k are model parameters, P(r|e
s
,e
o
) is the probability that relation
r is associated with the entity pair(e
s
,e
o
). In this way, we can reduce the number of parameters
from d
2
to d
2
/k. We use the binary cross entropy loss for training. During inference, we tune a
10
Figure 2.2: An artificial illustration of our proposed adaptive-thresholding loss. A TH class is
introduced to separate positive classes and negative classes: positive classes would have higher
probabilities than TH, and negative classes would have lower probabilities than TH.
global thresholdθ that maximizes evaluation metrics (F
1
score for RE) on the development set and
return r as an associated relation if P(r|e
s
,e
o
)>θ or return NA if no relation exists.
Our enhanced base model achieves near state-of-the-art performance in our experiments, signifi-
cantly outperforms existing BERT baselines.
2.4 Adaptive Thresholding
The RE classifier outputs the probability P(r|e
s
,e
o
) within the range[0,1], which needs thresholding
to be converted to relation labels. As the threshold neither has a closed-form solution nor is
differentiable, a common practice for deciding threshold is enumerating several values in the range
(0,1) and picking the one that maximizes the evaluation metrics (F
1
score for RE). However, the
model may have different confidence for different entity pairs or classes in which one global
threshold does not suffice. The number of relations varies (multi-label problem) and the models
may not be globally calibrated so that the same probability does not mean the same for all entity
pairs. This problem motivates us to replace the global threshold with a learnable, adaptive one,
which can reduce decision errors during inference.
For the convenience of explanation, we split the labels of entity pair T =(e
s
,e
o
) into two subsets:
positive classesP
T
and negative classesN
T
, which are defined as follows:
• positive classesP
T
⊆ R are the relations that exist between the entities in T . If T does not
express any relation,P
T
is empty.
• negative classesN
T
⊆ R are the relations that do not exist between the entities. If T does not
express any relation,N
T
=R.
If an entity pair is classified correctly, the logits of positive classes should be higher than the
threshold while those of negative classes should be lower. Here we introduce a threshold class TH,
which is automatically learned in the same way as other classes (see Eq.2.5). At the test time, we
return classes with higher logits than the TH class as positive classes or return NA if such classes do
not exist. This threshold class learns an entities-dependent threshold value. It is a substitute for the
global threshold and thus eliminates the need for tuning threshold on the development set.
11
To learn the new model, we need a special loss function that considers the TH class. We
design our adaptive-thresholding loss based on the standard categorical cross-entropy loss. The loss
function is broken down into two parts as shown below:
L
1
=− ∑
r∈P
T
log

exp(logit
r
)
∑
r
′
∈P
T
∪{TH}
exp(logit
r
′)
!
,
L
2
=− log

exp(logit
TH
)
∑
r
′
∈N
T
∪{TH}
exp(logit
r
′)
!
,
L =L
1
+L
2
.
The first part L
1
involves positive classes and the TH class. Since there may be multiple positive
classes, the total loss is calculated as the sum of categorical cross entropy losses on all positive
classes [52, 53].L
1
pushes the logits of all positive classes to be higher than the TH class. It is not
used if there is no positive label. The second partL
2
involves the negative classes and threshold
class. It is a categorical cross entropy loss with TH class being the true label. It pulls the logits of
negative classes to be lower than the TH class. Two parts are simply summed for the total loss.
The proposed adaptive-thresholding loss is illustrated in Figure 2.2. It obtains a large perfor-
mance gain to the global threshold in our experiments.
2.5 Localized Context Pooling
The logsumexp pooling (see Eq. 2.2) accumulates the embedding of all mentions for an entity across
the whole document and generates one embedding for this entity. The entity embedding from this
document-level global pooling is then used in the classification of all entity pairs. However, for
an entity pair, some context of the entities may not be relevant. For example, in Figure 2.1, the
second mention of John Stanistreet and its context are irrelevant to the entity pair (John Stanistreet,
Bendigo). Therefore, it is better to have a localized representation that only attends to the relevant
context in the document that is useful to decide the relation for this entity pair.
Therefore we propose localized context pooling, where we enhance the embedding of an entity
pair with an additional local context embedding that is related to both entities. In this work, since
we use pre-trained transformer-based models as the encoder, which has already learned token-level
dependencies by multi-head self-attention [39], we consider directly using their attention heads for
localized context pooling. This method transfers the well-learned dependencies from the pre-trained
language model without learning new attention layers from scratch.
Specifically, given a pre-trained multi-head attention matrix A∈R
H× l× l
, where A
i jk
represents
attention from token j to token k in the i
th
attention head, we first take the attention from the “*”
symbol as the mention-level attention, then average the attention over mentions of the same entity
to obtain entity-level attention A
E
i
∈R
H× l
, which denotes attention from the i
th
entity to all tokens.
12
… …
, …
BERT
Layer BERT
Layer − 1
Context
Pooling
Figure 2.3: Illustration of localized context pooling. Tokens are weighted averaged to form the
localized context c
(s,o)
of the entity pair(e
s
,e
o
). The weights of tokens are derived by multiplying
the attention weights of the subject entity e
s
and the object entity e
o
from the last transformer layer
so that only the tokens that are important to both entities (highlighted in light yellow) receive higher
weights.
Then given an entity pair(e
s
,e
o
), we locate the local context that is important to both e
s
and e
o
by
multiplying their entity-level attention, and obtain the localized context embedding c
(s,o)
by:
A
(s,o)
= A
E
s
· A
E
o
,
q
(s,o)
=
H
∑
i=1
A
(s,o)
i
,
a
(s,o)
= q
(s,o)
/1
⊺
q
(s,o)
,
c
(s,o)
= H
⊺
a
(s,o)
,
where H is the contextual embedding in Eq. 2.1. The localized context embedding is then fused into
the globally pooled entity embedding to obtain entity representations that are different for different
entity pairs, by modifying the original linear layer in Eq. 2.3 and Eq. 2.4 as follows:
z
(s,o)
s
= tanh

W
s
h
e
s
+ W
c
1
c
(s,o)

, (2.6)
z
(s,o)
o
= tanh

W
o
h
e
o
+ W
c
2
c
(s,o)

, (2.7)
where W
c
1
,W
c
2
∈R
d× d
are model parameters. The proposed localized context pooling is illustrated
in Figure 2.3. In experiments, we use the attention matrix from the last transformer layer.
13
Statistics DocRED CDR GDA
# Train 3053 500 23353
# Dev 1000 500 5839
# Test 1000 500 1000
# Relations 97 2 2
Avg.# entities per Doc. 19.5 7.6 5.4
Table 2.1: Statistics of the datasets in experiments.
Hyperparam DocRED CDR GDA
BERT RoBERTa SciBERT SciBERT
Batch size 4 4 4 16
# Epoch 30 30 30 10
lr for encoder 5e-5 3e-5 2e-5 2e-5
lr for classifier 1e-4 1e-4 1e-4 1e-4
Table 2.2: Hyper-parameters in training.
2.6 Experiments
2.6.1 Datasets
We evaluate our ATLOP model on three public document-level relation extraction datasets. The
dataset statistics are shown in Table 2.1.
• DocRED [17] is a large-scale crowdsourced dataset for document-level RE. It is constructed
from Wikipedia articles. DocRED consists of 3053 documents for training. For entity pairs
that express relation(s), about 7% of them have more than one relation label.
• CDR [45] is a human-annotated dataset in the biomedical domain. It consists of 500 doc-
uments for training. The task is to predict the binary interactions between Chemical and
Disease concepts.
• GDA [46] is a large-scale dataset in the biomedical domain. It consists of 29192 articles for
training. The task is to predict the binary interactions between Gene and Disease concepts.
We follow Christopoulou, Miwa & Ananiadou [32] to split the training set into an 80/20 split
as training and development sets.
2.6.2 Experiment Settings
Our model is implemented based on Huggingface’s Transformers [56]. We use cased BERT-
base [13] or RoBERTa-large [44] as the encoder on DocRED, and cased SciBERT [57] on CDR and
GDA. We use mixed-precision training [58] based on the Apex library
1
. Our model is optimized
1
https://github.com/NVIDIA/apex
14
Model Dev Test
Ign F
1
F
1
Ign F
1
F
1
Sequence-based Models
CNN [17] 41.58 43.45 40.33 42.26
BiLSTM [17] 48.87 50.94 48.78 51.06
Graph-based Models
BiLSTM-AGGCN [35] 46.29 52.47 48.89 51.45
BiLSTM-LSR [33] 48.82 55.17 52.15 54.18
BERT-LSR
BASE
[33] 52.43 59.00 56.97 59.05
Transformer-based Models
BERT
BASE
[42] - 54.16 - 53.20
BERT-TS
BASE
[42] - 54.42 - 53.92
HIN-BERT
BASE
[43] 54.29 56.31 53.70 55.60
CorefBERT
BASE
[54] 55.32 57.51 54.54 56.96
CorefRoBERTa
LARGE
[54] 57.35 59.43 57.90 60.25
Our Methods
BERT
BASE
(our implementation) 54.27± 0.28 56.39± 0.18 - -
BERT-E
BASE
56.51± 0.16 58.52± 0.19 - -
BERT-ATLOP
BASE
59.22± 0.15 61.09± 0.16 59.31 61.30
RoBERTa-ATLOP
LARGE
61.32± 0.14 63.18± 0.19 61.39 63.40
Table 2.3: Main results (%) on the development and test set of DocRED. We report the mean and
standard deviation of F
1
on the development set by conducting 5 runs of training using different
random seeds. We report the official test score of the best checkpoint on the development set.
Model CDR GDA
BRAN [29] 62.1 -
CNN [55] 62.3 -
EoG [32] 63.6 81.5
LSR [33] 64.8 82.2
SciBERT (our implementation) 65.1± 0.6 82.5± 0.3
SciBERT-E 65.9± 0.5 83.3± 0.3
SciBERT-ATLOP 69.4± 1.1 83.9± 0.2
Table 2.4: Test F
1
score (%) on CDR and GDA dataset. Our ATLOP model with the SciBERT
encoder outperforms the current SOTA results.
with AdamW [59] using learning rates∈{2e− 5,3e− 5,5e− 5,1e− 4}, with a linear warmup [60] for
the first 6% steps followed by a linear decay to 0. We apply dropout [61] between layers with rate
0.1, and clip the gradients of model parameters to a max norm of 1.0. We perform early stopping
based on the F
1
score on the development set. All hyper-parameters are tuned on the development
set. We list some of the hyper-parameters in Table 2.2.
15
For models that use a global threshold, we search threshold values from{0.1,0.2,...,0.9} and
pick the one that maximizes dev F
1
. All models are trained with 1 Tesla V100 GPU. For the
DocRED dataset, the training takes about 1 hour 45 minutes with BERT-base encoder and 3 hours
30 minutes with RoBERTa-large encoder. For CDR and GDA datasets, the training takes 20 minutes
and 3 hours 30 minutes with SciBERT encoder, respectively.
2.6.3 Main Results
We compare ATLOP with sequence-based models, graph-based models, and transformer-based
models on the DocRED dataset. The experiment results are shown in Table 2.3. Following Yao, Ye,
Li, Han, Lin, Liu, et al. [17], we use F
1
and Ign F
1
in evaluation. The Ign F
1
denotes the F
1
score
excluding the relational facts that are shared by the training and dev/test sets.
Sequence-based Models. These models use neural architectures such as CNN [62] and bidirectional
LSTM [63] to encode the entire document, then obtain entity embeddings and predict relations for
each entity pair with bilinear function.
Graph-based Models. These models construct document graphs by learning latent graph structures
of the document and perform inference with graph convolutional network [64]. We include two
state-of-the-art graph-based models, AGGCN [35] and LSR [33], for comparison. The result of
AGGCN is from the re-implementation by Nan, Guo, Sekulic & Lu [33].
Transformer-based Models. These models directly adapt pre-trained language models to document-
level RE without using graph structures. They can be further divided into pipeline models (BERT-
TS [42]), hierarchical models (HIN-BERT [43]), and pre-training methods (CorefBERT and Core-
fRoBERTa [54]). We also include the BERT baseline [42] and our re-implemented BERT baseline
in comparison.
We find that our re-implemented BERT baseline gets significantly better results than Wang,
Focke, Sylvester, Mishra & Wang [42], and outperforms the state-of-the-art RNN-based model
BiLSTM-LSR by 1.2%. It demonstrates that pre-trained language models can capture long-distance
dependencies among entities without explicitly using graph structures. After integrating other
techniques, our enhanced baseline BERT-E
BASE
achieves an F1 score of 58.52%, which is close to
the current state-of-the-art model BERT-LSR
BASE
. Our BERT-ATLOP
BASE
model further improves
the performance of BERT-E
BASE
by 2.6%, demonstrating the efficacy of the proposed two novel
techniques. Using RoBERTa-large as the encoder, our ALTOP model achieves an F1 score of
63.40%, which is a new state-of-the-art result on DocRED.
2.6.4 Results on Biomedical Datasets
Experiment results on two biomedical datasets are shown in Table 2.4. Verga, Strubell & McCallum
[29] and Nguyen & Verspoor [55] are both sequence-based models that use self-attention network
and CNN as the encoders, respectively. Christopoulou, Miwa & Ananiadou [32] and Nan, Guo,
Sekulic & Lu [33] use graph-based models that construct document graphs by heuristics or structured
attention, and perform inference with graph neural network. To our best knowledge, transformer-
based pre-trained language models have not been applied to document-level RE datasets in the
biomedical domain. In experiments, we replace the encoder with SciBERT, which is pre-trained
on multi-domain corpora of scientific publications. The SciBERT baseline already outperforms all
16
Model Ign F
1
F
1
BERT-ATLOP
BASE
59.22 61.09
− Adaptive Thresholding 58.32 60.20
− Localized Context Pooling 58.19 60.12
− Adaptive-Thresholding Loss 39.52 41.74
BERT-E
BASE
56.51 58.52
− Entity Marker 56.22 58.28
− Group Bilinear 55.51 57.54
− Logsumexp Pooling 55.35 57.40
Table 2.5: Ablation study of ATLOP on DocRED. We turn off different components of the model
one at a time. These ablation results show that both adaptive thresholding and localized context
pooling are effective. Logsumexp pooling and group bilinear both bring noticeable gain to the
baseline.
Strategy Dev F
1
Test F
1
Global Thresholding 60.14 60.62
Per-class Thresholding 61.73 60.35
Adaptive Thresholding 61.27 61.30
Table 2.6: Result of different thresholding strategies on DocRED. Our adaptive thresholding
consistently outperforms other strategies on the test set.
existing methods. Our SciBERT-ATLOP model further improves the F
1
score by 4.3% and 1.4% on
CDR and GDA, respectively, yielding new state-of-the-art results on these two datasets.
2.6.5 Ablation Study
To show the efficacy of our proposed techniques, we conduct two sets of ablation studies on ATLOP
and enhanced baseline, by turning off one component at a time. We observe that all components
contribute to model performance. The adaptive thresholding and localized context pooling are
equally important to model performance, leading to a drop of 0.89% and 0.97% in dev F
1
score
respectively when removed from ATLOP. Note that the adaptive thresholding only works when the
model is optimized with the adaptive-thresholding loss. Applying adaptive thresholding to models
trained with binary cross entropy results in dev F
1
of 41.74%.
For our enhanced baseline model BERT-E
BASE
, both group bilinear and logsumexp pooling lead
to about 1% increase in dev F
1
. We find the improvement from entity markers is minor ( 0.24% in
dev F
1
) but still use the technique in the model as it makes the derivation of mention embedding
and mention-level attention easier.
17
1-5 6-10 11-15 16-20 21-25 26-30 31-35
# of Entities per Document
50
55
60
65
70
75
dev F
1
(in %)
w/ LOP
w/o LOP
Figure 2.4: Dev F
1
score of documents with the different number of entities on DocRED. Our
localized context pooling achieves better results when the number of entities is larger than 5. The
improvement becomes more significant when the number of entities increases.
2.6.6 Analysis of Thresholding
Global thresholding does not consider the variations of model confidence in different classes or
instances, and thus yields suboptimal performance. One interesting question is whether we can
improve global thresholding by tuning different thresholds for different classes. To answer this
question, We try to tune different thresholds on different classes to maximize the dev F
1
score on
DocRED using the cyclic optimization algorithm [65]. Results are shown in Table 2.6. We find
that using per-class thresholding significantly improves the dev F
1
score to 61.73%, which is even
higher than the result of adaptive thresholding. However, this gain does not transfer to the test set.
The result of per-class thresholding is even worse than global thresholding. It indicates that tuning
per-class thresholding after training can lead to severe over-fitting to the development set. While our
adaptive thresholding technique learns the threshold in training, which can generalize to the test set.
2.6.7 Analysis of Context Pooling
To show that our localized context pooling (LOP) technique mitigates the multi-entity issue, we
divide the documents in the development set of DocRED into different groups by the number
of entities, and evaluate models trained with or without localized context pooling on each group.
Experiment results are shown in Figure 2.4. We observe that for both models, their performance gets
worse when the document contains more entities. The model w/ LOP consistently outperforms the
model w/o LOP except when the document contains very few entities (1 to 5), and the improvement
gets larger when the number of entities increases. However, the number of documents that only
contain 1 to 5 entities is very small (4 in the dev set), and the documents in DocRED contain 19
entities on average. Therefore our localized context pooling still improves the overall F
1
score
significantly. This indicates that the localized context pooling technique can capture related context
for entity pairs and thus alleviates the multi-entity problem.
We also visualize the context weights of the example in Figure 2.1. As shown in Figure 2.5,
our localized context pooling gives high weights to born and died, which are most relevant to both
18
John Stanistreet was an Australian politician. He wa
s born in Bendigo to legal manager John Jepson Stan
istreet and Maud McIlroy. (… 4 sentences …) In 195
5 John Stanistreet was elected to the Victorian Legisl
ative Assembly as the Liberal and Country Party mem
ber for Bendigo, but he was defeated in 1958. Stanistr
eet died in Bendigo in 1971.

Subject: John Stanistreet Object: Bendigo
Relation: place of birth; place of death
Figure 2.5: Context weights of an example from DocRED. We visualize the weight of context
tokens a
(s,o)
in localized context pooling. The model attends to the most relevant context born and
died for entity pair (John Stanistreet, Bendigo).
entities (John Stanistreet, Bendigo). These two tokens are also evidence for the two ground truth
relationships place of birth and place of death, respectively. Tokens like elected and politician
get much smaller weights because they are only related to the subject entity John Stanistreet. The
visualization demonstrates that the localized context can locate the context that is related to both
entities.
2.7 Related Work
Early research efforts on relation extraction concentrate on predicting the relationship between
two entities within a sentence. Various approaches including sequence-based methods [19, 26,
66], graph-based methods [27, 28, 35, 67], transformer-based methods [14, 15], and pre-training
methods [47, 68] have been shown effective in tackling this problem.
However, as large amounts of relationships are expressed by multiple sentences [17, 29], recent
work starts to explore document-level relation extraction. Most approaches on document-level RE
are based on document graphs, which were introduced by Quirk & Poon [69]. Specifically, they use
words as nodes and inner and inter-sentential dependencies (dependency structures, coreferences,
etc.) as edges. This document graph provides a unified way of extracting the features for entity pairs.
Later work extends the idea by improving neural architectures [29, 30, 49, 70, 71] or adding more
types of edges [32, 33]. In particular, Christopoulou, Miwa & Ananiadou [32] constructs nodes
of different granularities (sentence, mention, entity), connects them with heuristically generated
edges, and infers the relations with an edge-oriented model [72]. Nan, Guo, Sekulic & Lu [33]
treats the document graph as a latent variable and induces it by structured attention [31]. This work
also proposes a refinement mechanism to enable multi-hop information aggregation from the whole
document. Their LSR model achieved state-of-the-art performance on document-level RE.
There have also been models that directly apply pre-trained language models without introducing
document graphs, since edges such as dependency structures and coreferences can be automatically
learned by pre-trained language models [40, 41, 73, 74]. In particular, Wang, Focke, Sylvester,
Mishra & Wang [42] proposes a pipeline model that first predicts whether a relationship exists in
an entity pair and then predicts the specific relation types. Tang, Cao, Zhang, Cao, Fang, Wang,
19
et al. [43] proposes a hierarchical model that aggregates entity information from the entity level,
sentence level, and document level. Ye, Lin, Du, Liu, Sun & Liu [54] introduces a copy-based
training objective to pre-training, which enhances the model’s ability in capturing coreferential
information and brings noticeable gain on various NLP tasks that require coreferential reasoning.
However, none of the models focus on the multi-entity and multi-label problems, which are
among the key differences of document-level RE to its sentence-level RE counterpart. Our ATLOP
model deals with the two problems by two novel techniques: adaptive thresholding and localized
context pooling, and significantly outperforms existing models.
2.8 Conclusion
In this work, we propose the ATLOP model for document-level relation extraction, which features
two novel techniques: adaptive thresholding and localized context pooling. The adaptive threshold-
ing technique replaces the global threshold in multi-label classification with a learnable threshold
class that can decide the best threshold for each entity pair. The localized context pooling utilizes
pre-trained attention heads to locate relevant context for entity pairs and thus helps in alleviating
the multi-entity problem. Experiments on three public document-level relation extraction datasets
demonstrate that our ATLOP model significantly outperforms existing models and yields new
state-of-the-art results on all datasets.
20
Chapter 3
Noisy Label Learning
Recent information extraction approaches have relied on training deep neural models. However,
such models can easily overfit noisy labels and suffer from performance degradation. While it is
very costly to filter noisy labels in large learning resources, recent studies show that such labels take
more training steps to be memorized and are more frequently forgotten than clean labels, therefore
are identifiable in training. Motivated by such properties, we propose a simple co-regularization
framework for entity-centric information extraction, which consists of several neural models with
identical structures but different parameter initialization. These models are jointly optimized with
the task-specific losses and are regularized to generate similar predictions based on an agreement
loss, which prevents overfitting on noisy labels. Extensive experiments on two widely used but noisy
benchmarks for information extraction, TACRED and CoNLL03, demonstrate the effectiveness of
our framework.
3.1 Introduction
Deep neural models have achieved significant success on various information extraction (IE) tasks.
However, when training labels contain noise, deep neural models can easily overfit the noisy labels,
leading to severe performance degradation [75, 76]. Unfortunately, labeling on large corpora,
regardless of using human annotation [77] or automated heuristics [78], inevitably suffers from
labeling errors. This problem has even drastically affected widely used benchmarks, such as
CoNLL03 [79] and TACRED [19], where a notable portion of incorrect labels have been caused in
annotation and largely hindered the performance of SOTA systems [80, 81]. Hence, developing
a robust learning method that better tolerates noisy supervision represents an urged challenge for
emerging IE models.
So far, few research efforts have been made to developing noise-robust IE models, and existing
work mainly focuses on the weakly supervised or distantly supervised setting [82–85]. Most of such
methods typically depend on multi-instance learning that relies on bags of instances provided by
distant supervision [13, 22, 82] or require an additional clean and sufficiently large reference dataset
to develop a noise filtering model [86]. Accordingly, those methods may not be generally adapted to
supervised training settings, where the aforementioned auxiliary learning resources are not always
available. Particularly, CrossWeigh [87] is a representative work that denoises a natural language
dataset without using extra learning resources. This method trains multiple independent models
on different partitions of training data and downweighs instances on which the models disagree.
21
AAAB7nicbVBNSwMxEJ31s9avqkcvwSLUS9mIqMeiF48V7Ae0a8mm2TY0mw1JVihLf4QXD4p49fd489+YtnvQ1gcDj/dmmJkXKsGN9f1vb2V1bX1js7BV3N7Z3dsvHRw2TZJqyho0EYluh8QwwSVrWG4FayvNSBwK1gpHt1O/9cS04Yl8sGPFgpgMJI84JdZJLfWYVfDZpFcq+1V/BrRMcE7KkKPeK311+wlNYyYtFcSYDvaVDTKiLaeCTYrd1DBF6IgMWMdRSWJmgmx27gSdOqWPokS7khbN1N8TGYmNGceh64yJHZpFbyr+53VSG10HGZcqtUzS+aIoFcgmaPo76nPNqBVjRwjV3N2K6JBoQq1LqOhCwIsvL5PmeRVfVvH9Rbl2k8dRgGM4gQpguIIa3EEdGkBhBM/wCm+e8l68d+9j3rri5TNH8Afe5w+MiI8O
p
(1)
AAAB6HicbVBNS8NAEJ34WetX1aOXxSJ4KomIeix68diC/YA2lM120q7dbOLuRiihv8CLB0W8+pO8+W/ctjlo64OBx3szzMwLEsG1cd1vZ2V1bX1js7BV3N7Z3dsvHRw2dZwqhg0Wi1i1A6pRcIkNw43AdqKQRoHAVjC6nfqtJ1Sax/LejBP0IzqQPOSMGivVH3ulsltxZyDLxMtJGXLUeqWvbj9maYTSMEG17nhuYvyMKsOZwEmxm2pMKBvRAXYslTRC7WezQyfk1Cp9EsbKljRkpv6eyGik9TgKbGdEzVAvelPxP6+TmvDaz7hMUoOSzReFqSAmJtOvSZ8rZEaMLaFMcXsrYUOqKDM2m6INwVt8eZk0zyveZcWrX5SrN3kcBTiGEzgDD66gCndQgwYwQHiGV3hzHpwX5935mLeuOPnMEfyB8/kD3ZWM+w==
q
AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPRi8cW7Ae0oWy2k3btZhN2N0Io/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4bua3n1BpHssHkyXoR3QoecgZNVZqZP1yxa26c5BV4uWkAjnq/fJXbxCzNEJpmKBadz03Mf6EKsOZwGmpl2pMKBvTIXYtlTRC7U/mh07JmVUGJIyVLWnIXP09MaGR1lkU2M6ImpFe9mbif143NeGNP+EySQ1KtlgUpoKYmMy+JgOukBmRWUKZ4vZWwkZUUWZsNiUbgrf88ippXVS9q6rXuKzUbvM4inACp3AOHlxDDe6hDk1ggPAMr/DmPDovzrvzsWgtOPnMMfyB8/kD6bWNAw==
y
AAAB7nicbVBNSwMxEJ2tX7V+VT16CRahXsquiHosevEiVLAf0K4lm2bb0Gw2JFmhLPsjvHhQxKu/x5v/xrTdg7Y+GHi8N8PMvEBypo3rfjuFldW19Y3iZmlre2d3r7x/0NJxoghtkpjHqhNgTTkTtGmY4bQjFcVRwGk7GN9M/fYTVZrF4sFMJPUjPBQsZAQbK7XlY1q9O8365Ypbc2dAy8TLSQVyNPrlr94gJklEhSEca931XGn8FCvDCKdZqZdoKjEZ4yHtWipwRLWfzs7N0IlVBiiMlS1h0Ez9PZHiSOtJFNjOCJuRXvSm4n9eNzHhlZ8yIRNDBZkvChOOTIymv6MBU5QYPrEEE8XsrYiMsMLE2IRKNgRv8eVl0jqreRc17/68Ur/O4yjCERxDFTy4hDrcQgOaQGAMz/AKb450Xpx352PeWnDymUP4A+fzB7cwjyo=
p
(M)
AAAB7nicbVBNSwMxEJ2tX7V+VT16CRahXspuEfVY9OKxgv2Adi3ZNNuGZrMhyQpl6Y/w4kERr/4eb/4bs+0etPXBwOO9GWbmBZIzbVz32ymsrW9sbhW3Szu7e/sH5cOjto4TRWiLxDxW3QBrypmgLcMMp12pKI4CTjvB5DbzO09UaRaLBzOV1I/wSLCQEWys1JGPabV+PhuUK27NnQOtEi8nFcjRHJS/+sOYJBEVhnCsdc9zpfFTrAwjnM5K/URTickEj2jPUoEjqv10fu4MnVlliMJY2RIGzdXfEymOtJ5Gge2MsBnrZS8T//N6iQmv/ZQJmRgqyGJRmHBkYpT9joZMUWL41BJMFLO3IjLGChNjEyrZELzll1dJu17zLmve/UWlcZPHUYQTOIUqeHAFDbiDJrSAwASe4RXeHOm8OO/Ox6K14OQzx/AHzucPjg6PDw==
p
(2)
model 1
model 2
model M
…
…
aggregation of p
(1)
to p
(M)
ground-truth labels
back-propagation
mean aggregation
Figure 3.1: Illustration of our co-regularization framework. The base models are jointly optimized
with the task-specific loss from label y and an agreement loss, which regularizes the models to
generate similar predictions to the aggregated soft target probability q q q.
Though effective, a method of this kind requires training tens of redundant neural models, leading
to excessive computational overhead for large models. As far as we know, the problem of noisy
labels in supervised learning for IE tasks has not been well investigated.
We aim to develop a general denoising framework that can easily incorporate existing supervised
learning models for entity-centric IE tasks. Our method is motivated by studies [75, 88] showing
that noisy labels often have delayed learning curves, as incorrectly labeled instances are more likely
to contradict the inductive bias captured by the model. Hence, noisy label instances take a longer
time to be picked up by neural models and are frequently forgotten in later epochs. Therefore,
predictions by more than one model tend to disagree on such instances. Accordingly, we propose
a simple yet effective co-regularization framework to handle noisy training labels, as illustrated
in Figure 3.1. Our framework consists of two or more neural classifiers with identical structures
but different initialization. In training, all classifiers are optimized on the training data with the
task-specific loss and jointly regularized with regard to an agreement loss that is defined as the
Kullback–Leibler (KL) divergence among predicted probability distributions. Then for instances
where a classifier’s predictions disagree with labels, the agreement loss encourages the classifier to
give similar predictions to the other classifier(s) instead of the actual (possibly noisy) labels. In this
way, the framework prevents the incorporated classifiers from overfitting noisy labels.
We apply the framework to two important entity-centric IE tasks, named entity recognition
(NER) and relation extraction (RE). We conduct extensive experiments on two prevalent but
noisy benchmarks, CoNLL03 for NER and TACRED for RE, and apply the proposed learning
frameworks to train various models from prior studies for these two tasks. The results demonstrate
the effectiveness of our method in noise-robust training, leading to promising and consistent
performance improvement. We present contributions as follows:
• We propose a general co-regularization framework that can effectively learn supervised IE
models from noisy datasets without the need for any extra learning resources.
22
• We discuss in detail the different design strategies of the framework and the trade-off between
efficiency and effectiveness.
• Extensive experiments on NER and RE demonstrate that our framework yields promising
improvements on various SOTA models and outperforms existing denoising frameworks.
3.2 Method
We focus on developing a noise-robust learning framework that improves supervised models for
entity-centric IE tasks. In such tasks, (noisy) labels can be assigned to either individual tokens (NER)
or pairs of entities (RE) in natural language text. Specifically, D ={(x x x
i
,y
i
)}
N
i=1
is a noisily labeled
dataset, where each data instance consists of a lexical sequence or a context x x x, and a label y. y
is annotated either on tokens of x x x for NER or on a pair of entity mentions in x x x for RE. For some
instances inD, the labels are incorrect. Our objective is to learn a noise-robust model f with the
presence of such noisily labeled instances fromD without using external resources such as a clean
development dataset [86].
3.2.1 Learning Process
Our framework is motivated by the delayed learning curve of a neural model on noisy data, compared
with learning on clean data. On noisy data, neural models tend to fit easy and clean instances that
are more consistent with the well-represented patterns of data in early steps but need more steps
to capture noise [75]. Moreover, learned noisy examples tend to be frequently forgotten in later
epochs [88] since they conflict with the general inductive bias represented by the clean data majority.
Therefore, model prediction is likely to be consistent with the clean labels while is often inconsistent
or oscillates on noisy labels over different training epochs. As a result, labels that are different
from the model’s predictions in the later epochs of training are likely to be noisy and should be
down-weighted or rectified so as to reduce their impact on optimization.
The proposed framework incorporates several copies of a task-specific IE model with the
same architecture but different (random) parameter initialization. These IE models are jointly
optimized on the noisy dataset based on their task-specific losses as well as on an agreement
loss. During training, the predicted probability distributions from models are aggregated as a soft
target probability, which represents the models’ estimations of the true label. The agreement loss is
responsible for encouraging these models to generate similar predictions to the soft target probability.
In this learning process, models starting their training from varied initialization generate different
decision boundaries. By aggregating their predictions, the soft target probability can better separate
noisy labels from clean labels that have not yet been learned.
The learning process of our framework is described in Algorithm 1. It consists of M (M≥ 2) copies of the task-specific model, denoted { f
k
}
M
k=1
, with different initialization. Regarding
initialization, for models that are trained from scratch, all parameters are randomly initialized.
Otherwise, for those that are built upon pre-trained language models, only the parameters that
are external to the language models (e.g., those of a downstream softmax classifier) are randomly
initialized, while the pre-trained parameters are the same. Once initialized, our framework trains
those models in two phases. The first α% training steps undergo a warm-up phase, whereα is a
23
Algorithm 1: Learning Process
Input: DatasetD, hyperparameters T,α,M,γ.
Output: A trained model f .
Initialize M neural models{ f
k
}
M
k=1
.
for t= 1...T do
Sample a batchB fromD.
Calculate task losses{L
(k)
sup
}
M
k=1
.
L
T
=
1
M
∑
M
k=1
L
(k)
sup
if t <α%× T then // Warmup
Update model parameters w.r.t.L
T
.
else
Get the probability distribution of classes{p p p}
M
k=1
with M models.
Calculate the soft target probability q q q by Equation 3.1.
Calculate the agreement lossL
agg
by Equation 3.2 and Equation 3.3.
L =L
T
+γ·L
agg
.
Update model parameters w.r.t.L .
Return f
1
or the best-performing model.
hyperparameter. This phase seeks to help the model reach initial convergence on the task. When
a new batch comes in, we first calculate the task-specific training losses on M models{L
(k)
sup
}
M
k=1
and average them asL
T
, then update model parameters w.r.t.L
T
. After the warm-up phase, an
agreement lossL
agg
is further introduced to measure the distance from the predictions of M models
to the soft target probability q q q. Parameters are accordingly updated based on the joint lossL ,
encouraging the model to generate predictions that are consistent with both the training labels and
the soft target probability. The formalization of the loss function is described next (Section 3.2.2).
In the end, we can either use the model f
1
or select the best-performing model for inference.
3.2.2 Co-regularization Objective
In our framework, the influence of noisy labels in training is decreased by optimizing the agreement
loss. Specifically, given a batch of data instances B={(x x x
i
,y
i
)}
N
i=1
, we first feed the instances to M
incorporated models to get their predictions
n
{p p p
(k)
i
}
N
i=1
o
M
k=1
onB, where p p p∈R
C
is the predicted
probability distribution of C classes. Then we calculate the soft target probability q q q by averaging
the predictions:
q q q
i
=
1
M
M
∑
k=1
p p p
(k)
i
, (3.1)
24
which represents the models’ estimates of the true label. Finally, we calculate the agreement loss
L
agg
as the average KL divergence from q q q to each p p p
(k)
,k= 1,...,M:
d(q q q
i
||p p p
(k)
i
)=
C
∑
j=1
q q q
i j
log

q q q
i j
+ε
p p p
i j
+ε
!
, (3.2)
L
agg
=
1
MN
N
∑
i=1
M
∑
k=1
d(q q q
i
||p p p
(k)
i
), (3.3)
whereε is a small positive number to avoid division by zero. We can easily tell that the agreement
loss encourages the models to get similar predictions based on the same input. As the KL divergence
is non-negative, the agreement loss is minimized only when q q q
i
= p p p
(k)
i
for k= 1,...,M, which implies
that all p p p
(k)
i
should be equal because we use the average probability for q q q. We may also use other
aggregates for q q q as long as they satisfy that q q q
i
= p p p
(k)
i
,k= 1,...,M when all p p p
(k)
i
are equal so as to
maintain such property of the agreement loss. We consider the following alternatives for q q q:
• Average logits. Given the logits{l l l
(k)
i
}
M
k=1
of predicted probabilities on the M models, we
first average the logits l l l
i
=
1
M
∑
M
k=1
l l l
(k)
i
and then feed l l l
i
to a softmax function to get the soft
target probability q q q.
• Max-loss probability. A noise-robust model will disagree on noisy labels and produce large
training losses. Therefore, for each instance i in the batch, we assume the prediction p
∗ i
that
has the largest task-specific loss among the M models to be more reliable and use it as the
soft target probability for instance i.
In experiments, we observe that all aggregate functions generally achieve similar performance. We
present the results of different q q q in Section 3.4.6.
3.2.3 Joint Training
The main learning objective of our framework is then to optimize the joint lossL =L
T
+γL
agg
,
where γ is a positive hyperparameter andL
T
is the average of task-specific classification losses
n
L
(k)
sup
o
M
k=1
. For classification problems such as NER and RE, the task-specific loss is defined as
the following cross-entropy loss, where I I I denotes an indicator function:
L
sup
=− 1
N
N
∑
i=1
C
∑
j=1
I I I[y
i
= j]log p p p
i j
. (3.4)
N thereof is the number of tokens for NER and the number of sentences for RE.
The joint training can be interpreted as a “soft-pruning” scheme. For clean labels where the
models’ predictions are usually close to the labels, the agreement loss and its gradient are both
small, so they have a small impact on training. While for noisy labels where the model predictions
disagree with the training labels, the agreement loss incurs a large magnitude of gradients in training,
which prevents the model from overfitting the noisy labels.
25
Besides co-regularization, denoising may also be attempted by “hard-pruning” the noisy labels.
Small-loss selection [89, 90] assumes that instances with large task-specific loss are noisy and
excludes them from training. However, some clean label instances, especially those from long-tail
classes, can also have large task-specific losses and will be incorrectly pruned. While for the
frequent classes, some noisy instances can have smaller task-specific losses and fail to be identified.
Such errors can accumulate during training and may hinder model performance. In our framework,
as we use the agreement loss instead of hard pruning, such errors will not be easily propagated (see
Section 3.4.6).
3.3 Tasks
We evaluate our framework on two fundamental entity-centric IE tasks, namely RE and NER. Our
framework can incorporate any kind of neural model that is dedicated to either task. Particularly, we
adopt off-the-shelf SOTA models that are mainly based on Transformers. This section introduces
the two attempted tasks and the design of task-specific models.
Relation extraction. RE aims at identifying the relations between a pair of entities in a piece of text
from the given vocabulary of relations. Specifically, given a sentence x x x and two entities e
s
and e
o
,
identified as the subject and object entities respectively, the goal is to predict the relation between
e
s
and e
o
. Following Shi & Lin [15], we formulate this task as a sentence classification problem.
Accordingly, we first apply the entity masking technique [19] to the input sentence and replace the
subject and object entities with their named entity types. For example, a short sentence “Bill Gates
founded Microsoft” will become “[SUBJECT-PERSON] founded [OBJECT-ORGANIZATION]” after
entity masking. We then feed the sentence to the pre-trained language model and use a softmax
classifier on the representation of the [ CLS] token to predict the relation.
Named entity recognition. NER seeks to locate and classify named entities in text into pre-defined
categories. Following Devlin, Chang, Lee & Toutanova [13], we formulate the task as a token
classification problem. In detail, a Transformer-based language model first tokenizes an input
sentence into a sub-token sequence. To classify each token, the representation of its first sub-token
is sent into a softmax classifier. We use the BIO tagging scheme [91] and output the tag with the
maximum likelihood as the predicted label.
3.4 Experiment
In this section, we evaluate the proposed learning framework based on two (noisy) benchmark
datasets for the two entity-centric IE tasks (Section 3.4.1-Section 3.4.4). In addition, a noise
filtering analysis is presented to show how our framework prevents an incorporated neural model
from overfitting noisy training data (Section 3.4.5), along with a detailed ablation study about
configurations with varied model copies, alternative noise filtering strategies, target functions, and
different noise rates (Section 3.4.6).
26
Dataset # train # dev # test # classes % noise
TACRED 68124 22631 15509 42 6.62
CoNLL03 14041 3250 3453 9 5.38
Table 3.1: Data statistics of TACRED and CoNLL03.
3.4.1 Datasets
The experiments are conducted on TACRED [19] and CoNLL03 [79]. TACRED is a crowdsourced
dataset for relation extraction. A recent study by Alt, Gabryszak & Hennig [81] found a large
portion of examples to be mislabeled and rectified some incorrect labels in the development and test
sets. CoNLL03 is a human-annotated dataset for NER. Another study by Wang, Shang, Liu, Lu,
Liu & Han [87] found that in 5.38% of sentences in CoNLL03, at least one token is mislabeled.
Accordingly, Wang, Shang, Liu, Lu, Liu & Han [87] also relabeled the test set
1
. We summarize the
statistics of both datasets in Table 3.1. For all compared methods, we report the results on both the
original and relabeled evaluation sets.
3.4.2 Base Models
We evaluate our framework by incorporating the following SOTA models:
• C-GCN [28] is a graph-based model for RE. It prunes the dependency graph and applies
graph convolutional networks to get the representation of entities.
• BERT [13] is a Transformer-based language model that is pre-trained from large-scale text
corpora. BothBase andLarge versions of the model are considered in our experiments.
• LUKE [92] is a Transformer-based language model that is pre-trained on both large-scale
text corpora and knowledge graphs. It achieves SOTA performance on various entity-related
tasks, including RE and NER.
We report the performance of the base models trained with and without our co-regularization
framework. We also compare our framework to CrossWeigh [87], which is another noisy-label
learning framework. Specifically, CrossWeigh partitions the training set into equal-sized chunks,
reserves each chunk, and then trains several models on the rest ones. After training, the models
predict on the reserved chunk, and instances on which the models disagree are down-weighted.
In the end, the chunks are combined and used to train a new model for inference. Learning by
CrossWeigh is dependant on a high computation cost. Wang, Shang, Liu, Lu, Liu & Han [87] split
the CoNLL03 dataset into 10 chunks and train 3 models on each partition, resulting in a total of
number 30 models. We follow their settings and train 30 models on both TACRED and CoNLL03.
1
Note that neither of these two datasets comes with a relabeled (clean) training set. All models are still trained on
the original noisy training set.
27
Model Original Relabeled
Dev F
1
Test F
1
Dev F
1
Test F
1
C-GCN♣ [28] 67.2 66.7 74.9 74.6
C-GCN-CrossWeigh 67.8 67.4 75.6 75.7
C-GCN-CR 67.7 67.2 75.6 75.4
BERT
BASE
[13] 69.1 68.9 76.4 76.9
BERT
BASE
-CrossWeigh 71.3 70.8 79.2 79.1
BERT
BASE
-CR 71.5 71.1 79.9 80.0
BERT
LARGE
[13] 70.9 70.2 78.3 77.9
BERT
LARGE
-CrossWeigh 72.1 71.9 79.5 79.8
BERT
LARGE
-CR 73.1 73.0 81.3 82.0
LUKE♣ [92] 71.1 70.9 80.1 80.6
LUKE-CrossWeigh 71.0 71.6 80.4 81.6
LUKE-CR 71.8 72.4 81.9 83.1
Table 3.2: F
1
score (%) on the dev and test set of TACRED.♣ marks results obtained from the
originally released implementation. We report the median of F
1
on 5 runs of training using different
random seeds. For fair comparison, the CR results are reported based on the predictions from model
f
1
in our framework.
3.4.3 Model Configurations
For base models C-GCN [28] and LUKE [92], we rerun the officially released implementations
using the recommended hyperparameters in the original papers. We implement BERT
BASE
and
BERT
LARGE
based on Huggingface’s Transformers [93]. For CrossWeigh [87], we re-implement
this framework using those compared base models. All models are optimized with Adam [94] using
a learning rate of 6e− 5 for TACRED and that of 1e− 5 for CoNLL03, with a linear learning rate
decay to 0. The batch size is fixed as 64 for all models. We finetune the TACRED model for 5
epochs and the CoNLL03 model for 50 epochs. The best model checkpoint is chosen based on
the F
1
score on the development set. We tune γ from{1.0,2.0,5.0,10.0,20.0}, and tune α from
{10,30,50,70,90}. We report the median of F
1
of 5 runs using different random seeds.
For efficiency, we use the simplest setup of our framework with two model copies ( M= 2) in
the main experiments (Section 3.4.4). q q q is set as the average probability in the main experiment.
Performance with more model copies and alternative aggregates is later studied in Section 3.4.6.
3.4.4 Main Results
The experiment results on TACRED and CoNLL03 are reported in Table 3.2 and Table 3.3 respec-
tively, where methods incorporated in our learning framework are marked with “CR”. As stated, the
results are reported under the setup where M= 2. For a fair comparison, the results are reported
based on the predictions from model f
1
in the framework. On TACRED, our framework leads to an
absolute improvement of 2.5− 4.1% in F
1
on the relabeled test set for Transformer-based models,
and a relatively smaller gain (0.8% in F
1
) for C-GCN. In particular, our framework enhances the
28
Model Original Relabeled
Dev F
1
Test F
1
Test F
1
BERT
BASE
[13] 95.58 91.96 92.91
BERT
BASE
-CrossWeigh 95.65 92.15 93.03
BERT
BASE
-CR 95.87 92.53 93.48
BERT
LARGE
[13] 96.16 92.24 93.22
BERT
LARGE
-CrossWeigh 96.32 92.49 93.61
BERT
LARGE
-CR 96.59 92.82 94.04
LUKE♣ [92] 97.03 93.91 95.60
LUKE–CrossWeigh 97.09 93.98 95.75
LUKE-CR 97.21 94.22 95.88
Table 3.3: F
1
score (%) on the dev and test set of CoNLL03.♣ marks results obtained using the
originally released code.
SOTA method LUKE by 2.5% in F
1
, leading to a very promising F
1
score of 83.1%. On CoNLL03,
where the noise rate is smaller than TACRED, our framework leads to a performance gain of
0.28− 0.82% in F
1
on the relabeled test set. On both IE tasks, our framework also leads to a
consistent improvement on the original test set. Compared to CrossWeigh, except for C-GCN where
the results are similar, our framework consistently outperforms it by 0.9− 2.2% on TACRED and by
0.13− 0.45% on CoNLL03. Moreover, as our framework requires training M models concurrently
while CrossWeigh requires training redundant models (30 in experiments), the computation cost of
our co-regularization framework is much lower than CrossWeigh. In general, the results here show
the effectiveness and practicality of the proposed framework.
3.4.5 Noise Filtering Analysis
The main experiments show that our framework can improve the overall performance of models
trained with noisy labels. In this section, we further demonstrate how our framework prevents
overfitting on noisy labels. To do so, we extract the 2,526 noisy instances from the development
and test sets of TACRED where the relabeling by Alt, Gabryszak & Hennig [81] disagrees with the
original labels. Accordingly, we obtain a noisy set containing those examples with original labels
and a clean set with rectified labels. We train a relation classifier on the union of the training set
and the noisy set and then evaluate the model on the clean set. In this case, worse performance on
the clean set indicates more severe overfitting on noisy labels.
Figure 3.2 shows the results by C-GCN-CR and BERT
BASE
-CR on the clean set, where we
observe that: (1) Compared to the original base models (γ = 0.0), those trained with our framework
achieves higher F
1
scores, indicating improved robustness against the label noise; (2) Comparing
different base models, the large classifier BERT
BASE
is typically less noise-robust than a smaller
model like C-GCN, which explains why the performance gain from our framework is more notable
on BERT
BASE
; (3) For both models, the F
1
score first increases then decreases, consistent with the
delayed learning curves that the neural models have on noisy instances [75].
29
0 50k 100k
Step
0
20
40
60
F1(%)
C-GCN
0 2k 4k
Step
0
20
40
60
BERT
BASE
= 0.0 = 1.0 = 2.0 = 5.0
Figure 3.2: F
1
score (%) on the clean set of TACRED. Classifiers trained with our framework are
more noise-robust compared to baselines (γ = 0).
# Models 2 3 4
BERT
BASE
-CR 80.0 79.5 79.8
BERT
LARGE
-CR 82.0 82.4 82.7
Table 3.4: F
1
score (%) of using different number of models on the relabeled test set of TACRED.
3.4.6 Ablation Study
Using extra model copies. The main results show that using two copies of a model in the co-
regularization framework has already improved the performance by a remarkable margin. Intuitively,
more models may generate higher-quality soft target probabilities and thus further improve the
performance. We further show the performance on TACRED by incorporating more model copies.
We report the relabeled test F
1
on TACRED in Table 3.4. We observe that increasing the number of
copies does not necessarily lead to a notable increase in performance. On BERT
LARGE
, increasing
the number of model copies from 2 to 4 gradually improves the performance from 82.0% to 82.7%.
While on BERT
BASE
, increasing the number of model copies does not improve the performance.
We notice that the increased number of copies leads to a significant increase in the agreement loss
for BERT
BASE
, indicating that the copies of BERT
BASE
fail to reach a consensus based on the
same input. This may be due to the relatively small model capacity of BERT
BASE
. Overall, this
study shows that the optimal M is dependent on the models and needs to be tuned on the specific
task. Note that as the models can be trained in parallel, increasing the number of models does not
necessarily increase the training time, though being at the cost of more computational resources.
Alternative strategies for noise filtering. Besides co-regularization, we also experiment with other
noise-filtering strategies. Small-loss selection [89, 90, 95] prunes the instances with the largest
training losses in the training batches. This method is motivated by the fact that the noisy instances
take a longer time to be memorized and usually cause a large training loss. We further try another
30
Model Original Relabeled
Test F
1
Test F
1
BERT
BASE
68.9 76.9
BERT
BASE
+ Small-loss selection 68.7 76.6
BERT
BASE
+ Relabeling 69.0 77.7
Table 3.5: F
1
score (%) of alternative noise filtering strategies on the test set of TACRED. The best
results are achieved whenδ = 2% for both methods.
Functions Avg prob Avg logit Max-loss prob
BERT
BASE
-CR 80.0 79.9 79.4
BERT
LARGE
-CR 82.0 81.6 82.2
Table 3.6: F
1
score (%) of different functions for q q q on the relabeled test set of TACRED.
strategy named relabeling. Instead of pruning the large-loss training instances, we relabel them with
the most likely labels from model predictions.
We evaluate the two noise filtering strategies on TACRED using BERT
BASE
as the base model.
For both strategies, we prune/relabelδ
t
=δ· t
T
percent of examples with the largest training loss in
each training batch following Han, Yao, Yu, Niu, Xu, Hu, et al. [90], where t is the current number
of training steps, T is the total number of training steps, andδ is the maximum pruning/relabeling
rate. These hyperparameters are tuned on the development set. The training loss is defined as
the average task-specific loss of the M models, where we set M= 2 in consistent with the main
experiments (Section 3.4.4). We tryδ from{2%,5%,8%} and report the best results.
Results are shown in Table 3.5. We find that δ = 2% achieves the best performance for both
strategies. The small-loss selection strategy underperforms the base model without noise filtering.
Relabeling outperforms the base model slightly, but the improvements are lesser than the proposed
co-regularization method. We observe that these two strategies do not work well on imbalanced
datasets, mostly pruning or relabeling training examples from long-tail classes. Specifically, on
the TACRED dataset, where the NA class accounts for 80% of the total labels, only 20% pruned
labels are from NA while the remaining 80% are from other classes. It is because that the model’s
predictions will be biased towards the frequent classes on imbalanced datasets, therefore leading to
the large training loss on long-tail instances. Once pruned or relabeled, such long-tail instances are
excluded from training, causing further error propagation that can lead to more biased predictions.
Our framework, on the contrary, adopts an agreement loss instead of hard pruning or relabeling,
which reduces such error propagation.
Alternative aggregates for q q q. Besides the average probability, we evaluate two other aggregates
for q q q, i.e. the average logits and the max-loss probability (Section 3.2.2). This experiment is
conducted with M = 2. F
1
results on the relabeled TACRED test set (Table 3.6) suggest that
different aggregates generally achieve comparable performance,with a marginal difference of up to
0.6% in F
1
. Therefore, the default setup is suggested to be the average probability, which is easier
to implement.
31
Flipped labels (%) 10 30 50 70 90
BERT
BASE
74.2 70.8 62.9 48.6 0
BERT
BASE
-CrossWeigh 77.3 75.6 71.6 61.3 25.1
BERT
BASE
-CR 79.3 78.3 73.2 63.5 34.1
BERT
BASE
w/o flipped labels 76.5 74.9 72.9 70.8 57.4
Table 3.7: F
1
score (%) under different noise rates on the relabeled set of TACRED.
Performance under different noise rates. We further evaluate our framework on training data of
different noise rates. To do so, we create noisy training data by randomly flipping 10%, 30%, 50%,
70%, or 90% labels in the training set of TACRED. Then we use those synthetic noisy training sets
to train RE models and evaluate them on the relabeled test set of TACRED. We use BERT
BASE
as
the base model and report the median F
1
score of 5 trials. Results are given in Table 3.7, which show
that our co-regularization framework consistently outperforms both the base model and CrossWeigh
under different noise rates. The gain generally becomes larger as the noise rate increases. In
comparison to BERT
BASE
trained on the training sets where all flipped labels are removed, our
framework, even trained on synthetic noise, achieves comparable or better results when the noise
rates are below 50%.
3.5 Related Work
We discuss two lines of related work. Each has a large body of work which we can only provide as
a highly selected summary.
Distant supervision. Distant supervision [21] generates noisy training data with heuristics to
align unlabeled data with labels, whereas much effort has been devoted to reducing labeling noise.
Multi-instance learning [22, 96–98] creates bags of noisily labeled instances and assumes at least
one instance in each bag is correct, then it uses heuristics or auxiliary classifiers to select the correct
labels. However, such instance bags may not exist in a general supervised setting. Reinforcement
learning [86, 99, 100] and curricular learning [84, 89] methods use a clean validation set to obtain an
auxiliary model for noise filtering, while constructing a perfectly labeled validation set is expensive.
Our framework can learn noise-robust IE models without extra learning resources and can be easily
incorporated into existing supervised IE models.
Supervised learning with noisy labels. A deep neural network can memorize noisy labels, and
its generalizability will severely degrade when trained with noisy labels [76]. In computer vision,
much investigation has been conducted for supervised image classification with noise, producing
techniques such as robust loss functions [101, 102], noise filtering layers [103, 104], label re-
weighting [105, 106], robust regularization [61, 107, 108], and sample selection [89, 90, 109–111].
The robust loss functions and noise filtering layers require modifying model structures and may
not be easily adapted to IE models. The sample selection methods assume the data instances with
large training losses to be noisy and exclude them from training. However, some clean instances,
especially those from long-tail classes, can also have a large training loss and be wrongly pruned,
leading to propagated errors.
32
In NLP, few efforts have focused on learning with denoising. CrossWeigh [87], one label
re-weighting method, partitions the training data into multiple folds and trains multiple models
on each fold. Instances on which models disagree are regarded as noisy and down-weighted in
training. However, this method requires training many models and is computationally expensive.
Our framework only requires training several models concurrently, which is more computationally
efficient and achieves better performance. NetAb [112] assumes that noisy labels are created from
randomly flipping clean labels and uses a CNN to model the noise transition matrix [113]. However,
this assumption does not hold for real datasets, where the noise rate vary among data instances [114].
3.6 Conclusion
This chapter presents a co-regularization framework for learning supervised IE models from
noisy data. This framework consists of two or more identically structured models with different
initialization, which are encouraged to give similar predictions on the same inputs by optimizing
an agreement loss. On noisy examples where model predictions usually differ from the labels, the
agreement loss prevents the model from overfitting noisy labels. Experiments on NER and RE
benchmarks show that our framework yields promising improvements on various IE models. For
future work, we plan to extend the use of the proposed framework to other tasks such as event-centric
IE [115] and co-reference resolution [116].
33
Chapter 4
Out-of-distribution Detection
Pretrained Transformers achieve remarkable performance when training and test data are from the
same distribution. However, in real-world scenarios, the model often faces out-of-distribution (OOD)
instances that can cause severe semantic shift problems at inference time. Therefore, in practice, a
reliable model should identify such instances, and then either reject them during inference or pass
them over to models that handle another distribution. We develop an unsupervised OOD detection
method, in which only the in-distribution (ID) data are used in training. We propose to fine-tune
the Transformers with a contrastive loss, which improves the compactness of representations, such
that OOD instances can be better differentiated from ID ones. These OOD instances can then be
accurately detected using the Mahalanobis distance in the model’s penultimate layer. We experiment
with comprehensive settings and achieve near-perfect OOD detection performance, outperforming
baselines drastically. We further investigate the rationales behind the improvement, finding that
more compact representations through margin-based contrastive learning bring improvement.
4.1 Introduction
Many natural language classifiers are developed based on a closed-world assumption, i.e., the
training and test data are sampled from the same distribution. However, training data can rarely
capture the entire distribution. In real-world scenarios, out-of-distribution (OOD) instances, which
come from categories that are not known to the model, can often be present in inference phases.
These instances could be misclassified by the model into known categories with high confidence,
causing the semantic shift problem [117]. As a practical solution to this problem in real-world
applications, the model should detect such instances, and signal exceptions or transmit to models
handling other categories or tasks. Although pretrained Transformers [13] achieve remarkable
results when intrinsically evaluated on in-distribution (ID) data, recent work [1] shows that many of
these models fall short of detecting OOD instances.
Despite the importance, few attempts have been made for the problem of detecting OOD in NLP
tasks. One proposed method is to train a model on both the ID and OOD data and regularize the
model to produce lower confidence on OOD instances than ID ones [118, 119]. However, as the
OOD instances reside in an unbounded feature space, their distribution during inference is usually
unknown. Hence, it is hard to decide which OOD instances to use in training, let alone that they
may not be available in lots of scenarios. Another practiced method for OOD detection is to use the
maximum class probability as an indicator [1, 120], such that lower values indicate more probable
34
OOD instances. Though easy to implement, its OOD detection performance is far from perfection,
as prior studies [121, 122] show that OOD inputs can often get high probabilities as well.
We aim at improving the OOD detection ability of natural language classifiers, in particular,
the pretrained Transformers, which have been the backbones of many SOTA NLP systems. For
practical purposes, we adopt the setting where only ID data are available during task-specific
training. Moreover, we require that the model should maintain classification performance on the ID
task data. To this end, we propose a contrastive learning framework for unsupervised OOD detection,
which is composed of a contrastive loss and an OOD scoring function. Our contrastive loss aims
at increasing the discrepancy of the representations of instances from different classes in the task.
During training, instances belonging to the same class are regarded as pseudo-ID data while those
of different classes are considered mutually pseudo-OOD data. We hypothesize that increasing
inter-class discrepancies can help the model learn discriminative features for ID/OOD distinctions,
and therefore help detect true OOD data at inference. We study two versions of the contrastive
loss: a similarity-based contrastive loss [123–125] and a margin-based contrastive loss. The OOD
scoring function maps the representations of instances to OOD detection scores, indicating the
likelihood of an instance being OOD. We examine different combinations of contrastive losses
and OOD scoring functions, including maximum softmax probability, energy score, Mahalanobis
distance, and maximum cosine similarity. Particularly, we observe that OOD scoring based on the
Mahalanobis distance [126], when incorporated with the margin-based contrastive loss, generally
leads to the best OOD detection performance. The Mahalanobis distance is computed from the
penultimate layer
1
of Transformers by fitting a class-conditional multivariate Gaussian distribution.
The main contributions of this work are three-fold. First, we propose a contrastive learning
framework for unsupervised OOD detection, where we comprehensively study combinations of
different contrastive learning losses and OOD scoring functions. Second, extensive experiments
on various tasks and datasets demonstrate the significant improvement our method has made to
OOD detection for Transformers. Third, we provide a detailed analysis to reveal the importance of
different incorporated techniques, which also identifies further challenges for this emerging research
topic.
4.2 Related Work
Out-of-Distribution Detection. Determining whether an instance is OOD is critical for the safe
deployment of machine learning systems in the real world [127]. The main challenge is that the
distribution of OOD data is hard to estimate a priori. Based on the availability of OOD data, recent
methods can be categorized into supervised, self-supervised, and unsupervised ones. Supervised
methods train models on both ID and OOD data, where the models are expected to output a uniform
distribution over known classes on OOD data [118, 121, 128]. However, it is hard to assume the
presence of a large dataset that provides comprehensive coverage for OOD instances in practice.
Self-supervised methods [129] apply augmentation techniques to change certain properties of data
(e.g., through rotation of an image) and simultaneously learn an auxiliary model to predict the
property changes (e.g., the rotation angle). Such an auxiliary model is expected to have worse
generalization on OOD data which can in turn be identified by a larger loss. However, it is hard to
1
I.e., the input to the softmax layer.
35
define such transformations for natural language. Unsupervised methods use only ID data in training.
They detect OOD data based on the class probabilities [120, 122, 130, 131] or other latent space
metrics [126, 132]. Particularly, Vyas, Jammalamadaka, Zhu, Das, Kaul & Willke [133] randomly
split the training classes into two subsets and treat them as pseudo-ID and pseudo-OOD data,
respectively. They then train an OOD detector that requires the entropy of probability distribution
on pseudo-OOD data to be lower than pseudo-ID data. This process is repeated to obtain multiple
OOD detectors, and their ensemble is used to detect the OOD instances. This method conducts
OOD detection at the cost of high computational overhead in training redundant models and has the
limitation of not supporting the detection for binary classification tasks.
Though extensively studied for computer vision (CV), OOD detection has been overlooked in
NLP, and most prior works [118, 134, 135] require both ID and OOD data in training. Hendrycks,
Liu, Wallace, Dziedzic, Krishnan & Song [1] use the maximum softmax probability as the detection
score and show that pretrained Transformers exhibit better OOD detection performance than models
such as LSTM [36], while the performance is still imperfect. Our framework, as an unsupervised
OOD detection approach, significantly improves the OOD detection of Transformers only using ID
data.
Contrastive Learning. Recently, contrastive learning has received a lot of research attention. It
works by mapping instances of the same class into a nearby region and make instances of different
classes uniformly distributed [136]. Many efforts on CV [125, 137, 138] and NLP [139] incorporate
contrastive learning into self-supervised learning, which seeks to gather the representations of
different augmented views of the same instance and separate those of different instances. Prior work
on image classification [140, 141] shows that model trained with self-supervised contrastive learning
generates discriminative features for detecting distributional shifts. However, such methods heavily
rely on data augmentation of instances and are hard to be applied to NLP. Other efforts on CV [142]
and NLP [143] conduct contrastive learning in a supervised manner, which aims at embedding
instances of the same class closer and separating different classes. They show that models trained
with supervised contrastive learning exhibit better classification performance. To the best of our
knowledge, we are the first to introduce supervised contrastive learning to OOD detection. Such
a method does not rely on data augmentation, thus can be easily adapted to existing NLP models.
We also propose a margin-based contrastive objective that greatly outperforms standard supervised
contrastive losses.
4.3 Method
In this section, we first formally define the OOD detection problem (Section 4.3.1), then introduce
the overall framework (Section 4.3.2), and finally present the contrastive representation learning
and scoring functions (Section 4.3.3 and Section 4.3.4).
4.3.1 Problem Definition
We aim at improving the OOD detection performance of natural language classifiers that are
based on pretrained Transformers, using only ID data in the main-task training. Generally, the
out-of-distribution (OOD) instances can be defined as instances (x x x,y) sampled from an underlying
distribution other than the training distribution P(X
train
,Y
train
), whereX
train
andY
train
are the
36
Algorithm 2: Learning Process
Input: ID training setD
train
and ID validation setD
val
.
Output: A trained classifier and an OOD detector.
Initialize the pretrained Transformer M.
for t= 1...T do
Sample a batch fromD
train
.
Calculate the classification loss L
ce
.
Calculate the contrastive lossL
cont
as eitherL
scl
orL
margin
L =L
ce
+λL
cont
.
Update model parameters w.r.t.L .
if t % evaluation steps= 0 then
Fit the OOD detector onD
val
.
Evaluate both the classifier and OOD detector on D
val
.
Return the best model checkpoint.
training corpus and training label set, respectively. In this context, literature further divides OOD
data into those with semantic shift or non-semantic shift [117]. Semantic shift refers to the instances
that do not belong toY
train
. More specifically, instances with semantic shift may come from
unknown categories or irrelevant tasks. Therefore, the model is expected to detect and reject such
instances (or forward them to models handling other tasks), instead of mistakenly classifying them
intoY
train
. Non-semantic shift, on the other hand, refers to the instances that belong toY
train
but
are sampled from a distribution other thanX
train
, e.g., a different corpus. Though drawn from
OOD, those instances can be classified into Y
train
, thus can be accepted by the model. Hence, in
the context of this thesis, we primarily consider an instance (x x x,y) to be OOD if y / ∈Y
train
, i.e.,
exhibiting semantic shift, to be consistent with the problem settings of prior studies [1, 126, 131].
We hereby formally define the OOD detection task. Specifically, given a main task of natural
language classification (e.g., sentence classification, NLI, etc.), for an instance x x x to be classified,
our goal is to develop an auxiliary OOD scoring function f(x x x) :X →R. This function should
return a low score for an ID instance where y∈Y
train
, and a high score for an OOD instance where
y / ∈Y
train
(y is the underlying label for x x x and is unknown at inference). During inference, we can set
a threshold for the OOD score to filter out most OOD instances. This process involves a trade-off
between false negative and false positive and may be specific to the application. Meanwhile, we
expect that the OOD detection auxiliary should not negatively affect the performance of the main
task on ID data.
4.3.2 Framework Overview
Next, we introduce the formation of our contrastive learning framework for OOD detection. We
decompose OOD detection into two steps. The first step is contrastive representation learning,
where we focus on learning a representation spaceH where the distribution of ID and that of OOD
data are distinct. Accordingly, we need another function to map the representation to an OOD score.
This process is equivalent to expressing OOD detection as f(x x x)= g(h h h), where h h h∈H is the dense
representation of the input text x x x given by an encoder, g :H →R is a scoring function mapping the
representation to an OOD detection score. Using this decomposition, we can use different training
strategies for h h h and different functions for g, which are studies in the following sections.
37
AAACAHicbVA9SwNBEN2LXzF+nVpY2BwGwSrciahl0MbCIoL5gCSEvc0kWbK3d+zOieG4xr9iY6GIrT/Dzn/jXnKFJj4YeLw3w8w8PxJco+t+W4Wl5ZXVteJ6aWNza3vH3t1r6DBWDOosFKFq+VSD4BLqyFFAK1JAA19A0x9fZ37zAZTmobzHSQTdgA4lH3BG0Ug9+6ATUBwxKpLbtJd0EB4xYZCmPbvsVtwpnEXi5aRMctR69lenH7I4AIlMUK3bnhthN6EKOROQljqxhoiyMR1C21BJA9DdZPpA6hwbpe8MQmVKojNVf08kNNB6EvimMztXz3uZ+J/XjnFw2U24jGIEyWaLBrFwMHSyNJw+V8BQTAyhTHFzq8NGVFGGJrOSCcGbf3mRNE4r3nnFuzsrV6/yOIrkkByRE+KRC1IlN6RG6oSRlDyTV/JmPVkv1rv1MWstWPnMPvkD6/MH8u6XSQ==
L
ce
AAACGnicbVBNS8NAEN3Ur1q/qh69BIsgCCURUY9FLx48VLAf0JSy2U7bpZtN2J2IJeR3ePGvePGgiDfx4r9x2+agbR8MPN6bYWaeHwmu0XF+rNzS8srqWn69sLG5tb1T3N2r6zBWDGosFKFq+lSD4BJqyFFAM1JAA19Awx9ej/3GAyjNQ3mPowjaAe1L3uOMopE6RdcLKA4YFclt2kk8hEdMGKTpySI9lJimnWLJKTsT2PPEzUiJZKh2il9eN2RxABKZoFq3XCfCdkIVciYgLXixhoiyIe1Dy1BJA9DtZPJaah8ZpWv3QmVKoj1R/04kNNB6FPimc3ywnvXG4iKvFWPvsp1wGcUIkk0X9WJhY2iPc7K7XAFDMTKEMsXNrTYbUEUZmjQLJgR39uV5Uj8tu+dl9+6sVLnK4siTA3JIjolLLkiF3JAqqRFGnsgLeSPv1rP1an1Yn9PWnJXN7JN/sL5/AZHxoxY=
L
ce
+L
cont
AAACA3icbVBNS8NAEN3Ur1q/ot70EiyCp5KIqMeiF48V7Ac0IWy2m3bpZhN2J2IJAS/+FS8eFPHqn/Dmv3HT5qCtDwYe780wMy9IOFNg299GZWl5ZXWtul7b2Nza3jF39zoqTiWhbRLzWPYCrChngraBAae9RFIcBZx2g/F14XfvqVQsFncwSagX4aFgISMYtOSbB26EYUQwz3q5n7lAHyADiZnIc9+s2w17CmuROCWpoxIt3/xyBzFJIyqAcKxU37ET8DIsgRFO85qbKppgMsZD2tdU4IgqL5v+kFvHWhlYYSx1CbCm6u+JDEdKTaJAdxYXq3mvEP/z+imEl17GRJICFWS2KEy5BbFVBGINmKQE+EQTTCTTt1pkhCUmoGOr6RCc+ZcXSee04Zw3nNuzevOqjKOKDtEROkEOukBNdINaqI0IekTP6BW9GU/Gi/FufMxaK0Y5s4/+wPj8AZpgmMk=
X
train
AAACAXicbVDLSsNAFJ34rPUVdSO4GSyCq5KIqMuiLty1gn1AE8JkOmmHTh7M3IglxI2/4saFIm79C3f+jdM2C209cOFwzr3ce4+fCK7Asr6NhcWl5ZXV0lp5fWNza9vc2W2pOJWUNWksYtnxiWKCR6wJHATrJJKR0Bes7Q+vxn77nknF4+gORglzQ9KPeMApAS155r4TEhhQIrJO7mUOsAfI6vXrPPfMilW1JsDzxC5IBRVoeOaX04tpGrIIqCBKdW0rATcjEjgVLC87qWIJoUPSZ11NIxIy5WaTD3J8pJUeDmKpKwI8UX9PZCRUahT6unN8r5r1xuJ/XjeF4MLNeJSkwCI6XRSkAkOMx3HgHpeMghhpQqjk+lZMB0QSCjq0sg7Bnn15nrROqvZZ1b49rdQuizhK6AAdomNko3NUQzeogZqIokf0jF7Rm/FkvBjvxse0dcEoZvbQHxifP1sal3k=
X
OOD
Figure 4.1: Illustration of our proposed contrastive loss. The contrastive loss seeks to increase
the discrepancy of the representations for instances from different training classes, such that OOD
instances from unknown classes can be better differentiated.
The learning process of our framework is described in Algorithm 2. In the training phase, our
framework takes training and validation datasets that are both ID as input. The model is optimized
with both the (main task) classification loss and the contrastive loss on batches sampled from
ID training data. The best model is selected based on the ID validation data. Specifically, for a
distribution-based OOD scoring function such as the Mahalanobis distance, we first need to fit the
OOD detector on the ID validation data. We then evaluate the trained model on the ID validation
data, where a satisfactory model should have a low contrastive loss and preserve the classification
performance. In the end, our framework returns a classifier to handle the main task on ID data and
an OOD detector to identify OOD instances at inference.
4.3.3 Contrastive Representation Learning
In this section, we discuss how to learn distinctive representations for OOD detection. For better
OOD detection performance, the representation spaceH is supposed to minimize the overlap of
the representations of ID and OOD data. In a supervised setting where both ID and OOD data are
available in training, it would be easy to obtain suchH . For example, Dhamija, G¨ unther & Boult
[121] train the neural model on both ID and OOD data and require the magnitude of representations
of OOD instances to be smaller than ID representations. However, in real-world applications, the
distribution of OOD data is usually unknown beforehand. We thus tackle a more general problem
setting where the OOD data are assumed unavailable in training (unsupervised OOD detection,
introduced below).
In this unsupervised setup, though all training data used are ID, they may belong to different
classes. We leverage data of distinct classes to learn more discriminative features. Through a
contrastive learning objective, instances of the same class form compact clusters, while instances of
different classes are encouraged to live apart from each other beyond a certain margin, as illustrated
in Figure 4.1. The discriminative feature space is generalizable to OOD data, which ultimately leads
to better OOD detection performance in inference when encountering an unknown distribution. We
38
realize such a strategy using two alternatives of contrastive losses, i.e., the supervised contrastive
loss and the margin-based contrastive loss.
Supervised Contrastive Loss. Different from the contrastive loss used in self-supervised represen-
tation learning [125, 138] that compares augmented instances to other instances, our contrastive
loss contrasts instances to those from different ID classes. To give a more specific illustration of our
technique, we first consider the supervised contrastive loss [142, 143]. Specifically, for a multi-class
classification problem with C classes, given a batch of training instances{(x x x
i
,y
i
)}
M
i=1
, where x x x
i
is
the input text, y
i
is the ground-truth label, the supervised contrastive loss can be formulated as:
L
scl
=
M
∑
i=1
− 1
M|P(i)|
∑
p∈P(i)
log
e
z z z
⊺
i
z z z
p
/τ
∑
a∈A(i)
e
z z z
⊺
i
z z z
a
/τ
,
where A(i)={1,...,M}\{i} is the set of all anchor instances, P(i)={p∈ A(i) : y
i
= y
p
} is the
set of anchor instances from the same class as i, τ is a temperature hyper-parameter, z z z is the
L2-normalized[CLS] embedding before the softmax layer [142, 143]. The L2 normalization is for
avoiding huge values in the dot product, which may lead to unstable updates. In this case, this loss
is optimized to increase the cosine similarity of instance pairs if they are from the same class and
decrease it otherwise.
Margin-based Contrastive Loss. The supervised contrastive loss produces minimal gradients
when the similarity difference of positive and negative instances exceeds a certain point. However,
to better separate OOD instances, it is beneficial to enlarge the discrepancy between classes as much
as possible. Therefore, we propose another margin-based contrastive loss. It encourages the L2
distances of instances from the same class to be as small as possible, forming compact clusters,
and the L2 distances of instances from different classes to be larger than a margin. Our loss is
formulated as:
L
pos
=
M
∑
i=1
1
|P(i)|
∑
p∈P(i)
∥h h h
i
− h h h
p
∥
2
,
L
neg
=
M
∑
i=1
1
|N(i)|
∑
n∈N(i)
(ξ−∥ h h h
i
− h h h
n
∥
2
)
+
,
L
margin
=
1
dM

L
pos
+L
neg

.
Here N(i)={n∈ A(i) : y
i
̸= y
n
} is the set of anchor instances from other classes than y
i
, h h h∈R
d
is the unnormalized[CLS] embedding before the softmax layer,ξ is a margin, d is the number of
dimensions of h h h. As we do not use OOD data in training, it is hard to properly tune the margin.
Hence, we further incorporate an adaptive margin. Intuitively, distances between instances from the
same class should be smaller than those from different classes. Therefore, we define the margin as
the maximum distance between pairs of instances from the same class in the batch:
ξ =
M
max
i=1
max
p∈P(i)
∥h h h
i
− h h h
p
∥
2
.
39
We evaluate both contrastive losses in experiments. In training, the model is jointly optimized with
the cross-entropy classification loss L
ce
and the contrastive lossL
cont
:
L =L
ce
+λL
cont
,
where λ is a positive coefficient. We tune λ based on the contrastive loss and the classification
performance on the ID validation set, where a selected value for λ should achieve a smaller
contrastive loss while maintaining the classification performance.
4.3.4 OOD Scoring Functions
Next, we introduce the modeling of the OOD scoring function g. The goal of the scoring function g
is to map the representations of instances to OOD detection scores, where higher scores indicate
higher likelihoods for being OOD. In the following, we describe several choices of this scoring
function.
Maximum Softmax Probability (MSP). Hendrycks & Gimpel [131] use the maximum class
probability 1− max
C
j=1
p p p
j
among C training classes in the softmax layer as an OOD indicator. This
method has been widely adopted as a baseline for OOD detection [1, 117, 129, 131].
Energy Score (Energy). Liu, Wang, Owens & Li [132] interpret the softmax function as the ratio
of the joint probability inX × Y to the probability inX , and estimates the probability density of
inputs as:
g=− log
C
∑
j=1
exp(w w w
⊺
j
h h h),
where w w w
j
∈R
d
is the weight of the j
th
class in the softmax layer, h h h is the input to the softmax layer.
A higher g means lower probability density in ID classes and thus implies higher OOD likelihood.
Mahalanobis Distance (Maha). Lee, Lee, Lee & Shin [126] model the ID features with class-
conditional multivariate Gaussian distributions. It first fits the Gaussian distributions on the ID
validation setD
val
={(x x x
i
,y
i
)}
M
i=1
using the input representation h h h in the penultimate layer of model:
µ µ µ
j
=E
y
i
= j
[h h h
i
], j= 1,...,C,
Σ Σ Σ=E
h
h h h
i
− µ µ µ
y
i

h h h
i
− µ µ µ
y
i

⊺
i
,
where C is the number of classes, µ µ µ
j
is the mean vector of classes, andΣ Σ Σ is a shared covariance
matrix of all classes. Then, given an instance x x x during inference, it calculates the OOD detection
score as the minimum Mahalanobis distance among the C classes:
g=− C
min
j=1
(h h h− µ µ µ
j
)
⊺
Σ Σ Σ
+
(h h h− µ µ µ
j
),
whereΣ Σ Σ
+
is the pseudo-inverse ofΣ Σ Σ. The Mahalanobis distance calculates the probability density
of h h h in the Gaussian distribution.
40
Cosine Similarity can also be incorporated to consider the angular similarity of input representations.
To do so, the scoring function returns the OOD score as the maximum cosine similarity of h h h to
instances of the ID validation set:
g=− M
max
i=1
cos(h h h,h h h
(val)
i
).
The above OOD scoring functions, combined with options of contrastive losses, lead to different
variants of our framework. We evaluate each combination in experiments.
4.4 Experiments
This section presents experimental evaluations of the proposed OOD detection framework. We start
by describing experimental datasets and settings (Section 4.4.1 and 4.4.2), followed by detailed
results analysis and case studies (Section 4.4.3 to 4.4.5).
4.4.1 Datasets
Previous studies on OOD detection mostly focus on image classification, while few have been made
on natural language. Currently, there still lacks a well-established benchmark for OOD detection in
NLP. Therefore, we extend the selected datasets by Hendrycks, Liu, Wallace, Dziedzic, Krishnan &
Song [1] and propose a more extensive benchmark, where we use different pairs of NLP datasets as
ID and OOD data, respectively. The criterion for dataset selection is that the OOD instances should
not belong to ID classes. To ensure this, we refer to the label descriptions in datasets and manually
inspect samples of instances.
We use the following datasets as alternatives of ID data that correspond to three natural language
classification tasks:
• Sentiment Analysis. We include two datasets for this task. SST2 [144] and IMDB [145]
are both datasets for sentiment analysis, where the polarities of sentences are labeled either
positive or negative. For SST2, the train/validation/test splits are provided in the dataset. For
IMDB, we randomly sample 10% of the training instances as the validation set. Note that
both datasets belong to the same task and are not considered OOD to each other.
• Topic Classification. We use 20 Newsgroup [146], a dataset for topic classification con-
taining 20 classes. We randomly divide the whole dataset into an 80/10/10 split as the
train/validation/test set.
• Question Classification. TREC-10 [147] classifies questions based on the types of their
sought-after answers. We use its coarse version with 6 classes and randomly sample 10% of
the training instances as the validation set.
Moreover, for the above three tasks, any pair of datasets for different tasks can be regarded as
OOD to each other. Besides, following Hendrycks, Liu, Wallace, Dziedzic, Krishnan & Song [1],
we also select four additional datasets solely as the OOD data: concatenations of the premises and
respective hypotheses from two NLI datasets RTE [148–151] and MNLI [152], the English source
41
Dataset # train # dev # test # class
SST2 67349 872 1821 2
IMDB 22500 2500 25000 2
TREC-10 4907 545 500 6
20NG 15056 1876 1896 20
MNLI - - 19643 -
RTE - - 3000 -
Multi30K - - 2532 -
WMT16 - - 2999 -
Table 4.1: Statistics of the datasets.
AUROC↑ / FAR95↓ Avg SST2 IMDB TREC-10 20NG
w/oL
cont
MSP 94.1 / 35.0 88.9 / 61.3 94.7 / 40.6 98.1 / 7.6 94.6 / 30.5
Energy 94.0 / 34.7 87.7 / 63.2 93.9 / 49.5 98.0 / 10.4 96.5 / 15.8
Maha 98.5 / 7.3 96.9 / 18.3 99.8 / 0.7 99.0 / 2.7 98.3 / 7.3
Cosine 98.2 / 9.7 96.2 / 23.6 99.4 / 2.1 99.2 / 2.3 97.8 / 10.7
w/L
scl
L
scl
+ MSP 90.4 / 46.3 89.7 / 59.9 93.5 / 48.6 90.2 / 36.4 88.1 / 39.2
L
scl
+ Energy 90.5 / 43.5 88.5 / 64.7 92.8 / 50.4 90.3 / 32.2 90.2 / 26.8
L
scl
+ Maha 98.3 / 10.5 96.4 / 26.6 99.6 / 2.0 99.2 / 1.9 97.9 / 11.6
L
scl
+ Cosine 97.7 / 13.0 95.9 / 28.2 99.2 / 4.2 99.0 / 2.4 96.8 / 17.0
w/L
margin
L
margin
+ MSP 93.0 / 33.7 89.7 / 49.2 93.9 / 46.3 97.6 / 6.5 90.9 / 32.6
L
margin
+ Energy 93.9 / 31.0 89.6 / 48.8 93.4 / 52.1 98.4 / 4.6 94.1 / 18.6
L
margin
+ Maha 99.5 / 1.7 99.9 / 0.6 100 / 0 99.3 / 0.4 98.9 / 6.0
L
margin
+ Cosine 99.0 / 3.8 99.6 / 1.7 99.9 / 0.2 99.0 / 1.5 97.4 / 11.8
Table 4.2: OOD detection performance (in %) of RoBERTa
LARGE
trained on the four ID datasets.
Due to space limits, for each of the four training ID dataset, we report the macro average of AUROC
and FAR95 on all OOD datasets (check Appendix for full results). Results where the contrastive loss
improves OOD detection on both evaluation metrics are highlighted in green .“w/oL
cont
+MSP”
thereof is the method in Hendrycks, Liu, Wallace, Dziedzic, Krishnan & Song [1].
side of Machine Translation (MT) datasets English-German WMT16 [153] and Multi30K [154].
We take the test splits in those datasets as OOD instances in testing. Particularly, for MNLI, we
use both the matched and mismatched test sets. For Multi30K, we use the union of the flickr 2016
English test set, mscoco 2017 English test set, and filckr 2018 English test set as the test set. There
are several reasons for not using them as ID data: (1) WMT16 and Multi30K are MT datasets and
do not apply to a natural language classification problem. Therefore, we cannot train a classifier on
these two datasets. (2) The instances in NLI datasets are labeled either as entailment/non-entailment
for RTE or entailment/neural/contradiction for MNLI, which comprehensively covers all possible
relationships of two sentences. Therefore, it is hard to determine OOD instances for NLI datasets.
The statistics of the datasets are shown in Table 4.1.
42
L
c o n t
L
s c l
L
m a r g L n
AAACCHicbVC7TgMxEPSFVwivA0oKLBIkqnCXAigjaCgogkQeUhJFPsdJrPjsk70HRKcrafgVGgoQouUT6PgbnEcBgZFWGs3sancniAQ34HlfTmZhcWl5JbuaW1vf2Nxyt3dqRsWasipVQulGQAwTXLIqcBCsEWlGwkCwejC8GPv1W6YNV/IGRhFrh6QveY9TAlbquPt3xwoXWiGBASUiuUo7SQvYPSRUSUjTQsfNe0VvAvyX+DOSRzNUOu5nq6toHDIJVBBjmr4XQTshGjgVLM21YsMiQoekz5qWShIy004mj6T40Cpd3FPalgQ8UX9OJCQ0ZhQGtnN8sZn3xuJ/XjOG3lk74TKKgUk6XdSLBQaFx6ngLteMghhZQqjm9lZMB0QTCja7nA3Bn3/5L6mViv5J0b8u5cvnsziyaA8doCPko1NURpeogqqIogf0hF7Qq/PoPDtvzvu0NePMZnbRLzgf3whVmgA=
w/oL
cont
AAACBnicbVC7TsNAEDyHVwgvAyVCOpEgUQU7BVBG0FBQBIk8pDiKzpdzcsr5obs1EFmuaPgVGgoQouUb6PgbzokLCIy00mhmV7s7biS4Asv6MgoLi0vLK8XV0tr6xuaWub3TUmEsKWvSUISy4xLFBA9YEzgI1okkI74rWNsdX2R++5ZJxcPgBiYR6/lkGHCPUwJa6pv7d8e44vgERpSI5CrtJw6we0gUFWla6Ztlq2pNgf8SOydllKPRNz+dQUhjnwVABVGqa1sR9BIigVPB0pITKxYROiZD1tU0ID5TvWT6RooPtTLAXih1BYCn6s+JhPhKTXxXd2YHq3kvE//zujF4Z72EB1EMLKCzRV4sMIQ4ywQPuGQUxEQTQiXXt2I6IpJQ0MmVdAj2/Mt/SatWtU+q9nWtXD/P4yiiPXSAjpCNTlEdXaIGaiKKHtATekGvxqPxbLwZ77PWgpHP7KJfMD6+AVCVmQs=
w/L
scl
AAACCXicbVC7TsNAEDyHVwivACXNiQSJKtgpgDKChoIiSOQhJZZ1vlySU85n624NRJZbGn6FhgKEaPkDOv6GS+ICEkZaaTSzq90dPxJcg21/W7ml5ZXVtfx6YWNza3unuLvX1GGsKGvQUISq7RPNBJesARwEa0eKkcAXrOWPLid+644pzUN5C+OIuQEZSN7nlICRvCK+P8HlbkBgSIlIrlMv6QJ7gCQgasBlmpa9Ysmu2FPgReJkpIQy1L3iV7cX0jhgEqggWnccOwI3IQo4FSwtdGPNIkJHZMA6hkoSMO0m009SfGSUHu6HypQEPFV/TyQk0Hoc+KZzcrOe9ybif14nhv65m3AZxcAknS3qxwJDiCex4B5XjIIYG0Ko4uZWTIdEEQomvIIJwZl/eZE0qxXntOLcVEu1iyyOPDpAh+gYOegM1dAVqqMGougRPaNX9GY9WS/Wu/Uxa81Z2cw++gPr8wfChZpl
w/L
margin
Figure 4.2: Visualization of the representations for positive, negative instances in SST2 and OOD
ones. The discrepancy between ID and OOD representations is greater on representations obtained
withL
margin
.
4.4.2 Experimental Settings
Evaluation Protocol. We train the model on the training split of each of the four aforementioned
ID datasets in turn. In the inference phase, the respective test split of that dataset is used as ID test
data, while all the test splits of datasets from other tasks are treated as OOD test data.
We adopt two metrics that are commonly used for measuring OOD detection performance
in machine learning research [126, 131]: (1) AUROC is the area under the receiver operating
characteristic curve, which plots the true positive rate (TPR) against the false positive rate (FPR). A
higher AUROC value indicates better OOD detection performance, and a random guessing detector
corresponds to an AUROC of 50%. (2) FAR95 is the probability for a negative example (OOD)
to be mistakenly classified as positive (ID) when the TPR is 95%, in which case a lower value
indicates better performance. Both metrics are threshold-independent.
Compared Methods. We evaluate all configurations of contrastive losses and OOD scoring
functions. Those include 12 settings composed of 3 alternative setups for contrastive losses (L
scl
,
L
margin
or w/o a contrastive loss) and 4 alternatives of OOD scoring functions (MSP, the energy
score, Maha, or cosine similarity).
Model Configuration. We implement our framework upon Huggingface’s Transformers [155]
and build the text classifier based on RoBERTa
LARGE
[44] in the main experiment. All models are
optimized with Adam [94] using a learning rate of 1e− 5, with a linear learning rate decay towards 0.
We use a batch size of 32 and fine-tune the model for 10 epochs. When training the model on each
training split of a dataset, we use the respective validation split for both hyper-parameter tuning and
The hyper-parameters are tuned according to the classification performance and the contrastive loss
on the ID validation set. We find that τ = 0.3 andλ = 2 work well withL
scl
, whileλ = 2 work
well withL
margin
, and we apply them to all datasets.
4.4.3 Main Results
We hereby discuss the main results of the OOD detection performance. Note that the incorporation
of our OOD techniques does not lead to noticeable interference of the main-task performance, for
which an analysis is later given in Section 4.4.5.
43
AUROC↑ / FAR95↓ TREC-10 20NG
MSP 73.7 / 56.5 76.4 / 80.7
Maha 75.5 / 56.1 77.2 / 74.1
L
margin
+ MSP 64.1 / 66.4 74.6 / 82.0
L
margin
+ Maha 76.6 / 61.3 78.5 / 72.7
Table 4.3: Novel class detection performance.
The OOD detection results by different configurations of models are given in Table 4.2. For all
results, we report the average of 5 runs using different random seeds. Each model configuration is
reported with separate sets of results when being trained on different datasets, on top of which the
macro average performance is also reported. For settings withL
scl
andL
margin
, results better than
the baselines (w/o a contrastive loss) are marked as red. We observe that: (1) Among OOD detection
functions, the Mahalanobis distance performs the best on average and drastically outperforms the
MSP baseline used in Hendrycks, Liu, Wallace, Dziedzic, Krishnan & Song [1]. This is due to that
the Mahalanobis distance can better capture the distributional difference. (2) Considering models
trained on different ID datasets, the model variants withL
margin
have achieved near-perfect OOD
detection performance on SST2, IMDB, and TREC-10. While on the 20 Newsgroup dataset that
contains articles from multiple genres, there is still room for improvement. (3) Overall, The margin-
based contrastive loss (L
margin
) significantly improves OOD detection performance. Particularly, it
performs the best with the Mahalanobis distance, reducing the average FAR95 of Maha by 77%
from 7.3% to 1.7%. (4) The supervised contrastive loss (L
scl
) does not effectively improve OOD
detection in general. In many cases, its performance is even worse than the baseline.
4.4.4 Novel Class Detection
We further evaluate our framework in a more challenging setting of novel class detection. Given
a dataset containing multiple classes (≥ 3), We randomly reserve one class as OOD data while
treating others as ID data. We then train the model on the ID data and require it to identify OOD data
in inference. In this case, the OOD data are sampled from the same task corpus as the ID data, and
thus is much harder to be distinguished. We report the average performance of 5 trials in Table 4.3.
The results are consistent with the main results in general. The Mahalanobis distance consistently
outperforms consistently outperforms MSP, and theL
margin
achieves better performance except for
the FAR95 metric on the TREC-10 dataset. However, the performance gain is notably smaller than
that in the main experiments. Moreover, none of the compared methods achieve an AUROC score
of over 80%. This experiment shows that compared to detecting OOD instances from other tasks,
detecting OOD instances from similar corpora is much more challenging and remains room for
further investigation.
4.4.5 Analysis
Visualization of Representations. To help understand the increased OOD detection performance of
our method, we visualize the penultimate layer of the Transformer trained with different contrastive
44
Accuracy SST2 IMDB TREC-10 20NG
w/oL
cont
96.4 95.3 97.7 93.6
w/L
scl
96.3 95.3 97.4 93.4
w/L
margin
96.3 95.3 97.5 93.9
Table 4.4: Accuracy of the trained classifier.
AUROC↑ / FAR95↓ L1 Cosine L2
MSP 93.6 / 31.1 94.1 / 30.9 92.2 / 32.0
Energy 93.8 / 27.2 94.7 / 26.9 94.4 / 27.5
Maha 99.3 / 2.8 99.2 / 3.0 99.4 / 1.7
Cosine 98.1 / 10.9 98.8 / 5.3 99.0 / 3.9
Table 4.5: Average OOD detection performance of different distance metrics.
AUROC↑ / FAR95↓ Maha Maha +L
margin
BERT
BASE
95.7 / 21.5 98.4 / 8.1
BERT
LARGE
97.7 / 13.3 99.1 / 3.9
RoBERTa
BASE
98.4 / 9.3 99.6 / 2.0
RoBERTa
LARGE
98.5 / 7.3 99.4 / 1.7
Table 4.6: Average OOD detection performance of other pretrained Transformers.
losses. Specifically, we train the model on SST2 and visualize instances from the SST2 validation set
and OOD datasets using t-SNE [156], as shown in Figure 4.2. We observe that the representations
obtained withL
margin
can distinctly separate ID and OOD instances, such that ID and OOD clusters
see almost no overlap.
Main Task Performance. As stated in Section 4.3.1, the increased OOD detection performance
should not interfere with the classification performance on the main task. We evaluate the trained
classifier on the four ID datasets. The results are shown in Table 4.4. We observe that the
contrastive loss does not noticeably decrease the classification performance, nor does it increase the
performance, which differs from the observations by Gunel, Du, Conneau & Stoyanov [143].
Distance Metrics. Besides L2 distance, we further evaluate the L1 distance and the cosine distance
with the margin-based contrastive lossL
margin
. Results are shown in Table 4.5. Due to space
limitations, we only report the average OOD performance on the four ID datasets. We observe that
the three metrics achieve similar performance, and all outperform the baseline when using Maha as
the scoring function. Among them, L2 distance gets slightly better OOD detection performance.
Moreover,L
margin
significantly outperforms L
scl
when both use cosine as the distance metric. It
shows that their performance difference arises from the characteristics of the losses instead of the
metric.
OOD Detection by Other Transformers. We also evaluate the OOD detection ability of other
pretrained Transformers in Table 4.6 and report the average performance on the four ID datasets.
For BERT [13], we useλ = 0.2. We observe that: (1) Larger models have better OOD detection
45
ability. For both BERT and RoBERTa, the large versions offer better results than the base versions.
(2) Pretraining on diverse data improves OOD detection. RoBERTa, which uses more pretraining
corpora, outperforms BERT models. (3) The margin-based contrastive loss consistently improves
OOD detection on all encoders.
4.5 Conclusion
This work presents an unsupervised OOD detection framework for pretrained Transformers requir-
ing only ID data. We systematically investigate the combination of contrastive losses and scoring
functions, the two key components in our framework. In particular, we propose a margin-based
contrastive objective for learning compact representations, which, in combination with the Maha-
lanobis distance, achieves the best performance: near-perfect OOD detection on various tasks and
datasets. We further propose novel class detection as the future challenge for OOD detection.
46
Chapter 5
Resolving Knowledge Conflicts
Large language models (LLMs) encode parametric knowledge about world facts and have shown
remarkable performance in knowledge-driven NLP tasks. However, their reliance on parametric
knowledge may cause them to overlook contextual cues, leading to incorrect predictions in context-
sensitive NLP tasks (e.g., knowledge acquisition tasks). We seek to assess and enhance LLMs’
contextual faithfulness in two aspects: knowledge conflict and prediction with abstention. We
demonstrate that LLMs’ faithfulness can be significantly improved using carefully designed prompt-
ing strategies. In particular, we identify opinion-based prompts and counterfactual demonstrations
as the most effective methods. Opinion-based prompts reframe the context as a narrator’s state-
ment and inquire about the narrator’s opinions, while counterfactual demonstrations use instances
containing false facts to improve faithfulness in knowledge conflict situations. Neither technique
requires additional training. We conduct experiments on three datasets of two standard NLP tasks,
machine reading comprehension and relation extraction, and the results demonstrate significant
improvement in faithfulness to contexts.
5.1 Introduction
Large language models (LLMs; [157–160]) have made remarkable advances in solving various NLP
problems, particularly in (context-free) knowledge-driven tasks such as question answering [161,
162] and commonsense reasoning [163, 164]. Without external context, LLMs can answer factual
questions and achieve comparable results to supervised approaches [157, 158], indicating that LLMs
encode parametric knowledge about open-world facts.
Although parametric knowledge can be beneficial for knowledge-driven tasks, overly relying on
it can cause problems in context-specific NLP tasks. First, LLMs may encode misconceptions [165]
or obsolete facts [166–168], in which case we expect LLMs to update their predictions when
provided with relevant context. Second, when using LLMs for knowledge acquisition tasks such
as machine reading comprehension (MRC; [169, 170]) and information extraction (IE; [19, 79]),
LLMs should always extract the knowledge in context instead of relying solely on their parametric
knowledge. In such context-specific application scenarios, we expect LLMs to make decisions
faithful to the context and avoid simply parroting answers from pretraining. However, studies have
discovered that LLMs can overlook or ignore context [168, 171, 172], posing a significant challenge
for their application in these scenarios.
47
Context: Elon Musk is a business magnate and
investor. He is the owner and CEO of Twitter.
Question: Who is the CEO of Twitter?
Answer: Elon Musk
GPT-3.5: Jack Dorsey
Knowledge Conflict
Context: Bill Gates was born in Seattle,
Washington.
Question: Is Bill Gates the founder of Microsoft?
Answer: I don’t know
GPT-3.5: yes
Prediction with Abstention
Figure 5.1: Examples of knowledge conflict and prediction with abstention. LLMs may ignore the
provided context and make unfaithful predictions based on their parametric knowledge before Q4
2021.
We aim to investigate techniques for improving the faithfulness of LLMs in context-specific
NLP tasks. Conceptually, faithfulness is not simply about how much accuracy the model can
offer. Instead, it should concern the validity and reliability of its extraction process. Specifically,
when there is a concept or relation to extract, a faithful extractor should genuinely induce what is
described in the context but not give trivial guesses based on parametric knowledge or statistical
biases. Besides, when no known decision-related information is described in the context, the
model should selectively abstain from predicting. Accordingly, to provide a realistic assessment
of LLMs in terms of faithfulness, we narrow our focus to two sub-problems, namely entity-based
knowledge conflict [173] and prediction with abstention [174], examples of which are shown
in Figure 5.1. In cases of knowledge conflict, where the given context contains facts different from
the pretraining data, LLMs need to return the facts locally described in the context instead of the
globally memorized facts. For example, in Figure 5.1, text-davinci-003 identifies Jack Dorsey
instead of Elon Musk as the CEO of Twitter, based on its pretrained data before Q4 2021. In cases of
prediction with abstention, where the provided context does not provide information to answer the
questions, LLMs should abstain from making predictions and notify the users, rather than answering
the questions that become a trivial guess. For example, in Figure 5.1, when asked about the founder
of Microsoft based on an irrelevant context, LLMs should admit that, from here, they cannot infer
the answer.
We present various prompting strategies to improve the faithfulness of LLMs, including design-
ing effective prompts and choosing appropriate in-context demonstrations. We find that constraining
the scope of questions to the context by adding phrases (e.g., based on the given context) or natural
language instructions improve faithfulness in both facets. Particularly, we find that reformulating
the context and questions to opinion-based question-answering problems [175, 176], where the
48
context is expressed in terms of a narrator’s statement, and the question asks about this narrator’s
opinion, delivers the most gains. Additionally, we find that adding counterfactual demonstrations to
prompts improves faithfulness in the aspect of knowledge conflict, while using the original (factual)
demonstrations leads to limited or negative effects. Finally, combining both techniques delivers the
largest gain than using each one independently.
We evaluate our methods based on three datasets, including Re-TACRED [177] for relation
extraction, and natural questions [162] and RealTime QA [168] for MRC. We find that the proposed
strategies can greatly improve faithfulness, e.g., reducing the memorization ratio
1
of text-davinci-
003 from 35.2% to 3.0% on natural questions. Additionally, we evaluate our methods across LLMs
of different scales, finding that larger LLMs are more likely to update memorized answers than
smaller ones, both with and without the application of our methods.
5.2 Related Work
Knowledge conflicts. LLMs [157–159] have shown promising results in closed-book QA tasks,
indicating their ability to memorize facts about the world. However, as the world is constantly
evolving, memorized facts may become outdated [166–168], emphasizing the need to update LLMs’
predictions with new facts. To address this challenge, some studies [178–182] have explored ways
to identify and edit the facts stored in model parameters. For example, Meng, Sharma, Andonian,
Belinkov & Bau [182] identify the MLPs that have a causal effect on factual knowledge recall in
LLMs and edit the parameters in these critical MLPs to update the knowledge. Their method can
update up to thousands of memorized facts with little impairment to others. However, it remains
unclear whether memory editing methods allow sufficient capacity to encompass all new factual
knowledge. Another promising direction is to augment LLM prompting with external context
containing relevant knowledge [183–185]. Coupled with retrieval systems [186–188], such methods
have the potential to update LLMs with large amounts of new facts. However, such methods face the
challenge that LLMs may persist with the memorized facts and ignore the provided context [173].
To tackle this challenge, recent works [171, 189] finetune LLMs on counterfactual contexts, where
the original facts are replaced with counterfactual ones. They find that such finetuning processes can
effectively improve the LLMs’ utilization of contexts instead of relying solely on their parametric
knowledge. In this study, we propose a novel approach using prompting to improve context
faithfulness in LLMs without additional finetuning, which offers a more general and cost-effective
method for LLMs.
Prediction with abstention. Selective prediction with abstention [190–192] is an important problem
in trustworthy AI. When models are uncertain about their predictions, it is critical that they should
admit the uncertainty and notify users instead of returning incorrect predictions. Selective prediction
may be adopted in different scenarios, such as being on the model side where instances are close to
the decision boundary [193–195], or on the data side where instances are from different domains
to training [1, 131, 196]. In the scope of context-specific NLP, abstention is preferred when the
context is irrelevant to the question. For example, SQuAD 2.0 [174] introduces unanswerable
questions to extractive MRC, while Yatskar [197] finds it focused on questions of extreme confusion
and thus is less relevant to the focus of our study. CoQA [198] and QuAC [199] introduce
1
The percentage of times that LLMs return memorized answers versus answers in the context.
49
unanswerable questions to conversational question answering. RealTime QA [168] finds that
GPT-3 still generates outdated answers when provided with irrelevant documents. To address the
problem, Neeman, Aharoni, Honovich, Choshen, Szpektor & Abend [189] propose the answerability
augmentation where LLMs should predict Unanswerable when presented with an empty or randomly
sampled document. We tackle this problem with a part of our prompting method, which we find to
significantly enhance the LLMs’ ability to make selective predictions.
5.3 Method
We focus on context-specific NLP tasks. The input of these tasks is formulated as (c,q) for free-form
generation tasks, where c is the context and q is the question, or(c,q,o) for tasks with close decision
spaces (e.g., multi-choice tasks), where o is the set of decisions/choices. The desired output is
either a free-form text or a choice. We solve these tasks by prompting LLMs and study ways of
designing prompting templates and demonstrations that are dedicated to improving the faithfulness
of LLMs. Specifically, we find two proposed methods, opinion-based prompts and counterfactual
demonstrations, to be the most effective ones. Our methods only change the prompts without
finetuning the LLMs [171, 173, 189], targeting a more general and affordable solution.
5.3.1 Opinion-based Prompting
Given an input(c,q,o), we begin with the following base prompting template:
2
Base prompt
{c} Q:{q}? Options:{o} A:
Here, {} serves as a placeholder that is filled with specific inputs during prompting. We
investigate two types of prompting templates for context-specific NLP, namely opinion-based
prompts and instructed prompts. Opinion-based prompts transform original questions into opinion-
seeking questions, which naturally demand more attention to the context. Instructed prompts, on
the other hand, explicitly instruct LLMs to read the context by natural language. Details of these
templates are discussed in the remaining section.
Opinion-based prompts. We propose to transform the context to a narrator’s statement and the
question to enquire about the narrator’s opinion in this statement. Our approach is motivated by our
own cognitive process for answering different types of questions. When answering questions that
seek factual information, we can often rely on our own memory and answer without needing to refer
to the context, as these questions typically have only one correct answer. However, when questions
are seeking opinions from someone else (in this context, the narrator), it is important to comprehend
the narrator’s words before answering the questions, as opinions may vary greatly from person to
person. Besides, as opinions are inherently subjective and can be influenced by many factors such as
personal experiences and beliefs, opinion-seeking questions are sometimes difficult to answer solely
based on the narrator’s statement compared to a fact-seeking question that typically has definite and
verifiable answer(s). As a result, transforming factual questions into opinion-seeking questions can
2
Options only apply for multiple-choice tasks and are removed in free-form text generation tasks.
50
lead to more attention to the context, as memorized answers alone may not suffice. It also helps the
model more selectively predict under cases where contexts do not describe answers. Both factors
lead to improved faithfulness on LLMs. The opinion-based prompting template is as follows:
Opinion-based prompt
Bob said, ”{c}” Q:{q} in Bob’s opinion? Options:{o} A:
Throughout our experiments, we consistently use Bob to represent the narrator for the context,
although other names could be utilized as well.
Instructed prompts. We also explicitly instruct LLMs to read context by natural language. We
start by extending questions in prompts with attributive phrases such as “based on the given text”,
leading to the following attributed prompting template:
Attributed prompt
{c} Q:{q} based on the given text? Options:{o} A:
We also augment the prompts with natural language instructions. Since manually writing
instructions can be laborious and often fails to account for the compatibility between instructions
and LLMs, we leverage automatic prompt engineering (APE; [200]) to generate the prompts. Using
a few instances and their desired outputs as demonstrations, APE uses LLMs to automatically
generate candidate instructions and select the best one based on the results on a dev set. We then
use the following instruction-based prompting template:
Instruction-based prompt
Instruction:{Instruction} {c} Q:{q}? Options:{o} A:
Experiments show that all prompting templates perform better than the base prompting template.
Specifically, opinion-based prompts outperform instructed prompts in both knowledge conflict and
prediction with abstention facets, and combining these two prompting methods results in the most
significant improvements.
5.3.2 Counterfactual Demonstration
Using demonstrations is a standard way to perform few-shot inference on LLMs [157]. To enhance
the faithfulness of LLMs in knowledge conflict scenarios, previous works [171, 189] propose to
finetune the models using counterfactual instances, where the facts in the context are substituted
with false ones, and the model learns to update its predictions accordingly. Following this strategy,
we propose to use counterfactual instances as demonstrations for LLMs. To do this, we start with
a labeled set of counterfactual instances and a test instance and then use KATE [201] to retrieve
the most relevant counterfactual instances as demonstrations. We encode both the test instance and
counterfactual instances with RoBERTa
nli+sts-b
[44, 202] and select the top counterfactual instances
based on cosine similarity. As a part of our analysis, we also experimented with using the original
51
Method
MRC RE
p
s
↑ p
o
↓ M
R
↓ EM↑ p
s
↑ p
o
↓ M
R
↓ F
1
↑
Zero-shot
Base 59.0 32.1 35.2 6.2 73.9 21.5 22.5 81.0
Attr 71.9 14.4 16.6 29.6 72.4 23.6 24.6 80.0
Instr 74.2 16.0 17.7 27.1 75.8 15.6 17.1 81.6
Opin 79.4 9.8 11.0 24.9 76.0 19.6 20.5 82.9
Opin + Instr 79.1 7.9 9.1 48.6 79.4 15.0 15.9 84.7
Original
Base 43.3 49.4 53.3 35.1 76.2 19.8 20.6 83.3
Attr 54.1 37.7 41.0 45.5 76.5 19.7 20.5 83.7
Instr 54.6 37.7 40.8 45.8 77.3 18.4 19.2 84.2
Opin 60.6 28.7 32.1 51.1 76.8 18.4 19.3 83.8
Opin + Instr 64.7 26.8 29.3 53.8 78.2 17.1 17.9 84.9
Counter
Base 86.9 6.5 7.0 80.2 78.7 13.7 14.8 83.9
Attr 89.1 4.6 4.9 83.0 79.7 13.0 14.0 84.3
Instr 86.2 6.3 6.8 80.1 78.0 12.8 14.1 82.9
Opin 90.1 3.7 3.9 84.3 79.7 12.8 13.8 84.4
Opin + Instr 90.9 2.8 3.0 85.2 80.0 10.5 11.6 85.1
Table 5.1: Results (in %) in the knowledge conflict setting. The overall best results are highlighted
in bold. The best and the second best results in each setting are highlighted in green and orange ,
respectively.
(factual) instances as demonstrations but found this approach to underperform counterfactual
demonstrations and sometimes even zero-shot inference.
5.4 Experiments
This section presents our experimental setups (Section 5.4.1) for the evaluation of the proposed
methods concerning two aspects of faithfulness: knowledge conflict (Section 5.4.2) and prediction
with abstention (Section 5.4.3). We provide additional analysis (Section 5.4.4) on results across
different model sizes and results on the original datasets. We also show examples of prompts and
LLMs’ outputs in the case study (Section 5.4.5).
5.4.1 Experimental Setup
Our experiments are conducted using the InstructGPT model (text-davinci-003, 175B parameters).
We use the base prompt as our baseline, and compare it against the proposed prompting templates
in Section 5.3.1, including attributed prompt (ATTR), instruction-based prompt (INSTR), opinion-
based prompt (OPIN), and the combination of opinion-based prompt and instruction-based prompt
(OPIN + INSTR). We evaluate the effectiveness of these templates in both zero-shot and few-shot
settings (with demonstrations).
52
5.4.2 Knowledge Conflict
Datasets. We evaluate in the knowledge conflict setting using counterfactual datasets that contain
incorrect facts, which can conflict with what the LLM has memorized. We use two datasets
based on real-world texts: natural questions [162] for MRC and Re-TACRED [177] for relation
extraction (RE). To create counterfactuals, we adopt the framework proposed by Longpre, Perisetla,
Chen, Ramesh, DuBois & Singh [173], which modifies the context to support a counterfactual
answer. Specifically, for MRC, we follow Longpre, Perisetla, Chen, Ramesh, DuBois & Singh [173]
and replace the gold entity answer in the context with a randomly sampled entity of the same entity
type from the corpus. For RE, we first randomly sample a context that has the entity mentions of
the same type but different relations from the original one, and then insert the original entities into
the sampled context. In this scenario, a faithful LLM should update its prediction to the new answer
instead of returning the original one. Moreover, to measure LLMs’ ability to update answers, we
need to ensure that they have knowledge of the original answers in the first place. Therefore, we
only evaluate LLMs on a subset of instances where they can correctly predict the original answers
without additional contexts. This process left us with 2,773 test instances on natural questions and
4,353 test instances on Re-TACRED.
Task setup. We use the same set of evaluation metrics as Longpre, Perisetla, Chen, Ramesh, DuBois
& Singh [173]. Specifically, we measure the frequency that the LLMs’ predictions contain an
exact match of the original answers (p
o
) and the substituted answers (p
s
), after both predictions
and answers have been normalized by removing stop words and punctuation To assess the model’s
reluctance to update its prediction, we use the memorization ratio (M
R
), which is calculated as
M
R
=
p
o
p
o
+p
s
. A completely faithful LLM should have an M
R
of 0. We also report task-specific
metrics, including exact match (EM) for MRC and F
1
for RE. For EM, we also use normalized
predictions and answers, but the requirement is that the prediction and answer must be exactly the
same, rather than just containing the answer. We conduct experiments in three different settings:
zero-shot, demonstration using original instances, and demonstration using counterfactual instances.
We retrieve demonstrations from the original/counterfactual training set, and evaluate LLMs on
the counterfactual test set. In the few-shot setting, we use 16 demonstration instances. Since
text-davinci-003 has a maximum input length of 4,096 tokens, some instances are filtered out during
evaluation.
Results and discussion. The results in Table 5.1 demonstrate that the combination of OPIN +
INSTR prompting and counterfactual demonstrations is most effective. It results in a reduction
of 32.2% and 10.9% in M
R
for MRC and RE respectively when compared to the zero-shot base
prompts. We also find that opinion-based prompts generally perform better than other templates,
achieving the second-best results on 17 out of 24 metrics, indicating that LLMs are more faithful
to the context when answering opinion-seeking questions. Combining opinion-based prompts and
instruction-based prompts further improves faithfulness, with the best results obtained in 23 out of
24 metrics.
When it comes to few-shot settings, counterfactual demonstrations lead to further improved
performance. Using the original (factual) instances as demonstrations, on the other hand, leads to
limited effects or may even impair faithfulness in MRC. This finding suggests that demonstrations
do not always improve the generalization of LLMs’ inference, especially when they contain dataset
bias. In the MRC experiments, the natural questions dataset used is constructed based on Wikipedia,
which mainly consists of world facts. This potentially allows for a simplicity bias of LLMs where
53
Method
Acc↑ Brier↓
NoAns All All
Zero-shot
Base 30.6 68.5 29.4
Attr 65.3 84.4 14.6
Instr 81.6 91.7 7.7
Opin 83.3 92.6 6.6
Opin + Instr 87.8 94.4 5.2
Few-shot
Base 73.5 88.2 11.2
Attr 81.6 91.9 8.0
Instr 85.7 93.7 6.1
Opin 87.8 94.6 4.1
Opin + Instr 89.8 95.5 3.4
Table 5.2: Results (in %) on RealTime QA. The overall best results are highlighted in bold. The
best and the second best results in each setting are highlighted in green and orange , respectively.
As all prompts achieve perfect accuracy (100%) on the HasAns subset, it is not included in the
table.
questions can be answered without contexts. Therefore, our study suggests the importance of using
counterfactual demonstrations in knowledge conflict scenarios.
5.4.3 Prediction with Abstention
Datasets. As for the second aspect of faithfulness, we evaluate LLMs’ ability to selectively abstain
from making uncertain predictions based on irrelevant context. Since existing datasets such as
SQuAD 2.0 [174] generally contain questions with confusion [197] and are less related to our
problem setting, we curate our own evaluation data based on RealTime QA [168], a dataset that
inquires about novel information from June 2022 onwards. In this formulation, LLMs are presented
with a question and multiple choices, and they need to choose the correct answer based on several
retrieved documents. These documents were obtained using tools like Google custom search and
may not contain the answer to the question. To adapt this dataset to our setting, we added a new
“I don’t know” choice and relabeled the dataset. Instances where the retrieved documents do not
answer the question are relabeled to “I don’t know”. We used questions in the first six weeks of 2022
as the test set and randomly picked three questions of 2023 as demonstration instances. This process
results in a total of 113 test instances, including 63 answerable questions and 50 unanswerable ones.
Task setup. We calculate the probability of a choice as P(choice|prompt) followed by normalization
across all choices.
3
We take the choice with the largest probability as the prediction. We report
accuracy on the entire dataset (All), accuracy on the subset of questions that can be answered based
on retrieved documents (HasAns), and accuracy on questions that cannot be answered based on
retrieved documents (NoAns). The latter two metrics measure LLMs’ ability to extract answers
from context and their ability to abstain from making predictions when the context does not describe
3
We tried to scale the probability of choices for length or unconditional probability as done in Brown, Mann, Ryder,
Subbiah, Kaplan, Dhariwal, et al. [157], whereas we find both lead to worse results on RealTime QA.
54
the answer, respectively. Besides, we use the probability of “I don’t know” as LLM’s probability
estimation of whether the question can be answered. We use the Brier score to evaluate the accuracy
of the estimation, which measures the mean squared difference between the estimation and the true
binary outcome of answerability. We use three demonstrations for each instance in the few-shot
setting. As text-davinci-003 has a maximum input length of 4,096 tokens, some instances are
filtered out during evaluation.
Results and discussion. The results presented in Table 5.2 demonstrate that the OPIN + INSTR
prompt achieves the best results in both the zero-shot and few-shot settings, outperforming base
prompts by 57.2% and 16.3% in accuracy in the NoAns subset, respectively. It also reduces the Brier
score by 24.2% and 7.8% compared to base prompts in the two settings, respectively. The OPIN
prompt is the second best in terms of these metrics. These findings demonstrate that opinion-based
prompts can enhance the LLMs’ ability to make selective predictions. Furthermore, all proposed
prompting templates achieve better results on NoAns instances compared to base prompts, and
maintain perfect accuracy on HasAns subset, indicating their ability to improve LLMs’ selective
prediction without compromising performance on answerable instances. In addition, The use of
demonstrations consistently improves the LLMs’ ability to make selective predictions, as evidenced
by the lower Brier scores in the few-shot setting compared to the zero-shot setting.
5.4.4 Additional Analysis
Memorization by different sizes of LLMs. Figure 5.2 shows the memorization ratio M
R
across
different sizes of InstructGPTs under the zero-shot evaluation of natural questions.
4
Overall, OPIN
+ INSTR consistently outperforms other prompts across different model sizes. In the upper plot,
results are shown for filtered evaluation sets where the corresponding LLMs can correctly predict
the original answers without additional contexts, thereof the size of evaluation sets varies across
different LLMs.
5
We observe that M
R
generally decreases with increased model size, showing
that larger LLMs are better at updating memorized answers based on given contexts in knowledge
conflicts. However, the lower plot reveals that larger LLMs have more severe memorization on
the full (unfiltered) evaluation set. This is because larger LLMs can memorize more answers
than smaller ones, as evidenced by the number of instances in the filtered evaluation set where
larger LLMs have more instances. Our analysis suggests that while larger LLMs are better at
updating memorized answers, they still tend to have more memorization due to the larger number of
memorized answers. Therefore, we need to pay more attention when using larger LLMs in scenarios
with new or potentially conflicting knowledge.
Selective prediction by different sizes of LLMs. Figure 5.3 shows the Brier score across different
sizes of InstructGPTs under the zero-shot evaluation of RealTime QA. On smaller LLMs, opinion-
based prompt achieves similar or even higher Brier score than base prompts, indicating it does
not improve the selective prediction ability of LLMs. We hypothesize that this is because smaller
LLMs have inferior reading comprehension ability, resulting in uncertainty in many instances.
Opinion-based prompts change uncertain predictions of answerable questions to I don’t know,
4
The 0.3B, 1.3B, 6.7B models refer to text-ada-001, text-babbage-001, text-curie-001, respectively. We do not
perform few-shot evaluation as different sizes of LLMs have different maximum input lengths and can take different
numbers of demonstrations, thus hard to be compared to each other.
5
The sizes of the filtered evaluation sets, in the order of increased model sizes, are 121, 132, 756, and 2,773.
55
0
10
20
30
40
M
R
Filtered Evaluation Set
Base
Attr
Instr
Opin
Opin+Instr
0.3B 1.3B 6.7B 175B
Model size
0
10
20
30
40
M
R
Full Evaluation Set
Figure 5.2: Memorization ratios across different sizes of InstructGPTs, evaluated in the zero-shot
setting using natural questions.
56
0.3B 1.3B 6.7B 175B
Model size
0
10
20
30
40
50
Brier
RealTime QA
Base
Attr
Instr
Opin
Opin+Instr
Figure 5.3: Brier scores across different sizes of InstructGPTs, evaluated in the zero-shot setting of
RealTime QA.
Method p
o
↑ EM↑
Zero-shot
Base 92.1 11.1
Opin 91.3 25.2
Opin + Instr 90.5 57.2
Original
Base 93.2 77.8
Opin 92.7 78.7
Opin + Instr 93.9 80.1
Counter
Base 93.6 82.1
Opin 92.8 82.3
Opin + Instr 92.7 82.1
Table 5.3: Results (in %) on the filtered evaluation set of natural questions with original (factual)
contexts and answers.
which could lead to worse results. For other prompting templates, we do not observe a consistent
improvement across different LLMs either. This analysis shows that while the selective prediction
ability can be more easily activated by zero-shot prompting for LLMs such as text-davinci-003,
smaller LLMs may require dedicated adaptations such as calibration and finetuning to activate this
ability.
Results on original datasets. While our main experiments demonstrate the effectiveness of the
proposed methods in resolving knowledge conflicts, LLMs in real-world applications may also
see instances without knowledge conflicts. Therefore, we investigate how our methods affect
57
Knowledge Conflict Prediction with Abstention
Context (Counterfactual passage) The Super Bowl LI Half-
time show took place on February 5, 2017, at
NRG Stadium in Houston, Texas as part of Su-
per Bowl LI. The show was headlined by Bosco,
who performed a medley of her songs, including
newer material from her most recent studio album
Joanne.
Tara Connolly is senior gas campaigner at Global
Witness, an international NGO working towards a
more sustainable, just and equal planet. She has
over a decade of experience in EU energy policy.
The views expressed in this commentary are her
own.
Prompt Instruction: read the given information and answer
the corresponding question.
Instruction: answer a question based on the pro-
vided input-output pairs.
Bob said, “The Super Bowl ... album Joanne.” Bob said, “Tara Connolly ... are her own.”
Q: who performed the halftime show at Super
Bowl 51 in Bob’s opinion based on the given text?
Q: Mo Farah made public that he was trafficked
from which African country to the UK in Bob’s
opinion based on the given text?
Choices: Somaliland; Djibouti; Ethiopia; Soma-
lia; I don’t know
Base Lady Gaga✗ Somalia✗
Attr Lady Gaga✗ Somalia✗
Instr Lady Gaga✗ Somaliland✗
Opin Bosco✓ I don’t know✓
Instr + Opin Bosco✓ I don’t know✓
Answer Bosco I don’t know
Table 5.4: Examples of prompts and LLMs’ corresponding predictions. In the “Prompt” row, we
show and highlight the added parts from different prompting templates including attributed prompts,
instruction-based prompts, and opinion-based prompts.
inference when the memorized answers align with the given contexts. To do so, we evaluate LLMs
on the same set of filtered evaluation set used in the main results section (Section 5.4.2), but we
use the original contexts and answers instead of counterfactual ones. The results in Table 5.3
show that opinion-based prompts yield similar or better results in all settings. Furthermore, using
either counterfactual or original demonstrations does not significantly impact results on the original
(factual) dataset. This analysis reveals that our methods do not impair performance on instances
without knowledge conflicts.
5.4.5 Case Study
Table 5.4 shows some examples of prompts and the corresponding answers generated by text-
davinci-003. The left column of the table presents a knowledge conflict case where the original
answer, Lady Gaga, is replaced with a counterfactual answer, Bosco. When using base prompts,
LLM ignores the context and return the memorized answer Lady Gaga. However, using opinion-
based prompts and their combination with instructions leads to a more faithful response, with the
language model returning Bosco in the given context. The right column presents a scenario where
the retrieved context from Google search is irrelevant to the given question. In such cases, base
prompts still return a choice, leading to a potentially incorrect answer. However, opinion-based
prompts and their combination with instructions can abstain from making predictions and return
58
I don’t know. These examples demonstrate the effectiveness of proposed prompts in generating
context-faithful responses.
5.5 Conclusion
We focus on addressing the faithfulness issue of LLMs in context-specific NLP tasks, particu-
larly in scenarios with knowledge conflict and prediction with abstention. We propose that two
methods, opinion-based prompts and counterfactual demonstrations, are effective in improving
LLMs’ faithfulness to contexts. We evaluate our methods on three datasets of two tasks, namely
machine reading comprehension and relation extraction, and observed significant improvement in
faithfulness to contexts. Future work includes evaluating the effectiveness of proposed methods
on a broader range of NLP tasks such as open-domain QA and summarization, and studying other
techniques to improve faithfulness further.
59
Chapter 6
Dataset Annotation by Neural Rule Grounding
Deep neural models for relation extraction tend to be less reliable when perfectly labeled data is
limited, despite their success in label-sufficient scenarios. Instead of seeking more instance-level
labels from human annotators, here we propose to annotate frequent surface patterns to form
labeling rules. These rules can be automatically mined from large text corpora and generalized
via a soft rule matching mechanism. Prior works use labeling rules in an exact matching fashion,
which inherently limits the coverage of sentence matching and results in the low-recall issue. We
present a neural approach to ground rules for RE, named NERO, which jointly learns a relation
extraction module and a soft matching module. One can employ any neural relation extraction
models as the instantiation for the RE module. The soft matching module learns to match rules with
semantically similar sentences such that raw corpora can be automatically labeled and leveraged by
the RE module (in a much better coverage) as augmented supervision, in addition to the exactly
matched sentences. Extensive experiments and analysis on two public and widely-used datasets
demonstrate the effectiveness of the proposed NERO framework, comparing with both rule-based
and semi-supervised methods. Through user studies, we find that the time efficiency for a human
to annotate rules and sentences are similar (0.30 vs. 0.35 min per label). In particular, NERO’s
performance using 270 rules is comparable to the models trained using 3,000 labeled sentences,
yielding a 9.5x speedup. Moreover, NERO can predict for unseen relations at test time and provide
interpretable predictions.
6.1 Introduction
Relation extraction (RE) plays a key role in information extraction tasks and knowledge base
construction, which aims to identify the relation between two entities in a given sentence. For
example, given the sentence “Bill Gates is the founder of Microsoft” and an entity pair (“Bill Gates”,
“Microsoft”), a relation classifier is supposed to predict the relation of ORG:FOUNDED BY. Recent
advance in neural language processing has shown that neural models [19, 28, 203] gained great
success on this task, yielding state-of-the-art performance when a large amount of well-annotated
sentences are available. However, these supervised learning methods degrade dramatically when
the sentence-level labels are insufficient. This problem is partially solved by the use of distant
supervision based on knowledge bases (KBs) [204, 205]. They usually utilize an existing KB for
automatically annotating an unlabeled corpus with an over-simplified assumption — two entities
co-occurring in a sentence should be labeled as their relations in the KB regardless of their contexts.
60
→ ORG:FOUNDED_BY
(SUBJ-PER, ‘s child, OBJ-PER) → PER:CHILDREN
(SUBJ-PER, is known as, OBJ-PER) → PER:ALTERNATIVE NAMES
Rules
p
1
p
2
p
3
Corpus
s
1
s
2
s
3
Hard Matching
1.0
0.9
0.8
Matching Score
Microsoft was founded by Bill Gates in 1975.
(SUBJ-ORG, was founded by, OBJ-PER)
Microsoft was created by Bill Gates in 1975.
In 1975, Bill Gates launched Microsoft.
Figure 6.1: Current rule-based methods mostly rely on exact/hard matching to raw corpus and
suffer from limited coverage. For example, the rule body of p
3
only matches sentence s
1
but is also
similar to s
2
and s
3
, which express the same relation as p
3
. A “soft rule matching” mechanism is
desirable to make better use of the corpus for label generation.
While distant supervision automates the labeling process, it also introduces noisy labels due to
context-agnostic labeling, which can hardly be eliminated.
In addition to KBs, labeling rules are also important means of representing domain knowl-
edge [83], which can be automatically mined from large corpora [206, 207]. Labeling rules can
be seen as a set of pattern-based heuristics for matching a sentence to a relation. These rules are
much more accurate than using KBs as distant supervision. The traditional way of using such rules
is to perform exact string matching, and a sentence is either able or unable to be matched by a rule.
This kind of hard-matching method inherently limits the generalization of the rules for sentences
with similar semantics but dissimilar words, which consequently causes the low-recall problem and
data-insufficiency for training neural models. For example, in Fig. 6.1, the rule P
3
can only find S
1
as one hard-matched instance for training the model while other sentences (e.g., S
2
and S
3
) should
also be matched for they have a similar meaning.
Many attempts have been made to RE using labeling rules and unlabeled corpora, as summarized
in Fig. 6.2. Rule-based bootstrapping methods [208–210] extract relation instances from the raw
corpus by a pre-defined rule mining function (e.g., TF-IDF, CBOW) and expanding the rule set
in an iterative way to increase the coverage. However, they still make predictions by performing
hard-matching on rules, which is context-agnostic and suffers from the low-recall problem. Also,
their matching function is not learnable and thus has trouble in extracting semantically similar
sentences. Another typical option is data programming [83, 211], which aims to annotate the
corpus using rules with model fitting. It trains a neural RE model using the hard-matched sentences
by rules and then reduces the noise of rules with their proposed algorithms. However, it does
not consider the massive data that fail to be annotated by hard-matching. Self-training [212], as
a semi-supervised framework, does attempt to utilize unmatched sentences by using confident
predictions of a learned model. Then, they train the model again and again by iteratively generating
more confident predictions. However, it does not explicitly model the soft matching ability of rules
over unmatched sentences, making the generated labels noisy and unreliable.
61
(A) Bootstrapping (B) Data Programming

Rules
Corpus
Hard-Matching
Soft Rule
Matcher
Annotated
Sentences
Unannotated
Sentences
Model
Rules
Corpus
Hard-Matching
Model
Annotated
Sentences
(D) NERO
Annotated
Sentences
Unannotated
Sentences
Rules
Corpus
Extracting
Rules
Finding
Matches
Rules
Corpus
Extracting
Rules
Finding
Matches
Neural Rule Grounding
Model
Rules
Corpus
Hard-Matching
Denoising
Unannotated
Sentences
!"# Self-Training
Figure 6.2: Comparison between previous work and the proposed NERO framework. (A) Bootstrap-
ping. (B) Data Programming. (C) Self-Training. (D) NERO. The neural rule grounding process
enables the soft-matching between rules and unmatched sentences using the soft rule matcher.
62
We want to explicitly exploit labeling rules over unmatched sentences as supervision for training
better RE models, for which we propose a NEural Rule grOunding (NERO) framework. The NERO
framework has two major modules: a sentence-level relation classifier and a soft rule matcher. The
former aims to learn the neural representations of sentences and classify which relation it talks
about, which serves as the outcome of NERO. We first apply our collected rules on a raw corpus by
hard matching and use hard-matched sentences as the main training data for NERO. The unmatched
sentences are assigned with pseudo labels by the soft rule matcher, which is a learnable module
that produces matching scores for unmatched sentences with collected rules. The key intuition
behind our soft rule matcher is that the distances between rules and sentences can be modeled by
simple cosine computations in a new space transformed from their neural representations (e.g. word
embeddings), which can be learned by a contrastive loss. Jointly training the two modules reinforce
the quality of pseudo-labels and then improve the performance of the relation classification module.
Specifically, in a batch-mode training process, we use the up-to-date learned soft matcher to assign
all unmatched sentences with weighted labels, which further serve as an auxiliary learning objective
for the relation classifier module.
Extensive experiments on two public datasets, TACRED [19] and Semeval [213], demonstrate
that the proposed NERO consistently outperforms baseline methods using the same resources by
a large margin. To investigate the label efficiency of NERO, we further conduct a user study to
compare models trained with rules and labeled sentences, respectively, both created under the same
time constraint. Results show that to achieve similar performance, NERO requires about ten times
fewer human efforts in annotation time than traditional methods of gathering labels.
The contributions of our work are summarized as follows: (1) Our proposed method, NERO, is
among the first methods that can learn to generalize labeling rules for training an effective relation
extraction model, without relying on additional sentence-level labeled data; (2) We propose a
learnable rule matcher that can semantically ground rules for sentences with similar meaning that
are not able to be hard-matched. This method also provides interpretable predictions and supports
few-shot learning ability on unseen relations; (3) We conduct extensive experiments to demonstrate
the effectiveness of our framework in a low-resource setting and show careful investigations on the
efficiency of using human efforts.
6.2 Problem Formulation
We first introduce basic concepts and their notations and then present the problem definition as well
as the scope of the work.
Relation Extraction. Given a pair of entity strings (e
subj
, e
obj
), identified as the subject and
object entities respectively, the task of relation extraction (RE) is to predict (classify) the relation
r∈R∪{NONE} between them, whereR is a pre-defined set of relations of interest. Specifically,
here we focus on sentence-level RE [19, 213, 214], which aims to predict the relation of entity
mention pairs in a sentence s—i.e., identifying the label r∈R for an instance in the form of
(e
subj
,e
obj
;s)
1
and yielding the prediction(e
subj
,r,e
obj
;s).
Labeling Rules. As an alternative to instance-level human-annotated data (e.g., instance-label pairs),
labeling rules formalize human’s domain knowledge in a structured way, and can be either directly
1
We use “instance” and ”sentence” exchangeably in the rest of the chapter.
63
used for making predictions over test instances [209], or applied to generate labeled instances for
model learning [83]. Formally, an inductive labeling rule consists of a rule body and a rule head.
In the context of relation extraction, we consider the cases where the rule body is a textual pattern
p=[SUBJ-TYPE;c; OBJ-TYPE], in which c denotes a word sequence between the two entities in
a sentence (i.e., context), and SUBJ-TYPE and OBJ-TYPE specify the entity types of the subject
and object entities required by the rule body. An instance(e
subj
,e
obj
;s) is hard-matched by a rule
p if and only if p’s context c exactly match the context between e
subj
and e
obj
in s and the entity
types of e
subj
and e
obj
match p’s associated entity types. For example, sentence s
1
in Fig. 6.1 is
hard-matched by rule p
3
.
Such labeling rules can either be hand-written by domain experts, or automatically generated
from text corpora with a knowledge base that shares the relations of interests. In this work, we
adopt a hybrid approach (as detailed in Sec. 6.3.2): 1) we first extract surface patterns ( e.g. frequent
patterns, relational phrases, etc.) as candidate rule bodies from raw corpora (after running named
entity recognition tool) using pattern mining tools, and 2) ask human annotators to assign relation
labels (i.e., rule heads) to the rule bodies. Then, we apply these human-annotated rules over
instances from a raw corpus to generate labeled instances (as training data). Compared to directly
annotating instances, our study will show that annotating rules is much more label-efficient, as
one rule may match many instances in the corpus with reliable confidence, generating multiple
labeled instances. Moreover, annotating rules may require similar (or even less) time than annotating
instances since they are shorter in length, as shown in our later analysis.
Example 1 (Labeling Rule Generation) Given instances such as{(Bill Gates, Microsoft; “Bill
Gates founded Microsoft. ”), (Bill Gates, Steve Jobs; “Guests include Bill Gates and Steve Jobs. ”)},
the candidate patterns include{“SUBJ-PERSON founded OBJ-PERSON”, “SUBJ-PERSON, OBJ-
PERSON”} , and the rule set after human annotation includes{“SUBJ-PERSON founded OBJ-
PERSON” → FOUNDED BY}. The second pattern is filtered out by annotators due to its low
quality.
Problem Definition. We focus on sentence-level relation extraction using labeling rules created
by a semi-automatic method. Specifically, given a raw corpus consisting of a set of instances
S ={(e
i
subj
,e
i
obj
;s
i
)}
N
i=1
, our approach first extracts surface patterns {[SUBJ-TYPE;c; OBJ-TYPE]} as
candidate rule bodies, and then annotates them to generate labeling rulesP ={p
k
}
|P|
k=1
with p
k
denoting [SUBJ-TYPE
k
;c
k
; OBJ-TYPE
k
]→ r
k
and r
k
∈R∪{NONE}. Next, our framework aims
to leverage both the rulesP and the corpusS to learn an effective relation classifier f :S →
R∪{NONE} which can predict relation for new instances in unseen sentences.
In addition to predicting relations already observed in the training phase, we aim to predict
unseen relations solely based on new rules specified for the unseen relations at the test time— i.e., a
few-shot learning scenario. Such a generalization ability is desirable for domain adaptation. In this
setting, the model is trained on a corpusS using labeling rules defined on relation set R
c
. In the
testing phase, the model is expected to make predictions on unseen relation typesR
u
given some
corresponding labeling rules.
Our Focus. Instead of relying on pre-existing knowledge bases to conduct distant supervision,
we focus on the scenarios where labeling rules are relatively cheaper and faster to obtain by
automatically inducing from large text corpora (as compared to manually curating a KB from
scratch). We study neural rule grounding to integrate information from rules and corpora into
64
Microsoft was founded by Bill Gates in 1975.
ORG:FOUNDED_BY
Annotated Sentences
Mike was born March 26 , 1965 , in US.
Unannotated Sentences
1.Hard-matching
(SUBJ-ORG, was founded by, OBJ-PER)
ORG:FOUNDED_BY
Rules
Microsoft was founded by Bill Gates in 1975.
Mike was born March 26 , 1965 , in US.
Corpus
Mike was born March 26 , 1965 , in US.
Pseudo-labeled Sentences
PER:ORIGIN
2. Soft-matching
(SUBJ-PERSON, born in, OBJ-LOCATION)
PER:ORIGIN
3.RC Model
Mike was born US ··· SUBJ-PERSON born OBJ-LOCATION
···
in
Corpus Mike was born March 26 , 1965 , in US.
Microsoft was founded by Bill Gates in 1975.
→ ORG:FOUNDED_BY
(SUBJ-PERSON, born in, OBJ-LOCATION) → PER:ORIGIN
Rules (SUBJ-ORG, was founded by, OBJ-PER)
Hard Matching
Matched Sent. ℎ
Microsoft was founded by Bill Gates in 1975. ORG:FOUNDED_BY
Unmatched Sent. ℎ
Mike was born March 26 , 1965 , in US.
Soft Rules Matcher
Mike was born US ··· SUBJ -PERSON born OBJ -LOCATION
···
in
Pseudo-labeled Sent. {( s, p̂ )}
Mike was born March 26 , 1965 , in US. PER:ORIGIN
Relations
Relation Classifier
Soft Matching
ORG:FOUNDED_BY
PER:ORIGIN
…
…
( SRM )
(RC)
Iterative Joint Learning
Figure 6.3: Overview of the NERO framework. Each unmatched sentence is first annotated by the
soft rule matcher (SRM) to generate pseudo labels, and then fed into the relation classifier ( RC) to
update the model parameters. The whole framework is trained iteratively and jointly, based on
multiple loss functions as introduced in Sec. 6.3.5.
learning an effective relation extraction model that can generalize better than solely using rules in a
hard-matching manner. We mainly use surface pattern rules (e.g., similar to the ones used in [215])
and leave the study of more complex rules (e.g., regular expression) as future work.
6.3 Neural Rule Grounding (NERO)
This section introduces the important concepts in NERO, and provides details of the framework
design and learning objectives. We first present our key motivation, then give an overview of NERO
framework (Sec. 6.3.1) and introduce the details of each component (Secs. 6.3.2-6.3.4). Lastly, we
show how we jointly learn model parameters in NERO for sequence encoding and rule matching
(Sec. 6.3.5).
Motivation of Soft Rule Matching. Traditional methods of using labeling rules for relation
extraction are mainly based on exact string/pattern matching (i.e., hard-matching), where a sentence
can be annotated by a rule if and only if it has exactly the same surface form as the rule body.
While being easy to implement and yielding relatively high precision for the prediction, these hard-
matching methods only achieve sub-optimal performance because of the severe low recall issue (see
Fig. 6.1 for an example). Such a low coverage of using rules can only produce a minimal number of
“labeled” sentences for learning a RE model, which may be far from enough for data-hungry neural
models. To address this problem, we would like to study “soft-matching” for exploiting the rules
for relation extraction. A soft-rule matcher is proposed to semantically match a rule even when the
surfaces are different (e.g. “founded” v.s. “created”). Towards improving relation extraction with
soft-matching, we would like to learn a SRM such that we can assign pseudo-labels to sentences.
Further, we can use these pseudo-labeled sentences to learn a sentence-level relation classifier.
6.3.1 Framework Overview
The NERO framework consists of a relation classifier ( RC) and a soft rule matcher (SRM). The RC is
for learning neural representations for sequences of words, from either an instance or a rule body
(Sec. 6.3.3). The SRM gives a matching score to indicate the similarity between a rule body and
an instance (Sec. 6.3.4). Previous works [210] mostly adopt a fixed metric function (e.g., cosine
similarity measure between the sequence embeddings). In contrast, our SRM is a learnable neural
65
network and thus can be more effective in modeling the semantic similarity between sentences and
rule bodies. As shown in Fig. 6.3, we first apply our collected rules to a raw corpus S to perform
hard matching.S will then be divided into two subsets: “hard-matched sentences”S
matched
and
“unmatched sentences”S
unmatched
. Lastly, we useSRM and rules to iteratively generate pseudo labels
overS
unmatched
, while jointly learning RC and SRM with the pseudo labels to promote a common
representation space (Sec. 6.3.5).
6.3.2 Labeling Rule Generation
As introduced in Sec. 6.2, our candidate rule bodies (i.e., surface patterns) are automatically
extracted from the raw corpus. In this work, we adopt a simple yet effective pattern mining method.
Given a raw corpus, we first replace entities with entity type masks SUBJ/OBJ-NER, where
“NER” denotes entity type (same as the procedure in [19]). Then we pick the word sequences
between and including the two entities as candidate rules. So each candidate rule is just a short
phrase / sentence containing two masked entities. To reduce the annotation efforts and ensure the
popularity of rules, we convert all words to their root form using Porter stemming algorithm [216]
and only keep the rules whose stemmed form appears at least N times in the whole corpus. Finally,
we ask human annotators to select rules that indicate a relation and assign labels to them. This
process is similar to label a sentence with a pre-defined relation or NONE. In spite of the existence
of more complex rule extraction methods such as shortest dependency path [206, 217] and meta
patterns [207], we adopt frequent word sequences in this work because of their simplicity and high
readability. Our framework can also be adapted to other rule generation methods as long as they can
generate human-readable candidate rules. Regardless of their high generalization ability, labeling
rules may bias the label distribution, e.g. PER:TITLE counts for 19% in the official training data
while only counts for 5.5% in the matched sentences. This domain gap poses another challenge for
learning with rules.
6.3.3 Relation Classifier ( RC)
The relation classifier (named as RC) is the end product of our NERO framework, which can represent
a relation instance(e
1
,e
2
,s) into vector embeddings. This module is general and can use various
designs (e.g., attention CNN [218], attention RNN [19, 219], and GCN [28, 35]). In this work,
we stick with the LSTM + ATT model [219]. Given a sequence of n words, we first look up
their word embeddings as{x
t
}
n
t=1
. Then we use a bi-directional LSTM network [36] to obtain the
contextualized embeddings of each word{h
t
}
n
t=1
where h
t
∈R
d
h
. Finally, we employ an attention
layer [220] to get a sentence representation c:
{h
t
}
n
t=1
= BiLSTM({x
t
}
n
t=1
) (6.1)
α
t
=
exp(v
T
tanh(Ah
t
))
∑
n
t
′
=1
exp(v
T
tanh(Ah
t
′))
(6.2)
c=
n
∑
t=1
α
t
h
t
(6.3)
66
where A∈R
d
a
× d
h
, v∈R
d
a
are learnable model parameters for the attention mechanism. We can get
the relation type probability distribution by feeding W
rc
c into a SoftMax layer:
RC(s,e
subj
,e
obj
)= SoftMax(W
rc
c),
where W
rc
∈R
d
h
×| R|
is a matrix to learn. In other words, we are using the output vector of the RC
for modeling the conditional distribution of relationP
θ
RC
(r= i|s)=RC(s,e
subj
,e
obj
)[i], meaning that
i− th element of theRC output vector as the probability of the i-th relation for the input sentence s.
Theθ
rc
denotes the set of model parameters of RC that we need to learn, including A,v,W
rc
, and
the weights of the BiLSTM.
6.3.4 Soft Rule Matcher (SRM)
The core component of the NERO is the soft rule matcher (named to be SRM), which is defined as
a neural function for modeling the matching score between a sentence s and a rule pattern p. We
formulate the Soft Rule Matcher as a function SRM :(s, p)→[− 1,1]. As shown in Fig. 6.4, we
first map the sentence s and rule p into the same embedding space by the RC, and then apply a
distance metric on their embeddings to get the matching score ranging from− 1 to 1. Note that
a matching score of 1 indicates the sentence can be hard-matched by the rule. Because the rules
are typically short phrases, we use word-level attention – a weighted continuous bag-of-words
model to calculate matching scores. We explored more designs in later experiments and this part
can be further studied in future work. We represent a sentence and a rule pattern as the sequence
of the word embeddings,{x
s
t
}
n
t=1
and{x
p
t
}
m
t=1
, respectively. By applying the attention on word
embeddings instead of contextualized embedding generated by LSTM, we aim to reduce over-fitting
given the scarcity of rules per relation. Note that we are using a different set of parameters for this
word embedding-level attention (B and v). Specifically, we have
z
s
=
n
∑
t=1
exp(u
T
tanh(Bx
s
t
))
∑
n
t
′
=1
exp(u
T
tanh(Bx
s
t
′
))
x
s
t
, (6.4)
z
p
=
m
∑
t=1
exp(u
T
tanh(Bx
p
t
))
∑
m
t
′
=1
exp(u
T
tanh(Bx
p
t
′
))
x
p
t
, (6.5)
SRM(s, p)=
(Dz
s
)
T
(Dz
p
)
∥Dz
s
∥∥Dz
p
∥
, (6.6)
where z
s
, z
p
are the representations for the sentence s and the rule p respectively, and D is a trainable
diagonal matrix which denotes the importance of each dimension. Similar idea has been shown
effective in other tasks for capturing semantically close phrases, and it generalizes well when only
limited training data is available [221]. In practice, a sentence may be very long and contain much
irrelevant information. So for the sentence, we only keep the words between (and including) the
subject and object entities, which is a common practice in other rule-based RE systems [210, 217].
67
Mike
(subject)
was born US
(object)
···
z
s
x
1
s
x
1
s
x
2
s
x
2
s
x
3
s
x
3
s
SUBJ-PERSON born OBJ-LOCATION
z
p
x
1
p
x
1
p
x
2
p
x
2
p
x
3
p
x
3
p
···
x
n
s
x
n
s
x
n
s
in
x
4
p
x
4
p
SRM SRM
cosine
Dz
s
Dz
p
Figure 6.4: Detailed architecture of the soft rule matcher (SRM). The cosine similarity between two
embeddings indicates the degree of matching between rules and sentences.
6.3.5 Joint Module Learning
We now formulate the overall learning process of NERO in a low-resource learning scenario, where
we only have collected rules but no sentence-level labels at all. Given a raw corpusS (i.e. a
large set of sentences without human-annotated labels) and a set of inductive labeling rulesP,
we first exactly match every rule on S and get a set of hard-matched sentences namedS
matched
.
These hard-matched labels are pretty accurate
2
, which can serve the purpose as training examples.
The other unmatched sentences, denoted asS
unmatched
are also informative for our model, while
they are ignored by most prior works. Although not able to be exactly matched by any rules, the
unmatched sentences can semantically align the meaning of some rules and thus help improve the
RE performance. By applying the SRM upon them to create pseudo-labels, we can further utilize the
hidden useful information as the supervision for learning the final RE model. Apart from that, the
SRM is also optimized for improving the quality of matching in the meantime. Finally, supervision
signals fromS
matched
,S
unmatched
, andP together train the RE model in a joint learning schema.
Learning with Hard-Matched Sentences (S
matched
). As the primary goal of the learning, we
want to minimize the error of the RC in classifying relations in the sentences inS
matched
respect
to their matched relation labels. Given a hard-matched sentence s∈S
matched
and its associated
relation r
s
∈R, we aim to minimize the cross-entropy loss L
matched
as follows.
L
matched
(θ
RC
)=E
s∼ S
matched

− logP
θ
RC
(r= r
s
|s)

. (6.7)
Learning with Rules (P). To incorporate our collected rules for exploiting the unmatched
sentencesS
unmatched
as well, we propose two additional auxiliary tasks for imposing rules in the
RC and SRM. First, we can treat the rule body p∈P as a “sentence” and its rule head r
p
as its
associated label, thus forming a labeled instance(p,r
p
). With such rule-reformatted instances as
2
We conduct an experimental analysis on the quality of these matches in Sec. 6.5.
68
+
−
−
+
+
−
−
+
Contrastive Loss
Figure 6.5: The contrastive loss in learning the SRM, which increases the matching score of rules
with the same relation type and decrease the matching score otherwise.
training data, we aim to minimize the error for RC in classifying them as follows. This objective
helpsRC to explicitly reinforce the memory ofRC about the collected rules.
L
rules
(θ
RC
)=E
p∼ P

− logP
θ
RC
(r= r
p
|p)

(6.8)
More importantly, we present how we utilize the rules for learning the soft rule matcher (SRM)
such that we can exploit the unmatched sentences as distant supervision. Our key motivation here
is that a good matcher should be able to clustering rules with the same type, and thus we expect
theSRM to increase the distances between rules with different types and reduce the variance of the
distances between rules with the same types. Simply put, given two rules, their similarity score
should be high if they belong to the same relation type and low otherwise.
We use the contrastive loss [222] for this objective. Given a rule p∈P and its relation r
p
∈R,
we divide all the other rules with the same relation type asP
+
(p)={p
′
|r
p
′ = r
p
}, and the ones
with different types asP
− (p)={p
′
|r
p
′̸= r
p
}. The contrastive loss aims to pull the rule p closer to
its most dissimilar rule inP
+
(p) and in the meantime push it away from its most similar rule in
P
− (p). Mathematically, the loss is defined as follows:
L
clus
=E
p∼ P

max
p
i
∈P
+
(p)
dist
+
(p, p
i
)− min
p
j
∈P
− (p)
dist
− (p, p
j
)

, (6.9)
where the measurement of the most dissimilar same-type rule and most similar different-type rule
can be measured with the distances defined as below:
dist
+
(p, p
i
)= max

τ− SRM(p, p
i
), 0

2
,
dist
− (p, p
j
)= 1− max

SRM(p, p
j
), 0

2
.
τ is a hyper-parameter for avoiding collapse of the rules’ representations. Minimizing L
clus
pushes
the matching scores of rules that are of the same relation type up toτ and pull the matching score
down to 0 otherwise. Without labels about the alignment between rulesP and unmatched sentences
S
unmatched
, this objective can be seen as a secondary supervision for SRM to train the parameters
θ
SRM
usingP itself only by this contrastive loss.
Learning with Unmatched Sentences (S
unmatched
). We use theSRM to label unmatched sentences
S
unmatched
(see Fig. 6.3). As the labels generated by the SRM may be noisy, we propose to use
pseudo-labeling [223] with instance weighting [224] to alleviate the noise. Each unmatched
sentence is weighed by the matching score from theSRM. Specifically, for each unmatched sentence
69
s∈S
unmatched
, we first apply the SRM to compute its matching scores with each rule p∈P and
then assign the pseudo-label to the sentence s as the relation of the highest-score rule r
ˆ p
, where
ˆ p= arg max
p∈P
SRM(s, p).
The corresponding weight for a particular pseudo-labeled instance,(s
i
, ˆ p
i
), is the SoftMax over all
the sentences (in a mini-batch when training; see Sec. 6.4.1 for more details):
w
s
=
exp(σSRM(s, ˆ p
i
))
∑
s
′
∈S
unmatched
exp(σSRM(s
′
, ˆ p
j
))
, (6.10)
whereσ represents the temperature of the SoftMax function. The loss function for each batch of
unmatched sentences would be:
L
unmatched
(θ
RC
)=E
s∼ S
unmatched

− w
s
logP
θ
RC
(r= r
ˆ p
|s)

. (6.11)
Joint Optimization Objective. The whole framework is jointly trained under the overall loss
function:
L(θ
RC
,θ
SRM
)= L
matched
(θ
RC
)+
α· L
rules
(θ
RC
)+β· L
clus
(θ
SRM
)+γ· L
unmatched
(θ
RC
).
To sum up, the proposed NERO framework has two trainable components: a relation classifier
(RC) and a soft rule matcher (SRM) with their specific parameters ( θ
RC
and θ
SRM
). Our primary
task is to minimize the error of the RC on the hard-matched sentences (L
matched
). To exploit the
unmatched sentences with collected rules, we first let RC to explicitly learn the rules with the
objective of rule classification ( L
rules
) and also learn the SRM with the help of contrastive loss (L
clus
)
for clustering rules. We jointly learn the two modules by connecting them through pseudo-labeling
on the unmatched sentences with the SRM and then expect the RC has better performance on such
pseudo-labeled unmatched sentences as well (L
unmatched
).
6.4 Model Learning and Inference
In this section, we introduce the specific training details of our proposed framework, and how the
model conduct inference.
6.4.1 Parameter Learning of NERO
NERO starts with a raw corpusS and pre-defined relations R∪{NONE}. It first extracts candidate
rules fromS with pattern mining tools, then asks human annotators to select and label the rules to
get the rule setP. Before training, NERO does a quick pass ofS by hard-matching to split it into
hard-matched sentencesS
matched
and unmatched sentencesS
unmatched
. Both of them are utilized
to train the relation classifier RC in a joint-training manner. Algorithm 3 summarizes the overall
training procedure.
The joint training objective is efficiently optimized by batch-mode training. Specifically, in each
training step we randomly sample a batch of hard-matched sentencesB
m
fromS
matched
, and then
70
Algorithm 3: Optimization of NERO model
Input: A raw corpusS , pre-defined relations R∪{NONE}.
Output: A relation classifier f :S →R∪{NONE}.
Extract candidate rules fromS with pattern mining tools.
Ask human annotators to select and label the candidate rules to getP.
PartitionS intoS
matched
andS
unmatched
by hard-matching withP.
while L in Eq. 6.3.5 not converge do
Sample batchB
m
={(s
i
,r
i
)}
n
i=1
fromS
matched
.
Update L
matched
by Eq. 6.7.
Sample batchB
u
={s
j
}
m
j=1
fromS
unmatched
.
foreach s∈B
u
do
Find highest-scored rule ˆ p and pseudo label r
ˆ p
bySRM.
Update L
unmatched
by Eq. 6.12.
Update L
rules
by Eq. 6.8.
foreach p∈P do
CalculateSRM(p, p
′
) for each p
′
∈P−{ p}.
Update L
clus
.
L= L
matched
+α· L
rules
+β· L
clus
+γ· L
unmatched
.
Update model parameters w.r.t. L.
calculate the average cross-entropy loss in the batch by L
matched
. Similarly, we sampleB
u
from
S
unmatched
, and generate the pseudo-labels with the entire rule setP, and calculate normalized
weights for L
unmatched
by:
w
s
=
exp(σSRM(s, ˆ p
i
))
∑
s
′
∈B
u
exp(σSRM(s
′
, ˆ p
j
))
,
L
unmatched
(θ
RC
)=
1
|B
u
|
∑
s∈B
u

− w
s
logP
θ
RC
(r= r
ˆ p
|s)

.
(6.12)
The normalized instance weights ensure the scale of L
unmatched
to be stable across different steps.
Finally, we calculate L
rules
and L
clus
with the entireP. For L
rules
, the loss is averaged for all p∈P.
For L
clus
, we take all p∈P to calculate the contrastive loss w.r.t.P−{ p}, and average the losses.
We do not sampleP due to the relatively small number of rules and simple structure of the soft rule
matcher (SRM). We stop training when the joint training loss L converges and takeRC as the output.
Note that word vectors are also trainable parameters that are initialized by pre-trained embeddings.
6.4.2 Model Inference
Model inference in NERO aims at predicting the relation of a new sentence. Intuitively, we can
give out a prediction by passing the sentence through theRC and return the relation with the highest
probability as the prediction. As an alternative, we can also predict the relation using theSRM. Given
a sentence, we can first find the most similar rule and return its rule head as the prediction. This
alternative way of inference can be applied for predicting unseen relations given new rules in the
71
Dataset # Train / Dev / Test # Relations # Rules # matched Sent.
TACRED [19] 75,049 /25,763 / 18,659 42 270 1,630
SemEval [213] 7,199 /800 / 1,864 19 164 1,454
Table 6.1: Statistics for TACRED and SemEval datasets.
testing time. In experiments, the first method shows much better performance, since RC can capture
rich contextual information while theSRM cannot.
For predicting “NONE” examples, our model filters out the predictions that our model is least
certain
3
about. Specifically, we measure the uncertainty using the entropy of softmax distribution or
the similarity score produced bySRM.
6.5 Experiments
In this section, we first introduce the datasets and compared baseline methods. Then, we illustrate
the detailed setup and present extensive experiments with discussion and analysis. Finally, we
conduct user studies to investigate the efficiency of our proposed approach.
6.5.1 Data Preparation
We choose two public and widely-used sentence-level relation extraction datasets in our experiments
(see Tab. 6.1) as follows. For both datasets, we construct the rule set as illustrated in Section 6.3.2
4
.
• TACRED [19] (TAC relation extraction dataset) contains more than 100,000 sentences
categorized into 42 relation types. Among the sentences, 79.5% of the examples are labeled
as NONE. We construct 270 rules which 1,630 hard-matched sentences in the official training
data;
• SemEval 2010 Task 8 [213] contains about 10,000 sentences with 19 relation types, where
17.4% of the sentences are NONE. We construct 164 rules which hard-match 1,454 instances
in the official training data.
6.5.2 Compared Methods
Recall Sec. 6.3.5, we first apply hard-matching on the unlabeled sentences and partition them
into hard-matched sentences (S
matched
) and unmatched sentences (S
unmatched
). Our framework is
applicable in both training withS
matched
and training withS
matched
∪S
unmatched
. Thus, we compare
our models with rule-based methods, supervised methods and semi-supervised methods. For both
semi-supervised and supervised methods, we use hard-matched sentences as the training data for
comparing under the same setting. The experiments evaluate the ability to use rules for improving
RE.
3
A thresholdδ tuned on dev set is used.
4
We filter candidate patterns by requiring frequency ≥ 3.
72
Rule-based Baseline Methods. We apply the following models onS
matched
∪S
unmatched
: (1)
CBOW-GloVe adopts continuous bag-of-words [225] on GloVe embeddings [226] to represent a
sentence or rule body, which labels a sentence using its most similar rule (in cosine distance).
(2) BREDS [210] is a rule-based bootstrapping method originally designed for corpus-level RE.
Given some entity pairs as seeds, it alternates between extracting new rules and new entity pairs.
BREDS performs soft matching between rules and sentences using average word embeddings. In
our experiments, we apply the set of rules learned by BREDS to perform hard-matching over the
test sentences at the prediction time. (3) The Neural Rule Engine (NRE [227]) is an unsupervised
method of soft-matching. It generalizes given rules by first performing unigram matching for each
token in the rules using CNN, then accumulates the matching scores of all tokens along the parsed
tree structure. When used for prediction, NRE performs soft-matching on the given rules.
Supervised Baseline Methods. We apply the following methods onS
matched
: (1) PCNN [203]
represents each token using both word embeddings and positional embeddings. Convolution and
piecewise max pooling layer are performed to produce the sentence embedding. (2)LSTM+ATT
adopts bi-directional LSTM and attention mechanism [220] to produce a sentence embedding,
which is fed into a fully-connected layer and a softmax classifier to predict the relation. (3)
PA-LSTM [19] extends the LSTM model by incorporating positional information into attention
mechanism and achieved state-of-the-art performance on TACRED. (4)Data Programming [83,
211] denoises the conflicting rules by learning their accuracy and correlation structures. (5)
LSTM+ATT (S
matched
+P) extends the LSTM+ATT model by also using rules as training data. It
serves as the base model for all semi-supervised baselines.
Semi-Supervised Baseline Methods. To make fair comparisons with NERO, we also regard labeling
rules as training data (same to L
rules
) and use LSTM+ATT (S
matched
+P) as the base model. We
apply the following methods onS
matched
∪S
unmatched
∪P: (1) Pseudo-Labeling [223] trains the
network using labeled and unlabeled data simultaneously. Pseudo-Labels are created for unlabeled
data by picking up the class with the maximum predicted probability and are used as if they are
labeled data during training. (2) Self-Training [212] iteratively trains the model using the
labeled dataset and expands the labeled set using the most confident predictions among the unlabeled
set. This procedure stops when unlabeled data is exhausted. (3) Mean-Teacher [228] assumes
that data points with small differences should have similar outputs. We perturb each unlabeled
sentence using word dropout and regularize their outputs to be similar. (4) DualRE [214] jointly
trains a relation prediction and a retrieval module which mutually enhance each other by selecting
high-quality instances from unlabeled data.
Variants of NERO. (1)Nero w/o S
unmatched
removes the loss on unmatched sentences and only
useS
matched
andP in model training. It only performs hard-matching on rules. (2)Nero-SRM
has the same training objective as NERO, but uses the soft rule matcher (SRM) to make predictions.
Given a sentence, NERO-SRM finds the most similar rule and returns the rule head as the prediction.
It can be taken as a context-agnostic version of NERO (Sec. 6.4.2).
6.5.3 Experiment Settings
Implementation. We implement most baselines from the scratch using Tensorflow 1.10 [229]
except for those that have released their codes (likePA-LSTM andDualRE). We adapt the baselines
73
Method / Dataset TACRED SemEval
Precision Recall F
1
Precision Recall F
1
Rules 85.0 11.4 20.1 81.2 17.2 28.5
BREDS [210] 53.8 20.3 29.5 62.0 24.5 35.1
CBOW-GloVe 27.9 45.7 34.6 44.0 52.8 48.0
NRE [227] 65.2 17.2 27.2 78.6 18.5 30.0
PCNN [203] 44.5± 0.4 24.1± 2.8 31.1± 2.6 59.1± 1.4 43.0± 0.7 49.8± 0.5
LSTM+ATT 38.1± 2.7 39.6± 2.7 38.8± 2.4 64.5± 2.8 53.3± 2.8 58.2± 0.8
PA-LSTM [19] 39.8± 2.5 40.2± 2.0 39.0± 0.6 64.0± 3.6 54.2± 2.5 58.5± 0.6
Data Programming [83] 39.2± 1.3 40.1± 2.0 39.7± 0.9 61.8± 2.1 54.8± 1.1 58.1± 0.7
LSTM+ATT (S
matched
+P) 39.2± 1.7 45.5± 1.7 42.1± 0.9 63.4± 2.1 55.0± 0.3 58.8± 0.9
Pseudo-Labeling [223] 34.5± 4.1 37.4± 5.1 35.3± 0.8 59.4± 3.3 55.8± 2.1 57.4± 1.3
Self-Training [212] 37.8± 3.5 41.1± 3.1 39.2± 2.1 62.3± 2.0 53.0± 2.7 57.1± 1.0
Mean-Teacher [228] 46.0± 2.7 41.6± 2.2 43.6± 1.3 62.3± 1.5 54.5± 1.2 57.9± 0.5
DualRE [214] 40.2± 1.5 42.8± 2.0 41.7± 0.5 63.7± 2.8 54.6± 2.1 58.6± 0.8
NERO w/oS
unmatched
41.9± 1.8 44.3± 3.8 42.9± 1.4 61.4± 2.4 56.2± 1.9 58.6± 0.6
NERO-SRM 45.6± 2.2 45.2± 1.2 45.3± 1.0 54.8± 1.6 55.2± 2.0 54.9± 0.6
NERO 54.0± 1.8 48.9± 2.2 51.3± 0.6 66.0± 1.5 55.8± 0.9 60.5± 0.7
Table 6.2: Performance comparison (in %) of relation extraction on the TACRED and SemEval
datasets. We report the mean and standard deviation of the evaluation metrics by conducting 5 runs
of training and testing using different random seeds. We use LSTM+ATT (S
matched
+P) as the
base model for all semi-supervised baselines and our models.
to our setting. For supervised and semi-supervised baselines, we use the hard-matched sentences
(S
matched
) as the “labeled data” and the unmatched sentences (S
unmatched
) as the “unlabeled data”.
Training details. We use pre-trained Glove embeddings [226] to initialize the word embeddings and
fine-tune them during training. For NERO, We set the batch size to 50 for hard-matched sentences,
and 100 for unmatched sentences. For other baselines, we set the batch size to 50. To exclude
the possibility that NERO takes advantage of a larger batch size, we also tried a batch size of 150
for other baselines but it showed no difference in performance. To reduce over-fitting, we adopt
the entity masking technique [19] and replace the subject and object with SUBJ/OBJ–NER and
regard them as normal tokens. We use a two-layer bi-directional LSTM as the encoder. The hidden
dimension is 100 for the LSTM, and 200 for the attention layers. We setβ ,γ,τ to 0.05, 0.5, and 1.0
respectively. We setα to 1.0 for TACRED, and 0.1 for SemEval. The temperatureθ in instance
weighting is set to 10. We apply dropout [61] after the LSTM with a rate of 0.5. All models are
optimized using AdaGrad [230] with initial learning rate 0.5 and decay rate 0.95. For models that
require the prediction of NONE (including NERO and all supervised / semi-supervised baselines),
we select the threshold (see 6.4.2) from range [0.0, 1.0] based on the dev set.
6.5.4 Performance Comparison
Rule-based Models. We first compare the models that solely use rules, as shown in Table 6.2. Hard-
matching with rules achieves a high precision (85% on TACRED) but suffers from the low-recall
problem due to its failure in matching instances with similar meanings (only 11.4% on TACRED).
Bootstrapping methods (such as BREDS) and unsupervised soft-matching methods (such as NRE)
74
95 147 210 270
# of Rules on TACRED
32
34
36
38
40
42
44
46
48
50
52
54
F1 (%)
500
1000
2000
3000
4000
5000
LSTM + ATT(
a
+ ) Mean-Teacher NERO
Figure 6.6: Performance w.r.t. different number of rules and human-annotated labels on TACRED.
We show different models’ F1 scores and number of rules or labels used for training the correspond-
ing model.
manage to cover more instances (6% gain in recall on TACRED dataset), but they fail to capture
contextual information.
Models Trained onS
matched
. Neural networks are capable of fitting the hard-matched data, but
they suffer from over-fitting due to the small data size. Adding the pattern encoder loss ( L
rules
) and
the soft matching loss (L
clus
) boosts the performance by a huge margin, as indicated by LSTM+ATT
(S
matched
+P) and NERO w/oS
unmatched
). However, their performance is still far from satisfactory,
since they cannot utilize large amounts of unmatched sentences. Data programming does not bring
any improvement because our rule mining and labeling method rarely introduces conflicted labeling
rules. Hence, data programming with our rules is the same as LSTM+ATT.
Models Trained onS
matched
∪S
unmatched
∪P. Table 6.2 shows the results for methods that
incorporate unmatched data and rules. For methods that actively create labels for unmatched
data, namely pseudo-labeling, self-training and DualRE, their performance is even worse than the
supervised counterparts (LSTM+ATT (S
matched
+P)). It is because in a low-resource setting, the
model performance is so low that the created labels are very noisy. Mean-Teacher manages to
improve model performance on TACRED, but only by a small margin. Our proposed model, NERO,
is relatively indifferent to the noise due to the soft rule matcher, which learns directly from similar
textual sentences. This results in a significant gain in precision (14% on TACRED) and recall
(4% on TACRED) over the semi-supervised baselines. Compared to TACRED, the improvement
on SemEval is much smaller (only 1.7% in F1). We hypothesize it is because the sentences in
SemEval are quite short and contain very simple rule patterns. Thus, the soft-matched sentences
can hardly provide additional information. We also try to predict with the soft rule matcher, but the
performance is lower in F1 (45.3% on TACRED and 54.9% on SemEval) since it cannot capture
contextual information.
75
6.5.5 Performance Analysis
1. Performance on Various Amounts of Rules and Human Annotated Labels. To show that
our rules are more powerful than human-annotated labels, we show the performance of models
trained with different number of rules (30%, 50%, 70%, and 100%) and the number of labels
required to reach comparable performance using supervised baseline (LSTM+ATT). We random
sample subsets of rules w.r.t. each relation to avoid extreme cases where some relations are not
sampled at all. We conduct 5 runs of sampling and training and report the average performance.
As shown in Figure 6.6, NERO consistently outperforms other methods by a huge margin, and
its performance increases with the number of rules (from 42.3% to 51.3%). Mean-Teacher also
outperforms the base model especially when the number of rules is small, but its performance
does not increase much (from 41.3% to 43.6%). It shows that Mean-Teacher generalizes well in
low-resource setting but cannot fully utilize the knowledge in rules. In terms of label efficiency, one
rule in NERO is equivalent to 10 human-annotated labels in model performance. Even in the base
model LSTM+ATT (S
matched
+P), one rule is still equivalent to about 4 labels. It demonstrates
the superior generalization ability of rules over labels.
2. Performance on Various Amounts of the Raw Corpus. To test the robustness of our model,
we plot the performance curve in Fig. 6.7 when different amounts of raw corpus are available
on TACRED. Again, NERO outperforms all methods, and its F
1
score positively correlates to the
amount of available data (from 43.4% to 51.3%). We also observe similar phenomena in other
methods (all based on LSTM+ATT (S
matched
+P) model), which show that matching on rules
provides additional knowledge to model training. Just memorizing these rules with a neural model
leads to very bad performance. Also, we observe that with only 10% of data, our final model
(NERO) already outperforms the best supervised baseline NERO w/oS
unmatched
. This indicates that
the soft matching module can utilize the raw corpus and rules in a more data-efficient way.
3. Performance on Unseen Relations. To show that NERO has the capacity of predicting unseen
relations, we evaluate NERO on relations unseen in training. In the experiment, we randomly
sample 5 relations as unseen relations and repeat the experiment 10 times. In each time, we use the
same sets of unseen relations and the same random seed for all methods. We remove the sampled
relations (both sentences and rules) in training while keeping only them in testing. We sample
sentences of other relations as NONE relation in testing with the same ratio as the raw corpus (79.5%
for TACRED and 17.4% for SemEval). Moreover, since the special tokens SUBJ/OBJ-NER are
out-of-vocabulary in both Glove and BERT, we only use the words between two entities in encoding
and only use rules with the same entity types to the sentence in inference. To make predictions
by NERO, we find the most similar rule using soft rule matcher. Baseline models include exact
rule matching, cosine similarity of average Glove embedding, and cosine similarity of pre-trained
BERT embedding [13]. As shown in Table 6.3, our soft rule matcher achieves similar or better
performance compared to all baselines. It shows that instead of just memorizing the rules and
the corresponding relation types, the soft rule matcher also learns knowledge about how to better
perform rule matching and can even generalize to unseen relations.
6.5.6 Model Ablation Study
1. Effect of Different Loss Functions. To show the efficacy of our proposed losses ( L
clus
, L
unmatched
and L
rules
), we conduct ablation study on loss functions by removing one loss function at a time. As
76
10 30 50 70 100
Amount of Raw Corpus (%)
25
30
35
40
45
50
55
F1 (%)
Self-Training
Mean-Teacher
DualRE
NERO ( S
unmatched
)
NERO
Figure 6.7: Performance of different semi-supervised models trained using various amounts of
unlabeled sentences randomly sampled from the raw corpus on TACRED.
TACRED SemEval
Method P R F
1
P R F
1
Rule (exact match) 100 6.1 10.8 83.2 17.7 28.2
CBOW-GloVe 52.4 86.3 64.7 40.3 45.5 34.7
BERT-base (frozen) 66.2 76.8 69.5 37.8 33.2 35.3
NERO 61.4 80.5 68.9 43.0 54.1 45.5
Table 6.3: Performance on predicting unseen relations. NERO applies the learned soft rule matcher
on unseen relation rules to make predictions.
shown in Table 6.4, all three losses contribute to the final performance, while L
unmatched
helps the
most, which proves the effectiveness of soft-matching. When removing the contrastive loss L
clus
,
the F1 score drops from 51.3 to 46.4, which shows the effectiveness of a trainable SRM. The L
rules
loss also brings improvement, similar to including more training instances.
Objective Precision Recall F
1
L (ours) 54.0 48.9 51.3
− L
rules
50.0 47.7 49.0
− L
clus
50.9 43.0 46.4
− L
unmatched
41.9 44.3 42.9
Table 6.4: Ablation Study of Different Training Objectives on TACRED dataset. We remove each
loss term one at a time.
77
2. Effect of Different Soft-matching Models. Besides word-level attention, we also try other
soft matching mechanisms including CBOW-Glove, LSTM+ATT, and BERT-base (both frozen
and fine-tuned). All methods first summarize the sentences / rules into vector representations,
then calculate the matching scores by cosine similarity. We report their performance in Table 6.5.
We observe that our method achieves the best performance despite its simplicity. To our surprise,
BERT-base performs much worse than our method. In the experiment we find that BERT tends to
given high matching scores for all sentence-rule pairs so our framework cannot distinguish the false
pseudo-labels from reliable ones.
Objective Precision Recall F
1
CBOW-Glove 49.4 43.5 46.2
LSTM+ATT 56.2 46.0 50.6
BERT-base (frozen) 45.6 47.6 46.5
BERT-base (fine-tuned) 50.3 45.8 47.9
Word-level attention (ours) 54.0 48.9 51.3
Table 6.5: Ablation study of different soft-matching models for NERO on the TACRED dataset.
3. Sensitivity Analysis of Model Hyper-parameters. The most important hyper-parameters in
NERO areα,β,γ, andτ. We test the sensitivity of these parameters on TACRED. We adjust one
hyper-parameter at a time and remain the other three unchanged. The results are shown in Figure
6.8. We observe that forτ andβ , there exists a wide interval where the F1 score remains stable. For
α andγ, the F1 score hardly changes with the choice of their values. This is because that L
matched
,
L
clus
, and L
rules
quickly drop to 0 and L
unmatched
dominates the joint loss in the training phase, and
the impact ofγ is lessened by adaptive subgradient method (AdaGrad).
6.5.7 Case Study
Study on Label Efficiency. To test the label efficiency of rules and sentences in real scenarios, we
ask 5 college students in computer science to annotate frequent candidate rules (with frequency
≥ 3) and the unlabeled sentences, during a 40-minute period. For annotating sentences, they are
required to provide labels from the pre-defined relation set R∪ NONE. For annotating candidate
rules, they are required to filter out uninformative rules (e.g. “ SUBJ-PERSON and OBJ-PERSON”)
and assign labels to the remaining ones. For each candidate rule, they are further given an example
of hard-matched sentence to help understand the meaning and context of the rule. Then we use each
user’s annotations to train the RE model. We use LSTM+ATT for labeled sentences and NERO for
rules.
We report the average number of selected rules / labeled sentences and performance of trained
RE model every 5 minutes in Figure 6.10. We observe that getting rules is nearly as fast as getting
labels but is much more efficient. With the same annotation time, NERO can already learn a good
RE classifier (with 41.3% F1), while LSTM+ATT cannot learn anything from the labels. In just 40
minutes, NERO can achieve comparable performance to 1000 labels (refer to Figure 6.6), which
requires about 320 minutes to annotate.
Interpretability of Soft Rule Grounding. Fig. 6.9 gives an example of the similarity matrix,
showing that our soft matching module learns meaningful representations of rules for soft matching.
78
The left part of the figure presents 1) which words are more important for building attention-
pooled rule/sentence representations, and 2) the soft matching scores between the rule and three
sentences. We also show two concrete examples of attention matrix in the right part. For each
word-pair between a rule and a sentence, we visualize the cosine similarity matrix between word
representations learned with the contrastive loss. While the word “chairman” in our rule matches
perfectly with the same surface word in the sentence, it also has a high similarity score with “Chief”
and “Executive”.
6.6 Related Work
Relation Extraction. Traditional corpus-level RE systems take a rule-based approach, in which
rules are either handcrafted [231] or automatically learned from large corpora [209, 210]. Although
feature-based soft matching methods (e.g. TF-IDF, CBOW) are used in rule mining, the inference
(i.e, labeling new sentences) is still done by hard-matching, which leads to high precision but
low recall due to the low coverage of rules. While we use a trainable soft rule matcher, which is
more powerful than prior feature-based methods. And we apply soft-matching to both training
and inference phase in a consistent manner. Recent RE models [19, 203] successfully apply deep
neural networks with the help of large scale datasets such as SemEval [213] and TACRED [19].
However, their performance degrades heavily when labeled data is insufficient. Some work [211,
232] proposes to train a RE classifier from natural language explanations, which are first transformed
to labeling rules by a semantic parser then applied to raw corpora by hard-matching. But none of
0.1 0.3 0.5 0.7 0.91.0
40
42
44
46
48
50
52
54
F1
0.1 0.3 0.5 0.7 0.91.0
40
42
44
46
48
50
52
54
F1
0.1 0.3 0.5 0.7 0.91.0
40
42
44
46
48
50
52
54
F1
0.1 0.3 0.5 0.7 0.91.0
40
42
44
46
48
50
52
54
F1
Figure 6.8: Sensitivity analysis ofτ,α,β, andγ on TACRED. We report the mean and standard
deviation F1 score by conducting 5 runs of experiments using different random seeds.
79
Rule
0.85
0.65
0.41
Figure 6.9: Output visualization of SRM. Left: attention weights of words and the soft matching
scores between a rule and three sentences. Right: cosine similarity matrix between word embeddings
learned with the contrastive loss.
5 10 15 20 25 30 35 40
Annotation Time (min)
0
10
20
30
40
50
60
70
80
90
100
110
120
130
140
150
# of Annotation
Rule Sentence
0
5
10
15
20
25
30
35
40
45
F1 (%)
Sentence (LSTM + ATT)
Rule (NERO)
Figure 6.10: Study on label efficiency. Average number of rules / sentences labeled by annotators
(dashed line) are shown on the x-axis over the left-hand side; and the performance of models
trained with these corresponding labeled rules / sentences (solid line) are shown on the x-axis over
the right-hand side. We use NERO and LSTM+ATT as the base model for the labeling rules and
sentences, respectively.
them utilizes the data that cannot be matched by rules, which is of massive amount. This motivates
us to propose a framework that can fully utilize the information from both rules and raw corpora.
Semi-Supervised Learning. Our work is relevant to semi-supervised learning, if we consider
rule-annotated data as labeled among the corpus, e.g., self-training [212], mean-teacher [228], and
semi-supervised V AE [233]. In relation extraction, one line of work proposes to alleviate the low
recall problem in rule-based approach using bootstrapping framework [209] and revised pattern
matching [210, 234]. However, their settings are tailored for corpus-level RE. Another line of work
applies self-training framework in supervised learning models. However, These models turn out to
be ineffective in rule-labeled data due to potentially large difference in label distribution, and the
generated psuedo-labels may be quite noisy.
80
6.7 Conclusion
We proposed a novel framework, named NERO, for label-efficient relation extraction. We auto-
matically extracted frequent patterns from large raw corpora and asked users to annotate them to
form labeling rules. To increase the coverage of these rules, we further proposed the soft-matching
mechanism, where unlabeled sentences are annotated by their most semantically similar labeling
rules and weighted in training the RE model. Experiments on two public datasets proved the
effectiveness of our framework.
81
Chapter 7
Continual Contrastive Finetuning
Relation extraction (RE), which has relied on structurally annotated corpora for model training,
has been particularly challenging in low-resource scenarios and domains. Recent literature has
tackled low-resource RE by self-supervised learning, where the solution involves pretraining the
relation embedding by RE-based objective and finetuning on labeled data by classification-based
objective. However, a critical challenge to this approach is the gap in objectives, which prevents the
RE model from fully utilizing the knowledge in pretrained representations. We aim at bridging the
gap and propose to pretrain and finetune the RE model using consistent objectives of contrastive
learning. Since in this kind of representation learning paradigm, one relation may easily form
multiple clusters in the representation space, we further propose a multi-center contrastive loss
that allows one relation to form multiple clusters to better align with pretraining. Experiments on
two document-level RE datasets, BioRED and Re-DocRED, demonstrate the effectiveness of our
method. Particularly, when using 1% end-task training data, our method outperforms PLM-based
RE classifier by 10.5% and 6.1% on the two datasets, respectively.
7.1 Introduction
Relation extraction (RE) is a fundamental task in NLP. It aims to identify the relations among
entities in a given text from a predefined set of relations. While much effort has been devoted
to RE in supervised settings [19, 28, 235], RE is extremely challenging in high-stakes domains
such as biology and medicine, where annotated data are comparatively scarce due to overly high
annotation costs. Therefore, there is a practical and urgent need for developing low-resource RE
models without the reliance on large-scale end-task annotations.
To realize low-resource RE, previous work has focused on pretraining relation mention rep-
resentations on large corpora using RE-based pretraining objectives. Particularly, Baldini Soares,
FitzGerald, Ling & Kwiatkowski [236] propose a self-supervised matching-the-blanks (MTB)
objective that encourages the representations of pairs of relation mentions to be similar if they
contain the same entity pairs. Later work [237, 238] extends this idea by generating relation
mention pairs with distant supervision [239] and improves representation learning using contrastive
learning [124, 125, 240]. To adapt to training on RE annotations, these works finetune pretrained
relation mention representations on labeled data using classification-based objectives. Although
this paradigm produces better results compared to RE models initialized with pretrained language
82
models (PLMs), it creates a significant divergence between pretraining and finetuning objectives,
thus preventing the model from fully exploiting knowledge in pretrained representations.
We aim to bridge this gap in RE pretraining and finetuning. Our key idea is to use similar
objectives in pretraining and finetuning. First, we propose to continually finetune the representations
by contrastive learning, which encourages the representations of relation mentions corresponding to
the same relation to be similar. However, as pretraining and finetuning are conducted on different
tasks, relation mentions of the same relation can form multiple different clusters in the pretrained
representation, where standard supervised contrastive loss [142] may distort the representation
because of its underlying one-cluster assumption [241]. Therefore, we further propose a multi-center
contrastive loss (MCCL), which encourages a relation mention to be similar to only a subset of
mentions of the same relation type, allowing one relation to have multiple clusters. Second, we
propose to use k-nearest neighbors (kNN; [242, 243]) in inference, where predictions are made
based on most similar instances.
We focus our work on document-level RE [20, 244], which consists of both intra- and cross-
sentence relations. To the best of our knowledge, this work represents the first effort to explore
self-supervised pretraining for document-level RE. Unlike prior studies [237, 238], we do not use
distant supervision. Instead, we pretrain relation mention representations with an improved MTB
objective on unlabeled corpora, where we use contrastive learning to learn representations that suit
downstream RE. We then finetune the pretrained model on labeled data with MCCL. Experiments on
two datasets, BioRED [10] in the biomedical domain and Re-DocRED [245] in the general domain,
demonstrate that our pretraining and finetuning objectives significantly outperform baseline methods
in low-resource settings. Particularly, in the low-resource setting of using 1% of labeled data, our
method outperforms PLM-based classifiers by 10.5% and 6.1% on BioRED and Re-DocRED,
respectively. Based on our pretrained representations, MCCL outperforms classification-based
finetuning by 6.0% and 4.1%, respectively.
Our technical contributions are three-fold. First, we propose to pretrain the PLMs based on
our improved MTB objective and show that it significantly improves PLM performance in low-
resource document-level RE. Second, we present a technique that bridges the gap of learning
objectives between RE pretraining and finetuning with continual contrastive finetuning and kNN-
based inference, helping the RE model leverage pretraining knowledge. Third, we design a novel
MCCL finetuning objective, allowing one relation to form multiple different clusters, thus further
reducing the distributional gap between pretraining and finetuning.
7.2 Related Work
Document-level RE. Existing document-level RE models can be classified into graph-based and
sequence-based models. Graph-based models construct document graphs spanning across sentence
boundaries and use graph encoders such as the graph convolution network (GCN; [64]) to aggregate
information. Particularly, Quirk & Poon [246] build document graphs using words as nodes with
inner- and inter-sentence dependencies (e.g., syntactic dependencies, coreference, etc.) as edges.
Later work extends this idea by applying different network structures [244, 247] or introducing
other node types and edges [235, 248, 249]. On the other hand, sequence-based methods [250–252]
use PLMs to learn cross-sentence dependencies without using graph structures. Particularly, Zhou,
Huang, Ma & Huang [250] propose to enrich relation mention representation by localized context
83
pooling. Zhang, Chen, Xie, Deng, Tan, Chen, et al. [251] propose to model the inter-dependencies
between relation mentions by semantic segmentation [253]. In this work, we study a general method
of self-supervised RE. Therefore, our method is independent of the model architecture and can be
adapted to different RE models.
Low-resource RE. Labeled RE data may be scarce in real-world applications, especially in low-
resource and high-stakes domains such as finance and biomedicine. Much effort has been devoted
to training RE models in low-resource settings. Some work tackles low-resource RE by indirect
supervision, which solves RE by other tasks such as machine reading comprehension [254], textual
entailment [24], and abstractive summarization [255]. However, indirect supervision may not be
practical in high-stake domains, where annotated data for other tasks are also scarce. Other ef-
forts [236–238] improve low-resource RE by pretraining on large corpora with RE-based objectives.
Specifically, Baldini Soares, FitzGerald, Ling & Kwiatkowski [236] propose an MTB objective
that encourages the representations of relation mentions containing the same entity pairs to be
similar. Peng, Gao, Han, Lin, Li, Liu, et al. [237] propose to pretrain on distantly labeled corpora,
where they seek to represent the relation mentions of the same distant label to be similar. They also
introduce a contrastive learning based training objective to improve representation learning. Qin,
Lin, Takanobu, Liu, Li, Ji, et al. [238] further introduce an entity discrimination task and pretrain
the RE model on distantly labeled document corpora. We study self-supervised pretraining for
document-level RE. We study how to reduce the gap between pretraining and finetuning, which is
critical to bridge the training signals obtained in these two stages but has been overlooked in prior
work.
7.3 Method
In this work, we study a self-supervised approach for document-level RE. Given a document d
and a set of entities{e
i
}
N
i=1
, where each entity e
i
can have one or multiple entity mentions in the
document, document-level RE aims at predicting the relations of all entity pairs(e
s
,e
o
)
s,o∈{1,...,N}
from a predefined set of relationships R (including an NA class indicating no relation exists), where
e
s
and e
o
are the subject and object entities, respectively. In the self-supervised RE setting, we have
a large unlabeled document corpus for pretraining and a labeled RE dataset for finetuning. The
document corpus has been annotated with entity mentions and the associated named entity types
but no relations. Our goal is to train a document-level RE classifier, especially in the low-resource
setting.
Our training pipeline consists of two phases: pretraining and finetuning. In pretraining, we
use the (unlabeled) document corpus to pretrain the relation mention representations based on our
improved matching-the-blanks training objective (MTB; [236]), where the LM learns to decide
whether two relation mentions contain the same entity pairs or not, and the learning of representation
is enhanced with contrastive learning. In finetuning, we continue to train the pretrained model
on relation-labeled data using a multi-center contrastive loss (MCCL), which achieves better
performance than the traditional classifier paradigm due to the better-aligned learning objective
with pretraining. After training, we use k-nearest neighbor (kNN) inference that suits well the
contrastively finetuned model.
84
The rest of this section is organized as follows: we introduce the model architecture used in
both pretraining and finetuning in Section 7.3.1, the pretraining process in Section 7.3.2, finetuning
in Section 7.3.3, and inference in Section 7.3.4.
7.3.1 Model Architecture
Encoder. Given a document d=[x
1
,x
2
,...,x
l
], we first mark the spans of the entity mentions by
adding special entity markers [E] and [/E] to the start and the end of each mention. Then we encode
the document with a PLM to get the contextual embedding of textual tokens:
H H H =[h h h
1
,h h h
2
,...,h h h
l
]= PLM([x
1
,x
2
,...,x
l
]).
We take the contextual embedding of [E] at the last layer of the PLM as the embedding of
entity mentions. We accumulate the embedding of mentions corresponding to the same entity by
LogSumExp pooling [244] to get the entity embedding h h h
e
i
.
Relation Embedding. Given a relation mention t=(e
s
,e
o
,d), where e
s
and e
o
are the subject and
object entities, respectively, we represent the relation mention as:
z z z
t
= W W W
linear
h
h h h
e
s
,h h h
e
o
,c c c
(e
s
,e
o
)
i
.
Here h h h
e
s
,h h h
e
o
∈R
d
are the entity embedding vectors, c c c
e
s
,e
o
∈R
d
is the localized context encoding for
(e
s
,e
o
), W W W
linear
∈R
3d× d
is a linear projector. The localized context encoding is introduced by Zhou,
Huang, Ma & Huang [250] to derive the context embedding conditioned on a relation mention,
which finds the context that both the subject and object entities attend to. Specifically, denote the
multi-head attention in the last layer of PLM as A A A∈R
m× l× l
, where m is the number of attention
heads, l is the input length, we first take the attention scores from [E] as the attention from each
entity mention, then accumulate the attention of this entity mention by mean pooling to get the
entity-level attention A A A
(e
i
)
∈R
m× l
. Finally, we compute c c c
(e
s
,e
o
)
by:
A A A
(e
s
,e
o
)
= A A A
(e
s
)
⊙ A A A
(e
o
)
,
q q q
(e
s
,e
o
)
=
m
∑
i=1
A A A
(e
s
,e
o
)
i
,
a a a
(e
s
,e
o
)
= q q q
(e
s
,e
o
)
/1 1 1
⊺
q q q
(e
s
,e
o
)
,
c c c
(e
s
,e
o
)
= H H H
⊺
a a a
(e
s
,e
o
)
.
We introduce in the rest of the section how to pretrain and finetune the RE model based on the
relation embedding z z z
(e
s
,e
o
)
.
7.3.2 Pretraining
We pretrain the LM on the document corpus using the MTB objective. MTB is based on a simple
assumption that, in contrast to different entity pairs, it is more often for the same entity pair to be
connected with the same relation. The MTB objective transforms the similarity learning problem
85
into a pairwise binary classification problem: given two relation-describing utterances where entity
mentions are masked, the model classifies whether the entity pairs are the same or not. This
pretraining objective has shown effectiveness in several sentence-level RE datasets[18, 19, 23].
However, when it comes to document-level RE, Qin, Lin, Takanobu, Liu, Li, Ji, et al. [238]
have observed no improvement led by the vanilla MTB pretraining. Therefore, we replace the
pairwise binary classification with contrastive learning, which is adopted in later RE pretraining
works [237, 238] and can effectively learn from more positive and negative examples. Details of
training objectives are elaborated in the rest of the section.
Training objective. The overall goal of pretraining is to make the relation embedding of positive
pairs closer than that of negative pairs. We use the InfoNCE loss [124] to model this objective.
Given the documents in batch,P as the set of all positive relation mention pairs, andN
t
denote the
set of negative relation mentions w.r.t. relation mention t, the loss for relation mention pairs is
1
:
L
rel
=− 1
|P|
∑
t
i
,t
j
∈P
log
e
sim(z z z
t
i
,z z z
t
j
)/τ
Z
t
i
, (7.1)
Z
t
i
= e
sim(z z z
t
i
,z z z
t
j
)/τ
+
∑
t
k
∈N
t
i
e
sim(z z z
t
i
,z z z
t
k
)/τ
,
where sim(z z z
t
i
,z z z
t
j
) denotes the similarity between the relation mention t
i
and t
j
. Following Chen,
Kornblith, Norouzi & Hinton [125], we use cosine similarity as the similarity metric. Similar to
SimCSE [256], we further add a self-supervised contrastive loss that requires the same relation
embedding augmented by different dropout masks to be similar, thus encouraging the model to
learn more instance discriminative features that lead to less collapsed representations. Specifically,
denote the two relation embedding of t derived by different dropout masks as z z z
t
and ˆ z z z
t
, respectively,
the set of all relation mentions in the batch asT , and the set of relation mentions in positive pairs
asT
P
, the self-supervised loss is formulated as:
L
self
=− 1
|T
P
|
∑
t
i
∈T
P
log
e
sim(z z z
t
i
,ˆ z z z
t
i
)/τ
Z
t
i
,
Z
t
i
= e
sim(z z z
t
i
,ˆ z z z
t
i
)/τ
+
∑
t
k
∈T\{t
i
}
e
sim(z z z
t
i
,ˆ z z z
t
k
)/τ
.
Finally, we use a masked language model lossL
mlm
to adapt the LM to the document corpus. The
overall pretraining objective is:
L
pretrain
=L
rel
+L
self
+L
mlm
.
For faster convergence, we initialize our model with a PLM that is pretrained on a larger corpus,
and continually pretrain the PLM on the document corpus with our new pretraining objectives. We
use BERT [13] for the general domain and PubmedBERT [257] for the biomedical domain.
1
Similar to Baldini Soares, FitzGerald, Ling & Kwiatkowski [236], we randomly mask the entities in documents
with a probability of 0.7 to avoid shortcut learning.
86
Classifier BioRED Re-DocRED
One-cluster
Softmax 28.6 39.3
Nearest centroid 12.5 4.1
Multi-cluster
kNN 36.7 54.1
Table 7.1: Probing results (in F
1
) on the test set of BioRED and Re-DocRED.
7.3.3 Finetuning
After pretraining, we finetune the LM on labeled document-level RE datasets. In previous stud-
ies [236–238], pretraining and finetuning are conducted in processes with different learning ob-
jectives. Specifically, after using the pretrained weights to initialize a RE classifier, the model
is finetuned with a classification-based training objective. Based on our model architecture, a
straightforward finetuning method is to add a softmax classifier upon the relation embedding, for
which a cross-entropy loss for a batch of relation mentionsT is formulated as:
P
t
i
r
= softmax(W W W
r
z z z
t
i
+ b
r
),
L
ce
=− 1
|T|
∑
t
i
∈T
log(P
t
i
y
t
i
),
where y
t
is the ground-truth label for relation mention t, W W W
r
,b
r
are the weight and bias of the
classifier. Though this approach has shown improvements, it may produce sub-optimal outcomes
from MTB pretraining since it implicitly assumes that relation mentions corresponding to the same
relation are in the same cluster, while MTB pretraining may learn multiple clusters for a relation.
For example, the entity pairs (Honda Corp., Japan) and (Mount Fuji, Japan), although likely to be
expressed with the same relation country in documents, are likely to be in different clusters since
MTB views them as negative pairs due to different entity names. Therefore, we propose an MCCL
objective that can bridge these gaps. Next, we will discuss the distributional assumption of the
softmax classifier as well as supervised contrastive loss, then present our MCCL objective.
Distributional assumption. We conduct a probing analysis on the distribution of pretrained
representations to further justify the multi-cluster assumption. Specifically, we fix the weights of
the pretrained MTB model and fit different classifiers on top of it, including a softmax classifier,
nearest centroid classifier (both assuming one cluster for a relation), and kNN classifier (assuming
multiple clusters for a relation). We evaluate these classifiers on the test set. Results are shown in
Table 7.1. We find that kNN greatly outperforms other classifiers, showing that MTB pretraining
learns multiple clusters for a relation.
Therefore, to accommodate this multi-cluster assumption, we need to finetune the representations
with a training objective that suits multiple clusters for each relation. Besides using the softmax
classifier with cross-entropy loss, we also consider supervised contrastive loss (SupCon; [142, 143]),
which is shown to be more effective than softmax on NLP tasks. SupCon has a similar loss form to
InfoNCE in Eq. (7.1), except that it uses instances of the same/different relations as positive/negative
pairs. However, previous work [241] has shown that both softmax and SupCon are minimized when
87
the representations of each class collapse to the vertex of a regular simplex. In our case, this means
the relation embedding corresponding to the same relation in pretraining collapses to a single point,
which creates a distributional gap between pretraining and finetuning.
Training objective. We thereby propose the MCCL objective. Given relation mentionsT and sets
of relation mentions grouped by their relations{T
r
}
r∈R
, our loss is formulated as:
w
(t
i
,t
j
)
r
=
e
sim(z z z
t
i
,z z z
t
j
)/τ
1
∑
t
k
∈T
r
\{t
i
}
e
sim(z z z
t
i
,z z z
t
k
)/τ
1
,
s
t
i
r
=
∑
t
j
∈T
r
\{t
i
}
w
(t
i
,t
j
)
r
sim(z z z
t
i
,z z z
t
j
),
P
t
i
r
= softmax((s
t
i
r
+ b
r
)/τ
2
),
L
mccl
=− 1
|T|
∑
t
i
∈T
log(P
t
i
y
t
i
),
whereτ
1
andτ
2
are temperature hyperparameters, b
r
∈R is the classwise bias. The loss calculation
can be split into two parts. First, we calculate the similarity between t
i
and relation r, which is a
weighted average of the similarity between t
i
and t
j
∈T
r
such that a more similar t
j
has a larger
weight. Next, we use the cross-entropy loss to make the similarity of ground-truth relation larger
than others. In this way, MCCL only optimizes t
i
to be similar to a few closest relation mentions of
the ground-truth relation, and thus encourages multiple clusters in relation embedding.
Proxies. We use batched training for finetuning, where relation mentions in the current batch are
used to calculate MCCL. However, it is possible that a subset of relations inR, especially the
long-tail relations, are rare or missing in the current batch. WhenT
r
\{t
i
} is empty, s
t
i
r
and the
MCCL loss become undefined. To tackle this problem, we propose the use of proxies [258, 259].
We add one proxy vector p
r
for each relation r, which is a trainable parameter and associated with
an embedding z z z
p
r
. We incorporate the proxies into MCCL by changingT
r
toT
′
r
=T
r
∪{p
r
},
ensuring thatT
′
r
\{t
i
} is never empty in training and preventing MCCL from becoming undefined.
The proxies are randomly initialized and updated during training by backward propagation.
7.3.4 Inference
We use the classwise kNN [260] for inference, which predicts relations based on similarly repre-
sented instances and thus aligns with our contrastive finetuning objective. Given a new relation
mention to predict, we first find k most similar instances
2
in the training data of each relation
(including NA), then calculate the average cosine similarity of each relation s
avg
r
. Finally, the model
returns the relation with the maximum s
avg
r
+ b
r
for single-label prediction, and all relations with
higher s
avg
r
+ b
r
than NA for multi-label prediction. We use classwise kNN because it is more
suitable for RE datasets, where the label distribution is usually long-tailed [261].
2
Measured by cosine similarity. If a relation has fewer than k mentions in training data, we use all of its mentions.
88
7.4 Experiments
We evaluate our proposed method with a focus on low-resource RE (Sections 7.4.1-7.4.3), and
present detailed analyses (Section 7.4.4) and visualization (Section 7.4.5) to justify method design
choices.
7.4.1 Datasets
We conduct experiments with two document-level RE datasets. The BioRED dataset [10] is a
manually labeled single-label RE dataset in the biomedical domain. The relation mentions are
classified into 9 types (including an NA type indicating no relation). It has a training set consisting
of 400 documents, which we use in finetuning. For pretraining, we use the PubTator Central
corpus [262], which annotates the PubMed corpus with entity mentions and their named entity types.
The Re-DocRED dataset [245] is a multi-label large-scale dataset of the general domain. It is a
relabeled version of the DocRED dataset [20]. Re-DocRED addresses the incomplete annotation
issue of DocRED, where a large percentage of relation mentions are mislabeled as NA. The relation
mentions in Re-DocRED are classified into 97 types (incl. NA). It has a training set consisting of
3,053 documents, which we use in finetuning. For pretraining, we use the distantly labeled training
set provided by DocRED, which consists of 101,873 documents. We remove the relation labels and
use our improved MTB to pretrain the model.
7.4.2 Experimental Setup
Model configurations. We implement our models using Hugging Face Transformers [155]. We
use AdamW [263] in optimization with a weight decay of 0.01. During pretraining, we use a batch
size of 16, a learning rate of 5e-6, a temperature of 0.05, and epochs of 3 and 10 for BioRED and
DocRED, respectively. During finetuning, we use a batch size of 32, a learning rate of 5e-5, and
epochs of 100 and 30 for BioRED and DocRED, respectively. The temperatures in MCCL are set to
τ
1
=τ
2
= 0.2 for BioRED andτ
1
= 0.01,τ
2
= 0.03 for DocRED. We search k from{1,3,5,10,20}
for kNN using the development set
3
. We run experiments with Nvidia V100 GPUs.
Evaluation settings. In this work, in addition to the standard full-shot training, we consider
low-resource settings. To create each of the settings, we randomly sample a fixed proportion p%
of the relation mentions from the training set as our training data, and use the original test set for
evaluation. We use the same evaluation metrics as the original papers. We use micro-F
1
for BioRED,
and micro-F
1
and micro-F
1
-Ign for Re-DocRED. The micro-F
1
-Ign removes the relational facts in
the test set that have appeared in training.
Compared methods. We experiment with the following finetuning objectives: (1) Lazy learning,
which directly uses the pretrained embedding and training data to perform kNN without finetuning;
(2) Cross-entropy loss (CE), which adds a softmax classifier on top of PLM and uses cross-entropy
loss to finetune the model; (3) Supervised contrastive loss (SupCon); and (4) Multi-center
contrastive loss (MCCL). In inference, classwise kNN is used for all methods except for CE. Note
that as SupCon does not apply to multi-label scenarios, we only evaluate it on BioRED. For each
3
For low-resource setting with p% of training data, we sample p% of development data as the development set.
89
Encoder Objective 1% 5% 10% 100%
F
1
F
1
-Ign F
1
F
1
-Ign F
1
F
1
-Ign F
1
F
1
-Ign
PLM
Lazy 15.6 14.9 20.1 19.4 21.6 19.2 28.7 28.0
CE 40.3 38.9 54.1 52.6 61.3 60.3 70.9 69.4
MCCL 44.7 43.1 59.1 57.5 63.2 61.8 68.2 66.7
MTB
Lazy 35.2 34.4 44.7 43.4 47.3 46.2 54.1 52.9
CE 42.3 40.7 57.9 56.4 62.9 61.4 71.2 69.9
MCCL 46.4 44.5 59.7 58.2 63.8 62.1 69.3 67.9
Table 7.2: Results on the test set of Re-DocRED.
Encoder Objective 1% 5% 10% 100%
PLM
Lazy 14.5 17.6 18.8 28.3
CE 24.1 35.4 42.5 57.7
SupCon 20.0 30.9 38.0 52.2
MCCL 20.8 41.3 45.5 55.1
MTB
Lazy 24.3 28.4 34.4 36.7
CE 28.6 41.2 49.8 61.5
SupCon 24.4 29.1 31.4 43.1
MCCL 34.6 48.5 54.2 60.8
Table 7.3: F
1
on the test set of BioRED.
objective, we also evaluate the PLM before and after MTB pretraining. We use different PLMs as
the backbone of the model, namely PubmedBERT
BASE
for BioRED and BERT
BASE
for Re-DocRED,
which are pretrained on the biomedical and general domains, respectively.
7.4.3 Main Results
The results on the test sets of Re-DocRED and BioRED are shown in Table 7.2 and Table 7.3,
respectively. All results are averaged for five runs of training using different random seeds. Overall,
the combination of MTB and MCCL achieves the best performance in low-resource settings where
1%, 5%, and 10% of relation-labeled data are used. Further, when using the same MTB-based
representations, MCCL shows better results than CE in low-resource settings. It shows that in
low-resource settings, MCCL can better leverage the pretraining knowledge with a well-aligned
finetuning objective. However, this improvement diminishes when abundant labeled data are
available, as MCCL underperforms CE on both datasets with full training data on both datasets. In
addition, we observe that MTB pretraining consistently improves MCCL and CE on both datasets.
These results demonstrate the effectiveness of MTB pretraining for more precise document-level
RE with less needed end-task supervision.
Considering other training objectives, we observe that lazy learning produces meaningful results.
On both datasets, the results of lazy learning based on MTB with 10% of data are comparable to
finetuning with 1% of data. This shows that the relation embedding pretrained on unlabeled corpora
90
Pretraining Objective 1% 10% 100%
PLM 20.8 45.5 55.1
vanilla MTB 22.9 45.0 56.0
our MTB 34.6 54.2 60.8
w/oL
rel
21.0 47.1 56.7
w/oL
self
24.1 49.3 58.6
w/oL
mlm
32.9 50.2 58.2
Table 7.4: F
1
on the test set of BioRED with different pretraining objectives. We use MCCL in
finetuning.
contains knowledge that can be transferred to unseen relations. We also observe that SupCon using
kNN-based inference underperforms both CE and MCCL on BioRED, showing that its one-cluster
assumption hurts the knowledge transfer.
7.4.4 Ablation Study
Pretraining objectives. We analyze the effectiveness of our proposed pretraining losses in Sec-
tion 7.3.2. To do so, we pretrain the model with one loss removed at a time while keeping the
finetuning setup on BioRED fixed with the MCCL loss. The results are shown in Table 7.4. Overall,
we observe that all losses are effective. If we remove all proposed techniques and use the vanilla
MTB pretraining objective of binary pairwise classification, the results are only slightly better or
even worse. Among the techniques, removingL
rel
leads to the largest performance drop, showing
that MTB-based pretraining is critical to improve low-resource RE. RemovingL
self
also leads to a
large performance drop. It is becauseL
self
encourages the model to learn more discriminative fea-
tures that lead to less collapsed representations. Our finding aligns with recent studies in computer
vision [264, 265], showing that reducing collapsed representations with self-supervised contrastive
learning improves the transferability to downstream tasks.
Performance w.r.t. different temperatures. We discuss the impact of two temperatures in MCCL.
In MCCL, τ
1
controls the weighting of instances. With a very small τ
1
, each instance will only
form a cluster with its nearest neighbor in the batch, while with very largeτ
1
, instances of the same
relation will collapse to the same cluster. τ
2
controls the importance of hard instances, which is
also used in other contrastive losses (e.g.,τ in Eq. (7.1)). Wang & Liu [266] observe that smallτ
2
makes the model focus more on hard instances, while Khosla, Teterwak, Wang, Sarna, Tian, Isola,
et al. [142] observe that too smallτ
2
leads to numerical instability. We show the results of using
different temperatures in Figure 7.1, where we keep one temperature fixed and change the other.
Forτ
1
, we find that using large temperature harms the performance significantly, showing that our
multi-cluster assumption improves low-resource RE. Forτ
2
, we observe that both small and large
values impair the performance, which is aligned with prior observations.
Performance w.r.t. different amount of data. The main results show that MCCL outperforms
CE in the low-resource setting, while slightly underperforming CE when full training data is used.
We further evaluate MCCL and CE using different amounts of end-task data. We experiment on
BioRED and use the relation embedding pretrained with MTB. Results are shown in Figure 7.2. We
91
0.01 0.20.3 0.5 1.0
Temperature
15
20
25
30
35
40
F1
1
2
Figure 7.1: F
1
using different temperatures on 1% of BioRED.
1 3 5 10 20 50 100
% of Training Data
20
30
40
50
60
F1
MCCL
CE
Figure 7.2: F
1
under different percentages of BioRED training data.
observe that MCCL consistently outperforms CE by a large marge when less than 20% of training
data is used, while it performs similarly or worse than CE after that. It again demonstrates the
effectiveness of MCCL in low-resource RE. However, as the pretraining and finetuning are based
on different tasks, fully adapting the model to downstream data by CE results in similar or better
performance in data-sufficient scenarios.
7.4.5 Visualization
Figure 7.3 shows the t-SNE [156] projection of relation embedding finetuned with different ob-
jectives on BioRED. For clarity, we visualize the embedding of the four most frequent relations
in BioRED with different colors, including the NA class shown in grey. The visualization shows
92
Lazy CE
SupCon MCCL
Figure 7.3: Visualization of relation embedding finetuned with different objectives on BioRED. NA
instances are shown in grey.
that both CE and SupCon learn one cluster for each relation, while lazy learning and MCCL, as
expected, generate multiple small clusters for a relation. This observation indicates that MCCL can
better align with the pretraining objective, further explaining its better performance in low-resource
settings.
7.5 Conclusion
We study self-supervised learning for document-level RE. Our method conducts an improved
MTB pretraining objective that acquires cheap supervision signals from large corpora without
relation labels. To bridge the gap between pretraining and end-task finetuning, we propose a
continual contrastive finetuning objective, in contrast to prior studies that typically use classification-
based finetuning, and use kNN-based inference. As pretrained representation may form multi-
cluster representation, we further propose a multi-center contrastive loss that aligns well with the
nature of the pretrained representation. Extensive experiments on two document-level RE datasets
demonstrate the effectiveness of these key techniques in our method. Future work is adapting our
method to other tasks in information extraction, such as n-ary relation extraction, named entity
recognition, typing, and linking.
93
Chapter 8
Conclusion and Future Work
8.1 Summary
In this dissertation, we aim to build robust and generalizable knowledge acquisition systems. Our
contributions can be summarized into three aspects:
1. We propose a framework for extracting knowledge from long contexts such as documents.
2. We study the robustness of knowledge acquisition systems from multiple aspects, including
robustness to noisy training labels, out-of-distribution instances, and knowledge conflicts.
3. We investigate methods for building data-efficient knowledge acquisition systems. This
includes data annotation by neural rule grounding, as well as improving data efficiency
through continual contrastive finetuning.
More specifically, in Chapter 2, we proposed the ATLOP model for document-level relation
extraction, featuring two techniques, namely adaptive thresholding and localized context pooling.
Adaptive thresholding learned a threshold class that decides the best threshold for each entity pair.
Localized context pooling utilized attention in PLMs to locate salient contexts for entity pairs.
In Chapter 3, we proposed a co-regularization framework for learning from noisy training labels.
This framework involved multiple models with identical structures but different initializations,
optimizing an agreement loss to encourage them to give similar predictions on the same inputs.
On noisy examples, where model predictions often deviate from the labels, the agreement loss
prevented overfitting on noisy labels.
In Chapter 4, we presented an unsupervised OOD detection framework for pretrained Trans-
formers. We systematically investigated the combination of contrastive losses and scoring functions.
Specifically, we proposed a margin-based contrastive objective for learning compact representa-
tions that, when combined with the Mahalanobis distance, achieved near-perfect OOD detection
performance on various tasks and datasets.
In Chapter 5, we addressed the faithfulness issue of LLMs, particularly in scenarios with
knowledge conflict and prediction with abstention. We proposed two effective methods to enhance
LLMs’ faithfulness to contexts: opinion-based prompts and counterfactual demonstrations, which
significantly improve the faithfulness of LLMs to contexts.
In Chapter 6, we proposed the NERO framework for label-efficient relation extraction. The
framework involved automatically extracting frequent patterns from large raw corpora and requesting
94
users to create labeling rules. To improve the coverage of these rules, we proposed a soft-matching
mechanism, where unlabeled sentences are annotated using their most semantically similar labeling
rules and weighted when training the RE model.
Finally, in Chapter 7, we proposed a self-supervised framework to utilize unlabeled data in
knowledge acquisition. We pretrained the model using an improved MTB pretraining objective. To
address the gap between pretraining and end-task finetuning, we proposed a continual contrastive
finetuning objective and kNN-based inference. We further proposed a multi-center contrastive loss
that aligns well with the nature of the pretrained representation.
This dissertation thoroughly addresses several challenges of knowledge acquisition in real-world
applications. The methods proposed in this research support a wide range of applications, including
but not limited to constructing knowledge bases, handling knowledge-intensive NLP tasks such as
question answering, fact verification, and dialogue, as well as biomedical and clinical applications
such as in-silico biomedical research and precision medicine.
8.2 Future Directions
We discuss several promising future directions.
• Universal knowledge acquisition. Most current knowledge acquisition systems are designed
for specific tasks (such as entities, events, etc.) or domains (such as general, biomedical,
finance, etc.). Unifying these domains and tasks offers two significant advantages. First,
many knowledge acquisition tasks require common information, such as syntactic features
like dependencies, which are essential for both entity [28] and event-centric relation extrac-
tion [267]. In addition, the outputs of entity or event detection tasks are typically used as
inputs for relation extraction tasks. As such, unifying different knowledge acquisition tasks
can improve the ability to learn and utilize shared cross-task information. Second, by jointly
learning across multiple domains, a knowledge acquisition system can learn transferable
features that generalize to different domains, resulting in improved domain generalization
ability. As a result, universal knowledge acquisition has the potential to perform any task in
any distribution in a zero-shot manner, which can significantly benefit real-world applications,
particularly in high-stakes domains.
• Cross-document knowledge acquisition. In addition to relations within a single document,
some relations may require complex reasoning over multiple documents. As highlighted
in a recent quantitative study [268], over 57.6% of relational facts in Wikidata involve
subject and object entities that do not appear in the same Wikipedia document. Therefore,
cross-document knowledge acquisition represents a more practical direction for developing
practical knowledge acquisition systems. This problem also poses new research challenges,
such as retrieving salient contexts from multiple documents and performing multi-step cross-
document relation reasoning.
• Fact-checking for large language models. Large language models such as ChatGPT are
increasingly popular in natural language processing. However, their responses may contain
factual errors despite being seemingly plausible [269]. Therefore, fact-checking the outputs
of these models is critical for ensuring AI safety. To achieve this, knowledge acquisition
95
systems can be used to convert the models’ responses into structured formats, making them
easier to align and verify against existing knowledge bases. This approach can enhance
the accuracy of language models and reduce the potential harm caused by misinformation,
ultimately advancing the development of responsible and reliable AI systems.
96
Bibliography
1. Hendrycks, D., Liu, X., Wallace, E., Dziedzic, A., Krishnan, R. & Song, D. Pretrained
Transformers Improve Out-of-Distribution Robustness in Proceedings of the 58th Annual
Meeting of the Association for Computational Linguistics (Online, 2020), 2744–2751.
2. Zhang, Y ., Dai, H., Kozareva, Z., Smola, A. & Song, L. Variational reasoning for question
answering with knowledge graph in Proceedings of the AAAI conference on artificial
intelligence 32 (2018).
3. Huang, X., Zhang, J., Li, D. & Li, P. Knowledge graph embedding based question answering
in Proceedings of the twelfth ACM international conference on web search and data mining
(2019), 105–113.
4. Shiralkar, P., Flammini, A., Menczer, F. & Ciampaglia, G. L. Finding streams in knowledge
graphs to support fact checking in 2017 IEEE International Conference on Data Mining
(ICDM) (2017), 859–864.
5. Zhou, J., Han, X., Yang, C., Liu, Z., Wang, L., Li, C., et al. GEAR: Graph-based Evidence
Aggregating and Reasoning for Fact Verification in Proceedings of the 57th Annual Meeting
of the Association for Computational Linguistics (Association for Computational Linguistics,
Florence, Italy, 2019), 892–901. doi:10.18653/v1/P19-1085.
6. Eric, M., Krishnan, L., Charette, F. & Manning, C. D. Key-Value Retrieval Networks for
Task-Oriented Dialogue in Proceedings of the 18th Annual SIGdial Meeting on Discourse
and Dialogue (Association for Computational Linguistics, Saarbr¨ ucken, Germany, 2017),
37–49. doi:10.18653/v1/W17-5506.
7. Yang, S., Zhang, R. & Erfani, S. GraphDialog: Integrating Graph Knowledge into
End-to-End Task-Oriented Dialogue Systems in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP) (Association for
Computational Linguistics, Online, 2020), 1878–1888.
doi:10.18653/v1/2020.emnlp-main.147.
8. Herrero-Zazo, M., Segura-Bedmar, I., Martınez, P. & Declerck, T. The DDI corpus: An
annotated corpus with pharmacological substances and drug–drug interactions. Journal of
biomedical informatics 46, 914–920 (2013).
97
9. Singhal, A., Simmons, M. & Lu, Z. Text mining genotype-phenotype relationships from
biomedical literature for database curation and precision medicine. PLoS computational
biology 12, e1005017 (2016).
10. Luo, L., Lai, P.-T., Wei, C.-H., Arighi, C. N. & Lu, Z. BioRED: a rich biomedical relation
extraction dataset. Briefings in Bioinformatics 23, bbac282 (2022).
11. Lv, X., Guan, Y ., Yang, J. & Wu, J. Clinical relation extraction with deep learning.
International Journal of Hybrid Information Technology 9, 237–248 (2016).
12. Wei, Q., Ji, Z., Si, Y ., Du, J., Wang, J., Tiryaki, F., et al. Relation extraction from clinical
narratives using pre-trained language models in AMIA annual symposium proceedings 2019
(2019), 1236.
13. Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep
Bidirectional Transformers for Language Understanding in Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational Linguistics:
Human Language Technologies, Volume 1 (Long and Short Papers) (Minneapolis,
Minnesota, 2019), 4171–4186.
14. Alt, C., H¨ ubner, M. & Hennig, L. Improving Relation Extraction by Pre-trained Language
Representations in AKBC (2019).
15. Shi, P. & Lin, J. Simple BERT Models for Relation Extraction and Semantic Role Labeling.
ArXiv abs/1904.05255 (2019).
16. Zhou, W. & Chen, M. An Improved Baseline for Sentence-level Relation Extraction in
Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for
Computational Linguistics and the 12th International Joint Conference on Natural
Language Processing (Volume 2: Short Papers) (Association for Computational Linguistics,
Online only, 2022), 161–168.
17. Yao, Y ., Ye, D., Li, P., Han, X., Lin, Y ., Liu, Z., et al. DocRED: A Large-Scale
Document-Level Relation Extraction Dataset in ACL (2019).
18. Hendrickx, I., Kim, S. N., Kozareva, Z., Nakov, P.,
´
O S´ eaghdha, D., Pad´ o, S., et al.
SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations between Pairs of
Nominals in Proceedings of the 5th International Workshop on Semantic Evaluation
(Association for Computational Linguistics, Uppsala, Sweden, 2010), 33–38.
19. Zhang, Y ., Zhong, V ., Chen, D., Angeli, G. & Manning, C. D. Position-aware Attention and
Supervised Data Improve Slot Filling in Proceedings of the 2017 Conference on Empirical
Methods in Natural Language Processing (Association for Computational Linguistics,
Copenhagen, Denmark, 2017), 35–45. doi:10.18653/v1/D17-1004.
98
20. Yao, Y ., Ye, D., Li, P., Han, X., Lin, Y ., Liu, Z., et al. DocRED: A Large-Scale
Document-Level Relation Extraction Dataset in Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics (Association for Computational Linguistics,
Florence, Italy, 2019), 764–777. doi:10.18653/v1/P19-1074.
21. Mintz, M., Bills, S., Snow, R. & Jurafsky, D. Distant supervision for relation extraction
without labeled data in Proceedings of the Joint Conference of the 47th Annual Meeting of
the ACL and the 4th International Joint Conference on Natural Language Processing of the
AFNLP (Association for Computational Linguistics, Suntec, Singapore, 2009), 1003–1011.
22. Zeng, D., Liu, K., Chen, Y . & Zhao, J. Distant Supervision for Relation Extraction via
Piecewise Convolutional Neural Networks in Proceedings of the 2015 Conference on
Empirical Methods in Natural Language Processing (Association for Computational
Linguistics, Lisbon, Portugal, 2015), 1753–1762.
23. Han, X., Zhu, H., Yu, P., Wang, Z., Yao, Y ., Liu, Z., et al. FewRel: A Large-Scale Supervised
Few-Shot Relation Classification Dataset with State-of-the-Art Evaluation in Proceedings of
the 2018 Conference on Empirical Methods in Natural Language Processing (Association
for Computational Linguistics, Brussels, Belgium, 2018), 4803–4809.
doi:10.18653/v1/D18-1514.
24. Sainz, O., Lopez de Lacalle, O., Labaka, G., Barrena, A. & Agirre, E. Label Verbalization
and Entailment for Effective Zero and Few-Shot Relation Extraction in Proceedings of the
2021 Conference on Empirical Methods in Natural Language Processing (Association for
Computational Linguistics, Online and Punta Cana, Dominican Republic, 2021), 1199–1212.
doi:10.18653/v1/2021.emnlp-main.92.
25. Patterson, S. E., Liu, R., Statz, C. M., Durkin, D., Lakshminarayana, A. & Mockus, S. M.
The clinical trial landscape in oncology and connectivity of somatic mutational profiles to
targeted therapies. Human genomics 10, 1–13 (2016).
26. Zeng, D., Liu, K., Lai, S., Zhou, G. & Zhao, J. Relation Classification via Convolutional
Deep Neural Network in COLING (2014).
27. Miwa, M. & Bansal, M. End-to-End Relation Extraction using LSTMs on Sequences and
Tree Structures in ACL (2016).
28. Zhang, Y ., Qi, P. & Manning, C. D. Graph Convolution over Pruned Dependency Trees
Improves Relation Extraction in Proceedings of the 2018 Conference on Empirical Methods
in Natural Language Processing (Association for Computational Linguistics, Brussels,
Belgium, 2018), 2205–2215. doi:10.18653/v1/D18-1244.
29. Verga, P., Strubell, E. & McCallum, A. Simultaneously Self-Attending to All Mentions for
Full-Abstract Biological Relation Extraction in NAACL-HLT (2018).
99
30. Peng, N., Poon, H., Quirk, C., Toutanova, K. & Yih, W.-t. Cross-Sentence N-ary Relation
Extraction with Graph LSTMs. Transactions of the Association for Computational
Linguistics 5, 101–115 (2017).
31. Liu, Y . & Lapata, M. Learning Structured Text Representations. Transactions of the
Association for Computational Linguistics 6, 63–75 (2018).
32. Christopoulou, F., Miwa, M. & Ananiadou, S. Connecting the Dots: Document-level Neural
Relation Extraction with Edge-oriented Graphs in EMNLP-IJCNLP (2019).
33. Nan, G., Guo, Z., Sekulic, I. & Lu, W. Reasoning with Latent Structure Refinement for
Document-Level Relation Extraction in ACL (2020).
34. Liang, X., Shen, X., Feng, J., Lin, L. & Yan, S. Semantic Object Parsing with Graph LSTM
in ECCV (2016).
35. Guo, Z., Zhang, Y . & Lu, W. Attention Guided Graph Convolutional Networks for Relation
Extraction in ACL (2019).
36. Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural computation 9,
1735–1780 (1997).
37. Chung, J., G¨ ulc ¸ehre, C ¸ ., Cho, K. & Bengio, Y . Empirical Evaluation of Gated Recurrent
Neural Networks on Sequence Modeling. ArXiv abs/1412.3555 (2014).
38. Khandelwal, U., He, H., Qi, P. & Jurafsky, D. Sharp Nearby, Fuzzy Far Away: How Neural
Language Models Use Context in ACL (2018).
39. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. Attention
is All you Need in NeurIPS (2017).
40. Clark, K., Khandelwal, U., Levy, O. & Manning, C. D. What Does BERT Look at? An
Analysis of BERT’s Attention in BlackboxNLP workshop (2019).
41. Tenney, I., Das, D. & Pavlick, E. BERT Rediscovers the Classical NLP Pipeline in ACL
(2019).
42. Wang, H., Focke, C., Sylvester, R., Mishra, N. & Wang, W. W. J. Fine-tune Bert for
DocRED with Two-step Process. ArXiv abs/1909.11898 (2019).
43. Tang, H., Cao, Y ., Zhang, Z., Cao, J., Fang, F., Wang, S., et al. HIN: Hierarchical Inference
Network for Document-Level Relation Extraction in PAKDD (2020).
44. Liu, Y ., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., et al. Roberta: A robustly optimized
bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019).
100
45. Li, J., Sun, Y ., Johnson, R. J., Sciaky, D., Wei, C.-H., Leaman, R., et al. BioCreative V CDR
task corpus: a resource for chemical disease relation extraction in Database (2016).
46. Wu, Y ., Luo, R., Leung, H. C., Ting, H.-F. & Lam, T.-W. RENET: A Deep Learning
Approach for Extracting Gene-Disease Associations from Literature in RECOMB (2019).
47. Soares, L. B., FitzGerald, N., Ling, J. & Kwiatkowski, T. Matching the Blanks:
Distributional Similarity for Relation Learning in ACL (2019).
48. Wang, H., Tan, M., Yu, M., Chang, S., Wang, D., Xu, K., et al. Extracting
Multiple-Relations in One-Pass with Pre-Trained Transformers in ACL (2019).
49. Jia, R., Wong, C. & Poon, H. Document-Level N-ary Relation Extraction with Multiscale
Representation Learning in NAACL-HLT (2019).
50. Zheng, H., Fu, J., Zha, Z.-J. & Luo, J. Learning Deep Bilinear Transformation for
Fine-grained Image Representation in NeurIPS (2019).
51. Tang, Y ., Huang, J., Wang, G., He, X. & Zhou, B. Orthogonal Relation Transforms with
Graph Context Modeling for Knowledge Graph Embedding in ACL (2020).
52. Menon, A., Rawat, A., Reddi, S. & Kumar, S. Multilabel reductions: what is my loss
optimising? in NeurIPS (2019).
53. Reddi, S. J., Kale, S., Yu, F., Holtmann-Rice, D., Chen, J. & Kumar, S. Stochastic Negative
Mining for Learning with Large Output Spaces in AISTATS (2019).
54. Ye, D., Lin, Y ., Du, J., Liu, Z., Sun, M. & Liu, Z. Coreferential Reasoning Learning for
Language Representation in EMNLP (2020).
55. Nguyen, D. Q. & Verspoor, K. Convolutional neural networks for chemical-disease relation
extraction are improved with character-based word embeddings in BioNLP workshop
(2018).
56. Wolf, T., Debut, L., Sanh, V ., Chaumond, J., Delangue, C., Moi, A., et al. HuggingFace’s
Transformers: State-of-the-art Natural Language Processing. ArXiv, arXiv–1910 (2019).
57. Beltagy, I., Lo, K. & Cohan, A. SciBERT: A Pretrained Language Model for Scientific Text
in EMNLP-IJCNLP (2019).
58. Micikevicius, P., Narang, S., Alben, J., Diamos, G., Elsen, E., Garc´ ıa, D., et al. Mixed
Precision Training in ICLR (2018).
59. Loshchilov, I. & Hutter, F. Decoupled Weight Decay Regularization in 7th International
Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9,
2019 (OpenReview.net, 2019).
101
60. Goyal, P., Doll´ ar, P., Girshick, R. B., Noordhuis, P., Wesolowski, L., Kyrola, A., et al.
Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. ArXiv abs/1706.02677
(2017).
61. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a
simple way to prevent neural networks from overfitting. The journal of machine learning
research 15, 1929–1958 (2014).
62. Goodfellow, I., Bengio, Y ., Courville, A. & Bengio, Y . Deep learning (MIT press
Cambridge, 2016).
63. Schuster, M. & Paliwal, K. Bidirectional recurrent neural networks. IEEE Trans. Signal
Process. 45, 2673–2681 (1997).
64. Kipf, T. N. & Welling, M. Semi-Supervised Classification with Graph Convolutional
Networks in International Conference on Learning Representations (ICLR) (2017).
65. Fan, R.-E. & Lin, C.-J. A study on threshold selection for multi-label classification.
Department of Computer Science, National Taiwan University, 1–23 (2007).
66. Wang, L., Cao, Z., De Melo, G. & Liu, Z. Relation Classification via Multi-Level Attention
CNNs in ACL (2016).
67. Wu, F., Zhang, T., Souza, A., Fifty, C., Yu, T. & Weinberger, K. Q. Simplifying Graph
Convolutional Networks in ICML (2019).
68. Zhang, Z., Han, X., Liu, Z., Jiang, X., Sun, M. & Liu, Q. ERNIE: Enhanced Language
Representation with Informative Entities in ACL (2019).
69. Quirk, C. & Poon, H. Distant Supervision for Relation Extraction beyond the Sentence
Boundary in EACL (2017).
70. Song, L., Zhang, Y ., Wang, Z. & Gildea, D. N-ary Relation Extraction using Graph-State
LSTM in EMNLP (2018).
71. Gupta, P., Rajaram, S., Sch¨ utze, H. & Runkler, T. Neural Relation Extraction Within and
Across Sentence Boundaries in AAAI (2019).
72. Christopoulou, F., Miwa, M. & Ananiadou, S. A Walk-based Model on Entity Graphs for
Relation Extraction in ACL (2018).
73. Vig, J. & Belinkov, Y . Analyzing the Structure of Attention in a Transformer Language
Model in BlackboxNLP workshop (2019).
74. Hewitt, J. & Manning, C. D. A Structural Probe for Finding Syntax in Word Representations
in NAACL-HLT (2019).
102
75. Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M. S., et al. A Closer
Look at Memorization in Deep Networks in Proceedings of the 34th International
Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017
(eds Precup, D. & Teh, Y . W.) 70 (PMLR, 2017), 233–242.
76. Zhang, C., Bengio, S., Hardt, M., Recht, B. & Vinyals, O. Understanding deep learning
requires rethinking generalization in 5th International Conference on Learning
Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track
Proceedings (OpenReview.net, 2017).
77. Raykar, V . C., Yu, S., Zhao, L. H., Valadez, G. H., Florin, C., Bogoni, L., et al. Learning
from crowds. Journal of Machine Learning Research 11 (2010).
78. Song, Y ., Wang, C., Zhang, M., Sun, H. & Yang, Q. Spectral Label Refinement for Noisy and
Missing Text Labels in Proceedings of the Twenty-Ninth AAAI Conference on Artificial
Intelligence, January 25-30, 2015, Austin, Texas, USA (eds Bonet, B. & Koenig, S.) (AAAI
Press, 2015), 2972–2978.
79. Sang, E. T. K. & De Meulder, F. Introduction to the CoNLL-2003 Shared Task:
Language-Independent Named Entity Recognition in Proceedings of CoNLL-2003,
Edmonton, Canada (2003), 142–145.
80. Reiss, F., Xu, H., Cutler, B., Muthuraman, K. & Eichenberger, Z. Identifying Incorrect
Labels in the CoNLL-2003 Corpus in Proceedings of the 24th Conference on Computational
Natural Language Learning (Association for Computational Linguistics, Online, 2020),
215–226.
81. Alt, C., Gabryszak, A. & Hennig, L. TACRED Revisited: A Thorough Evaluation of the
TACRED Relation Extraction Task in Proceedings of the 58th Annual Meeting of the
Association for Computational Linguistics (Association for Computational Linguistics,
Online, 2020), 1558–1569.
82. Surdeanu, M., Tibshirani, J., Nallapati, R. & Manning, C. D. Multi-instance Multi-label
Learning for Relation Extraction in Proceedings of the 2012 Joint Conference on Empirical
Methods in Natural Language Processing and Computational Natural Language Learning
(Association for Computational Linguistics, Jeju Island, Korea, 2012), 455–465.
83. Ratner, A. J., De Sa, C. M., Wu, S., Selsam, D. & R´ e, C. Data programming: Creating large
training sets, quickly in Advances in neural information processing systems (2016),
3567–3575.
84. Huang, Y . & Du, J. Self-Attention Enhanced CNNs and Collaborative Curriculum Learning
for Distantly Supervised Relation Extraction in Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing and the 9th International Joint
Conference on Natural Language Processing (EMNLP-IJCNLP) (Association for
Computational Linguistics, Hong Kong, China, 2019), 389–398.
103
85. Mayhew, S., Chaturvedi, S., Tsai, C.-T. & Roth, D. Named Entity Recognition with Partially
Annotated Training Data in Proceedings of the 23rd Conference on Computational Natural
Language Learning (CoNLL) (Association for Computational Linguistics, Hong Kong,
China, 2019), 645–655.
86. Qin, P., Xu, W. & Wang, W. Y . Robust Distant Supervision Relation Extraction via Deep
Reinforcement Learning in Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers) (Association for Computational
Linguistics, Melbourne, Australia, 2018), 2137–2147.
87. Wang, Z., Shang, J., Liu, L., Lu, L., Liu, J. & Han, J. CrossWeigh: Training Named Entity
Tagger from Imperfect Annotations in Proceedings of the 2019 Conference on Empirical
Methods in Natural Language Processing and the 9th International Joint Conference on
Natural Language Processing (EMNLP-IJCNLP) (Association for Computational
Linguistics, Hong Kong, China, 2019), 5154–5163.
88. Toneva, M., Sordoni, A., des Combes, R. T., Trischler, A., Bengio, Y . & Gordon, G. J. An
Empirical Study of Example Forgetting during Deep Neural Network Learning in 7th
International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA,
May 6-9, 2019 (OpenReview.net, 2019).
89. Jiang, L., Zhou, Z., Leung, T., Li, L. & Fei-Fei, L. MentorNet: Learning Data-Driven
Curriculum for Very Deep Neural Networks on Corrupted Labels in Proceedings of the 35th
International Conference on Machine Learning, ICML 2018, Stockholmsm¨ assan, Stockholm,
Sweden, July 10-15, 2018 (eds Dy, J. G. & Krause, A.) 80 (PMLR, 2018), 2309–2318.
90. Han, B., Yao, Q., Yu, X., Niu, G., Xu, M., Hu, W., et al. Co-teaching: Robust training of
deep neural networks with extremely noisy labels in Advances in Neural Information
Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018,
NeurIPS 2018, December 3-8, 2018, Montr´ eal, Canada (eds Bengio, S., Wallach, H. M.,
Larochelle, H., Grauman, K., Cesa-Bianchi, N. & Garnett, R.) (2018), 8536–8546.
91. Ramshaw, L. & Marcus, M. Text Chunking using Transformation-Based Learning in Third
Workshop on Very Large Corpora (1995).
92. Yamada, I., Asai, A., Shindo, H., Takeda, H. & Matsumoto, Y . LUKE: Deep Contextualized
Entity Representations with Entity-aware Self-attention in Proceedings of the 2020
Conference on Empirical Methods in Natural Language Processing (EMNLP) (Association
for Computational Linguistics, Online, 2020), 6442–6454.
93. Wolf, T., Debut, L., Sanh, V ., Chaumond, J., Delangue, C., Moi, A., et al. Transformers:
State-of-the-Art Natural Language Processing in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing: System Demonstrations (Association
for Computational Linguistics, Online, 2020), 38–45.
104
94. Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization in International
Conference for Learning Representations (2015).
95. Lee, J. & Chung, S.-Y . Robust training with ensemble consensus in International Conference
on Learning Representations (2020).
96. Lin, Y ., Shen, S., Liu, Z., Luan, H. & Sun, M. Neural Relation Extraction with Selective
Attention over Instances in Proceedings of the 54th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers) (Association for Computational
Linguistics, Berlin, Germany, 2016), 2124–2133.
97. Ji, G., Liu, K., He, S. & Zhao, J. Distant Supervision for Relation Extraction with
Sentence-Level Attention and Entity Descriptions in Proceedings of the Thirty-First AAAI
Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA
(eds Singh, S. P. & Markovitch, S.) (AAAI Press, 2017), 3060–3066.
98. Liu, T., Wang, K., Chang, B. & Sui, Z. A Soft-label Method for Noise-tolerant Distantly
Supervised Relation Extraction in Proceedings of the 2017 Conference on Empirical
Methods in Natural Language Processing (Association for Computational Linguistics,
Copenhagen, Denmark, 2017), 1790–1795.
99. Yang, K., He, L., Dai, X.-y., Huang, S. & Chen, J. Exploiting Noisy Data in Distant
Supervision Relation Classification in Proceedings of the 2019 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long and Short Papers) (Association for Computational
Linguistics, Minneapolis, Minnesota, 2019), 3216–3225.
100. Wang, Z., Wen, R., Chen, X., Huang, S.-L., Zhang, N. & Zheng, Y . Finding Influential
Instances for Distantly Supervised Relation Extraction. ArXiv abs/2009.09841 (2020).
101. Zhang, Z. & Sabuncu, M. R. Generalized Cross Entropy Loss for Training Deep Neural
Networks with Noisy Labels in Advances in Neural Information Processing Systems 31:
Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018,
December 3-8, 2018, Montr´ eal, Canada (eds Bengio, S., Wallach, H. M., Larochelle, H.,
Grauman, K., Cesa-Bianchi, N. & Garnett, R.) (2018), 8792–8802.
102. Wang, Y ., Ma, X., Chen, Z., Luo, Y ., Yi, J. & Bailey, J. Symmetric Cross Entropy for Robust
Learning With Noisy Labels in 2019 IEEE/CVF International Conference on Computer
Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019 (IEEE, 2019),
322–330.
103. Sukhbaatar, S., Bruna, J., Paluri, M., Bourdev, L. D. & Fergus, R. Training Convolutional
Networks with Noisy Labels in ICLR (2015).
105
104. Goldberger, J. & Ben-Reuven, E. Training deep neural-networks using a noise adaptation
layer in 5th International Conference on Learning Representations, ICLR 2017, Toulon,
France, April 24-26, 2017, Conference Track Proceedings (OpenReview.net, 2017).
105. Wang, R., Liu, T. & Tao, D. Multiclass learning with partially corrupted labels. IEEE
transactions on neural networks and learning systems 29, 2568–2580 (2017).
106. Chang, H.-S., Learned-Miller, E. & McCallum, A. Active Bias: Training More Accurate
Neural Networks by Emphasizing High Variance Samples in NeurIPS (2017).
107. Krogh, A. & Hertz, J. A. A simple weight decay can improve generalization in Advances in
neural information processing systems (1992), 950–957.
108. M¨ uller, R., Kornblith, S. & Hinton, G. When does label smoothing help? in Proceedings of
the 33rd International Conference on Neural Information Processing Systems (2019),
4694–4703.
109. Malach, E. & Shalev-Shwartz, S. Decoupling ”when to update” from ”how to update” in
Advances in Neural Information Processing Systems 30: Annual Conference on Neural
Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA (eds
Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V . N.,
et al.) (2017), 960–970.
110. Yu, X., Han, B., Yao, J., Niu, G., Tsang, I. W. & Sugiyama, M. How does Disagreement
Help Generalization against Label Corruption? in Proceedings of the 36th International
Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California,
USA (eds Chaudhuri, K. & Salakhutdinov, R.) 97 (PMLR, 2019), 7164–7173.
111. Wei, H., Feng, L., Chen, X. & An, B. Combating Noisy Labels by Agreement: A Joint
Training Method with Co-Regularization in 2020 IEEE/CVF Conference on Computer
Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 (IEEE,
2020), 13723–13732.
112. Wang, H., Liu, B., Li, C., Yang, Y . & Li, T. Learning with Noisy Labels for Sentence-level
Sentiment Classification in Proceedings of the 2019 Conference on Empirical Methods in
Natural Language Processing and the 9th International Joint Conference on Natural
Language Processing (EMNLP-IJCNLP) (Association for Computational Linguistics, Hong
Kong, China, 2019), 6286–6292.
113. Wang, R., Liu, T. & Tao, D. Multiclass Learning With Partially Corrupted Labels. IEEE
Transactions on Neural Networks and Learning Systems 29, 2568–2580 (2018).
114. Cheng, J., Liu, T., Ramamohanarao, K. & Tao, D. Learning with Bounded Instance and
Label-dependent Label Noise in Proceedings of the 37th International Conference on
Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event 119 (PMLR, 2020),
1789–1799.
106
115. Chen, M., Zhang, H., Ning, Q., Li, M., Ji, H., McKeown, K., et al. Event-Centric Natural
Language Processing in Proceedings of the 59th Annual Meeting of the Association for
Computational Linguistics (2021).
116. Peng, H., Khashabi, D. & Roth, D. Solving Hard Coreference Problems in Proceedings of
the 2015 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies (2015), 809–819.
117. Hsu, Y .-C., Shen, Y ., Jin, H. & Kira, Z. Generalized ODIN: Detecting Out-of-Distribution
Image Without Learning From Out-of-Distribution Data. 2020 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 10948–10957 (2020).
118. Hendrycks, D., Mazeika, M. & Dietterich, T. Deep Anomaly Detection with Outlier
Exposure in International Conference on Learning Representations (2018).
119. Larson, S., Mahendran, A., Peper, J. J., Clarke, C., Lee, A., Hill, P., et al. An Evaluation
Dataset for Intent Classification and Out-of-Scope Prediction in Proceedings of the 2019
Conference on Empirical Methods in Natural Language Processing and the 9th International
Joint Conference on Natural Language Processing (EMNLP-IJCNLP) (2019), 1311–1316.
120. Shu, L., Xu, H. & Liu, B. DOC: Deep Open Classification of Text Documents in
Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing
(Association for Computational Linguistics, Copenhagen, Denmark, 2017), 2911–2916.
121. Dhamija, A., G¨ unther, M. & Boult, T. Reducing Network Agnostophobia in NeurIPS (2018).
122. Liang, S., Li, Y . & Srikant, R. Enhancing The Reliability of Out-of-distribution Image
Detection in Neural Networks in International Conference on Learning Representations
(2018).
123. Sohn, K. Improved deep metric learning with multi-class n-pair loss objective in
Proceedings of the 30th International Conference on Neural Information Processing Systems
(2016), 1857–1865.
124. Oord, A. v. d., Li, Y . & Vinyals, O. Representation learning with contrastive predictive
coding. arXiv preprint arXiv:1807.03748 (2018).
125. Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive
learning of visual representations in International conference on machine learning (2020),
1597–1607.
126. Lee, K., Lee, K., Lee, H. & Shin, J. A simple unified framework for detecting
out-of-distribution samples and adversarial attacks in Proceedings of the 32nd International
Conference on Neural Information Processing Systems (2018), 7167–7177.
107
127. Amodei, D., Olah, C., Steinhardt, J., Christiano, P. F., Schulman, J. & Man´ e, D. Concrete
Problems in AI Safety. ArXiv abs/1606.06565 (2016).
128. Lee, K., Lee, H., Lee, K. & Shin, J. Training Confidence-calibrated Classifiers for Detecting
Out-of-Distribution Samples in International Conference on Learning Representations
(2018).
129. Bergman, L. & Hoshen, Y . Classification-Based Anomaly Detection for General Data in
International Conference on Learning Representations abs/2005.02359 (2020).
130. Bendale, A. & Boult, T. Towards Open Set Deep Networks. 2016 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 1563–1572 (2016).
131. Hendrycks, D. & Gimpel, K. A baseline for detecting misclassified and out-of-distribution
examples in neural networks in International Conference on Learning Representations
(2017).
132. Liu, W., Wang, X., Owens, J. & Li, Y . Energy-based Out-of-distribution Detection in
Advances in Neural Information Processing Systems 33 (2020).
133. Vyas, A., Jammalamadaka, N., Zhu, X., Das, D., Kaul, B. & Willke, T. L.
Out-of-distribution detection using an ensemble of self supervised leave-out classifiers in
Proceedings of the European Conference on Computer Vision (ECCV) (2018), 550–564.
134. Kim, J.-K. & Kim, Y .-B. Joint Learning of Domain Classification and Out-of-Domain
Detection with Dynamic Class Weighting for Satisficing False Acceptance Rates. Proc.
Interspeech 2018, 556–560 (2018).
135. Tan, M., Yu, Y ., Wang, H., Wang, D., Potdar, S., Chang, S., et al. Out-of-Domain Detection
for Low-Resource Text Classification Tasks in Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing and the 9th International Joint
Conference on Natural Language Processing (EMNLP-IJCNLP) (Hong Kong, China, 2019),
3566–3572.
136. Wang, T. & Isola, P. Understanding contrastive representation learning through alignment
and uniformity on the hypersphere in International Conference on Machine Learning (2020),
9929–9939.
137. Misra, I. & Maaten, L. v. d. Self-supervised learning of pretext-invariant representations in
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
(2020), 6707–6717.
138. He, K., Fan, H., Wu, Y ., Xie, S. & Girshick, R. Momentum contrast for unsupervised visual
representation learning in Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition (2020), 9729–9738.
108
139. Giorgi, J., Nitski, O., Wang, B. & Bader, G. DeCLUTR: Deep Contrastive Learning for
Unsupervised Textual Representations in Proceedings of the 59th Annual Meeting of the
Association for Computational Linguistics and the 11th International Joint Conference on
Natural Language Processing (Volume 1: Long Papers) (Association for Computational
Linguistics, Online, 2021), 879–895.
140. Tack, J., Mo, S., Jeong, J. & Shin, J. CSI: Novelty Detection via Contrastive Learning on
Distributionally Shifted Instances in 34th Conference on Neural Information Processing
Systems (NeurIPS) 2020 (2020).
141. Winkens, J., Bunel, R., Roy, A. G., Stanforth, R., Natarajan, V ., Ledsam, J. R., et al.
Contrastive training for improved out-of-distribution detection. arXiv preprint
arXiv:2007.05566 (2020).
142. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y ., Isola, P., et al. Supervised Contrastive
Learning. Advances in Neural Information Processing Systems 33 (2020).
143. Gunel, B., Du, J., Conneau, A. & Stoyanov, V . Supervised Contrastive Learning for
Pre-trained Language Model Fine-tuning in International Conference for Learning
Representations (2021).
144. Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., et al. Recursive Deep
Models for Semantic Compositionality Over a Sentiment Treebank in Proceedings of the
2013 Conference on Empirical Methods in Natural Language Processing (Association for
Computational Linguistics, Seattle, Washington, USA, 2013), 1631–1642.
145. Maas, A. L., Daly, R. E., Pham, P. T., Huang, D., Ng, A. Y . & Potts, C. Learning Word
Vectors for Sentiment Analysis in Proceedings of the 49th Annual Meeting of the Association
for Computational Linguistics: Human Language Technologies (Association for
Computational Linguistics, Portland, Oregon, USA, 2011), 142–150.
146. Lang, K. in Machine Learning Proceedings 1995 331–339 (Elsevier, 1995).
147. Li, X. & Roth, D. Learning question classifiers in COLING 2002: The 19th International
Conference on Computational Linguistics (2002).
148. Dagan, I., Glickman, O. & Magnini, B. The PASCAL Recognising Textual Entailment
Challenge in MLCW (2005).
149. Bar-Haim, R., Dagan, I., Dolan, B., Ferro, L., Giampiccolo, D. & Magnini, B. The Second
PASCAL Recognising Textual Entailment Challenge in (2006).
150. Giampiccolo, D., Magnini, B., Dagan, I. & Dolan, W. The Third PASCAL Recognizing
Textual Entailment Challenge in ACL-PASCAL@ACL (2007).
109
151. Bentivogli, L., Clark, P., Dagan, I. & Giampiccolo, D. The Sixth PASCAL Recognizing
Textual Entailment Challenge in TAC (2009).
152. Williams, A., Nangia, N. & Bowman, S. A Broad-Coverage Challenge Corpus for Sentence
Understanding through Inference in Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long Papers) (Association for Computational Linguistics, New
Orleans, Louisiana, 2018), 1112–1122.
153. Bojar, O., Chatterjee, R., Federmann, C., Graham, Y ., Haddow, B., Huck, M., et al. Findings
of the 2016 conference on machine translation in Proceedings of the First Conference on
Machine Translation: Volume 2, Shared Task Papers (2016), 131–198.
154. Elliott, D., Frank, S., Sima’an, K. & Specia, L. Multi30K: Multilingual English-German
Image Descriptions in Proceedings of the 5th Workshop on Vision and Language
(Association for Computational Linguistics, Berlin, Germany, 2016), 70–74.
155. Wolf, T., Debut, L., Sanh, V ., Chaumond, J., Delangue, C., Moi, A., et al. Transformers:
State-of-the-Art Natural Language Processing in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing: System Demonstrations (Online, 2020),
38–45.
156. Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. Journal of machine learning
research 9 (2008).
157. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., et al. Language
models are few-shot learners. Advances in neural information processing systems 33,
1877–1901 (2020).
158. Wei, J., Bosma, M., Zhao, V ., Guu, K., Yu, A. W., Lester, B., et al. Finetuned Language
Models are Zero-Shot Learners in International Conference on Learning Representations
(2022).
159. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., et al. Palm:
Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022).
160. Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y ., Fedus, W., et al. Scaling
instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022).
161. Joshi, M., Choi, E., Weld, D. & Zettlemoyer, L. TriviaQA: A Large Scale Distantly
Supervised Challenge Dataset for Reading Comprehension in Proceedings of the 55th
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
(Association for Computational Linguistics, Vancouver, Canada, 2017), 1601–1611.
doi:10.18653/v1/P17-1147.
110
162. Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., et al.
Natural Questions: A Benchmark for Question Answering Research. Transactions of the
Association for Computational Linguistics 7, 452–466. doi:10.1162/tacl a 00276 (2019).
163. Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., et al. Think you
have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint
arXiv:1803.05457 (2018).
164. Mihaylov, T., Clark, P., Khot, T. & Sabharwal, A. Can a Suit of Armor Conduct Electricity?
A New Dataset for Open Book Question Answering in Proceedings of the 2018 Conference
on Empirical Methods in Natural Language Processing (Association for Computational
Linguistics, Brussels, Belgium, 2018), 2381–2391. doi:10.18653/v1/D18-1260.
165. Lin, S., Hilton, J. & Evans, O. TruthfulQA: Measuring How Models Mimic Human
Falsehoods in Proceedings of the 60th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers) (Association for Computational Linguistics, Dublin,
Ireland, 2022), 3214–3252. doi:10.18653/v1/2022.acl-long.229.
166. Lazaridou, A., Kuncoro, A., Gribovskaya, E., Agrawal, D., Liska, A., Terzi, T., et al. Mind
the gap: Assessing temporal generalization in neural language models. Advances in Neural
Information Processing Systems 34, 29348–29363 (2021).
167. Liska, A., Kocisky, T., Gribovskaya, E., Terzi, T., Sezener, E., Agrawal, D., et al.
Streamingqa: A benchmark for adaptation to new knowledge over time in question
answering models in International Conference on Machine Learning (2022), 13604–13622.
168. Kasai, J., Sakaguchi, K., Takahashi, Y ., Bras, R. L., Asai, A., Yu, X., et al. RealTime QA:
What’s the Answer Right Now? arXiv preprint arXiv:2207.13332 (2022).
169. Clark, C., Lee, K., Chang, M.-W., Kwiatkowski, T., Collins, M. & Toutanova, K. BoolQ:
Exploring the Surprising Difficulty of Natural Yes/No Questions in Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational Linguistics:
Human Language Technologies, Volume 1 (Long and Short Papers) (Association for
Computational Linguistics, Minneapolis, Minnesota, 2019), 2924–2936.
doi:10.18653/v1/N19-1300.
170. Rajpurkar, P., Zhang, J., Lopyrev, K. & Liang, P. SQuAD: 100,000+ Questions for Machine
Comprehension of Text in Proceedings of the 2016 Conference on Empirical Methods in
Natural Language Processing (Association for Computational Linguistics, Austin, Texas,
2016), 2383–2392. doi:10.18653/v1/D16-1264.
171. Li, D., Rawat, A. S., Zaheer, M., Wang, X., Lukasik, M., Veit, A., et al. Large Language
Models with Controllable Working Memory. arXiv preprint arXiv:2211.05110 (2022).
172. Si, C., Gan, Z., Yang, Z., Wang, S., Wang, J., Boyd-Graber, J., et al. Prompting gpt-3 to be
reliable in International Conference on Learning Representations (2023).
111
173. Longpre, S., Perisetla, K., Chen, A., Ramesh, N., DuBois, C. & Singh, S. Entity-Based
Knowledge Conflicts in Question Answering in Proceedings of the 2021 Conference on
Empirical Methods in Natural Language Processing (Association for Computational
Linguistics, Online and Punta Cana, Dominican Republic, 2021), 7052–7063.
doi:10.18653/v1/2021.emnlp-main.565.
174. Rajpurkar, P., Jia, R. & Liang, P. Know What You Don’t Know: Unanswerable Questions for
SQuAD in Proceedings of the 56th Annual Meeting of the Association for Computational
Linguistics (Volume 2: Short Papers) (Association for Computational Linguistics,
Melbourne, Australia, 2018), 784–789. doi:10.18653/v1/P18-2124.
175. Gupta, M., Kulkarni, N., Chanda, R., Rayasam, A. & Lipton, Z. C. AmazonQA: A
Review-Based Question Answering Task in International Joint Conference on Artificial
Intelligence (2019).
176. Bjerva, J., Bhutani, N., Golshan, B., Tan, W.-C. & Augenstein, I. SubjQA: A Dataset for
Subjectivity and Review Comprehension in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP) (Association for
Computational Linguistics, Online, 2020), 5480–5494.
doi:10.18653/v1/2020.emnlp-main.442.
177. Stoica, G., Platanios, E. A. & P´ oczos, B. Re-tacred: Addressing shortcomings of the tacred
dataset in Proceedings of the AAAI Conference on Artificial Intelligence 35 (2021),
13843–13850.
178. Zhu, C., Rawat, A. S., Zaheer, M., Bhojanapalli, S., Li, D., Yu, F., et al. Modifying
memories in transformer models. arXiv preprint arXiv:2012.00363 (2020).
179. De Cao, N., Aziz, W. & Titov, I. Editing Factual Knowledge in Language Models in
Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing
(Association for Computational Linguistics, Online and Punta Cana, Dominican Republic,
2021), 6491–6506. doi:10.18653/v1/2021.emnlp-main.522.
180. Mitchell, E., Lin, C., Bosselut, A., Finn, C. & Manning, C. D. Fast Model Editing at Scale in
International Conference on Learning Representations (2022).
181. Meng, K., Bau, D., Andonian, A. J. & Belinkov, Y . Locating and editing factual associations
in gpt in Advances in Neural Information Processing Systems (2022).
182. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y . & Bau, D. Mass-editing memory in a
transformer in International Conference on Learning Representations (2023).
183. Lazaridou, A., Gribovskaya, E., Stokowiec, W. & Grigorev, N. Internet-augmented language
models through few-shot prompting for open-domain question answering. arXiv preprint
arXiv:2203.05115 (2022).
112
184. Izacard, G., Lewis, P., Lomeli, M., Hosseini, L., Petroni, F., Schick, T., et al. Few-shot
learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299
(2022).
185. Khattab, O., Santhanam, K., Li, X. L., Hall, D., Liang, P., Potts, C., et al.
Demonstrate-Search-Predict: Composing retrieval and language models for
knowledge-intensive NLP. arXiv preprint arXiv:2212.14024 (2022).
186. Karpukhin, V ., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., et al. Dense Passage
Retrieval for Open-Domain Question Answering in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP) (Association for
Computational Linguistics, Online, 2020), 6769–6781.
doi:10.18653/v1/2020.emnlp-main.550.
187. Santhanam, K., Khattab, O., Saad-Falcon, J., Potts, C. & Zaharia, M. ColBERTv2: Effective
and Efficient Retrieval via Lightweight Late Interaction in Proceedings of the 2022
Conference of the North American Chapter of the Association for Computational Linguistics:
Human Language Technologies (Association for Computational Linguistics, Seattle, United
States, 2022), 3715–3734. doi:10.18653/v1/2022.naacl-main.272.
188. Gao, L. & Callan, J. Unsupervised Corpus Aware Language Model Pre-training for Dense
Passage Retrieval in Proceedings of the 60th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers) (Association for Computational
Linguistics, Dublin, Ireland, 2022), 2843–2853. doi:10.18653/v1/2022.acl-long.203.
189. Neeman, E., Aharoni, R., Honovich, O., Choshen, L., Szpektor, I. & Abend, O. DisentQA:
Disentangling Parametric and Contextual Knowledge with Counterfactual Question
Answering. arXiv preprint arXiv:2211.05655 (2022).
190. Chow, C. On optimum recognition error and reject tradeoff. IEEE Transactions on
information theory 16, 41–46 (1970).
191. Fumera, G. & Roli, F. Support vector machines with embedded reject option in Pattern
Recognition with Support Vector Machines: First International Workshop, SVM 2002
Niagara Falls, Canada, August 10, 2002 Proceedings (2002), 68–82.
192. Cortes, C., DeSalvo, G. & Mohri, M. Learning with rejection in Algorithmic Learning
Theory: 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016,
Proceedings 27 (2016), 67–82.
193. Gal, Y . & Ghahramani, Z. Dropout as a bayesian approximation: Representing model
uncertainty in deep learning in international conference on machine learning (2016),
1050–1059.
113
194. Lakshminarayanan, B., Pritzel, A. & Blundell, C. Simple and scalable predictive uncertainty
estimation using deep ensembles. Advances in neural information processing systems 30
(2017).
195. Xin, J., Tang, R., Yu, Y . & Lin, J. The Art of Abstention: Selective Prediction and Error
Regularization for Natural Language Processing in Proceedings of the 59th Annual Meeting
of the Association for Computational Linguistics and the 11th International Joint Conference
on Natural Language Processing (Volume 1: Long Papers) (Association for Computational
Linguistics, Online, 2021), 1040–1051. doi:10.18653/v1/2021.acl-long.84.
196. Zhou, W., Liu, F. & Chen, M. Contrastive Out-of-Distribution Detection for Pretrained
Transformers in Proceedings of the 2021 Conference on Empirical Methods in Natural
Language Processing (Association for Computational Linguistics, Online and Punta Cana,
Dominican Republic, 2021), 1100–1111. doi:10.18653/v1/2021.emnlp-main.84.
197. Yatskar, M. A Qualitative Comparison of CoQA, SQuAD 2.0 and QuAC in Proceedings of
the 2019 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)
(Association for Computational Linguistics, Minneapolis, Minnesota, 2019), 2318–2323.
doi:10.18653/v1/N19-1241.
198. Reddy, S., Chen, D. & Manning, C. D. CoQA: A Conversational Question Answering
Challenge. Transactions of the Association for Computational Linguistics 7, 249–266.
doi:10.1162/tacl a 00266 (2019).
199. Choi, E., He, H., Iyyer, M., Yatskar, M., Yih, W.-t., Choi, Y ., et al. QuAC: Question
Answering in Context in Proceedings of the 2018 Conference on Empirical Methods in
Natural Language Processing (Association for Computational Linguistics, Brussels,
Belgium, 2018), 2174–2184. doi:10.18653/v1/D18-1241.
200. Zhou, Y ., Muresanu, A. I., Han, Z., Paster, K., Pitis, S., Chan, H., et al. Large language
models are human-level prompt engineers. arXiv preprint arXiv:2211.01910 (2022).
201. Liu, J., Shen, D., Zhang, Y ., Dolan, B., Carin, L. & Chen, W. What Makes Good In-Context
Examples for GPT-3? in Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd
Workshop on Knowledge Extraction and Integration for Deep Learning Architectures
(Association for Computational Linguistics, Dublin, Ireland and Online, 2022), 100–114.
doi:10.18653/v1/2022.deelio-1.10.
202. Reimers, N. & Gurevych, I. Sentence-BERT: Sentence Embeddings using Siamese
BERT-Networks in Proceedings of the 2019 Conference on Empirical Methods in Natural
Language Processing and the 9th International Joint Conference on Natural Language
Processing (EMNLP-IJCNLP) (Association for Computational Linguistics, Hong Kong,
China, 2019), 3982–3992. doi:10.18653/v1/D19-1410.
114
203. Zeng, D., Liu, K., Chen, Y . & Zhao, J. Distant supervision for relation extraction via
piecewise convolutional neural networks in Proceedings of the 2015 Conference on
Empirical Methods in Natural Language Processing (2015), 1753–1762.
204. Mintz, M., Bills, S., Snow, R. & Jurafsky, D. Distant supervision for relation extraction
without labeled data in Proceedings of the Joint Conference of the 47th Annual Meeting of
the ACL and the 4th International Joint Conference on Natural Language Processing of the
AFNLP: Volume 2-Volume 2 (2009), 1003–1011.
205. Surdeanu, M., Tibshirani, J., Nallapati, R. & Manning, C. D. Multi-instance multi-label
learning for relation extraction in Proceedings of the 2012 joint conference on empirical
methods in natural language processing and computational natural language learning
(2012), 455–465.
206. Nakashole, N., Weikum, G. & Suchanek, F. PATTY: a taxonomy of relational patterns with
semantic types in Proceedings of the 2012 Joint Conference on Empirical Methods in
Natural Language Processing and Computational Natural Language Learning (2012),
1135–1145.
207. Jiang, M., Shang, J., Cassidy, T., Ren, X., Kaplan, L. M., Hanratty, T. P., et al. Metapad:
Meta pattern discovery from massive text corpora in Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (2017), 877–886.
208. Jones, R., McCallum, A., Nigam, K. & Riloff, E. Bootstrapping for text learning tasks in
IJCAI-99 Workshop on Text Mining: Foundations, Techniques and Applications 1 (1999).
209. Agichtein, E. & Gravano, L. Snowball: Extracting relations from large plain-text collections
in Proceedings of the fifth ACM conference on Digital libraries (2000), 85–94.
210. Batista, D. S., Martins, B. & Silva, M. J. Semi-supervised bootstrapping of relationship
extractors with distributional semantics in Proceedings of the 2015 Conference on Empirical
Methods in Natural Language Processing (2015), 499–504.
211. Hancock, B., Varma, P., Wang, S., Bringmann, M., Liang, P. & R´ e, C. Training Classifiers
with Natural Language Explanations. Proceedings of the 56th Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long Papers).
doi:10.18653/v1/p18-1175 (2018).
212. Rosenberg, C., Hebert, M. & Schneiderman, H. Semi-supervised self-training of object
detection models (2005).
213. Hendrickx, I., Kim, S. N., Kozareva, Z., Nakov, P.,
´
O S´ eaghdha, D., Pad´ o, S., et al.
Semeval-2010 task 8: Multi-way classification of semantic relations between pairs of
nominals in Proceedings of the Workshop on Semantic Evaluations: Recent Achievements
and Future Directions (2009), 94–99.
115
214. Lin, H., Yan, J., Qu, M. & Ren, X. Learning Dual Retrieval Module for Semi-supervised
Relation Extraction in The Web Conference (2019).
215. Roth, B., Barth, T., Wiegand, M., Singh, M. & Klakow, D. Effective slot filling based on
shallow distant supervision methods. arXiv preprint arXiv:1401.1158 (2014).
216. Willett, P. The Porter stemming algorithm: then and now. Program 40, 219–223 (2006).
217. Qu, M., Ren, X., Zhang, Y . & Han, J. Weakly-supervised Relation Extraction by
Pattern-enhanced Embedding Learning in Proceedings of the 2018 World Wide Web
Conference on World Wide Web (2018), 1257–1266.
218. Wang, L., Cao, Z., de Melo, G. & Liu, Z. Relation Classification via Multi-Level Attention
CNNs in Proceedings of the 54th Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers) (Association for Computational Linguistics, Berlin,
Germany, 2016), 1298–1307. doi:10.18653/v1/P16-1123.
219. Zhou, P., Shi, W., Tian, J., Qi, Z., Li, B., Hao, H., et al. Attention-Based Bidirectional Long
Short-Term Memory Networks for Relation Classification in Proceedings of the 54th Annual
Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)
(Association for Computational Linguistics, Berlin, Germany, 2016), 207–212.
doi:10.18653/v1/P16-2034.
220. Bahdanau, D., Cho, K. & Bengio, Y . Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473 (2014).
221. Yu, M. & Dredze, M. Learning composition models for phrase embeddings. Transactions of
the Association for Computational Linguistics 3, 227–242 (2015).
222. Neculoiu, P., Versteegh, M. & Rotaru, M. Learning text similarity with siamese recurrent
networks in Proceedings of the 1st Workshop on Representation Learning for NLP (2016),
148–157.
223. Lee, D.-H. Pseudo-label: The simple and efficient semi-supervised learning method for deep
neural networks in (2013).
224. Jiang, J. & Zhai, C. Instance weighting for domain adaptation in NLP in Proceedings of the
45th annual meeting of the association of computational linguistics (2007), 264–271.
225. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S. & Dean, J. Distributed representations of
words and phrases and their compositionality in Advances in neural information processing
systems (2013), 3111–3119.
226. Pennington, J., Socher, R. & Manning, C. Glove: Global vectors for word representation in
Proceedings of the 2014 conference on empirical methods in natural language processing
(EMNLP) (2014), 1532–1543.
116
227. Li, S., Xu, H. & Lu, Z. Generalize Symbolic Knowledge With Neural Rule Engine. arXiv
preprint arXiv:1808.10326 (2018).
228. Tarvainen, A. & Valpola, H. Mean teachers are better role models: Weight-averaged
consistency targets improve semi-supervised deep learning results in Advances in neural
information processing systems (2017), 1195–1204.
229. Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro,
et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems Software
available from tensorflow.org. 2015.
230. Duchi, J., Hazan, E. & Singer, Y . Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research 12, 2121–2159 (2011).
231. Hearst, M. A. Automatic acquisition of hyponyms from large text corpora in Proceedings of
the 14th conference on Computational linguistics-Volume 2 (1992), 539–545.
232. Srivastava, S., Labutov, I. & Mitchell, T. M. Joint Concept Learning and Semantic Parsing
from Natural Language Explanations in EMNLP (2017).
233. Xu, W., Sun, H., Deng, C. & Tan, Y . Variational autoencoder for semi-supervised text
classification in Thirty-First AAAI Conference on Artificial Intelligence (2017).
234. Gupta, P., Roth, B. & Sch¨ utze, H. Joint bootstrapping machines for high confidence relation
extraction. arXiv preprint arXiv:1805.00254 (2018).
235. Nan, G., Guo, Z., Sekulic, I. & Lu, W. Reasoning with Latent Structure Refinement for
Document-Level Relation Extraction in Proceedings of the 58th Annual Meeting of the
Association for Computational Linguistics (Association for Computational Linguistics,
Online, 2020), 1546–1557. doi:10.18653/v1/2020.acl-main.141.
236. Baldini Soares, L., FitzGerald, N., Ling, J. & Kwiatkowski, T. Matching the Blanks:
Distributional Similarity for Relation Learning in Proceedings of the 57th Annual Meeting
of the Association for Computational Linguistics (Association for Computational Linguistics,
Florence, Italy, 2019), 2895–2905. doi:10.18653/v1/P19-1279.
237. Peng, H., Gao, T., Han, X., Lin, Y ., Li, P., Liu, Z., et al. Learning from Context or Names?
An Empirical Study on Neural Relation Extraction in Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP) (Association for
Computational Linguistics, Online, 2020), 3661–3672.
doi:10.18653/v1/2020.emnlp-main.298.
117
238. Qin, Y ., Lin, Y ., Takanobu, R., Liu, Z., Li, P., Ji, H., et al. ERICA: Improving Entity and
Relation Understanding for Pre-trained Language Models via Contrastive Learning in
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics
and the 11th International Joint Conference on Natural Language Processing (Volume 1:
Long Papers) (Association for Computational Linguistics, Online, 2021), 3350–3363.
doi:10.18653/v1/2021.acl-long.260.
239. Mintz, M., Bills, S., Snow, R. & Jurafsky, D. Distant supervision for relation extraction
without labeled data in Proceedings of the Joint Conference of the 47th Annual Meeting of
the ACL and the 4th International Joint Conference on Natural Language Processing of the
AFNLP (Association for Computational Linguistics, Suntec, Singapore, 2009), 1003–1011.
240. Hadsell, R., Chopra, S. & LeCun, Y . Dimensionality reduction by learning an invariant
mapping in 2006 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR’06) 2 (2006), 1735–1742.
241. Graf, F., Hofer, C., Niethammer, M. & Kwitt, R. Dissecting supervised constrastive learning
in International Conference on Machine Learning (2021), 3821–3830.
242. Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L. & Lewis, M. Generalization through
Memorization: Nearest Neighbor Language Models in International Conference on
Learning Representations (2020).
243. Khandelwal, U., Fan, A., Jurafsky, D., Zettlemoyer, L. & Lewis, M. Nearest Neighbor
Machine Translation in International Conference on Learning Representations (2021).
244. Jia, R., Wong, C. & Poon, H. Document-Level N-ary Relation Extraction with Multiscale
Representation Learning in Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies,
Volume 1 (Long and Short Papers) (Association for Computational Linguistics, Minneapolis,
Minnesota, 2019), 3693–3704. doi:10.18653/v1/N19-1370.
245. Tan, Q., Xu, L., Bing, L. & Ng, H. T. Revisiting DocRED - Addressing the Overlooked
False Negative Problem in Relation Extraction. arXiv preprint arXiv:2205.12696 (2022).
246. Quirk, C. & Poon, H. Distant Supervision for Relation Extraction beyond the Sentence
Boundary in Proceedings of the 15th Conference of the European Chapter of the Association
for Computational Linguistics: Volume 1, Long Papers (Association for Computational
Linguistics, Valencia, Spain, 2017), 1171–1182.
247. Peng, N., Poon, H., Quirk, C., Toutanova, K. & Yih, W.-t. Cross-Sentence N-ary Relation
Extraction with Graph LSTMs. Transactions of the Association for Computational
Linguistics 5, 101–115. doi:10.1162/tacl a 00049 (2017).
118
248. Christopoulou, F., Miwa, M. & Ananiadou, S. Connecting the Dots: Document-level Neural
Relation Extraction with Edge-oriented Graphs in Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing and the 9th International Joint
Conference on Natural Language Processing (EMNLP-IJCNLP) (Association for
Computational Linguistics, Hong Kong, China, 2019), 4925–4936.
doi:10.18653/v1/D19-1498.
249. Zeng, S., Xu, R., Chang, B. & Li, L. Double Graph Based Reasoning for Document-level
Relation Extraction in Proceedings of the 2020 Conference on Empirical Methods in Natural
Language Processing (EMNLP) (Association for Computational Linguistics, Online, 2020),
1630–1640. doi:10.18653/v1/2020.emnlp-main.127.
250. Zhou, W., Huang, K., Ma, T. & Huang, J. Document-level relation extraction with adaptive
thresholding and localized context pooling in Proceedings of the AAAI conference on
artificial intelligence 35 (2021), 14612–14620.
251. Zhang, N., Chen, X., Xie, X., Deng, S., Tan, C., Chen, M., et al. Document-level Relation
Extraction as Semantic Segmentation in IJCAI (2021).
252. Tan, Q., He, R., Bing, L. & Ng, H. T. Document-Level Relation Extraction with Adaptive
Focal Loss and Knowledge Distillation in Findings of the Association for Computational
Linguistics: ACL 2022 (Association for Computational Linguistics, Dublin, Ireland, 2022),
1672–1681. doi:10.18653/v1/2022.findings-acl.132.
253. Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image
segmentation in International Conference on Medical image computing and
computer-assisted intervention (2015), 234–241.
254. Levy, O., Seo, M., Choi, E. & Zettlemoyer, L. Zero-Shot Relation Extraction via Reading
Comprehension in Proceedings of the 21st Conference on Computational Natural Language
Learning (CoNLL 2017) (Association for Computational Linguistics, Vancouver, Canada,
2017), 333–342. doi:10.18653/v1/K17-1034.
255. Lu, K., Hsu, I.-H., Ma, M. D., Zhou, W. & Chen, M. Summarization as Indirect Supervision
for Relation Extraction in Findings of ACL: EMNLP (2022).
256. Gao, T., Yao, X. & Chen, D. SimCSE: Simple Contrastive Learning of Sentence Embeddings
in Proceedings of the 2021 Conference on Empirical Methods in Natural Language
Processing (Association for Computational Linguistics, Online and Punta Cana, Dominican
Republic, 2021), 6894–6910. doi:10.18653/v1/2021.emnlp-main.552.
257. Gu, Y ., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu, X., et al. Domain-specific
language model pretraining for biomedical natural language processing. ACM Transactions
on Computing for Healthcare (HEALTH) 3, 1–23 (2021).
119
258. Movshovitz-Attias, Y ., Toshev, A., Leung, T. K., Ioffe, S. & Singh, S. No fuss distance
metric learning using proxies in Proceedings of the IEEE International Conference on
Computer Vision (2017), 360–368.
259. Zhu, J., Wang, Z., Chen, J., Chen, Y .-P. P. & Jiang, Y .-G. Balanced Contrastive Learning for
Long-Tailed Visual Recognition in Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition (2022), 6908–6917.
260. Christobel, Y . A. & Sivaprakasam, P. A new classwise k nearest neighbor (CKNN) method
for the classification of diabetes dataset. International Journal of Engineering and Advanced
Technology 2, 396–200 (2013).
261. Zhang, N., Deng, S., Sun, Z., Wang, G., Chen, X., Zhang, W., et al. Long-tail Relation
Extraction via Knowledge Graph Embeddings and Graph Convolution Networks in
Proceedings of the 2019 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short
Papers) (Association for Computational Linguistics, Minneapolis, Minnesota, 2019),
3016–3025. doi:10.18653/v1/N19-1306.
262. Wei, C.-H., Allot, A., Leaman, R. & Lu, Z. PubTator central: automated concept annotation
for biomedical full text articles. Nucleic acids research 47, W587–W593 (2019).
263. Loshchilov, I. & Hutter, F. Decoupled Weight Decay Regularization in International
Conference on Learning Representations (2018).
264. Islam, A., Chen, C.-F. R., Panda, R., Karlinsky, L., Radke, R. & Feris, R. A broad study on
the transferability of visual representations with contrastive learning in Proceedings of the
IEEE/CVF International Conference on Computer Vision (2021), 8845–8855.
265. Chen, M., Fu, D. Y ., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., et al. Perfectly
balanced: Improving transfer and robustness of supervised contrastive learning in
International Conference on Machine Learning (2022), 3090–3122.
266. Wang, F. & Liu, H. Understanding the behaviour of contrastive loss in Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition (2021), 2495–2504.
267. McClosky, D., Surdeanu, M. & Manning, C. Event Extraction as Dependency Parsing in
Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics:
Human Language Technologies (Association for Computational Linguistics, Portland,
Oregon, USA, 2011), 1626–1635.
268. Yao, Y ., Du, J., Lin, Y ., Li, P., Liu, Z., Zhou, J., et al. CodRED: A Cross-Document Relation
Extraction Dataset for Acquiring Knowledge in the Wild in Proceedings of the 2021
Conference on Empirical Methods in Natural Language Processing (Association for
Computational Linguistics, Online and Punta Cana, Dominican Republic, 2021), 4452–4472.
doi:10.18653/v1/2021.emnlp-main.366.
120
269. Borji, A. A categorical archive of ChatGPT failures. arXiv preprint arXiv:2302.03494
(2023).
121

Abstract (if available)

Abstract Knowledge acquisition involves identifying and understanding concepts that are expressed in text, and then inferring and consolidating their relationships. With the ever-increasing volumes of digital text generated every day, the need for automated knowledge acquisition is becoming more crucial. However, real-world knowledge acquisition systems face more complex scenarios where knowledge is often expressed implicitly or within long contexts, and require models that are robust and able to handle these complexities. Additionally, data annotation for knowledge acquisition can be expensive, particularly in high-stakes domains, necessitating the development of efficient methods that minimize human involvement. In this dissertation, we propose solutions to build robust and generalizable knowledge acquisition systems, addressing challenges in: (i) extracting knowledge from long contexts efficiently and effectively; (ii) improving system robustness in various scenarios including noisy annotations, out-of-distribution instances, and knowledge conflicts; and (iii) data-efficient knowledge acquisition, utilizing cheap resources such as labeling rules or unlabeled corpora.
In the first part of this thesis, we study the extraction of knowledge from long contexts such as documents. To handle the long-context problem, we introduce a novel pretrained-attention-guided neural structure that finds context that both entities attend to in an entity pair. We show that our model achieves state-of-the-art performance in both general and biomedical domains.
In the second part of this thesis, our focus is on enhancing the robustness of knowledge acquisition systems. To start with, we investigate the mitigation of noisy labels during training. Since the data used for knowledge acquisition is often sourced from crowdsourcing or distant supervision, it is common for the data to be imperfect. We explore a co-regularization based framework to address this issue effectively. Moreover, many knowledge acquisition systems are typically developed for specialized domains or relations, making them susceptible to erroneous predictions on instances from other domains or relations. To overcome this challenge, we propose an out-of-distribution detection framework to identify such instances during inference and either alert users or redirect them to other dedicated models for improved accuracy. Another challenge is that the world is continually evolving, and the knowledge acquired by these systems may become outdated. We explore techniques for updating the model's predicted knowledge using context retrieval and opinion-based prompts.
Finally, we study how to build knowledge acquisition systems with minimal human effort. Our approach involves weak data creation, where we can generate a large amount of labeled data by utilizing labeling a few rules developed by humans. Then, we investigate enhancing the data efficiency of knowledge acquisition systems by self-supervised learning. To accomplish this, we propose a framework that pretrains a model using a large unlabeled corpus, which can then be fine-tuned with a small number of labeled instances. Additionally, we propose a novel continual contrastive finetuning framework that aims to minimize the discrepancy between the pretraining and fine-tuning stages.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Parametric and semi-parametric methods for knowledge acquisition from text

PDF

Emphasizing the importance of data and evaluation in the era of large language models

PDF

Countering problematic content in digital space: bias reduction and dynamic content adaptation

PDF

Multi-modal preconditioned inference of commonsense knowledge

PDF

Learning controllable data generation for scalable model training

PDF

Towards more occlusion-robust deep visual object tracking

PDF

Deep representations for shapes, structures and motion

PDF

Externalized reasoning in language models for scalable and trustworthy AI

PDF

Identifying and mitigating safety risks in language models

PDF

Efficient and accurate object extraction from scanned maps by leveraging external data and learning representative context

PDF

Behavior understanding from speech under constrained conditions: exploring sparse networks, transfer and unsupervised learning

PDF

Learning logical abstractions from sequential data

PDF

Towards more human-like cross-lingual transfer learning

PDF

Generating and utilizing machine explanations for trustworthy NLP

PDF

Fast and label-efficient graph representation learning

PDF

Towards generalized event understanding in text via generative models

PDF

Robust and adaptive online decision making

PDF

Green knowledge graph completion and scalable generative content delivery

PDF

Common ground reasoning for communicative agents

PDF

Inferring mobility behaviors from trajectory datasets

Asset Metadata

Creator Zhou, Wenxuan (author)

Core Title Robust and generalizable knowledge acquisition from text

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Degree Conferral Date 2023-08

Publication Date 08/03/2023

Defense Date 05/01/2023

Publisher University of Southern California. Libraries (digital)

Tag deep learning,information extraction,natural language processing,OAI-PMH Harvest

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Chen, Muhao (committee chair), Itti, Laurent (committee member), Jia, Robin (committee member), May, Jonathan (committee member), Sun, Tianshu (committee member)

Creator Email zhouwenx@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113296263

Unique identifier UC113296263

Identifier etd-ZhouWenxua-12189.pdf (filename)

Legacy Identifier etd-ZhouWenxua-12189

Document Type Dissertation

Rights Zhou, Wenxuan

Internet Media Type application/pdf

Type texts

Source 20230803-usctheses-batch-1078 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA

Repository Email cisadmin@lib.usc.edu