University of Southern California Dissertations and Theses

Machine learning paradigms for behavioral coding

(USC Thesis Other)

Machine learning paradigms for behavioral coding

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content MACHINE LEARNING PARADIGMS FOR BEHAVIORAL CODING
by
James Gibson
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)
August 2018
Copyright 2018 James Gibson
Contents
Contents ii
List of Tables iv
List of Figures v
Acknowledgements vi
Abstract viii
1 Introduction 1
1.1 Human Communicative Behaviors . . . . . . . . . . . . . . . . . . . 1
1.2 Behavioral Coding in Psychotherapy . . . . . . . . . . . . . . . . . 1
1.3 Behavioral Signal Processing . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Machine Learning Paradigms . . . . . . . . . . . . . . . . . . . . . 4
1.4.1 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.2 Attention Learning . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.3 Multi-label Learning . . . . . . . . . . . . . . . . . . . . . . 5
1.4.4 Multi-task Learning . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.5 Context Learning . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Notation Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Behavioral Coding Data Description 10
2.1 Motivational Interviewing Corpus . . . . . . . . . . . . . . . . . . . 10
2.1.1 Motivational Interviewing Skill Code . . . . . . . . . . . . . 11
2.2 Cognitive Behavioral Therapy Corpus . . . . . . . . . . . . . . . . . 14
2.2.1 Cognitive Therapy Rating System . . . . . . . . . . . . . . . 14
2.3 Training and Testing Data Splits . . . . . . . . . . . . . . . . . . . 16
3 A Deep Learning Approach for Behavioral Coding 18
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 End-to-end Deep Learning Network . . . . . . . . . . . . . . . . . . 18
ii
3.3 Model Implementation and Training . . . . . . . . . . . . . . . . . 22
3.4 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . 23
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4 Attention Networks for Behavioral Coding 26
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Word Attention Network . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.1 Attention-based LSTM Network . . . . . . . . . . . . . . . . 28
4.3 Model Implementation and Training . . . . . . . . . . . . . . . . . 30
4.4 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . 31
4.4.1 Analysis of Attention Weights . . . . . . . . . . . . . . . . . 32
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5 Multi-label Multi-task Learning for Behavioral Coding 37
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Multi-label Learning for Behavioral Coding . . . . . . . . . . . . . . 37
5.2.1 Multi-label Learning with Deep Neural Networks . . . . . . 39
5.2.2 Multi-label Sample Weights . . . . . . . . . . . . . . . . . . 39
5.3 Multi-task Learning for Behavioral Coding . . . . . . . . . . . . . . 41
5.4 Learning with Context . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.5 Model Implementation and Training . . . . . . . . . . . . . . . . . 43
5.6 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . 44
5.6.1 Multi-label learning . . . . . . . . . . . . . . . . . . . . . . . 44
5.6.2 Multi-task Learning . . . . . . . . . . . . . . . . . . . . . . . 45
5.6.3 Turn Context Learning . . . . . . . . . . . . . . . . . . . . . 46
5.6.4 Multi-label Multi-task Learning with Context . . . . . . . . 47
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6 Conclusions and Future Work 52
6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.2 Potential Applications of Automatic Behavioral Coding . . . . . . . 53
6.3 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.3.1 Automatic Behavioral Coding in the `Wild' . . . . . . . . . . 54
6.3.2 Transfer Learning and Domain Adaptation . . . . . . . . . . 54
6.3.3 Beyond Behavioral Coding . . . . . . . . . . . . . . . . . . . 55
Reference List 56
iii
List of Tables
1.1 Notation reference. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 MISC Categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Session-level behavior codes dened by the CTRS manual . . . . . . 16
2.3 MI and CBT Data: Session, turn, and word counts in training/testing
splits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Example of Therapist-Client Interaction. . . . . . . . . . . . . . . . 17
3.1 MISC8 code grouping and counts in the dataset. . . . . . . . . . . . 20
3.2 MISC Prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Empathy Prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1 MISC11 code grouping and counts in the dataset. . . . . . . . . . . 28
4.2 MISC Prediction (F1-measure). . . . . . . . . . . . . . . . . . . . . 32
4.3 Words with the highest average attention for each behavior category. 35
5.1 Comparison of single-label, multi-label,and multi-label multi-task
systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Multi-label Learning with Context . . . . . . . . . . . . . . . . . . 50
5.3 Multi-task Learning with Context . . . . . . . . . . . . . . . . . . . 51
iv
List of Figures
2.1 Empathy distribution of the MI dataset . . . . . . . . . . . . . . . 12
2.2 CTRS behavioral code distributions . . . . . . . . . . . . . . . . . 15
3.1 Proposed System Overview. . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Example of k-hot encoding of MISC behaviors . . . . . . . . . . . . 21
4.1 Diagram of LSTM with attention . . . . . . . . . . . . . . . . . . . 30
4.2 (a) Example of attention for RES and QUO codes in counselor turn
and (b) POS and NEG codes in client turn. . . . . . . . . . . . . . 33
5.1 Diagram of encoding and decoding networks . . . . . . . . . . . . . 38
5.2 Overview of single task and multitask systems . . . . . . . . . . . . 42
v
Acknowledgements
I would like to thank everyone who made this dissertation possible through their
support and encouragement over the years.
Firstly, I would like to thank my advisor Professor Shrikanth Narayanan for
giving me the opportunity to join the Signal Analysis and Interpretation Lab and
for supporting and guiding me throughout my time as a Ph.D. student. I would
like to thank Professor Panayiotis Georgiou who was always available for words of
encouragement and perspective. Thanks to everyone who served on my qualifying
and defense committees including Professors Antonio Ortega, Keith Jenkins, Yan
Liu, and Gayla Margolin for your valuable insights and feedback.
I would like to thank all my colleagues, collaborators, and friends in the Signal
Analysis and Interpretation Lab, including: Rahul Gupta, Theodora Chaspari, and
Dogan Can who started this journey with me; Danny Bone for always being avail-
able on Gchat for guidance through the Ph.D. woes; Colin Vaz, Pavlos Papadopou-
los, and Naveen Kumar who were always up for pool and a beer; Franky Romero
who I had the privilege of mentoring; Abe Kazemzadeh and Jangwon Kim with
whom I shared my rst oce; Nassos Katsamanis and Matt Black with whom I
co-authored my rst publication; and all the SAILers not mentioned here by name
with whom I have collaborated and been inspired by over the years.
vi
I would like to thank Angela for being there through the ups and downs of
completing my Ph.D. And a special thanks to my parents, Ray and Kathy, and
my sisters, Jennifer, Megan, and Shannon, for their unwavering love and support
throughout my life.
vii
Abstract
Modeling human behavior in conversational interactions is a complex and chal-
lenging task. This dissertation posits that by employing modeling and learning
techniques which better re
ect the complexities of human interaction, behaviors
can be better predicted and understood. We propose several machine learning
approaches for modeling these complex and important phenomena. Specically,
we investigate machine learning paradigms including deep, attentional, multi-label,
multi-task, and hierarchical learning models to achieve this aim. We discuss how
these learning techniques re
ect important aspects of understanding human inter-
actions.
Using behavioral coding in psychotherapy as an expository domain, we exper-
imentally validate the proposed approaches. For this purpose, we describe two
datasets of therapist-client therapy interactions, from the domains of motivational
interviewing and cognitive behavior therapy, and the specied behavioral con-
structs of interest in each. We demonstrate that by incorporating and merging the
proposed learning approaches we achieve increased performance in the behavioral
coding task and discuss the insights gained from these approaches. We discuss the
potential implications and applications of this work for automating and augment-
ing behavioral coding and understanding.
viii
Chapter 1
Introduction
1.1 Human Communicative Behaviors
Human communication is inherently complex and multifaceted. Understanding
and modeling behaviors occurring in these complex interactions is a challenging
task. Conversations involve multiple speakers, each with their own respective com-
munication styles, both continuously deciphering and responding to one another's
behaviors. Traditionally, eorts to examine and codify these behaviors required
manual observation, a process referred to as behavioral coding [1]. This process
involves establishing and dening behaviors of interest, training annotators to be
able to identify these behaviors with high agreement, and manual observations of
human interaction from domains of interest. This process is typically costly and
time consuming. Developments in signal processing and machine learning aord
an opportunity to both provide tools which automate and augment this process
and possibly enable deeper understanding of these important phenomena.
1.2 Behavioral Coding in Psychotherapy
In psychotherapy, the therapist seeks to understand and help address the behaviors
and concerns of the client. There are a variety of behaviors employed in this process
that vary according to the type of therapy, the aims of the therapy, and the training
and skill level of the therapist. Researchers have suggested that for a variety of
1
symptoms the type of therapy may not signicantly aect outcomes [2]. Thus,
it can be assumed that there are some underlying mechanisms at work that are
common across psychotherapy approaches.
In psychotherapy research, behavioral coding is used for identifying and codify-
ing the behaviors which are most relevant to the aims of therapy [1]. The objective
of this procedure is to dene clear and broadly applicable behavioral `codes' which
represent target behavioral constructs that are of interest to a particular study or
line of inquiry. Behavioral observation and coding is common practice in many
subelds of psychology including diagnosing autism [3], family and marital obser-
vational studies [4, 5], and several forms of psychotherapy [6, 7]. Because manual
behavioral coding is costly and time-intensive, there is an opportunity for the
development of methodologies aimed at automating aspects of this process.
In this work, we use data from two distinct, yet intrinsically related, psychother-
apy approaches to serve as example domains in which we evaluate machine learn-
ing paradigms for behavioral coding. Motivational Interviewing (MI) is a client
centered approach to therapy that aims to promote behavior change in clients
by exploring and resolving ambivalence. Cognitive Behavior Therapy (CBT), in
contrast to motivational interviewing, is focused on developing coping strategies
aimed at decreasing symptoms. Both therapies are goal-oriented, evidence based
practices which are concerned with client behavior change. So, despite dierences
in approach, there is signicant overlap in philosophical orientation and employed
techniques in the domains.
2
1.3 Behavioral Signal Processing
There is signicant ongoing eort to model human behavior in conversational inter-
actions in the emerging eld of behavioral signal processing (BSP) [8]. BSP is the
development and application of signal processing and machine learning tools for
aiding behavioral sciences research and translation most notably in mental and
behavioral health domains. A common approach in BSP is to develop signal-
derived representations and use these representations with appropriate machine
learning methods to correlate with or predict domain relevant behavioral codes.
There have been numerous works aimed at using human interaction data to help
automate and inform the behavioral observation and coding process in domains
such as marital therapy [9], motivational interviewing [10,11], and autism diagnosis
[12]. Additionally, there have been many features explored in these works including
acoustic and prosodic speech features [9,13,14], lexical and semantic features [15,
16], automatically derived lexical features [17], and visual features [18]. While these
works oer many insights into human behavior and automated behavioral coding,
they all fall into a traditional supervised learning paradigm. That is, each sample
feature vector has an associated label and a classier, typically a linear support
vector machine, attempts to learn a mapping from feature to label. This is a well
examined approach and oers reasonably good prediction results in many cases.
However, there are many aspects of human interaction which are not explicitly
addressed in this paradigm.
3
1.4 Machine Learning Paradigms
Machine learning is a rapidly developing eld and therefore the underlying under-
standing of successful techniques and terms of art are still evolving. We will present
several machine learning paradigms and discuss their potential to aid in the auto-
matic behavioral coding problem. While each paradigm oers some unique aspects
that dierentiate it from the others and warrant its investigation there are many
commonalities in philosophical perspective which enable their co-application. For
this reason, many of the systems which we propose in this work will merge aspects
from more than one of these paradigms into a single model.
1.4.1 Deep Learning
Deep learning is a machine learning approach which allows for multiple levels of
processing and layers of abstraction [19]. Deep neural networks have been proposed
and evaluated across most if not all traditional machine learning domains. One of
their primary strengths is their
exibility in enabling architectures which process
data at a variety of granularities and learn complex representations.
1.4.2 Attention Learning
One major criticism of deep neural networks is limited model interpretability.
Attention-based recurrent neural networks (RNNs) have gained popularity for their
ability to select the appropriate context in sequence data while also giving inter-
pretability towards which elements in the sequence lend most importance to the
predictions made by the models. They've shown promise for a number of problems
including machine translation [20], image classication [21], describing multimedia
content [22], and document classication [23].
4
In behavioral coding, there are many specic cues which may indicate the
presence of a particular behavior. For some behaviors certain salient words may
be cues which annotators attend to. For example, words such as `why' and `how'
indicating a question is being asked. For other behaviors, such as gestalt session
level behaviors, it is less clear on what specic aspects of the interaction prompt the
resultant behavior ratings. Therefore a system which learns to attune to particular
words or expressions of importance may both benet modeling and understanding
of behaviors.
1.4.3 Multi-label Learning
Multi-label learning is a learning paradigm in which each sample is associated with
several, possibly related, labels [24]. Such a framework allows for a model to learn
more general features because they must be relevant to multiple targets. Also,
this approach allows a model to account for relationships between labels which
can be especially useful for predicting less frequent labels. Multi-label learning
has been explored for a wide variety of applications including functional genomics,
text categorization, and scene classication [24, 25] and for a variety of classiers
including K-nearest neighbors [26], support vector machines [27], and deep neural
networks [28].
In behavioral coding, each code attempts to capture a distinct behavior of
interest. However many of these behaviors are fundamentally related; for example
open questioning and re
ective listening are considered the skills of a well trained
motivational interviewer, whereas confrontations are not. In this sense behavioral
coding is a problem with multiple interrelated outputs which motivates investigat-
ing a multi-label learning approach.
5
1.4.4 Multi-task Learning
Multi-task learning is a machine learning paradigm in which a single model is
optimized for more than one task [29]. Such a model can share part or all of its
architecture save for the outputs which are dedicated to specic tasks. These tasks
are often related, allowing the model to key in on features of general importance.
This approach also allows for a model to experience more data even though the
labels of each sample may not be available for all tasks. This framework has
shown promise in a variety of domains including text categorization [30], head
pose estimation [31], emotion recognition [32], and distance speech recognition
[33]. Recently, Liu et al., have proposed an adversarial training approach for
multi-task networks using a network consisting of shared and private layers where
the an adversarial loss is used to force the shared layers to learn task invariant
features [34].
Data from many behavioral coding domains are of a sensitive and private
nature. For this reason it is often dicult to obtain such data. Thus paradigms like
multi-task learning which can learn shared representations across related domains
allow for inclusion of data from corpora, even if these corpora do not share identical
types of interaction and behavioral coding schemes.
1.4.5 Context Learning
Hierarchical learning is a paradigm which takes advantage of hierarchies existing in
data or labels. For example, in the case of document classication, each document
is comprised of multiple sentences, which are each in turn made up of multiple
words. In this case, the architecture can be formulated with layers that learn a
representation from word to sentence followed by layers responsible for learning
a mapping from sentence to document label [23, 35]. In addition to document
6
classication, other notable application of hierarchical learning include sequence
generation [36], image classication [37], and sentiment analysis [38].
With respect to human behavioral coding, there exists many natural hierarchies
including sessions which are comprised of speaker turns which are in turn comprised
of speaker verbal and non-verbal expressions as well as the behaviors which are
expressed and at times coded at the utterance and session levels. Therefore it is
important to incorporate contextual information, whether across words or turns,
to learn representations which re
ect the nature of these interactions.
1.5 Related Work
While the majority of system proposed have been traditional fully supervised learn-
ing approaches, prior to this work, we proposed Multiple Instance Learning (MIL)
for behavioral coding in couples therapy [39]. MIL is a semi-supervised learn-
ing paradigm, in which several samples share a single label, thus an MIL system
attempts to learn a many-to-one mapping between samples and labels. In the case
of behavioral coding, this mapping is between the multiple turns (samples) to ses-
sion level behavioral codes (labels). In the present work, the proposed approaches
can be thought of as a type of multiple instance learning. However, the samples
in this work are treated as sequences rather than independent observations and
therefore better re
ect the temporal nature of human interactions.
Using a sequential model (a Conditional Random Field) to predict sequences
of utterance level behavioral codes in psychotherapy was rst proposed in [40].
This work also proposed using dialog acts as a proxy for utterance level behavioral
codes and demonstrated that using dialog acts for predicting session level behav-
iors achieved competitive performance to using carefully dened and annotated
7
utterance level behaviors. The rst application of neural networks to behavioral
coding was proposed in [41]. The authors propose a recursive neural network for
deriving an utterance representation and use a Maximum Entropy Markov Model
(MEMM) to perform detection of client change talk and sustain talk at the indi-
vidual utterance level. Recurrent neural networks (RNNs) were rst proposed for
behavioral coding in [42] and [43]. In [42], the authors compare Long short term
memory (LSTMs) and Gated Recurrent Units (GRUs), two varieties of RNNs,
for predicting utterance level behavioral codes from word embeddings. In [43],
LSTMs are used for encoding turn context from turn embeddings for predicting
utterance level behaviors which is subsequently used as the lower layers of a deeper
system that predicts sessions level empathy in psychotherapy interactions. This
work will be described in greater depth as a part of the present work. Recently,
in [44], the authors proposed using multimodal word-level based LSTMs trained
with prosodic and lexical features for predicting utterance level codes in moti-
vational interviewing sessions. In [45], the authors compare several lexical and
semantic feature representations for predicting session level behaviors in cognitive
behavioral therapy sessions.
1.6 Notation Reference
We present an overview of the notation used throughout this work in table 1.1.
8
Table 1.1: Notation reference.
Symbol Meaning
i session index
j turn index
k word index
N number of sessions
M
i
number of turns in session
K number of words in turn
L number of labels in multi-label set
C turn context
w
ij
turn word sequence
w
ijk
word embedding vector
h
ijk
hidden state of word encoder
X
ij
turn vector representation
r
ij
speaker role
X
C
ij
set of turn context vectors
g
ij
hidden state of context encoder
V
ij
turn context vector representation
y
ij
turn label
Y
ij
turn multi-label set
z
i
session label
Z
i
session multi-label set
s
ij
sample weight
T
ij
task relevance of the ij
th
sample
9
Chapter 2
Behavioral Coding Data
Description
2.1 Motivational Interviewing Corpus
Motivational interviewing (MI) is a client-centered, goal-oriented therapy, which
focuses on resolving clients' ambivalence towards their problems in order to moti-
vate behavior change [6]. Motivational interviewing (MI) is a client-centered, goal-
oriented therapy, which focuses on resolving clients' ambivalence towards their
problems in order to motivate behavior change [6]. This is achieved by the coun-
selor exhibiting and eliciting certain desired behaviors, while avoiding certain unde-
sired behaviors. Desirable counselor behaviors include re
ective listening and ask-
ing open-ended questions, with the purpose of motivating the client to give reasons
and make commitments towards behavior change. Undesirable counselor behav-
iors include confrontation and advising the client without their permission. Several
studies suggest that the counselor's skill in adhering to the therapy goals in
uence
client outcomes [46, 47]. Furthermore there is considerable interest in how these
measures directly relate to the spoken language of the counselor and client [48,49].
The Motivational Interviewing Corpus consists of over 1,700 psychotherapy
sessions conducted as part of six independent clinical trials. All of the six trials
focused of motivational interviewing for addressing various forms of addiction,
10
including alcohol (ARC, ESPSB, ESB21, CTT), marijuana (iCHAMP), and poly-
drug abuse (HMCBI) [48, 50]. The CTT and HMCBI consist of both real and
standardized patients. Standardized patients are actors portraying patients with
relevant addiction issues for the purpose of therapist training. A subset of the
sessions (N=345) were manually transcribed and subsequently segmented at the
utterance level. The utterances then received behavioral codes according to the
Motivational Interviewing Skill Code (MISC) manual [51]. Of the 345, eight were
discarded due to errors or inconsistencies in transcription or behavioral coding,
resulting in 337 sessions being considered for the present work.
2.1.1 Motivational Interviewing Skill Code
The MISC manual denes session-level and utterance-level behavioral codes for
evaluating motivational interviewing therapy sessions. One session-level gestalt
code of particular importance is empathy. It is dened as \the extent to which the
therapist understands and/or makes an eort to grasp the clients perspective".
Expression of empathy helps therapists to establish a rapport with their patients,
which leads to more eective interventions and ultimately more desirable outcomes
[46,47,52]. Is is rated on a 1-5 likert scale (1 `low' empathy to 5 `high' empathy). A
therapist receiving a rating of 4 or higher is considered `procient'. Other session-
level gestalt behavioral codes, e.g., MI spirit, are not considered in this work as they
have changed in various versions of the MISC manual. We show the distributions
of the empathy ratings in Figure 2.1.
The MISC manual denes 28 utterance-level behaviors dened in the manual:
19 therapist and 9 client. We show the utterance-level MISC codes and their
occurrences in the MI dataset in Table 3.2. Codes of especial importance for
evaluating therapist ecacy are open and closed questions and simple and complex
11
1234567
empathy
0
10
20
30
40
50
60
70
80
90
100
count
Figure 2.1: Empathy distribution of the MI dataset
re
ections, as these are considered essential elements of motivational interviewing.
The ratio of open to closed questions, complex to simple re
ections, and re
ections
to questions are all considered important indicators of a motivational interviewer's
skill.
For client utterances, the statements of most interest are ones where the client
is discussing change of the target behavior of interest. Statements which include
commitment, reasoning, or taking steps towards behavior change are referred to
as `change talk' and considered desirable behaviors to be elicited from the patient.
Statements which include commitment, reasoning, or taking steps towards contin-
uing negative behaviors are referred to as `sustain talk'. Change talk and sustain
talk are denoted by `+' and `-', respectively, in Table 3.2. Elicitation of these
12
Table 2.1: MISC Categories
Code Category Count
Counselor
ADP Advise with permission 105
ADW Advise w/o permission 598
AF Arm 1649
CO Confront 187
DI Direct 134
EC Emphasize Control 133
FA Facilitate 16296
FI Filler 157
GI Giving Information 15748
QUC Closed Question 5276
QUO Open Question 4562
RCP Raise Concern with permission 4
RCW Raise Concern w/o permission 42
REC Complex Re
ection 4703
RES Simple Re
ection 6354
RF Reframe 19
ST Structure 1223
SU Support 642
WA Warn 65
Client
C+/C- Commitment 111/21
FN Follow/Neutral 47491
R+/R- Reason 3278/2828
O+/O- Other 1788/1638
TS+/TS- Taking Steps 133/51
statements are an important aspect of motivational interviewing as the therapist
will help the client explore ambivalence, i.e., commitments, reasoning, and taking
steps towards or changing or sustaining the targeted behaviors of the therapy.
13
2.2 Cognitive Behavioral Therapy Corpus
Cognitive behavioral therapy is a problem-focused action-oriented psychotherapy
that focuses on developing strategies for clients to change dysfunctional cognitions
and behaviors [53]. Treatment is based on the therapist's conceptualization and
understanding of patients' beliefs and patterns of behavior. The therapist then
develops strategies with and assigns tasks to the patient which are aimed at mod-
ifying dysfunctional cognitions to bring about enduring changes in behavior and
aect.
The Cognitive Behavioral Therapy Corpus consists of over 5,000 audio record-
ings of therapists conducting cognitive behavioral therapy sessions [7]. More than
2,000 of these sessions have received manual behavioral coding according to the
Cognitive Therapy Rating System (CTRS) manual [54]. Of the behaviorally coded
sessions 100 were chosen for manual transcription. The sessions chosen for tran-
scription were the sessions which received the 50 highest and 50 lowest total CTRS
ratings (sum across 11 behavioral code ratings). Eight of these sessions are not con-
sidered due to formatting/transcription quality issues, leaving 92 sessions included
70 therapists to be considered in this work.
2.2.1 Cognitive Therapy Rating System
The CTRS denes 11 session-level behavioral codes, which are each scored on a 7
point likert scale (0 `poor' to 6 'excellent'). Figure 2.2 shows the distributions of
the CTRS codes in the dataset.
We pose this as a binary prediction task by assigning codes greater or equal
to 3 as `high' and those less than 3 as `low'. The choice of a rating 3 three as
the cut-o point is motivated by this score being the threshold for `satifactory'
14
0123456
agenda
0
10
20
30
count
0123456
application of technique
0
10
20
30
count
0123456
collaboration
0
5
10
15
20
25
count
0123456
feedback
0
10
20
30
count
0123456
guideddiscovery
0
10
20
30
count
0123456
homework
0
10
20
30
count
0123456
interpersonal
0
10
20
30
count
0123456
key cognition behavior
0
5
10
15
20
25
count
0123456
pacing and time
0
5
10
15
20
25
count
0123456
strategy for change
0
5
10
15
20
25
count
0123456
understanding
0
5
10
15
20
25
count
10 20 30 40 50
totalCTRS
0
5
10
15
20
25
count
Figure 2.2: CTRS behavioral code distributions
presence of the target behavior [53]. These are given in table 2.2 with the ratio
of `high' to `low' labels for each behavioral code. All the dened codes re
ect
therapist behaviors in the session, there are no codes which re
ect client behaviors.
They can be associated into a few broad categories including: management and
structure of the session (setting an agenda (AG), pacing and time (PT), assigning
homework (HW)), the aspects of the therapist-client relationship (interpersonal
relationship (IP), collaboration (CO), understanding (UN)), and conceptualization
of the clients' presented concerns and approaches for addressing them (providing
feedback (FB), guided discovery (GD), focusing on key cognitions and behaviors
(KC), developing a strategy for change (SC), application of cognitive-behavioral
techniques (AT)) [45].
15
Table 2.2: Session-level behavior codes dened by the CTRS manual
Abbr. CTRS Code `high'/`low'
AG agenda 47/45
AT application of cognitive-behavioral techniques 44/48
CO collaboration 62/30
FB feedback 46/46
GD guided discovery 48/44
HW homework 43/49
IP interpersonal eectiveness 82/10
KC focusing on key cognitions and behaviors 48/44
PT pacing and ecient use of time 51/41
SC strategy for change 46/46
UN understanding 71/21
2.3 Training and Testing Data Splits
In Table 2.3, we give an overview of the number of subjects, sessions, turns, and
spoken words for the counselors and clients in the MI and CBT datasets. These
are separated into `training' and `testing' splits which will be used for empirical
evaluation in subsequent sections. For both datasets there is an approximately 2:1
ratio of training to testing data (228:109 and 62:30 in the MI and CBT corpora,
respectively). Some counselors appear in multiple sessions, so all the sessions from
a given counselor are assigned to the same split to ensure speaker independent
training/testing splits. All standardized patient sessions in the MI dataset are
assigned to the training set. This is done because there are only three unique
standardized patient stories, so the language is likely very similar between many
of these sessions.
16
Table 2.3: MI and CBT Data: Session, turn, and word counts in training/testing
splits.
Subject Sessions Turns Words
MI
Counselor 228/109 28.7K/13.9K 579K/248K
Client 228/109 28.6K/13.6K 563K/269K
CBT
Counselor 62/30 11.3k/4.5k 180k/77.4k
Client 62/30 11.5k/4.7k 215k/109k
Table 2.4: Example of Therapist-Client Interaction.
turn role text
MISC
behaviors
1 T hi roger welcome my name is bill and i'm your counselor FI, GI
2 C hi okay FN
3 T so you've been here about ten days how's it going for you? RES, QUO
4 C i think in some ways i'm doing pretty well but i'm having
R+, R-
some concerns about whether this is where i need to be
.
.
.
17
Chapter 3
A Deep Learning Approach for
Behavioral Coding
3.1 Introduction
In [43], we proposed the rst deep learning approach for predicting both utterance
and session level behavioral codes. In this work we proposed a deep learning system
which uses the manually annotated local behavioral codes to train a recurrent
neural network (RNN) which learns a mapping from the client/counselor dialogue
to these local behavioral acts. Subsequently, this network is used to initialize
the lower layers of a deep network for predicting the global counselor empathy
rating. We show that this method outperforms training the DNN end-to-end in a
single stage and also outperforms a baseline neural network model that attempts to
predict empathy ratings directly from text without modeling turn level behavioral
dynamics.
3.2 End-to-end Deep Learning Network
We learn word embeddings to represent the language use of the counselor and
client using a continuous bag of words model [55]. Each word, is represented by
18
an d-dimensional dense vector, w. Each turn in a session is represented by the
average of the word vectors belonging to the words in that turn, i.e,
X
ij
=
1
jw
ij
j
X
w2w
ij
w; (3.1)
wherew
ij
is the set of words in turnj in sessioni. Every session is now represented
by a sequence of turn vectors, X
i
=fX
ij
g
M
i
j=1
, where M
i
is the number of turns in
session i. An additional indicator variable is appended to each vector to identify
whether the turn belongs to the counselor or client.
H
i
t
¯
Y
i
Mean Pooling Empathy Prediction
X
i
t
Feature
Transformation
Feature
Transformation
Feature
Transformation
Turn Dynamics
(RNN)
Turn Dynamics
(RNN)
Turn Dynamics
(RNN)
Local Behavior
Encoding
Local Behavior
Encoding
Local Behavior
Encoding
X
i
t 1
X
i
t+1
H
i
t 1
H
i
t+1
{
{
1 2
··· ···
ˆ z
i ˜
Y
i
t
˜
Y
i
t+1
˜
Y
i
t 1
´
X
i
t 1
´
X
i
t
´
X
i
t+1
Figure 3.1: Proposed System Overview.
Each turn has an associated L-dimensional k-hot target local behavior vector,
Y
ij
, representing all k MISC codes that occur in that turn. We use either L = 8
orL = 28 depending on the cardinality of the MISC code set we are working with
(see Figure 3.1). An example of k-hot MISC8 vector is given in Figure 3.2.
We train the proposed system in two stages: the rst maps from the turn
vectors,x
i
, to the k-hot representation of the local behaviors, Y
i
; the second maps
from the estimated local behaviors to the binarized session empathy score,Z
i
. We
refer to the rst stage as the encoder as it maps from the larger dimensional feature
representation to a low dimensional representation of local behaviors. We refer to
19
Table 3.1: MISC8 code grouping and counts in the dataset.
MISC8 MISC28 Count
Counselor
FA Facilitate 14659
GI Giving information 11880
QUC Closed question 6850
QUO Open question 5602
REC Complex re
ection 5825
RES Simple re
ection 8508
COU
MI adherent: Arm; Reframe;
6,236
Emphasize control; Support; Filler;
Advice with permission; Structure;
Raise concern with permission
MI non-adherent: Confront; Direct;
Advice without permission; Warn;
Raise concern without permission
Client
CLI
Follow/Neutral
48,283
Change talk: positive
Reasons; Commitments;
Taking steps; Other
Sustain talk: negative
Reasons; Commitments;
Taking steps; Other
the second stage as the decoder as it attempts to predict the session empathy score
from the estimated local behavioral encodings.
In the rst stage, the turn vector sequences are input to a dense feedforward
layer to allow for feature interactions. The transformed turn vectors,

X
ij
(d+1-
dimensional), are then input to an RNN which learns the dynamics of the turn
vector sequence. The output of the RNN,H
ij
(d+1-dimensional), is input to a feed-
forward layer which maps to the local behavioral encoding, Y
ij
, an L-dimensional
vector. We train this stage in a supervised manner using the reference local behav-
ior vectors as the multi-label targets. We use a sigmoid activation function on each
20
Client: I wouldn't mind
coming here for treatment
but I don't want to go to
one of those places where
everyone sits around crying
and complaining all day.
Counselor: You don't want
that. So you are kind of
wondering what it would be
like here.
t=0
t=1
CLI
RES, REC
0
0
0
0
0
0
0
1
0
0
0
0
1
1
0
0
Figure 3.2: Example of k-hot encoding of MISC behaviors
output of the encoding layer and the system is trained to minimize the average of
the binary cross-entropy between the reference and predicted outputs as given by:
E
1
=
N
X
i=1
M
i
X
j=1
L
X
l=1
Y
ij
(l) log

^
Y
ij
(l)

+ (1Y
ij
(l)) log

1
^
Y
ij
(l)

;
(3.2)
where,
^
Y
ij
=(
1
(X
ij
));
(3.3)
and
1
is encoder. We use long short-term memory (LSTM) RNNs [56] to address
the vanishing gradient problem that arises while training traditional RNNs [57].
In the second stage, we take the output of the rst stage prior to the sigmoid
activation,
~
Y
ij
=
1
(X
i
), as input. This encoding layer is averaged across turns
and input to a feedforward layer that predicts the session's empathy ratingZ
i
. We
use the sigmoid activation function and binary cross-entropy, as given by:
E
2
=
N
X
i=1
Z
i
log(
^
Z
i
) + (1Z
i
) log(1
^
Z
i
): (3.4)
where,
^
Z
i
=(
2
(
~
Y
i
));
(3.5)
21

2
is the decoder and
~
Y
i
=f
~
Y
ij
g
M
i
j=1
. We train the proposed system end-to-end,
i.e. back-propagating the error from the empathy prediction layer back to the turn
vector input layer. An overview of the proposed system is shown in Figure 3.1.
As a baseline model, we train a feedforward neural network with X
i
as input
to predict the empathy scores Z
i
. This model has a single fully connected layer
(X
ij
!

X
ij
) followed by averaging across all session turns (

X
i
=
1
M
i
P
M
i
j=1

X
ij
),
which is then input to the output layer, with sigmoid activation (

X
i
!
^
Z
i
).
We also predict empathy using the reference local behavioral representations,
Y
i
using the same network topology as the baseline, i.e., (Y
ij
!

Y
ij
!

Y
i
!
^
Z
i
).
This serves to show how much information we can gain about the annotated global
behavior from the annotated local behaviors.
3.3 Model Implementation and Training
We use the word2vec software to learn 300-dimensional word embedding vectors
from the counselor and client transcripts [55,58]. All neural network congurations
are learned using Keras [59] with Theano [60] as the back-end.
We perform z-normalization on all features using the empirical mean and stan-
dard deviation of the training set. The turns from training sessions were segmented
into sequences of 10 turns which overlapped by 50%. The data was shued and
divided into batches of size 32. Ten percent of the training data was randomly
assigned to a validation set for training. During each training epoch the batches
were shued to avoid overtting to any particular batch. All layer weights were
initialized with Glorot uniform initialization [61]. Ten percent dropout was applied
to the output of each layer to avoid overtting. The system was trained in 100
epochs (training was terminated early if the validation loss did not improve in
22
three consecutive epochs) with the model from the epoch giving the minimum loss
on the validation set being retained. The training procedure was optimized using
the ADAM algorithm [62].
3.4 Experiments and Results
In Table 3.2, we show recall, precision, and the f1-score of the rst stage, averaged
over the result of each output target. The MISC8 representation is much more
robustly predicted as it is a simplied version which only focuses on learning the
most frequently occurring codes.
Table 3.2: MISC Prediction.
code recall precision f1-score
MISC8 0.617 0.675 0.643
MISC28 0.228 0.348 0.258
In Table 3.3, we show unweighted average recall (UAR) of predicting the global
behavior, empathy. In this table, `baseline' refers to the model that directly predicts
empathy from the turn vectors and `reference' refers to the model that predicts
empathy from the reference MISC labels assigned to each turn. The proposed
system was trained in two stages as described in the Section 3.2. This is referred
to as `pre-training' in the table. For reference we also include the results from
training the full system end-to-end in a single stage, i.e., back-propagating the
error from the empathy prediction layer back to the turn vector input layer without
rst learning the mapping from the turn vectors to local behaviors.
The reference model using all 28 MISC codes produced the highest UAR
(79.6%) of all the prediction models. This demonstrates that the local behavioral
codes carry important information about the global empathy code. The MISC8
23
Table 3.3: Empathy Prediction.
model L UAR (%)
baseline N/A 71.8
reference
8 73.6
28 79.6
proposed system w/o pre-training
8 65.0
28 62.9
proposed system w pre-training
8 78.6
28 72.9
reference model also gave better performance than the baseline model (73.6% vs.
71.8%), so while it does not make full use of the local codes the reduced set still
carries important information for the empathy prediction task.
The proposed system, when trained in a single stage, fails to match the per-
formance of the shallow baseline system (65.0/62.9% vs. 71.8%). This result
suggests that there is not enough data to train a system of this depth without
any supervision of the intermediary layers. When pre-training the encoder of the
proposed system, we see improved performance over the baseline for both the 8
and 28 dimensional encoding layers (78.6% and 72.9%, respectively). Interestingly,
we see a reversed order of performance between the 8 and 28 dimensional encoding
layers compared to the predictions of the reference model. This is likely due to the
diculty of making accurate predictions of the 28 MISC codes including due to
increased data sparsity issues. The relatively stronger performance of predicting
the MISC8 labels gives a better initialization for the deep network. By back-
propagating the error through the entire system with the rst stage initialization,
the deep system with 8 dimensional encoding layer achieves better performance
than the reference system with reference MISC8 input (78.6% vs. 73.6%). This is
most likely due to the deep system being allowed to learn from both the turn level
language dynamics as well as local behavioral acts.
24
3.5 Conclusions
In this chapter, we presented a deep neural network system to predict counselor
empathy from MI session transcripts. We demonstrated that by training the system
in two stages, using local behavioral acts as supervision for the rst stage, we are
able to outperform a baseline shallow neural network predicting counselor empathy
rating directly from input turn vectors.
25
Chapter 4
Attention Networks for
Behavioral Coding
4.1 Introduction
The words spoken by therapists and clients in psychotherapy are of vital impor-
tance as they are the primary mode of communication. Prior work suggests there
are many key words and phrases which indicate behaviors of interest in psychother-
apy [15, 63, 64]. In [65], we proposed word-level LSTMs with an attention mech-
anism for predicting utterance level behaviors in MI sessions. This model maps
between word sequences from speaker turns to MISC behaviors present in the turns
while also learning to identify words of interest in the sequences.
4.2 Word Attention Network
We consider each speaker turn from the sessions to be a sequence of words, w =
fw
1
;w
2
; ;w
K
g. Because turns can contain multiple utterances, they may also
have multiple behavior labels assigned to them (e.g., a turn can contain both
a question and a re
ection). So the label, y, for a particular behavioral code
corresponds to the presence or absence of that behavior in the turn. Note, this is
a distinct approach from [42], where there is a one-to-one correspondence between
utterances and behavioral labels. This choice is made to facilitate fully automated
26
prediction, where segmenting by turn is a more well dened task than segmenting
by utterance.
We follow the procedure of Xiao et al. [42] and group the most infrequent of
these codes into composite groups resulting in 11 target labels: 8 therapist and
3 client. The non-grouped therapist codes are facilitate (FA), giving informa-
tion (GI), close and open questions (QUC/QUO), simple and complex re
ections
(RES/REC). The remainder of the codes are grouped into MI adherent (MIA),
i.e., behaviors that adhere to the spirit of the MI treatment, and MI non-adherent
(MIN), those which are inconsistent with MI. Client codes are follow/neutral (FN),
which covers the majority of the client utterances. This code indicates that the
client made a statement that was neutral towards changing the targeted behavior
of the therapy. The remainder of the client codes are grouped into positive `change
talk' (POS) or negative `sustain talk' (NEG) behaviors. Change talk is a state-
ment that re
ects a client's reasoning, commitments, or steps towards behavior
change. Sustain talk re
ects the opposite. We show the individual and grouped
MISC codes and their occurrences in Table 4.1
We rst learn a word embedding vector representation of the words used by
the counselors and clients in the motivational interviewing data. Each word, w, is
represented by an d-dimensional vector, w (d=300). The word embedding vectors
are input to the rst layer of the network. The rst layer is a single feed-forward
layer used to ne tune the word vector representation:
v
ijk
=W
v
w
ijk
+b
v
;
(4.1)
for the k
th
word in the j
th
turn in the i
th
session.
27
Table 4.1: MISC11 code grouping and counts in the dataset.
Group MISC Code Count
Counselor
FA Facilitate 14659
GI Giving information 11880
QUC Closed question 6850
QUO Open question 5602
REC Complex re
ection 5825
RES Simple re
ection 8508
MIA
MI adherent: Arm; Reframe;
5072
Emphasize control; Support; Filler;
Advice with permission; Structure;
Raise concern with permission
MIN
MI non-adherent: Confront; Direct;
1164 Advice without permission; Warn;
Raise concern without permission
Client
FN Follow/Neutral 37937
POS
Change talk: positive
5681 Reasons; Commitments;
Taking steps; Other
NEG
Sustain talk: negative
4665
Reasons; Commitments;
Taking steps; Other
4.2.1 Attention-based LSTM Network
The ne tuned feature vectors are then fed into a forward LSTM, resulting in the
hidden state vectors, h
jt
, i.e.,
fh
j1
;h
j2
; ;h
jK
g = LSTM(fv
j1
;v
j2
; ;v
jK
g):
(4.2)
These hidden state vectors represent the words and their left context. Subsequently
these representations are fed into the word attention mechanism.
28
The word attention mechanism consumes the hidden state vectors from the
LSTM and feeds them to a single layer feedforward network that projects them
into 1 dimension and applies the `tanh' activation function, according to,
u
ijk
= tanh(W
u
h
ijk
+b
u
):
(4.3)
These weights represent the estimated importance of the hidden vector at time,
t. Subsequently, the weights are input to the softmax function for normalization
across time, i.e.,

ijk
=
exp(u
ijk
)
P
t
exp(u
ijk
)
; (4.4)
giving the relative importance of each hidden vector with respect to the other
vectors from that turn. The hidden vectors are then combined via a weighted
average according to their relative importance,
X
ij
=
X
t

ijk
h
ijk
:
(4.5)
The nal stage of the network is a single feedforward layer with a sigmoid
activation to convert the weighted average vector into the probability of a particular
behavioral being present in that turn, i.e.,
^ y
ij
=(W
y
X
ij
+b
y
):
(4.6)
The network is trained to minimize the binary cross-entropy using the ADAM
algorithm [62]. We show an overview of the proposed attention-based LSTM in
Figure 4.1. For reference, we present the LSTM network without the attention
mechanism, which in eect means
jt
=
1
K
;8t2f1; ;Kg.
29
↵ ij1
AAAB83icbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bjw4jGCeUCyhN7JbDJmdnaYmRXCkt/w4kERr/6MN//GSbIHTSxoKKq66e6KlODG+v63t7a+sbm1Xdop7+7tHxxWjo5bJs00ZU2ailR3IjRMcMmallvBOkozTCLB2tH4dua3n5g2PJUPdqJYmOBQ8phTtE7q9VCoEfZz/hhM+5WqX/PnIKskKEgVCjT6la/eIKVZwqSlAo3pBr6yYY7acirYtNzLDFNIxzhkXUclJsyE+fzmKTl3yoDEqXYlLZmrvydyTIyZJJHrTNCOzLI3E//zupmNb8KcS5VZJuliUZwJYlMyC4AMuGbUiokjSDV3txI6Qo3UupjKLoRg+eVV0rqsBX4tuL+q1v0ijhKcwhlcQADXUIc7aEATKCh4hld48zLvxXv3Phata14xcwJ/4H3+AApTkZ0= AAAB83icbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bjw4jGCeUCyhN7JbDJmdnaYmRXCkt/w4kERr/6MN//GSbIHTSxoKKq66e6KlODG+v63t7a+sbm1Xdop7+7tHxxWjo5bJs00ZU2ailR3IjRMcMmallvBOkozTCLB2tH4dua3n5g2PJUPdqJYmOBQ8phTtE7q9VCoEfZz/hhM+5WqX/PnIKskKEgVCjT6la/eIKVZwqSlAo3pBr6yYY7acirYtNzLDFNIxzhkXUclJsyE+fzmKTl3yoDEqXYlLZmrvydyTIyZJJHrTNCOzLI3E//zupmNb8KcS5VZJuliUZwJYlMyC4AMuGbUiokjSDV3txI6Qo3UupjKLoRg+eVV0rqsBX4tuL+q1v0ijhKcwhlcQADXUIc7aEATKCh4hld48zLvxXv3Phata14xcwJ/4H3+AApTkZ0= AAAB83icbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bjw4jGCeUCyhN7JbDJmdnaYmRXCkt/w4kERr/6MN//GSbIHTSxoKKq66e6KlODG+v63t7a+sbm1Xdop7+7tHxxWjo5bJs00ZU2ailR3IjRMcMmallvBOkozTCLB2tH4dua3n5g2PJUPdqJYmOBQ8phTtE7q9VCoEfZz/hhM+5WqX/PnIKskKEgVCjT6la/eIKVZwqSlAo3pBr6yYY7acirYtNzLDFNIxzhkXUclJsyE+fzmKTl3yoDEqXYlLZmrvydyTIyZJJHrTNCOzLI3E//zupmNb8KcS5VZJuliUZwJYlMyC4AMuGbUiokjSDV3txI6Qo3UupjKLoRg+eVV0rqsBX4tuL+q1v0ijhKcwhlcQADXUIc7aEATKCh4hld48zLvxXv3Phata14xcwJ/4H3+AApTkZ0= AAAB83icbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bjw4jGCeUCyhN7JbDJmdnaYmRXCkt/w4kERr/6MN//GSbIHTSxoKKq66e6KlODG+v63t7a+sbm1Xdop7+7tHxxWjo5bJs00ZU2ailR3IjRMcMmallvBOkozTCLB2tH4dua3n5g2PJUPdqJYmOBQ8phTtE7q9VCoEfZz/hhM+5WqX/PnIKskKEgVCjT6la/eIKVZwqSlAo3pBr6yYY7acirYtNzLDFNIxzhkXUclJsyE+fzmKTl3yoDEqXYlLZmrvydyTIyZJJHrTNCOzLI3E//zupmNb8KcS5VZJuliUZwJYlMyC4AMuGbUiokjSDV3txI6Qo3UupjKLoRg+eVV0rqsBX4tuL+q1v0ijhKcwhlcQADXUIc7aEATKCh4hld48zLvxXv3Phata14xcwJ/4H3+AApTkZ0=
↵ ij2
AAAB83icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeCF48V7Ac0oUy2m3btJll2N0IJ/RtePCji1T/jzX/jts1BWx8MPN6bYWZeKAXXxnW/ndLG5tb2Tnm3srd/cHhUPT7p6DRTlLVpKlLVC1EzwRPWNtwI1pOKYRwK1g0nt3O/+8SU5mnyYKaSBTGOEh5xisZKvo9CjnGQ88fGbFCtuXV3AbJOvILUoEBrUP3yhynNYpYYKlDrvudKE+SoDKeCzSp+pplEOsER61uaYMx0kC9unpELqwxJlCpbiSEL9fdEjrHW0zi0nTGasV715uJ/Xj8z0U2Q80RmhiV0uSjKBDEpmQdAhlwxasTUEqSK21sJHaNCamxMFRuCt/ryOuk06p5b9+6vak23iKMMZ3AOl+DBNTThDlrQBgoSnuEV3pzMeXHenY9la8kpZk7hD5zPHwvYkZ4= AAAB83icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeCF48V7Ac0oUy2m3btJll2N0IJ/RtePCji1T/jzX/jts1BWx8MPN6bYWZeKAXXxnW/ndLG5tb2Tnm3srd/cHhUPT7p6DRTlLVpKlLVC1EzwRPWNtwI1pOKYRwK1g0nt3O/+8SU5mnyYKaSBTGOEh5xisZKvo9CjnGQ88fGbFCtuXV3AbJOvILUoEBrUP3yhynNYpYYKlDrvudKE+SoDKeCzSp+pplEOsER61uaYMx0kC9unpELqwxJlCpbiSEL9fdEjrHW0zi0nTGasV715uJ/Xj8z0U2Q80RmhiV0uSjKBDEpmQdAhlwxasTUEqSK21sJHaNCamxMFRuCt/ryOuk06p5b9+6vak23iKMMZ3AOl+DBNTThDlrQBgoSnuEV3pzMeXHenY9la8kpZk7hD5zPHwvYkZ4= AAAB83icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeCF48V7Ac0oUy2m3btJll2N0IJ/RtePCji1T/jzX/jts1BWx8MPN6bYWZeKAXXxnW/ndLG5tb2Tnm3srd/cHhUPT7p6DRTlLVpKlLVC1EzwRPWNtwI1pOKYRwK1g0nt3O/+8SU5mnyYKaSBTGOEh5xisZKvo9CjnGQ88fGbFCtuXV3AbJOvILUoEBrUP3yhynNYpYYKlDrvudKE+SoDKeCzSp+pplEOsER61uaYMx0kC9unpELqwxJlCpbiSEL9fdEjrHW0zi0nTGasV715uJ/Xj8z0U2Q80RmhiV0uSjKBDEpmQdAhlwxasTUEqSK21sJHaNCamxMFRuCt/ryOuk06p5b9+6vak23iKMMZ3AOl+DBNTThDlrQBgoSnuEV3pzMeXHenY9la8kpZk7hD5zPHwvYkZ4= AAAB83icbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeCF48V7Ac0oUy2m3btJll2N0IJ/RtePCji1T/jzX/jts1BWx8MPN6bYWZeKAXXxnW/ndLG5tb2Tnm3srd/cHhUPT7p6DRTlLVpKlLVC1EzwRPWNtwI1pOKYRwK1g0nt3O/+8SU5mnyYKaSBTGOEh5xisZKvo9CjnGQ88fGbFCtuXV3AbJOvILUoEBrUP3yhynNYpYYKlDrvudKE+SoDKeCzSp+pplEOsER61uaYMx0kC9unpELqwxJlCpbiSEL9fdEjrHW0zi0nTGasV715uJ/Xj8z0U2Q80RmhiV0uSjKBDEpmQdAhlwxasTUEqSK21sJHaNCamxMFRuCt/ryOuk06p5b9+6vak23iKMMZ3AOl+DBNTThDlrQBgoSnuEV3pzMeXHenY9la8kpZk7hD5zPHwvYkZ4=
↵ ij3
AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQq6LHgxWMF+wFNKJPtpl272Sy7G6GE/g0vHhTx6p/x5r9x2+agrQ8GHu/NMDMvkpxp43nfTmltfWNzq7xd2dnd2z+oHh61dZopQlsk5anqRqgpZ4K2DDOcdqWimEScdqLx7czvPFGlWSoezETSMMGhYDEjaKwUBMjlCPs5e7yc9qs1r+7N4a4SvyA1KNDsV7+CQUqyhApDOGrd8z1pwhyVYYTTaSXINJVIxjikPUsFJlSH+fzmqXtmlYEbp8qWMO5c/T2RY6L1JIlsZ4JmpJe9mfif18tMfBPmTMjMUEEWi+KMuyZ1ZwG4A6YoMXxiCRLF7K0uGaFCYmxMFRuCv/zyKmlf1H2v7t9f1RpeEUcZTuAUzsGHa2jAHTShBQQkPMMrvDmZ8+K8Ox+L1pJTzBzDHzifPw1dkZ8= AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQq6LHgxWMF+wFNKJPtpl272Sy7G6GE/g0vHhTx6p/x5r9x2+agrQ8GHu/NMDMvkpxp43nfTmltfWNzq7xd2dnd2z+oHh61dZopQlsk5anqRqgpZ4K2DDOcdqWimEScdqLx7czvPFGlWSoezETSMMGhYDEjaKwUBMjlCPs5e7yc9qs1r+7N4a4SvyA1KNDsV7+CQUqyhApDOGrd8z1pwhyVYYTTaSXINJVIxjikPUsFJlSH+fzmqXtmlYEbp8qWMO5c/T2RY6L1JIlsZ4JmpJe9mfif18tMfBPmTMjMUEEWi+KMuyZ1ZwG4A6YoMXxiCRLF7K0uGaFCYmxMFRuCv/zyKmlf1H2v7t9f1RpeEUcZTuAUzsGHa2jAHTShBQQkPMMrvDmZ8+K8Ox+L1pJTzBzDHzifPw1dkZ8= AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQq6LHgxWMF+wFNKJPtpl272Sy7G6GE/g0vHhTx6p/x5r9x2+agrQ8GHu/NMDMvkpxp43nfTmltfWNzq7xd2dnd2z+oHh61dZopQlsk5anqRqgpZ4K2DDOcdqWimEScdqLx7czvPFGlWSoezETSMMGhYDEjaKwUBMjlCPs5e7yc9qs1r+7N4a4SvyA1KNDsV7+CQUqyhApDOGrd8z1pwhyVYYTTaSXINJVIxjikPUsFJlSH+fzmqXtmlYEbp8qWMO5c/T2RY6L1JIlsZ4JmpJe9mfif18tMfBPmTMjMUEEWi+KMuyZ1ZwG4A6YoMXxiCRLF7K0uGaFCYmxMFRuCv/zyKmlf1H2v7t9f1RpeEUcZTuAUzsGHa2jAHTShBQQkPMMrvDmZ8+K8Ox+L1pJTzBzDHzifPw1dkZ8= AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQq6LHgxWMF+wFNKJPtpl272Sy7G6GE/g0vHhTx6p/x5r9x2+agrQ8GHu/NMDMvkpxp43nfTmltfWNzq7xd2dnd2z+oHh61dZopQlsk5anqRqgpZ4K2DDOcdqWimEScdqLx7czvPFGlWSoezETSMMGhYDEjaKwUBMjlCPs5e7yc9qs1r+7N4a4SvyA1KNDsV7+CQUqyhApDOGrd8z1pwhyVYYTTaSXINJVIxjikPUsFJlSH+fzmqXtmlYEbp8qWMO5c/T2RY6L1JIlsZ4JmpJe9mfif18tMfBPmTMjMUEEWi+KMuyZ1ZwG4A6YoMXxiCRLF7K0uGaFCYmxMFRuCv/zyKmlf1H2v7t9f1RpeEUcZTuAUzsGHa2jAHTShBQQkPMMrvDmZ8+K8Ox+L1pJTzBzDHzifPw1dkZ8=
↵ ijK
AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQi6LHgRfBSwX5AE8pku2nXbjbL7kYooX/DiwdFvPpnvPlv3LY5aOuDgcd7M8zMiyRn2njet1NaW9/Y3CpvV3Z29/YPqodHbZ1mitAWSXmquhFqypmgLcMMp12pKCYRp51ofDPzO09UaZaKBzORNExwKFjMCBorBQFyOcJ+zh7vpv1qzat7c7irxC9IDQo0+9WvYJCSLKHCEI5a93xPmjBHZRjhdFoJMk0lkjEOac9SgQnVYT6/eeqeWWXgxqmyJYw7V39P5JhoPUki25mgGellbyb+5/UyE1+HORMyM1SQxaI4465J3VkA7oApSgyfWIJEMXurS0aokBgbU8WG4C+/vEraF3Xfq/v3l7WGV8RRhhM4hXPw4QoacAtNaAEBCc/wCm9O5rw4787HorXkFDPH8AfO5w8x1ZG3 AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQi6LHgRfBSwX5AE8pku2nXbjbL7kYooX/DiwdFvPpnvPlv3LY5aOuDgcd7M8zMiyRn2njet1NaW9/Y3CpvV3Z29/YPqodHbZ1mitAWSXmquhFqypmgLcMMp12pKCYRp51ofDPzO09UaZaKBzORNExwKFjMCBorBQFyOcJ+zh7vpv1qzat7c7irxC9IDQo0+9WvYJCSLKHCEI5a93xPmjBHZRjhdFoJMk0lkjEOac9SgQnVYT6/eeqeWWXgxqmyJYw7V39P5JhoPUki25mgGellbyb+5/UyE1+HORMyM1SQxaI4465J3VkA7oApSgyfWIJEMXurS0aokBgbU8WG4C+/vEraF3Xfq/v3l7WGV8RRhhM4hXPw4QoacAtNaAEBCc/wCm9O5rw4787HorXkFDPH8AfO5w8x1ZG3 AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQi6LHgRfBSwX5AE8pku2nXbjbL7kYooX/DiwdFvPpnvPlv3LY5aOuDgcd7M8zMiyRn2njet1NaW9/Y3CpvV3Z29/YPqodHbZ1mitAWSXmquhFqypmgLcMMp12pKCYRp51ofDPzO09UaZaKBzORNExwKFjMCBorBQFyOcJ+zh7vpv1qzat7c7irxC9IDQo0+9WvYJCSLKHCEI5a93xPmjBHZRjhdFoJMk0lkjEOac9SgQnVYT6/eeqeWWXgxqmyJYw7V39P5JhoPUki25mgGellbyb+5/UyE1+HORMyM1SQxaI4465J3VkA7oApSgyfWIJEMXurS0aokBgbU8WG4C+/vEraF3Xfq/v3l7WGV8RRhhM4hXPw4QoacAtNaAEBCc/wCm9O5rw4787HorXkFDPH8AfO5w8x1ZG3 AAAB83icbVBNS8NAEJ3Ur1q/qh69BIvgqSQi6LHgRfBSwX5AE8pku2nXbjbL7kYooX/DiwdFvPpnvPlv3LY5aOuDgcd7M8zMiyRn2njet1NaW9/Y3CpvV3Z29/YPqodHbZ1mitAWSXmquhFqypmgLcMMp12pKCYRp51ofDPzO09UaZaKBzORNExwKFjMCBorBQFyOcJ+zh7vpv1qzat7c7irxC9IDQo0+9WvYJCSLKHCEI5a93xPmjBHZRjhdFoJMk0lkjEOac9SgQnVYT6/eeqeWWXgxqmyJYw7V39P5JhoPUki25mgGellbyb+5/UyE1+HORMyM1SQxaI4465J3VkA7oApSgyfWIJEMXurS0aokBgbU8WG4C+/vEraF3Xfq/v3l7WGV8RRhhM4hXPw4QoacAtNaAEBCc/wCm9O5rw4787HorXkFDPH8AfO5w8x1ZG3
h
ijK
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPBi+Clgv2ANpTNdtKu3WzC7kYooT/CiwdFvPp7vPlv3LQ5aOuDgcd7M8zMCxLBtXHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFcMWi0WsugHVKLjEluFGYDdRSKNAYCeY3OR+5wmV5rF8MNME/YiOJA85o8ZKnfEg4493s0G15tbdOcgq8QpSgwLNQfWrP4xZGqE0TFCte56bGD+jynAmcFbppxoTyiZ0hD1LJY1Q+9n83Bk5s8qQhLGyJQ2Zq78nMhppPY0C2xlRM9bLXi7+5/VSE177GZdJalCyxaIwFcTEJP+dDLlCZsTUEsoUt7cSNqaKMmMTqtgQvOWXV0n7ou65de/+stZwizjKcAKncA4eXEEDbqEJLWAwgWd4hTcncV6cd+dj0Vpyiplj+APn8wdoNY+L AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPBi+Clgv2ANpTNdtKu3WzC7kYooT/CiwdFvPp7vPlv3LQ5aOuDgcd7M8zMCxLBtXHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFcMWi0WsugHVKLjEluFGYDdRSKNAYCeY3OR+5wmV5rF8MNME/YiOJA85o8ZKnfEg4493s0G15tbdOcgq8QpSgwLNQfWrP4xZGqE0TFCte56bGD+jynAmcFbppxoTyiZ0hD1LJY1Q+9n83Bk5s8qQhLGyJQ2Zq78nMhppPY0C2xlRM9bLXi7+5/VSE177GZdJalCyxaIwFcTEJP+dDLlCZsTUEsoUt7cSNqaKMmMTqtgQvOWXV0n7ou65de/+stZwizjKcAKncA4eXEEDbqEJLWAwgWd4hTcncV6cd+dj0Vpyiplj+APn8wdoNY+L AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPBi+Clgv2ANpTNdtKu3WzC7kYooT/CiwdFvPp7vPlv3LQ5aOuDgcd7M8zMCxLBtXHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFcMWi0WsugHVKLjEluFGYDdRSKNAYCeY3OR+5wmV5rF8MNME/YiOJA85o8ZKnfEg4493s0G15tbdOcgq8QpSgwLNQfWrP4xZGqE0TFCte56bGD+jynAmcFbppxoTyiZ0hD1LJY1Q+9n83Bk5s8qQhLGyJQ2Zq78nMhppPY0C2xlRM9bLXi7+5/VSE177GZdJalCyxaIwFcTEJP+dDLlCZsTUEsoUt7cSNqaKMmMTqtgQvOWXV0n7ou65de/+stZwizjKcAKncA4eXEEDbqEJLWAwgWd4hTcncV6cd+dj0Vpyiplj+APn8wdoNY+L AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPBi+Clgv2ANpTNdtKu3WzC7kYooT/CiwdFvPp7vPlv3LQ5aOuDgcd7M8zMCxLBtXHdb6e0tr6xuVXeruzs7u0fVA+P2jpOFcMWi0WsugHVKLjEluFGYDdRSKNAYCeY3OR+5wmV5rF8MNME/YiOJA85o8ZKnfEg4493s0G15tbdOcgq8QpSgwLNQfWrP4xZGqE0TFCte56bGD+jynAmcFbppxoTyiZ0hD1LJY1Q+9n83Bk5s8qQhLGyJQ2Zq78nMhppPY0C2xlRM9bLXi7+5/VSE177GZdJalCyxaIwFcTEJP+dDLlCZsTUEsoUt7cSNqaKMmMTqtgQvOWXV0n7ou65de/+stZwizjKcAKncA4eXEEDbqEJLWAwgWd4hTcncV6cd+dj0Vpyiplj+APn8wdoNY+L
P
t
h
ij1
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMKtjZYBG8sI5gHJEfY2k2TN3t6xuyeEI629jYUiauffY+d/4+ZRaOIHCz++b4admTCRwljf//Zya+sbm1v57cLO7t7+QfGw1DBxqjnWeSxj3QqZQSkU1q2wEluJRhaFEpvh6GqaN+9RGxGrWztOMIjYQIm+4Mw6qznsZuKOTrrFsl/xZyKrQBdQrpY+Ht4BoNYtfnV6MU8jVJZLZkyb+okNMqat4BInhU5qMGF8xAbYdqhYhCbIZuNOyKlzeqQfa/eUJTP3d0fGImPGUegqI2aHZjmbmv9l7dT2L4NMqCS1qPj8o34qiY3JdHfSExq5lWMHjGvhZiV8yDTj1l2o4I5Al1dehcZ5hfoVekPLVR/mysMxnMAZULiAKlxDDerAYQSP8AwvXuI9ea/e27w05y16juCPvM8fFIaRhw== AAAB7nicbZBLSwMxFIXv+Kz1VSuu3ASL4KpMBNFlwY3LCvYB7VAyaaaNzWRCkhHK0K0rN25EFBF3/h53/hvTx0JbDwQ+zrmX3HtDJbixvv/tLS2vrK6t5zbym1vbO7uFvWLdJKmmrEYTkehmSAwTXLKa5VawptKMxKFgjXBwOc4bd0wbnsgbO1QsiElP8ohTYp3V6HcyfotHnULJL/sToUXAMyhVih/3z2cHD9VO4avdTWgaM2mpIMa0sK9skBFtORVslG+nhilCB6THWg4liZkJssm4I3TsnC6KEu2etGji/u7ISGzMMA5dZUxs38xnY/O/rJXa6CLIuFSpZZJOP4pSgWyCxrujLteMWjF0QKjmblZE+0QTat2F8u4IeH7lRaiflrFfxte4VPFhqhwcwhGcAIZzqMAVVKEGFAbwCC/w6invyXvz3qelS96sZx/+yPv8ASdiklU= AAAB7nicbZBLSwMxFIXv+Kz1VSuu3ASL4KpMBNFlwY3LCvYB7VAyaaaNzWRCkhHK0K0rN25EFBF3/h53/hvTx0JbDwQ+zrmX3HtDJbixvv/tLS2vrK6t5zbym1vbO7uFvWLdJKmmrEYTkehmSAwTXLKa5VawptKMxKFgjXBwOc4bd0wbnsgbO1QsiElP8ohTYp3V6HcyfotHnULJL/sToUXAMyhVih/3z2cHD9VO4avdTWgaM2mpIMa0sK9skBFtORVslG+nhilCB6THWg4liZkJssm4I3TsnC6KEu2etGji/u7ISGzMMA5dZUxs38xnY/O/rJXa6CLIuFSpZZJOP4pSgWyCxrujLteMWjF0QKjmblZE+0QTat2F8u4IeH7lRaiflrFfxte4VPFhqhwcwhGcAIZzqMAVVKEGFAbwCC/w6invyXvz3qelS96sZx/+yPv8ASdiklU= AAAB7nicbVA9SwNBEJ2LXzF+RS1tFoNgFW5ttAzYWEYwH5AcYW+zl6zZ2zt254Rw5EfYWChi6++x89+4Sa7QxAcDj/dmmJkXpkpa9P1vr7SxubW9U96t7O0fHB5Vj0/aNskMFy2eqMR0Q2aFklq0UKIS3dQIFodKdMLJ7dzvPAljZaIfcJqKIGYjLSPJGTqpMx7k8pHOBtWaX/cXIOuEFqQGBZqD6ld/mPAsFhq5Ytb2qJ9ikDODkisxq/QzK1LGJ2wkeo5qFgsb5ItzZ+TCKUMSJcaVRrJQf0/kLLZ2GoeuM2Y4tqveXPzP62UY3QS51GmGQvPloihTBBMy/50MpREc1dQRxo10txI+ZoZxdAlVXAh09eV10r6qU79O72mt4RdxlOEMzuESKFxDA+6gCS3gMIFneIU3L/VevHfvY9la8oqZU/gD7/MHP8OPbg==
h
ij2
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMJdGlMGbCwjmAckR9jb7CVr9vaO3TkhHGntbSwUUTv/Hjv/GzePQhM/WPjxfTPszASJFAZd99vJbWxube/kdwt7+weHR8XjUsvEqWa8yWIZ605ADZdC8SYKlLyTaE6jQPJ2ML6a5e17ro2I1S1OEu5HdKhEKBhFa7VH/UzcVaf9YtmtuHORdfCWUK6XPh7eAaDRL371BjFLI66QSWpM13MT9DOqUTDJp4VeanhC2ZgOedeiohE3fjYfd0rOrTMgYaztU0jm7u+OjEbGTKLAVkYUR2Y1m5n/Zd0Uw5qfCZWkyBVbfBSmkmBMZruTgdCcoZxYoEwLOythI6opQ3uhgj2Ct7ryOrSqFc+teDdeue7CQnk4hTO4AA8uoQ7X0IAmMBjDIzzDi5M4T86r87YozTnLnhP4I+fzBxYLkYg= AAAB7nicbZC7SwNBEMbn4ivGV4xY2SwGwSrcBUTLgI1lBPOAJIS9zV6yZm/v2J0TwpHWysZGRBGx8++x879x8yg08YOFH983w86MH0th0HW/nczK6tr6RnYzt7W9s7uX3y/UTZRoxmsskpFu+tRwKRSvoUDJm7HmNPQlb/jDy0neuOPaiEjd4CjmnZD2lQgEo2itxqCbitvyuJsvuiV3KrIM3hyKlcLH/fPZ4UO1m/9q9yKWhFwhk9SYlufG2EmpRsEkH+faieExZUPa5y2LiobcdNLpuGNyYp0eCSJtn0IydX93pDQ0ZhT6tjKkODCL2cT8L2slGFx0UqHiBLlis4+CRBKMyGR30hOaM5QjC5RpYWclbEA1ZWgvlLNH8BZXXoZ6ueS5Je/aK1ZcmCkLR3AMp+DBOVTgCqpQAwZDeIQXeHVi58l5c95npRln3nMAf+R8/gAo55JW AAAB7nicbZC7SwNBEMbn4ivGV4xY2SwGwSrcBUTLgI1lBPOAJIS9zV6yZm/v2J0TwpHWysZGRBGx8++x879x8yg08YOFH983w86MH0th0HW/nczK6tr6RnYzt7W9s7uX3y/UTZRoxmsskpFu+tRwKRSvoUDJm7HmNPQlb/jDy0neuOPaiEjd4CjmnZD2lQgEo2itxqCbitvyuJsvuiV3KrIM3hyKlcLH/fPZ4UO1m/9q9yKWhFwhk9SYlufG2EmpRsEkH+faieExZUPa5y2LiobcdNLpuGNyYp0eCSJtn0IydX93pDQ0ZhT6tjKkODCL2cT8L2slGFx0UqHiBLlis4+CRBKMyGR30hOaM5QjC5RpYWclbEA1ZWgvlLNH8BZXXoZ6ueS5Je/aK1ZcmCkLR3AMp+DBOVTgCqpQAwZDeIQXeHVi58l5c95npRln3nMAf+R8/gAo55JW AAAB7nicbVA9SwNBEJ3zM8avqKXNYhCswl0aLQM2lhHMByRH2NvMJWt2947dPSEc+RE2ForY+nvs/Ddukis08cHA470ZZuZFqeDG+v63t7G5tb2zW9or7x8cHh1XTk7bJsk0wxZLRKK7ETUouMKW5VZgN9VIZSSwE01u537nCbXhiXqw0xRDSUeKx5xR66TOeJDzx/psUKn6NX8Bsk6CglShQHNQ+eoPE5ZJVJYJakwv8FMb5lRbzgTOyv3MYErZhI6w56iiEk2YL86dkUunDEmcaFfKkoX6eyKn0pipjFynpHZsVr25+J/Xy2x8E+ZcpZlFxZaL4kwQm5D572TINTIrpo5Qprm7lbAx1ZRZl1DZhRCsvrxO2vVa4NeC+6Da8Is4SnAOF3AFAVxDA+6gCS1gMIFneIU3L/VevHfvY9m64RUzZ/AH3ucPQUiPbw==
h
ij3
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMKdFrEM2FhGMA9IjrC32UvW7O0du3NCONLa21goonb+PXb+N24ehSZ+sPDj+2bYmQkSKQy67reTW1vf2NzKbxd2dvf2D4qHpaaJU814g8Uy1u2AGi6F4g0UKHk70ZxGgeStYHQ1zVv3XBsRq1scJ9yP6ECJUDCK1moNe5m4u5j0imW34s5EVsFbQLlW+nh4B4B6r/jV7ccsjbhCJqkxHc9N0M+oRsEknxS6qeEJZSM64B2Likbc+Nls3Ak5tU6fhLG2TyGZub87MhoZM44CWxlRHJrlbGr+l3VSDC/9TKgkRa7Y/KMwlQRjMt2d9IXmDOXYAmVa2FkJG1JNGdoLFewRvOWVV6F5XvHcinfjlWsuzJWHYziBM/CgCjW4hjo0gMEIHuEZXpzEeXJenbd5ac5Z9BzBHzmfPxeQkYk= AAAB7nicbZC7SwNBEMbn4ivGV4xY2SwGwSrcKaJlwMYygnlAEsLeZi9Zs7d37M4J4UhrZWMjoojY+ffY+d+4eRSa+MHCj++bYWfGj6Uw6LrfTmZpeWV1Lbue29jc2t7J7xZqJko041UWyUg3fGq4FIpXUaDkjVhzGvqS1/3B5Tiv33FtRKRucBjzdkh7SgSCUbRWvd9Jxe3pqJMvuiV3IrII3gyK5cLH/fPZ/kOlk/9qdSOWhFwhk9SYpufG2E6pRsEkH+VaieExZQPa402LiobctNPJuCNyZJ0uCSJtn0IycX93pDQ0Zhj6tjKk2Dfz2dj8L2smGFy0U6HiBLli04+CRBKMyHh30hWaM5RDC5RpYWclrE81ZWgvlLNH8OZXXoTaSclzS961Vyy7MFUWDuAQjsGDcyjDFVSgCgwG8Agv8OrEzpPz5rxPSzPOrGcP/sj5/AEqbJJX AAAB7nicbZC7SwNBEMbn4ivGV4xY2SwGwSrcKaJlwMYygnlAEsLeZi9Zs7d37M4J4UhrZWMjoojY+ffY+d+4eRSa+MHCj++bYWfGj6Uw6LrfTmZpeWV1Lbue29jc2t7J7xZqJko041UWyUg3fGq4FIpXUaDkjVhzGvqS1/3B5Tiv33FtRKRucBjzdkh7SgSCUbRWvd9Jxe3pqJMvuiV3IrII3gyK5cLH/fPZ/kOlk/9qdSOWhFwhk9SYpufG2E6pRsEkH+VaieExZQPa402LiobctNPJuCNyZJ0uCSJtn0IycX93pDQ0Zhj6tjKk2Dfz2dj8L2smGFy0U6HiBLli04+CRBKMyHh30hWaM5RDC5RpYWclrE81ZWgvlLNH8OZXXoTaSclzS961Vyy7MFUWDuAQjsGDcyjDFVSgCgwG8Agv8OrEzpPz5rxPSzPOrGcP/sj5/AEqbJJX AAAB7nicbVA9SwNBEJ3zM8avqKXNYhCswq0WWgZsLCOYD0iOsLfZS9bs7R27c0I48iNsLBSx9ffY+W/cJFdo4oOBx3szzMwLUyUt+v63t7a+sbm1Xdop7+7tHxxWjo5bNskMF02eqMR0QmaFklo0UaISndQIFodKtMPx7cxvPwljZaIfcJKKIGZDLSPJGTqpPern8vFq2q9U/Zo/B1kltCBVKNDoV756g4RnsdDIFbO2S/0Ug5wZlFyJabmXWZEyPmZD0XVUs1jYIJ+fOyXnThmQKDGuNJK5+nsiZ7G1kzh0nTHDkV32ZuJ/XjfD6CbIpU4zFJovFkWZIpiQ2e9kII3gqCaOMG6ku5XwETOMo0uo7EKgyy+vktZljfo1ek+rdb+IowSncAYXQOEa6nAHDWgChzE8wyu8ean34r17H4vWNa+YOYE/8D5/AELNj3A=
w
ij1
AAAB7nicbZC7SwNBEMbnfMb4irG0WQyCVbi10TJgYxnBPCA5wt5mkqzZ2zt295RwpLW3sVBE7fx77Pxv3DwKTfxg4cf3zbAzEyZSGOv7397K6tr6xmZuK7+9s7u3Xzgo1k2cao41HstYN0NmUAqFNSusxGaikUWhxEY4vJzkjTvURsTqxo4SDCLWV6InOLPOatx3MnFLx51CyS/7U5FloHMoVYofD+8AUO0UvtrdmKcRKsslM6ZF/cQGGdNWcInjfDs1mDA+ZH1sOVQsQhNk03HH5MQ5XdKLtXvKkqn7uyNjkTGjKHSVEbMDs5hNzP+yVmp7F0EmVJJaVHz2US+VxMZksjvpCo3cypEDxrVwsxI+YJpx6y6Ud0egiysvQ/2sTP0yvaalig8z5eAIjuEUKJxDBa6gCjXgMIRHeIYXL/GevFfvbVa64s17DuGPvM8fK5yRlg== AAAB7nicbZBLSwMxFIXv1Fetr1px5SZYBFdlIoguC25cVnDaQjuUTJppYzOZIckoZejWlRs3IoqIO3+PO/+N6WOhrQcCH+fcS+69QSK4Nq777eSWlldW1/LrhY3Nre2d4m6pruNUUebRWMSqGRDNBJfMM9wI1kwUI1EgWCMYXIzzxi1Tmsfy2gwT5kekJ3nIKTHWatx1Mn6DR51i2a24E6FFwDMoV0sf98+n+w+1TvGr3Y1pGjFpqCBat7CbGD8jynAq2KjQTjVLCB2QHmtZlCRi2s8m447QkXW6KIyVfdKgifu7IyOR1sMosJURMX09n43N/7JWasJzP+MySQ2TdPpRmApkYjTeHXW5YtSIoQVCFbezItonilBjL1SwR8DzKy9C/aSC3Qq+wuWqC1Pl4QAO4RgwnEEVLqEGHlAYwCO8wKuTOE/Om/M+Lc05s549+CPn8wc+eJJk AAAB7nicbZBLSwMxFIXv1Fetr1px5SZYBFdlIoguC25cVnDaQjuUTJppYzOZIckoZejWlRs3IoqIO3+PO/+N6WOhrQcCH+fcS+69QSK4Nq777eSWlldW1/LrhY3Nre2d4m6pruNUUebRWMSqGRDNBJfMM9wI1kwUI1EgWCMYXIzzxi1Tmsfy2gwT5kekJ3nIKTHWatx1Mn6DR51i2a24E6FFwDMoV0sf98+n+w+1TvGr3Y1pGjFpqCBat7CbGD8jynAq2KjQTjVLCB2QHmtZlCRi2s8m447QkXW6KIyVfdKgifu7IyOR1sMosJURMX09n43N/7JWasJzP+MySQ2TdPpRmApkYjTeHXW5YtSIoQVCFbezItonilBjL1SwR8DzKy9C/aSC3Qq+wuWqC1Pl4QAO4RgwnEEVLqEGHlAYwCO8wKuTOE/Om/M+Lc05s549+CPn8wc+eJJk AAAB7nicbVA9SwNBEJ3zM8avqKXNYhCswq2NlgEbywjmA5Ij7G3mkjV7e8funhKO/AgbC0Vs/T12/hs3yRWa+GDg8d4MM/PCVApjff/bW1vf2NzaLu2Ud/f2Dw4rR8ctk2SaY5MnMtGdkBmUQmHTCiuxk2pkcSixHY5vZn77EbURibq3kxSDmA2ViARn1kntp34uHui0X6n6NX8OskpoQapQoNGvfPUGCc9iVJZLZkyX+qkNcqat4BKn5V5mMGV8zIbYdVSxGE2Qz8+dknOnDEiUaFfKkrn6eyJnsTGTOHSdMbMjs+zNxP+8bmaj6yAXKs0sKr5YFGWS2ITMficDoZFbOXGEcS3crYSPmGbcuoTKLgS6/PIqaV3WqF+jd7Ra94s4SnAKZ3ABFK6gDrfQgCZwGMMzvMKbl3ov3rv3sWhd84qZE/gD7/MHVtmPfQ==
w
ij2
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMJdGi0DNpYRzAOSI+xt9pI1e3vH7pwSjrT2NhaKqJ1/j53/jZtHoYkfLPz4vhl2ZoJECoOu++3k1tY3Nrfy24Wd3b39g+JhqWniVDPeYLGMdTughkuheAMFSt5ONKdRIHkrGF1O89Yd10bE6gbHCfcjOlAiFIyitVr3vUzcVie9YtmtuDORVfAWUK6VPh7eAaDeK351+zFLI66QSWpMx3MT9DOqUTDJJ4VuanhC2YgOeMeiohE3fjYbd0JOrdMnYaztU0hm7u+OjEbGjKPAVkYUh2Y5m5r/ZZ0Uwws/EypJkSs2/yhMJcGYTHcnfaE5Qzm2QJkWdlbChlRThvZCBXsEb3nlVWhWK55b8a69cs2FufJwDCdwBh6cQw2uoA4NYDCCR3iGFydxnpxX521emnMWPUfwR87nDy0hkZc= AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIW7gGgZsLGMYB6QHGFvs0nW7O0du3NKONJa2diIKCJ2/j12/jduHoUmfrDw4/tm2JkJYikMuu63s7S8srq2ntnIbm5t7+zm9vI1EyWa8SqLZKQbATVcCsWrKFDyRqw5DQPJ68HgYpzXb7k2IlLXOIy5H9KeEl3BKFqrftdOxU1p1M4V3KI7EVkEbwaFcv7j/vn04KHSzn21OhFLQq6QSWpM03Nj9FOqUTDJR9lWYnhM2YD2eNOioiE3fjoZd0SOrdMh3Ujbp5BM3N8dKQ2NGYaBrQwp9s18Njb/y5oJds/9VKg4Qa7Y9KNuIglGZLw76QjNGcqhBcq0sLMS1qeaMrQXytojePMrL0KtVPTconflFcouTJWBQziCE/DgDMpwCRWoAoMBPMILvDqx8+S8Oe/T0iVn1rMPf+R8/gA//ZJl AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIW7gGgZsLGMYB6QHGFvs0nW7O0du3NKONJa2diIKCJ2/j12/jduHoUmfrDw4/tm2JkJYikMuu63s7S8srq2ntnIbm5t7+zm9vI1EyWa8SqLZKQbATVcCsWrKFDyRqw5DQPJ68HgYpzXb7k2IlLXOIy5H9KeEl3BKFqrftdOxU1p1M4V3KI7EVkEbwaFcv7j/vn04KHSzn21OhFLQq6QSWpM03Nj9FOqUTDJR9lWYnhM2YD2eNOioiE3fjoZd0SOrdMh3Ujbp5BM3N8dKQ2NGYaBrQwp9s18Njb/y5oJds/9VKg4Qa7Y9KNuIglGZLw76QjNGcqhBcq0sLMS1qeaMrQXytojePMrL0KtVPTconflFcouTJWBQziCE/DgDMpwCRWoAoMBPMILvDqx8+S8Oe/T0iVn1rMPf+R8/gA//ZJl AAAB7nicbVBNT8JAEJ3iF+IX6tHLRmLiibRc9EjixSMmFkigIdtlCyvbbbM71ZCGH+HFg8Z49fd489+4QA8KvmSSl/dmMjMvTKUw6LrfTmljc2t7p7xb2ds/ODyqHp+0TZJpxn2WyER3Q2q4FIr7KFDybqo5jUPJO+HkZu53Hrk2IlH3OE15ENOREpFgFK3UeRrk4qExG1Rrbt1dgKwTryA1KNAaVL/6w4RlMVfIJDWm57kpBjnVKJjks0o/MzylbEJHvGepojE3Qb44d0YurDIkUaJtKSQL9fdETmNjpnFoO2OKY7PqzcX/vF6G0XWQC5VmyBVbLooySTAh89/JUGjOUE4toUwLeythY6opQ5tQxYbgrb68TtqNuufWvTuv1nSLOMpwBudwCR5cQRNuoQU+MJjAM7zCm5M6L86787FsLTnFzCn8gfP5A1hej34=
w
ij3
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMKdFloGbCwjmAckR9jbzCVr9vaO3T0lHGntbSwUUTv/Hjv/GzePQhM/WPjxfTPszASJ4Nq47reTW1ldW9/Ibxa2tnd294r7pYaOU8WwzmIRq1ZANQousW64EdhKFNIoENgMhpeTvHmHSvNY3phRgn5E+5KHnFFjreZ9N+O3Z+NusexW3KnIMnhzKFdLHw/vAFDrFr86vZilEUrDBNW67bmJ8TOqDGcCx4VOqjGhbEj72LYoaYTaz6bjjsmxdXokjJV90pCp+7sjo5HWoyiwlRE1A72YTcz/snZqwgs/4zJJDUo2+yhMBTExmexOelwhM2JkgTLF7ayEDaiizNgLFewRvMWVl6FxWvHcinftlasuzJSHQziCE/DgHKpwBTWoA4MhPMIzvDiJ8+S8Om+z0pwz7zmAP3I+fwAuppGY AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIU7RbQM2FhGMA9IQtjb7CVr9vaO3TklHGmtbGxEFBE7/x47/xs3j0ITP1j48X0z7Mz4sRQGXffbWVhcWl5Zzaxl1zc2t7ZzO/mqiRLNeIVFMtJ1nxouheIVFCh5Pdachr7kNb9/Mcprt1wbEalrHMS8FdKuEoFgFK1Vu2un4uZk2M4V3KI7FpkHbwqFUv7j/vl076Hczn01OxFLQq6QSWpMw3NjbKVUo2CSD7PNxPCYsj7t8oZFRUNuWul43CE5tE6HBJG2TyEZu787UhoaMwh9WxlS7JnZbGT+lzUSDM5bqVBxglyxyUdBIglGZLQ76QjNGcqBBcq0sLMS1qOaMrQXytojeLMrz0P1uOi5Re/KK5RcmCgD+3AAR+DBGZTgEspQAQZ9eIQXeHVi58l5c94npQvOtGcX/sj5/AFBgpJm AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIU7RbQM2FhGMA9IQtjb7CVr9vaO3TklHGmtbGxEFBE7/x47/xs3j0ITP1j48X0z7Mz4sRQGXffbWVhcWl5Zzaxl1zc2t7ZzO/mqiRLNeIVFMtJ1nxouheIVFCh5Pdachr7kNb9/Mcprt1wbEalrHMS8FdKuEoFgFK1Vu2un4uZk2M4V3KI7FpkHbwqFUv7j/vl076Hczn01OxFLQq6QSWpMw3NjbKVUo2CSD7PNxPCYsj7t8oZFRUNuWul43CE5tE6HBJG2TyEZu787UhoaMwh9WxlS7JnZbGT+lzUSDM5bqVBxglyxyUdBIglGZLQ76QjNGcqBBcq0sLMS1qOaMrQXytojeLMrz0P1uOi5Re/KK5RcmCgD+3AAR+DBGZTgEspQAQZ9eIQXeHVi58l5c94npQvOtGcX/sj5/AFBgpJm AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0n0oMeCF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jNs1BWx8MPN6bYWZekEhh0HW/ndLa+sbmVnm7srO7t39QPTxqmzjVjLdYLGPdDajhUijeQoGSdxPNaRRI3gkmN3O/88i1EbG6x2nC/YiOlAgFo2ilztMgEw+Xs0G15tbdHGSVeAWpQYHmoPrVH8YsjbhCJqkxPc9N0M+oRsEkn1X6qeEJZRM64j1LFY248bP83Bk5s8qQhLG2pZDk6u+JjEbGTKPAdkYUx2bZm4v/eb0Uw2s/EypJkSu2WBSmkmBM5r+TodCcoZxaQpkW9lbCxlRThjahig3BW355lbQv6p5b9+68WsMt4ijDCZzCOXhwBQ24hSa0gMEEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH1njj38=
w
ijK
AAAB7nicbZC7SwNBEMbn4ivGV4ylzWIQrMKdjZYBG8EmgnlAcoS9zV6yZm/v2J1TwpHW3sZCEbXz77Hzv3HzKDTxg4Uf3zfDzkyQSGHQdb+d3Mrq2vpGfrOwtb2zu1fcLzVMnGrG6yyWsW4F1HApFK+jQMlbieY0CiRvBsOLSd6849qIWN3gKOF+RPtKhIJRtFbzvpuJ26txt1h2K+5UZBm8OZSrpY+HdwCodYtfnV7M0ogrZJIa0/bcBP2MahRM8nGhkxqeUDakfd62qGjEjZ9Nxx2TY+v0SBhr+xSSqfu7I6ORMaMosJURxYFZzCbmf1k7xfDcz4RKUuSKzT4KU0kwJpPdSU9ozlCOLFCmhZ2VsAHVlKG9UMEewVtceRkapxXPrXjXXrnqwkx5OIQjOAEPzqAKl1CDOjAYwiM8w4uTOE/Oq/M2K805854D+CPn8wdTHpGw AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIU7QbQM2Ag2EcwDkhD2NnPJmr29Y3dPCUdaKxsbEUXEzr/Hzv/GzaPQxA8WfnzfDDszfiy4Nq777SwsLi2vrGbWsusbm1vbuZ18VUeJYlhhkYhU3acaBZdYMdwIrMcKaegLrPn981Feu0WleSSvzSDGVki7kgecUWOt2l075TeXw3au4Bbdscg8eFMolPIf988new/ldu6r2YlYEqI0TFCtG54bm1ZKleFM4DDbTDTGlPVpFxsWJQ1Rt9LxuENyaJ0OCSJlnzRk7P7uSGmo9SD0bWVITU/PZiPzv6yRmOCslXIZJwYlm3wUJIKYiIx2Jx2ukBkxsECZ4nZWwnpUUWbshbL2CN7syvNQPS56btG78golFybKwD4cwBF4cAoluIAyVIBBHx7hBV6d2Hly3pz3SemCM+3ZhT9yPn8AZfqSfg== AAAB7nicbZC7SwNBEMbnfMb4ihErm8UgWIU7QbQM2Ag2EcwDkhD2NnPJmr29Y3dPCUdaKxsbEUXEzr/Hzv/GzaPQxA8WfnzfDDszfiy4Nq777SwsLi2vrGbWsusbm1vbuZ18VUeJYlhhkYhU3acaBZdYMdwIrMcKaegLrPn981Feu0WleSSvzSDGVki7kgecUWOt2l075TeXw3au4Bbdscg8eFMolPIf988new/ldu6r2YlYEqI0TFCtG54bm1ZKleFM4DDbTDTGlPVpFxsWJQ1Rt9LxuENyaJ0OCSJlnzRk7P7uSGmo9SD0bWVITU/PZiPzv6yRmOCslXIZJwYlm3wUJIKYiIx2Jx2ukBkxsECZ4nZWwnpUUWbshbL2CN7syvNQPS56btG78golFybKwD4cwBF4cAoluIAyVIBBHx7hBV6d2Hly3pz3SemCM+3ZhT9yPn8AZfqSfg== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0m86LHgRfBSwX5AG8pmu2nXbjZhd6KU0B/hxYMiXv093vw3btMctPXBwOO9GWbmBYkUBl332ymtrW9sbpW3Kzu7e/sH1cOjtolTzXiLxTLW3YAaLoXiLRQoeTfRnEaB5J1gcj33O49cGxGre5wm3I/oSIlQMIpW6jwNMvFwOxtUa27dzUFWiVeQGhRoDqpf/WHM0ogrZJIa0/PcBP2MahRM8lmlnxqeUDahI96zVNGIGz/Lz52RM6sMSRhrWwpJrv6eyGhkzDQKbGdEcWyWvbn4n9dLMbzyM6GSFLlii0VhKgnGZP47GQrNGcqpJZRpYW8lbEw1ZWgTqtgQvOWXV0n7ou65de/OqzXcIo4ynMApnIMHl9CAG2hCCxhM4Ble4c1JnBfn3flYtJacYuYY/sD5/AF+W4+X
X
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeAF48RzAOSJcxOJskkszPLTK8QlvyDFw+KePV/vPk3TpI9aGJBQ1HVTXdXlEhh0fe/vcLG5tb2TnG3tLd/cHhUPj5pWp0axhtMS23aEbVcCsUbKFDydmI4jSPJW9Hkbu63nrixQqtHnCY8jOlQiYFgFJ3UbPcyMZ71yhW/6i9A1kmQkwrkqPfKX92+ZmnMFTJJre0EfoJhRg0KJvms1E0tTyib0CHvOKpozG2YLa6dkQun9MlAG1cKyUL9PZHR2NppHLnOmOLIrnpz8T+vk+LgNsyESlLkii0XDVJJUJP566QvDGcop45QZoS7lbARNZShC6jkQghWX14nzatq4FeDh+tKzc/jKMIZnMMlBHADNbiHOjSAwRie4RXePO29eO/ex7K14OUzp/AH3ucPtpyPJg== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeAF48RzAOSJcxOJskkszPLTK8QlvyDFw+KePV/vPk3TpI9aGJBQ1HVTXdXlEhh0fe/vcLG5tb2TnG3tLd/cHhUPj5pWp0axhtMS23aEbVcCsUbKFDydmI4jSPJW9Hkbu63nrixQqtHnCY8jOlQiYFgFJ3UbPcyMZ71yhW/6i9A1kmQkwrkqPfKX92+ZmnMFTJJre0EfoJhRg0KJvms1E0tTyib0CHvOKpozG2YLa6dkQun9MlAG1cKyUL9PZHR2NppHLnOmOLIrnpz8T+vk+LgNsyESlLkii0XDVJJUJP566QvDGcop45QZoS7lbARNZShC6jkQghWX14nzatq4FeDh+tKzc/jKMIZnMMlBHADNbiHOjSAwRie4RXePO29eO/ex7K14OUzp/AH3ucPtpyPJg== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeAF48RzAOSJcxOJskkszPLTK8QlvyDFw+KePV/vPk3TpI9aGJBQ1HVTXdXlEhh0fe/vcLG5tb2TnG3tLd/cHhUPj5pWp0axhtMS23aEbVcCsUbKFDydmI4jSPJW9Hkbu63nrixQqtHnCY8jOlQiYFgFJ3UbPcyMZ71yhW/6i9A1kmQkwrkqPfKX92+ZmnMFTJJre0EfoJhRg0KJvms1E0tTyib0CHvOKpozG2YLa6dkQun9MlAG1cKyUL9PZHR2NppHLnOmOLIrnpz8T+vk+LgNsyESlLkii0XDVJJUJP566QvDGcop45QZoS7lbARNZShC6jkQghWX14nzatq4FeDh+tKzc/jKMIZnMMlBHADNbiHOjSAwRie4RXePO29eO/ex7K14OUzp/AH3ucPtpyPJg== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeAF48RzAOSJcxOJskkszPLTK8QlvyDFw+KePV/vPk3TpI9aGJBQ1HVTXdXlEhh0fe/vcLG5tb2TnG3tLd/cHhUPj5pWp0axhtMS23aEbVcCsUbKFDydmI4jSPJW9Hkbu63nrixQqtHnCY8jOlQiYFgFJ3UbPcyMZ71yhW/6i9A1kmQkwrkqPfKX92+ZmnMFTJJre0EfoJhRg0KJvms1E0tTyib0CHvOKpozG2YLa6dkQun9MlAG1cKyUL9PZHR2NppHLnOmOLIrnpz8T+vk+LgNsyESlLkii0XDVJJUJP566QvDGcop45QZoS7lbARNZShC6jkQghWX14nzatq4FeDh+tKzc/jKMIZnMMlBHADNbiHOjSAwRie4RXePO29eO/ex7K14OUzp/AH3ucPtpyPJg==
ˆ y
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix48VjBfkATyma7adduNmF3IoTQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFqRQGXffbWVvf2NzaruxUd/f2Dw5rR8cdk2Sa8TZLZKJ7ITVcCsXbKFDyXqo5jUPJu+HkduZ3n7g2IlEPmKc8iOlIiUgwilby/THFIp8OCvE4HdTqbsOdg6wSryR1KNEa1L78YcKymCtkkhrT99wUg4JqFEzyadXPDE8pm9AR71uqaMxNUMxvnpJzqwxJlGhbCslc/T1R0NiYPA5tZ0xxbJa9mfif188wugkKodIMuWKLRVEmCSZkFgAZCs0ZytwSyrSwtxI2ppoytDFVbQje8surpHPZ8NyGd39Vb7plHBU4hTO4AA+uoQl30II2MEjhGV7hzcmcF+fd+Vi0rjnlzAn8gfP5A8BakhQ= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix48VjBfkATyma7adduNmF3IoTQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFqRQGXffbWVvf2NzaruxUd/f2Dw5rR8cdk2Sa8TZLZKJ7ITVcCsXbKFDyXqo5jUPJu+HkduZ3n7g2IlEPmKc8iOlIiUgwilby/THFIp8OCvE4HdTqbsOdg6wSryR1KNEa1L78YcKymCtkkhrT99wUg4JqFEzyadXPDE8pm9AR71uqaMxNUMxvnpJzqwxJlGhbCslc/T1R0NiYPA5tZ0xxbJa9mfif188wugkKodIMuWKLRVEmCSZkFgAZCs0ZytwSyrSwtxI2ppoytDFVbQje8surpHPZ8NyGd39Vb7plHBU4hTO4AA+uoQl30II2MEjhGV7hzcmcF+fd+Vi0rjnlzAn8gfP5A8BakhQ= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix48VjBfkATyma7adduNmF3IoTQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFqRQGXffbWVvf2NzaruxUd/f2Dw5rR8cdk2Sa8TZLZKJ7ITVcCsXbKFDyXqo5jUPJu+HkduZ3n7g2IlEPmKc8iOlIiUgwilby/THFIp8OCvE4HdTqbsOdg6wSryR1KNEa1L78YcKymCtkkhrT99wUg4JqFEzyadXPDE8pm9AR71uqaMxNUMxvnpJzqwxJlGhbCslc/T1R0NiYPA5tZ0xxbJa9mfif188wugkKodIMuWKLRVEmCSZkFgAZCs0ZytwSyrSwtxI2ppoytDFVbQje8surpHPZ8NyGd39Vb7plHBU4hTO4AA+uoQl30II2MEjhGV7hzcmcF+fd+Vi0rjnlzAn8gfP5A8BakhQ= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix48VjBfkATyma7adduNmF3IoTQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFqRQGXffbWVvf2NzaruxUd/f2Dw5rR8cdk2Sa8TZLZKJ7ITVcCsXbKFDyXqo5jUPJu+HkduZ3n7g2IlEPmKc8iOlIiUgwilby/THFIp8OCvE4HdTqbsOdg6wSryR1KNEa1L78YcKymCtkkhrT99wUg4JqFEzyadXPDE8pm9AR71uqaMxNUMxvnpJzqwxJlGhbCslc/T1R0NiYPA5tZ0xxbJa9mfif188wugkKodIMuWKLRVEmCSZkFgAZCs0ZytwSyrSwtxI2ppoytDFVbQje8surpHPZ8NyGd39Vb7plHBU4hTO4AA+uoQl30II2MEjhGV7hzcmcF+fd+Vi0rjnlzAn8gfP5A8BakhQ=
attn
LSTM
pred
···
···
Figure 4.1: Diagram of LSTM with attention
As a point of comparison, we use a feedforward neural network (FFNN) trained
with word vectors using a bag-of-words approach. The ne tuned feature vectors,
v
ijk
, are averaged over each turn to derive the turn vector representation,

X
ij
=
1
K
P
k
v
ijk
, which is then fed to a single feedforward layer with a sigmoid activation
for prediction:
y
ij
=(W
y

X
ij
+b
y
):
(4.7)
4.3 Model Implementation and Training
We use the word2vec software to train the word embeddings [55]. Keras, with
Theano as the back end, is used to implement the neural networks [59, 60]. All
30
hidden layers have the same dimension as the input feature vector. For regular-
ization, 10% dropout is applied to all hidden layers. All turns are assumed to
be 70 words in length, shorter turns are padded with zeros and longer turns are
truncated (less than 5% of turns in the data required truncation). Ten percent of
the training data is randomly chosen as a validation set. Training is performed
with a maximum of 100 epochs and the training process is terminated early if the
validation loss fails to improve after two consecutive epochs and only the weights
from the epoch with the lowest validation loss are saved.
4.4 Experiments and Results
In Table 4.2, we show the F1-measure (the harmonic mean of precision and recall)
for the baseline FFNN and LSTMs with and without attention (LSTMa and LSTM,
respectively). For all behavioral codes, the attention-based LSTM gives better
prediction than the baseline bag-of-words feedforward neural network baseline.
This illustrates the importance of the relationship between words that is captured
in the hidden states of the LSTMs. Adding attention to the traditional LSTM gives
an improvement for all behavioral codes but MI adherent (counselor) and Negative
(client). The highest relative improvements when using attention are given for RES
(+2.8%) and MIN (+7.7%) for counselor categories and POS (+1.9%) for client.
This may be a result of subjects using more salient words in the expression of these
behavior categories.
In Figures ?? and ?? we show examples of counselor and client turns and how
the attention-based LSTM assigns attention over the turn. In the counselor turn
(Figure ??) there is a simple re
ection followed by and open question. The RES
attention mechanism focuses the earlier portions of the turn which corresponds to
31
Table 4.2: MISC Prediction (F1-measure).
code FFNN LSTM LSTMa
Counselor
FA 0.898 0.902 0.904
GI 0.697 0.733 0.742
QUC 0.550 0.704 0.708
QUO 0.667 0.816 0.824
REC 0.521 0.559 0.562
RES 0.443 0.533 0.548
MIA 0.533 0.598 0.592
MIN 0.261 0.259 0.279
Client
FN 0.946 0.952 0.953
POS 0.452 0.463 0.472
NEG 0.406 0.421 0.419
avg 0.580 0.631 0.637
the re
ective portion, \also mentioned that you're feeling a little bit lost", while
the QUO attention mechanism ignores the early portion of the turn and focuses
on the question at the end, \what do you mean by that". Figure ?? shows an
example of the POS and NEG attention for a client turn. This turn begins with
a negative reason, \i mean it's not ideal", followed by a positive reason, \i need
to do this". The NEG attention focuses on the rst part of the turn. The POS
attention attributes some weight to the rst part of the turn but applies much
more attention to the positive portion of the turn.
4.4.1 Analysis of Attention Weights
In Table 4.3, we show the words which received the highest average attention for
each behavioral code. Facilitate (FA) is an utterance by a counselor that is used to
acknowledge what the client is saying and to indicate for them to keep going, hence
words like `yeah', `right', and `good' are attuned to for predicting this behavior.
32
and
you
also
mentioned
that
youre
feeling
a
little
bit
lost
can
you
enlighten
me
a
little
bit
on
what
you
mean
by
lost
RES attention
QUO attention
i
mean
it's
not
ideal
but
i
need
to
do
this
so
POS attention
NEG attention
Figure 4.2: (a) Example of attention for RES and QUO codes in counselor turn
and (b) POS and NEG codes in client turn.
Giving information (GI) is an utterance where the counselor explains something
or provides feedback. Interestingly, the word `mirror' is identied as important
because during many sessions the counselors explain to their clients that there is
a one way mirror in the counseling room. Closed questions (QUC) are questions
that prompt a specic fact as a response, whereas open questions (QUO) invite
the client's perspective. The word `what' draws attention for both QUC and QUO
as it is indicative of a question. Words in closed questions which receive attention
derive from questions about specics, e.g., \have you tried..." or \has anyone told
you...". In open questions, the words that receive attention come from questions
such as \how does that t..." or \what are your thoughts about...".
Re
ections are a key aspect of motivational interviewing. They are a way for the
counselor to express understanding of the client's perspective. Simple re
ections
33
(RES) are a restatement of something the client has said, while complex re
ections
(REC) add meaning or emphasis to something the client has expressed. Words
that receive attention in simple re
ections are specic words that were used in
statements by the client such as `pot' or `husband'. In complex re
ections, words
receiving attention are more likely to be less concrete in nature such as `scary'
or `felt'. Because MI adherent (MIA) and non-adherent (MIN) categories are
groups of dierent behavior codes, they provide a challenge for attention. The MI
adherent category is one where attention did not help the LSTM in prediction.
This is essentially an `other' category where any utterance that is not specically
one of the aforementioned behavioral codes but is also not a non-adherent behavior.
The client code follow/neutral (FN) is a catch all for utterances that are neither
indicative of or counter to behavior change. Because of the nature of this code,
attention is not particularly useful, however it does try to take advantage of some
quirks in the data. For example, the word `island' appears as an important word
for this code because there are a few sessions in which long island iced teas are dis-
cussed at length. The POS and NEG categories combine behavioral codes relating
to client statements about behavior change. The attention mechanism identies
words such as `control' and `important' which indicate commitment towards con-
fronting problems as well as words such as `drive' and `black' which stem from
statements about ramications of substance use like driving under the in
uence
and blacking out. Attention also identies words that justify continued substance
use such as `helps' and `relax' and words that indicate unwillingness to admit the
existence of an issue such as `dunno' and `problem'.
34
Table 4.3: Words with the highest average attention for each behavior category.
code top words
Counselor
FA
hmm, yeah, great, wow,
cool, right, oh, good
GI
yup, include, mirror, thank, i'm,
resources, yeah, participating
QUC
tried, questions, anything, anybody
ever, what, heard, think
QUO
t, does, what, happened
thoughts, else, acted, do
REC
main, wanted, changes, scary,
necessarily, could, hand, felt
RES
pot, husband, mentioned,
month, sounds, thirty, earlier, hour
MIA
glad, cool, excellent, dicult,
packet, we'll, tough, great
MIN
help, these, give, could,
try, able, keep, kids
Client
FN
accurate, students, percent, assume,
heavy, island, that'd, interesting
POS
control, drive, reason, us,
important, black, quit, ten
NEG
relax, dunno, problem, helps
smoked, eh, prefer, mostly
4.5 Conclusions
In this chapter, we presented an approach to predicting counselor and client behav-
iors in addiction counseling using attention-based LSTMs. We demonstrated that
this approach out performs a bag-of-words approach and that attention boosts per-
formance in some situations while providing additional insight into the data which
is not aorded by non-attentional LSTMs. The attention weights help identify
35
words that are salient in the speaker turns with respect to the behaviors present
in these turns.
36
Chapter 5
Multi-label Multi-task Learning
for Behavioral Coding
5.1 Introduction
In conversation, participants' behaviors unfold and evolve over time, occurring in
both brief and extended time scales. These behaviors are often co-occurring and
intrinsically related to one another. Additionally, many of these communicative
behaviors are shared across many human interaction domains. We propose a multi-
label multi-task system that accounts for turn context to address these phenomena
while also taking advantage of data from related domains. We demonstrate this
system oers behavioral code prediction performance gains with respect to single-
label and single-task systems.
5.2 Multi-label Learning for Behavioral Coding
In this work, we assume thei
th
session,i2f1; 2;:::;Ng, to be comprised of a series
of turns, j2f1; 2;:::;M
i
g, which are comprised of word sequences, w
ij
. A turn is
dened as all words which are spoken in a speaker homogenous region, the speaker
having a particular role, r
ij
, as either the therapist or client in our expository
domain. Each word sequence is represented by a series of word embedding vectors,
w
ij
= [w
ij1
;w
ij2
;:::;w
ijK
]. The word sequences are input to the word encoder,
37
consisting of a bidirectional long short term memory (BiLSTM) network [56]. The
resulting hidden states of the BiLSTM are averaged, giving a vector representation
of the j
th
turn, X
ij
. A visualization of the word encoder is given in gure 5.1a.
w
j1
AAAB7HicbVBNTwIxEJ3FL8Qv1KOXRmLiiewaE/RG9OIRExdIYEO6pQuFbrtpuxqy4T948aDGqz/Im//GAntQ8CWTvLw3k5l5YcKZNq777RTW1jc2t4rbpZ3dvf2D8uFRU8tUEeoTyaVqh1hTzgT1DTOcthNFcRxy2grHtzO/9UiVZlI8mElCgxgPBIsYwcZKzadeNvKmvXLFrbpzoFXi5aQCORq98le3L0kaU2EIx1p3PDcxQYaVYYTTaambappgMsYD2rFU4JjqIJtfO0VnVumjSCpbwqC5+nsiw7HWkzi0nTE2Q73szcT/vE5qoqsgYyJJDRVksShKOTISzV5HfaYoMXxiCSaK2VsRGWKFibEBlWwI3vLLq8S/qF5XvfvLSv0mT6MIJ3AK5+BBDepwBw3wgcAInuEV3hzpvDjvzseiteDkM8fwB87nDwL2jvM= AAAB7HicbVBNTwIxEJ3FL8Qv1KOXRmLiiewaE/RG9OIRExdIYEO6pQuFbrtpuxqy4T948aDGqz/Im//GAntQ8CWTvLw3k5l5YcKZNq777RTW1jc2t4rbpZ3dvf2D8uFRU8tUEeoTyaVqh1hTzgT1DTOcthNFcRxy2grHtzO/9UiVZlI8mElCgxgPBIsYwcZKzadeNvKmvXLFrbpzoFXi5aQCORq98le3L0kaU2EIx1p3PDcxQYaVYYTTaambappgMsYD2rFU4JjqIJtfO0VnVumjSCpbwqC5+nsiw7HWkzi0nTE2Q73szcT/vE5qoqsgYyJJDRVksShKOTISzV5HfaYoMXxiCSaK2VsRGWKFibEBlWwI3vLLq8S/qF5XvfvLSv0mT6MIJ3AK5+BBDepwBw3wgcAInuEV3hzpvDjvzseiteDkM8fwB87nDwL2jvM= AAAB7HicbVBNTwIxEJ3FL8Qv1KOXRmLiiewaE/RG9OIRExdIYEO6pQuFbrtpuxqy4T948aDGqz/Im//GAntQ8CWTvLw3k5l5YcKZNq777RTW1jc2t4rbpZ3dvf2D8uFRU8tUEeoTyaVqh1hTzgT1DTOcthNFcRxy2grHtzO/9UiVZlI8mElCgxgPBIsYwcZKzadeNvKmvXLFrbpzoFXi5aQCORq98le3L0kaU2EIx1p3PDcxQYaVYYTTaambappgMsYD2rFU4JjqIJtfO0VnVumjSCpbwqC5+nsiw7HWkzi0nTE2Q73szcT/vE5qoqsgYyJJDRVksShKOTISzV5HfaYoMXxiCSaK2VsRGWKFibEBlWwI3vLLq8S/qF5XvfvLSv0mT6MIJ3AK5+BBDepwBw3wgcAInuEV3hzpvDjvzseiteDkM8fwB87nDwL2jvM= AAAB7HicbVBNTwIxEJ3FL8Qv1KOXRmLiiewaE/RG9OIRExdIYEO6pQuFbrtpuxqy4T948aDGqz/Im//GAntQ8CWTvLw3k5l5YcKZNq777RTW1jc2t4rbpZ3dvf2D8uFRU8tUEeoTyaVqh1hTzgT1DTOcthNFcRxy2grHtzO/9UiVZlI8mElCgxgPBIsYwcZKzadeNvKmvXLFrbpzoFXi5aQCORq98le3L0kaU2EIx1p3PDcxQYaVYYTTaambappgMsYD2rFU4JjqIJtfO0VnVumjSCpbwqC5+nsiw7HWkzi0nTE2Q73szcT/vE5qoqsgYyJJDRVksShKOTISzV5HfaYoMXxiCSaK2VsRGWKFibEBlWwI3vLLq8S/qF5XvfvLSv0mT6MIJ3AK5+BBDepwBw3wgcAInuEV3hzpvDjvzseiteDkM8fwB87nDwL2jvM=
h
j1
AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG9FLx4rmLbQhrLZbtptN7thdyOU0P/gxYOKV3+QN/+N2zQHbX0w8Hhvhpl5YcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRS8tUEeoTyaXqhFhTzgT1DTOcdhJFcRxy2g4nd3O//USVZlI8mmlCgxgPBYsYwcZKrVE/G3uzfrXm1t0caJV4BalBgWa/+tUbSJLGVBjCsdZdz01MkGFlGOF0VumlmiaYTPCQdi0VOKY6yPJrZ+jMKgMUSWVLGJSrvycyHGs9jUPbGWMz0sveXPzP66Ymug4yJpLUUEEWi6KUIyPR/HU0YIoSw6eWYKKYvRWREVaYGBtQxYbgLb+8SvyL+k3de7isNW6LNMpwAqdwDh5cQQPuoQk+EBjDM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AevvjuQ= AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG9FLx4rmLbQhrLZbtptN7thdyOU0P/gxYOKV3+QN/+N2zQHbX0w8Hhvhpl5YcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRS8tUEeoTyaXqhFhTzgT1DTOcdhJFcRxy2g4nd3O//USVZlI8mmlCgxgPBYsYwcZKrVE/G3uzfrXm1t0caJV4BalBgWa/+tUbSJLGVBjCsdZdz01MkGFlGOF0VumlmiaYTPCQdi0VOKY6yPJrZ+jMKgMUSWVLGJSrvycyHGs9jUPbGWMz0sveXPzP66Ymug4yJpLUUEEWi6KUIyPR/HU0YIoSw6eWYKKYvRWREVaYGBtQxYbgLb+8SvyL+k3de7isNW6LNMpwAqdwDh5cQQPuoQk+EBjDM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AevvjuQ= AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG9FLx4rmLbQhrLZbtptN7thdyOU0P/gxYOKV3+QN/+N2zQHbX0w8Hhvhpl5YcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRS8tUEeoTyaXqhFhTzgT1DTOcdhJFcRxy2g4nd3O//USVZlI8mmlCgxgPBYsYwcZKrVE/G3uzfrXm1t0caJV4BalBgWa/+tUbSJLGVBjCsdZdz01MkGFlGOF0VumlmiaYTPCQdi0VOKY6yPJrZ+jMKgMUSWVLGJSrvycyHGs9jUPbGWMz0sveXPzP66Ymug4yJpLUUEEWi6KUIyPR/HU0YIoSw6eWYKKYvRWREVaYGBtQxYbgLb+8SvyL+k3de7isNW6LNMpwAqdwDh5cQQPuoQk+EBjDM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AevvjuQ= AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG9FLx4rmLbQhrLZbtptN7thdyOU0P/gxYOKV3+QN/+N2zQHbX0w8Hhvhpl5YcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRS8tUEeoTyaXqhFhTzgT1DTOcdhJFcRxy2g4nd3O//USVZlI8mmlCgxgPBYsYwcZKrVE/G3uzfrXm1t0caJV4BalBgWa/+tUbSJLGVBjCsdZdz01MkGFlGOF0VumlmiaYTPCQdi0VOKY6yPJrZ+jMKgMUSWVLGJSrvycyHGs9jUPbGWMz0sveXPzP66Ymug4yJpLUUEEWi6KUIyPR/HU0YIoSw6eWYKKYvRWREVaYGBtQxYbgLb+8SvyL+k3de7isNW6LNMpwAqdwDh5cQQPuoQk+EBjDM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AevvjuQ=
BiLSTM AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw==
w
ij1
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew/etF+uuFV3DrJKvJxUIEejX/7qDWKWRlwhk9SYrucm6GdUo2CST0u91PCEsjEd8q6likbc+Nn83Ck5s8qAhLG2pZDM1d8TGY2MmUSB7YwojsyyNxP/87ophld+JlSSIldssShMJcGYzH4nA6E5QzmxhDIt7K2EjaimDG1CJRuCt/zyKmldVD236t1dVurXeRxFOIFTOAcPalCHW2hAExiM4Rle4c1JnBfn3flYtBacfOYY/sD5/AFdM4+S AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew/etF+uuFV3DrJKvJxUIEejX/7qDWKWRlwhk9SYrucm6GdUo2CST0u91PCEsjEd8q6likbc+Nn83Ck5s8qAhLG2pZDM1d8TGY2MmUSB7YwojsyyNxP/87ophld+JlSSIldssShMJcGYzH4nA6E5QzmxhDIt7K2EjaimDG1CJRuCt/zyKmldVD236t1dVurXeRxFOIFTOAcPalCHW2hAExiM4Rle4c1JnBfn3flYtBacfOYY/sD5/AFdM4+S AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew/etF+uuFV3DrJKvJxUIEejX/7qDWKWRlwhk9SYrucm6GdUo2CST0u91PCEsjEd8q6likbc+Nn83Ck5s8qAhLG2pZDM1d8TGY2MmUSB7YwojsyyNxP/87ophld+JlSSIldssShMJcGYzH4nA6E5QzmxhDIt7K2EjaimDG1CJRuCt/zyKmldVD236t1dVurXeRxFOIFTOAcPalCHW2hAExiM4Rle4c1JnBfn3flYtBacfOYY/sD5/AFdM4+S AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew/etF+uuFV3DrJKvJxUIEejX/7qDWKWRlwhk9SYrucm6GdUo2CST0u91PCEsjEd8q6likbc+Nn83Ck5s8qAhLG2pZDM1d8TGY2MmUSB7YwojsyyNxP/87ophld+JlSSIldssShMJcGYzH4nA6E5QzmxhDIt7K2EjaimDG1CJRuCt/zyKmldVD236t1dVurXeRxFOIFTOAcPalCHW2hAExiM4Rle4c1JnBfn3flYtBacfOYY/sD5/AFdM4+S
w
ij2
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew+1ab9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5uVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmf1OBkJzhnJiCWVa2FsJG1FNGdqESjYEb/nlVdKqVT236t1dVOrXeRxFOIFTOAcPLqEOt9CAJjAYwzO8wpuTOC/Ou/OxaC04+cwx/IHz+QNeuI+T AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew+1ab9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5uVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmf1OBkJzhnJiCWVa2FsJG1FNGdqESjYEb/nlVdKqVT236t1dVOrXeRxFOIFTOAcPLqEOt9CAJjAYwzO8wpuTOC/Ou/OxaC04+cwx/IHz+QNeuI+T AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew+1ab9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5uVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmf1OBkJzhnJiCWVa2FsJG1FNGdqESjYEb/nlVdKqVT236t1dVOrXeRxFOIFTOAcPLqEOt9CAJjAYwzO8wpuTOC/Ou/OxaC04+cwx/IHz+QNeuI+T AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKoMeiF48V7Ae0oWy2m3btZhN2J0oJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GNzO//ci1EbG6x0nC/YgOlQgFo2il9lM/Ew+1ab9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5uVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmf1OBkJzhnJiCWVa2FsJG1FNGdqESjYEb/nlVdKqVT236t1dVOrXeRxFOIFTOAcPLqEOt9CAJjAYwzO8wpuTOC/Ou/OxaC04+cwx/IHz+QNeuI+T
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D
h
ij1
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbbTbt2swm7E6GE/ggvHhTx6u/x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlsslrHuBtRwKRRvoUDJu4nmNAok7wST29zvPHFtRKwecJpwP6IjJULBKFqpMx5k4tGbDao1t+7OQVaJV5AaFGgOql/9YczSiCtkkhrT89wE/YxqFEzyWaWfGp5QNqEj3rNU0YgbP5ufOyNnVhmSMNa2FJK5+nsio5Ex0yiwnRHFsVn2cvE/r5dieO1nQiUpcsUWi8JUEoxJ/jsZCs0ZyqkllGlhbyVsTDVlaBOq2BC85ZdXSfui7rl17/6y1rgp4ijDCZzCOXhwBQ24gya0gMEEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH0Ydj4M= AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbbTbt2swm7E6GE/ggvHhTx6u/x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlsslrHuBtRwKRRvoUDJu4nmNAok7wST29zvPHFtRKwecJpwP6IjJULBKFqpMx5k4tGbDao1t+7OQVaJV5AaFGgOql/9YczSiCtkkhrT89wE/YxqFEzyWaWfGp5QNqEj3rNU0YgbP5ufOyNnVhmSMNa2FJK5+nsio5Ex0yiwnRHFsVn2cvE/r5dieO1nQiUpcsUWi8JUEoxJ/jsZCs0ZyqkllGlhbyVsTDVlaBOq2BC85ZdXSfui7rl17/6y1rgp4ijDCZzCOXhwBQ24gya0gMEEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH0Ydj4M= AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbbTbt2swm7E6GE/ggvHhTx6u/x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlsslrHuBtRwKRRvoUDJu4nmNAok7wST29zvPHFtRKwecJpwP6IjJULBKFqpMx5k4tGbDao1t+7OQVaJV5AaFGgOql/9YczSiCtkkhrT89wE/YxqFEzyWaWfGp5QNqEj3rNU0YgbP5ufOyNnVhmSMNa2FJK5+nsio5Ex0yiwnRHFsVn2cvE/r5dieO1nQiUpcsUWi8JUEoxJ/jsZCs0ZyqkllGlhbyVsTDVlaBOq2BC85ZdXSfui7rl17/6y1rgp4ijDCZzCOXhwBQ24gya0gMEEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH0Ydj4M= AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUDbbTbt2swm7E6GE/ggvHhTx6u/x5r9x0+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlsslrHuBtRwKRRvoUDJu4nmNAok7wST29zvPHFtRKwecJpwP6IjJULBKFqpMx5k4tGbDao1t+7OQVaJV5AaFGgOql/9YczSiCtkkhrT89wE/YxqFEzyWaWfGp5QNqEj3rNU0YgbP5ufOyNnVhmSMNa2FJK5+nsio5Ex0yiwnRHFsVn2cvE/r5dieO1nQiUpcsUWi8JUEoxJ/jsZCs0ZyqkllGlhbyVsTDVlaBOq2BC85ZdXSfui7rl17/6y1rgp4ijDCZzCOXhwBQ24gya0gMEEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH0Ydj4M=
h
ij2
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI9FLx4r2A9oQ9lsN+3azSbsToQS+iO8eFDEq7/Hm//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1gNOE+xEdKREKRtFKnfEgE4+12aBccavuAmSdeDmpQI7moPzVH8YsjbhCJqkxPc9N0M+oRsEkn5X6qeEJZRM64j1LFY248bPFuTNyYZUhCWNtSyFZqL8nMhoZM40C2xlRHJtVby7+5/VSDK/9TKgkRa7YclGYSoIxmf9OhkJzhnJqCWVa2FsJG1NNGdqESjYEb/XlddKuVT236t1fVRo3eRxFOINzuAQP6tCAO2hCCxhM4Ble4c1JnBfn3flYthacfOYU/sD5/AFHoo+E AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI9FLx4r2A9oQ9lsN+3azSbsToQS+iO8eFDEq7/Hm//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1gNOE+xEdKREKRtFKnfEgE4+12aBccavuAmSdeDmpQI7moPzVH8YsjbhCJqkxPc9N0M+oRsEkn5X6qeEJZRM64j1LFY248bPFuTNyYZUhCWNtSyFZqL8nMhoZM40C2xlRHJtVby7+5/VSDK/9TKgkRa7YclGYSoIxmf9OhkJzhnJqCWVa2FsJG1NNGdqESjYEb/XlddKuVT236t1fVRo3eRxFOINzuAQP6tCAO2hCCxhM4Ble4c1JnBfn3flYthacfOYU/sD5/AFHoo+E AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI9FLx4r2A9oQ9lsN+3azSbsToQS+iO8eFDEq7/Hm//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1gNOE+xEdKREKRtFKnfEgE4+12aBccavuAmSdeDmpQI7moPzVH8YsjbhCJqkxPc9N0M+oRsEkn5X6qeEJZRM64j1LFY248bPFuTNyYZUhCWNtSyFZqL8nMhoZM40C2xlRHJtVby7+5/VSDK/9TKgkRa7YclGYSoIxmf9OhkJzhnJqCWVa2FsJG1NNGdqESjYEb/XlddKuVT236t1fVRo3eRxFOINzuAQP6tCAO2hCCxhM4Ble4c1JnBfn3flYthacfOYU/sD5/AFHoo+E AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mKUI9FLx4r2A9oQ9lsN+3azSbsToQS+iO8eFDEq7/Hm//GbZuDtj4YeLw3w8y8IJHCoOt+O4WNza3tneJuaW//4PCofHzSNnGqGW+xWMa6G1DDpVC8hQIl7yaa0yiQvBNMbud+54lrI2L1gNOE+xEdKREKRtFKnfEgE4+12aBccavuAmSdeDmpQI7moPzVH8YsjbhCJqkxPc9N0M+oRsEkn5X6qeEJZRM64j1LFY248bPFuTNyYZUhCWNtSyFZqL8nMhoZM40C2xlRHJtVby7+5/VSDK/9TKgkRa7YclGYSoIxmf9OhkJzhnJqCWVa2FsJG1NNGdqESjYEb/XlddKuVT236t1fVRo3eRxFOINzuAQP6tCAO2hCCxhM4Ble4c1JnBfn3flYthacfOYU/sD5/AFHoo+E
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D
w
ij
AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY=
h
ijK
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF8FLBfsBbSib7aZdu9mE3YlQQn+EFw+KePX3ePPfuG1z0NYHA4/3ZpiZFyRSGHTdb6ewtr6xuVXcLu3s7u0flA+PWiZONeNNFstYdwJquBSKN1Gg5J1EcxoFkreD8c3Mbz9xbUSsHnCScD+iQyVCwShaqT3qZ+LxbtovV9yqOwdZJV5OKpCj0S9/9QYxSyOukElqTNdzE/QzqlEwyaelXmp4QtmYDnnXUkUjbvxsfu6UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMMrPxMqSZErtlgUppJgTGa/k4HQnKGcWEKZFvZWwkZUU4Y2oZINwVt+eZW0LqqeW/XuLyv16zyOIpzAKZyDBzWowy00oAkMxvAMr/DmJM6L8+58LFoLTj5zDH/gfP4AbZ+PnQ== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF8FLBfsBbSib7aZdu9mE3YlQQn+EFw+KePX3ePPfuG1z0NYHA4/3ZpiZFyRSGHTdb6ewtr6xuVXcLu3s7u0flA+PWiZONeNNFstYdwJquBSKN1Gg5J1EcxoFkreD8c3Mbz9xbUSsHnCScD+iQyVCwShaqT3qZ+LxbtovV9yqOwdZJV5OKpCj0S9/9QYxSyOukElqTNdzE/QzqlEwyaelXmp4QtmYDnnXUkUjbvxsfu6UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMMrPxMqSZErtlgUppJgTGa/k4HQnKGcWEKZFvZWwkZUU4Y2oZINwVt+eZW0LqqeW/XuLyv16zyOIpzAKZyDBzWowy00oAkMxvAMr/DmJM6L8+58LFoLTj5zDH/gfP4AbZ+PnQ== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF8FLBfsBbSib7aZdu9mE3YlQQn+EFw+KePX3ePPfuG1z0NYHA4/3ZpiZFyRSGHTdb6ewtr6xuVXcLu3s7u0flA+PWiZONeNNFstYdwJquBSKN1Gg5J1EcxoFkreD8c3Mbz9xbUSsHnCScD+iQyVCwShaqT3qZ+LxbtovV9yqOwdZJV5OKpCj0S9/9QYxSyOukElqTNdzE/QzqlEwyaelXmp4QtmYDnnXUkUjbvxsfu6UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMMrPxMqSZErtlgUppJgTGa/k4HQnKGcWEKZFvZWwkZUU4Y2oZINwVt+eZW0LqqeW/XuLyv16zyOIpzAKZyDBzWowy00oAkMxvAMr/DmJM6L8+58LFoLTj5zDH/gfP4AbZ+PnQ== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF8FLBfsBbSib7aZdu9mE3YlQQn+EFw+KePX3ePPfuG1z0NYHA4/3ZpiZFyRSGHTdb6ewtr6xuVXcLu3s7u0flA+PWiZONeNNFstYdwJquBSKN1Gg5J1EcxoFkreD8c3Mbz9xbUSsHnCScD+iQyVCwShaqT3qZ+LxbtovV9yqOwdZJV5OKpCj0S9/9QYxSyOukElqTNdzE/QzqlEwyaelXmp4QtmYDnnXUkUjbvxsfu6UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMMrPxMqSZErtlgUppJgTGa/k4HQnKGcWEKZFvZWwkZUU4Y2oZINwVt+eZW0LqqeW/XuLyv16zyOIpzAKZyDBzWowy00oAkMxvAMr/DmJM6L8+58LFoLTj5zDH/gfP4AbZ+PnQ==
w
ijK
AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi+Clgv2ANpTNdtOu3WzC7kQpoT/CiwdFvPp7vPlv3LY5aOuDgcd7M8zMCxIpDLrut1NYWV1b3yhulra2d3b3yvsHTROnmvEGi2Ws2wE1XArFGyhQ8naiOY0CyVvB6Hrqtx65NiJW9zhOuB/RgRKhYBSt1HrqZeLhdtIrV9yqOwNZJl5OKpCj3it/dfsxSyOukElqTMdzE/QzqlEwySelbmp4QtmIDnjHUkUjbvxsdu6EnFilT8JY21JIZurviYxGxoyjwHZGFIdm0ZuK/3mdFMNLPxMqSZErNl8UppJgTKa/k77QnKEcW0KZFvZWwoZUU4Y2oZINwVt8eZk0z6qeW/Xuziu1qzyOIhzBMZyCBxdQgxuoQwMYjOAZXuHNSZwX5935mLcWnHzmEP7A+fwBhLWPrA== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi+Clgv2ANpTNdtOu3WzC7kQpoT/CiwdFvPp7vPlv3LY5aOuDgcd7M8zMCxIpDLrut1NYWV1b3yhulra2d3b3yvsHTROnmvEGi2Ws2wE1XArFGyhQ8naiOY0CyVvB6Hrqtx65NiJW9zhOuB/RgRKhYBSt1HrqZeLhdtIrV9yqOwNZJl5OKpCj3it/dfsxSyOukElqTMdzE/QzqlEwySelbmp4QtmIDnjHUkUjbvxsdu6EnFilT8JY21JIZurviYxGxoyjwHZGFIdm0ZuK/3mdFMNLPxMqSZErNl8UppJgTKa/k77QnKEcW0KZFvZWwoZUU4Y2oZINwVt8eZk0z6qeW/Xuziu1qzyOIhzBMZyCBxdQgxuoQwMYjOAZXuHNSZwX5935mLcWnHzmEP7A+fwBhLWPrA== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi+Clgv2ANpTNdtOu3WzC7kQpoT/CiwdFvPp7vPlv3LY5aOuDgcd7M8zMCxIpDLrut1NYWV1b3yhulra2d3b3yvsHTROnmvEGi2Ws2wE1XArFGyhQ8naiOY0CyVvB6Hrqtx65NiJW9zhOuB/RgRKhYBSt1HrqZeLhdtIrV9yqOwNZJl5OKpCj3it/dfsxSyOukElqTMdzE/QzqlEwySelbmp4QtmIDnjHUkUjbvxsdu6EnFilT8JY21JIZurviYxGxoyjwHZGFIdm0ZuK/3mdFMNLPxMqSZErNl8UppJgTKa/k77QnKEcW0KZFvZWwoZUU4Y2oZINwVt8eZk0z6qeW/Xuziu1qzyOIhzBMZyCBxdQgxuoQwMYjOAZXuHNSZwX5935mLcWnHzmEP7A+fwBhLWPrA== AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi+Clgv2ANpTNdtOu3WzC7kQpoT/CiwdFvPp7vPlv3LY5aOuDgcd7M8zMCxIpDLrut1NYWV1b3yhulra2d3b3yvsHTROnmvEGi2Ws2wE1XArFGyhQ8naiOY0CyVvB6Hrqtx65NiJW9zhOuB/RgRKhYBSt1HrqZeLhdtIrV9yqOwNZJl5OKpCj3it/dfsxSyOukElqTMdzE/QzqlEwySelbmp4QtmIDnjHUkUjbvxsdu6EnFilT8JY21JIZurviYxGxoyjwHZGFIdm0ZuK/3mdFMNLPxMqSZErNl8UppJgTKa/k77QnKEcW0KZFvZWwoZUU4Y2oZINwVt8eZk0z6qeW/Xuziu1qzyOIhzBMZyCBxdQgxuoQwMYjOAZXuHNSZwX5935mLcWnHzmEP7A+fwBhLWPrA==
1
K
P
k
h
ijk
AAACA3icbVDLSsNAFL2pr1pfUXe6GSyCbkriRpdFQQQ3FewDmhAm00k7ZvJgZiKUEHDjr7hxoYhbf8KdOz/F6WOhrQcuHM65l3vv8VPOpLKsL6O0sLi0vFJeraytb2xumds7LZlkgtAmSXgiOj6WlLOYNhVTnHZSQXHkc9r2w4uR376nQrIkvlXDlLoR7scsYAQrLXnmnhMITHK7yK8L5Mgs8kI08HJ2FxaeWbVq1hhonthTUq0fW5ffANDwzE+nl5AsorEiHEvZta1UuTkWihFOi4qTSZpiEuI+7Woa44hKNx//UKBDrfRQkAhdsUJj9fdEjiMph5GvOyOsBnLWG4n/ed1MBWduzuI0UzQmk0VBxpFK0CgQ1GOCEsWHmmAimL4VkQHWoSgdW0WHYM++PE9aJzXbqtk3drV+DhOUYR8O4AhsOIU6XEEDmkDgAZ7gBV6NR+PZeDPeJ60lYzqzC39gfPwAcGGZhw== AAACA3icbVDLSsNAFJ34rPUVdaebwSLopmTc6LIqiOCmgn1AE8JkOmnHziRhZiKUEHDjZ7h140IRt/6EO//AtV/g9LHQ1gMXDufcy733BAlnSjvOpzUzOze/sFhYKi6vrK6t2xubdRWnktAaiXksmwFWlLOI1jTTnDYTSbEIOG0EvbOB37ilUrE4utb9hHoCdyIWMoK1kXx72w0lJhnKs8scuioVfg92/Yzd9HLfLjllZwg4TdCYlCoHzvnX98NJ1bc/3HZMUkEjTThWqoWcRHsZlpoRTvOimyqaYNLDHdoyNMKCKi8b/pDDPaO0YRhLU5GGQ/X3RIaFUn0RmE6BdVdNegPxP6+V6vDYy1iUpJpGZLQoTDnUMRwEAttMUqJ53xBMJDO3QtLFJhRtYiuaENDky9OkflhGThldoVLlFIxQADtgF+wDBI5ABVyAKqgBAu7AI3gGL9a99WS9Wm+j1hlrPLMF/sB6/wHn0Jtf AAACA3icbVDLSsNAFJ34rPUVdaebwSLopmTc6LIqiOCmgn1AE8JkOmnHziRhZiKUEHDjZ7h140IRt/6EO//AtV/g9LHQ1gMXDufcy733BAlnSjvOpzUzOze/sFhYKi6vrK6t2xubdRWnktAaiXksmwFWlLOI1jTTnDYTSbEIOG0EvbOB37ilUrE4utb9hHoCdyIWMoK1kXx72w0lJhnKs8scuioVfg92/Yzd9HLfLjllZwg4TdCYlCoHzvnX98NJ1bc/3HZMUkEjTThWqoWcRHsZlpoRTvOimyqaYNLDHdoyNMKCKi8b/pDDPaO0YRhLU5GGQ/X3RIaFUn0RmE6BdVdNegPxP6+V6vDYy1iUpJpGZLQoTDnUMRwEAttMUqJ53xBMJDO3QtLFJhRtYiuaENDky9OkflhGThldoVLlFIxQADtgF+wDBI5ABVyAKqgBAu7AI3gGL9a99WS9Wm+j1hlrPLMF/sB6/wHn0Jtf AAACA3icbVDLSgMxFM34rPU16k43wSK4KhM3uiy6EdxUsA/oDEMmzbSxSWZIMkIZBtz4K25cKOLWn3Dn35i2s9DWAxcO59zLvfdEKWfaeN63s7S8srq2Xtmobm5t7+y6e/ttnWSK0BZJeKK6EdaUM0lbhhlOu6miWEScdqLR1cTvPFClWSLvzDilgcADyWJGsLFS6B76scIkR0V+U0BfZyIcwWGYs/tREbo1r+5NARcJKkkNlGiG7pffT0gmqDSEY617yEtNkGNlGOG0qPqZpikmIzygPUslFlQH+fSHAp5YpQ/jRNmSBk7V3xM5FlqPRWQ7BTZDPe9NxP+8XmbiiyBnMs0MlWS2KM44NAmcBAL7TFFi+NgSTBSzt0IyxDYUY2Or2hDQ/MuLpH1WR14d3aJa47KMowKOwDE4BQicgwa4Bk3QAgQ8gmfwCt6cJ+fFeXc+Zq1LTjlzAP7A+fwBFXqXyA==
H
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeglxwjmAckS5idzCaTzGOZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEs6M9f1vr7CxubW9U9wt7e0fHB6Vj09aRqWa0CZRXOlOhA3lTNKmZZbTTqIpFhGn7WhyP/fbT1QbpuSjnSY0FHgoWcwItk5q1fsZG8/65Ypf9RdA6yTISQVyNPrlr95AkVRQaQnHxnQDP7FhhrVlhNNZqZcammAywUPadVRiQU2YLa6doQunDFCstCtp0UL9PZFhYcxURK5TYDsyq95c/M/rpja+DTMmk9RSSZaL4pQjq9D8dTRgmhLLp45gopm7FZER1phYF1DJhRCsvrxOWlfVwK8GD9eV2l0eRxHO4BwuIYAbqEEdGtAEAmN4hld485T34r17H8vWgpfPnMIfeJ8/o3aPKA== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeglxwjmAckS5idzCaTzGOZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEs6M9f1vr7CxubW9U9wt7e0fHB6Vj09aRqWa0CZRXOlOhA3lTNKmZZbTTqIpFhGn7WhyP/fbT1QbpuSjnSY0FHgoWcwItk5q1fsZG8/65Ypf9RdA6yTISQVyNPrlr95AkVRQaQnHxnQDP7FhhrVlhNNZqZcammAywUPadVRiQU2YLa6doQunDFCstCtp0UL9PZFhYcxURK5TYDsyq95c/M/rpja+DTMmk9RSSZaL4pQjq9D8dTRgmhLLp45gopm7FZER1phYF1DJhRCsvrxOWlfVwK8GD9eV2l0eRxHO4BwuIYAbqEEdGtAEAmN4hld485T34r17H8vWgpfPnMIfeJ8/o3aPKA== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeglxwjmAckS5idzCaTzGOZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEs6M9f1vr7CxubW9U9wt7e0fHB6Vj09aRqWa0CZRXOlOhA3lTNKmZZbTTqIpFhGn7WhyP/fbT1QbpuSjnSY0FHgoWcwItk5q1fsZG8/65Ypf9RdA6yTISQVyNPrlr95AkVRQaQnHxnQDP7FhhrVlhNNZqZcammAywUPadVRiQU2YLa6doQunDFCstCtp0UL9PZFhYcxURK5TYDsyq95c/M/rpja+DTMmk9RSSZaL4pQjq9D8dTRgmhLLp45gopm7FZER1phYF1DJhRCsvrxOWlfVwK8GD9eV2l0eRxHO4BwuIYAbqEEdGtAEAmN4hld485T34r17H8vWgpfPnMIfeJ8/o3aPKA== AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMeglxwjmAckS5idzCaTzGOZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEs6M9f1vr7CxubW9U9wt7e0fHB6Vj09aRqWa0CZRXOlOhA3lTNKmZZbTTqIpFhGn7WhyP/fbT1QbpuSjnSY0FHgoWcwItk5q1fsZG8/65Ypf9RdA6yTISQVyNPrlr95AkVRQaQnHxnQDP7FhhrVlhNNZqZcammAywUPadVRiQU2YLa6doQunDFCstCtp0UL9PZFhYcxURK5TYDsyq95c/M/rpja+DTMmk9RSSZaL4pQjq9D8dTRgmhLLp45gopm7FZER1phYF1DJhRCsvrxOWlfVwK8GD9eV2l0eRxHO4BwuIYAbqEEdGtAEAmN4hld485T34r17H8vWgpfPnMIfeJ8/o3aPKA==
(a) Word encoder
BiLSTM AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw== AAAB7XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SmDLExkIhYr4gOcLeZi9Zs7d77O4J4ch/sLFQxNb/Y+e/cZNcoYkPBh7vzTAzL4g508Z1v53cxubW9k5+t7C3f3B4VDw+aWuZKEJbRHKpugHWlDNBW4YZTruxojgKOO0Ek+u533miSjMpmmYaUz/CI8FCRrCxUrvObh+ad4NiyS27C6B14mWkBBkag+JXfyhJElFhCMda9zw3Nn6KlWGE01mhn2gaYzLBI9qzVOCIaj9dXDtDF1YZolAqW8Kghfp7IsWR1tMosJ0RNmO96s3F/7xeYsKqnzIRJ4YKslwUJhwZieavoyFTlBg+tQQTxeytiIyxwsTYgAo2BG/15XXSrpQ9t+zdV0q1ehZHHs7gHC7BgyuowQ00oAUEHuEZXuHNkc6L8+58LFtzTjZzCn/gfP4A1WaOnw==
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D
X
C
ij
AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KokIeiz24rGCbQNtLZvtpl272cTdiVBC/4QXD4p49e9489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ2V1bX1js7BV3N7Z3dsvHRw2TZxqxhsslrH2A2q4FIo3UKDkfqI5jQLJW8GoNvVbT1wbEas7HCe8G9GBEqFgFK3k+71MPEzua71S2a24M5Bl4uWkDDnqvdJXpx+zNOIKmaTGtD03wW5GNQom+aTYSQ1PKBvRAW9bqmjETTeb3Tshp1bpkzDWthSSmfp7IqORMeMosJ0RxaFZ9Kbif147xfCqmwmVpMgVmy8KU0kwJtPnSV9ozlCOLaFMC3srYUOqKUMbUdGG4C2+vEya5xXPrXi3F+XqdR5HAY7hBM7Ag0uowg3UoQEMJDzDK7w5j86L8+58zFtXnHzmCP7A+fwB/1mP7Q== AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KokIeiz24rGCbQNtLZvtpl272cTdiVBC/4QXD4p49e9489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ2V1bX1js7BV3N7Z3dsvHRw2TZxqxhsslrH2A2q4FIo3UKDkfqI5jQLJW8GoNvVbT1wbEas7HCe8G9GBEqFgFK3k+71MPEzua71S2a24M5Bl4uWkDDnqvdJXpx+zNOIKmaTGtD03wW5GNQom+aTYSQ1PKBvRAW9bqmjETTeb3Tshp1bpkzDWthSSmfp7IqORMeMosJ0RxaFZ9Kbif147xfCqmwmVpMgVmy8KU0kwJtPnSV9ozlCOLaFMC3srYUOqKUMbUdGG4C2+vEya5xXPrXi3F+XqdR5HAY7hBM7Ag0uowg3UoQEMJDzDK7w5j86L8+58zFtXnHzmCP7A+fwB/1mP7Q== AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KokIeiz24rGCbQNtLZvtpl272cTdiVBC/4QXD4p49e9489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ2V1bX1js7BV3N7Z3dsvHRw2TZxqxhsslrH2A2q4FIo3UKDkfqI5jQLJW8GoNvVbT1wbEas7HCe8G9GBEqFgFK3k+71MPEzua71S2a24M5Bl4uWkDDnqvdJXpx+zNOIKmaTGtD03wW5GNQom+aTYSQ1PKBvRAW9bqmjETTeb3Tshp1bpkzDWthSSmfp7IqORMeMosJ0RxaFZ9Kbif147xfCqmwmVpMgVmy8KU0kwJtPnSV9ozlCOLaFMC3srYUOqKUMbUdGG4C2+vEya5xXPrXi3F+XqdR5HAY7hBM7Ag0uowg3UoQEMJDzDK7w5j86L8+58zFtXnHzmCP7A+fwB/1mP7Q== AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KokIeiz24rGCbQNtLZvtpl272cTdiVBC/4QXD4p49e9489+4bXPQ1gcDj/dmmJkXJFIYdN1vZ2V1bX1js7BV3N7Z3dsvHRw2TZxqxhsslrH2A2q4FIo3UKDkfqI5jQLJW8GoNvVbT1wbEas7HCe8G9GBEqFgFK3k+71MPEzua71S2a24M5Bl4uWkDDnqvdJXpx+zNOIKmaTGtD03wW5GNQom+aTYSQ1PKBvRAW9bqmjETTeb3Tshp1bpkzDWthSSmfp7IqORMeMosJ0RxaFZ9Kbif147xfCqmwmVpMgVmy8KU0kwJtPnSV9ozlCOLaFMC3srYUOqKUMbUdGG4C2+vEya5xXPrXi3F+XqdR5HAY7hBM7Ag0uowg3UoQEMJDzDK7w5j86L8+58zFtXnHzmCP7A+fwB/1mP7Q==
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D x
i(j C)
AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahHiy7Iuix2IvHCvYD26Vk02wbm02WJCuWpf/CiwdFvPpvvPlvTNs9aOuDgcd7M8zMC2LOtHHdbye3srq2vpHfLGxt7+zuFfcPmlomitAGkVyqdoA15UzQhmGG03asKI4CTlvBqDb1W49UaSbFnRnH1I/wQLCQEWysdP/US1n54ax2OukVS27FnQEtEy8jJchQ7xW/un1JkogKQzjWuuO5sfFTrAwjnE4K3UTTGJMRHtCOpQJHVPvp7OIJOrFKH4VS2RIGzdTfEymOtB5Hge2MsBnqRW8q/ud1EhNe+SkTcWKoIPNFYcKRkWj6PuozRYnhY0swUczeisgQK0yMDalgQ/AWX14mzfOK51a824tS9TqLIw9HcAxl8OASqnADdWgAAQHP8ApvjnZenHfnY96ac7KZQ/gD5/MHr7SQQQ== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahHiy7Iuix2IvHCvYD26Vk02wbm02WJCuWpf/CiwdFvPpvvPlvTNs9aOuDgcd7M8zMC2LOtHHdbye3srq2vpHfLGxt7+zuFfcPmlomitAGkVyqdoA15UzQhmGG03asKI4CTlvBqDb1W49UaSbFnRnH1I/wQLCQEWysdP/US1n54ax2OukVS27FnQEtEy8jJchQ7xW/un1JkogKQzjWuuO5sfFTrAwjnE4K3UTTGJMRHtCOpQJHVPvp7OIJOrFKH4VS2RIGzdTfEymOtB5Hge2MsBnqRW8q/ud1EhNe+SkTcWKoIPNFYcKRkWj6PuozRYnhY0swUczeisgQK0yMDalgQ/AWX14mzfOK51a824tS9TqLIw9HcAxl8OASqnADdWgAAQHP8ApvjnZenHfnY96ac7KZQ/gD5/MHr7SQQQ== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahHiy7Iuix2IvHCvYD26Vk02wbm02WJCuWpf/CiwdFvPpvvPlvTNs9aOuDgcd7M8zMC2LOtHHdbye3srq2vpHfLGxt7+zuFfcPmlomitAGkVyqdoA15UzQhmGG03asKI4CTlvBqDb1W49UaSbFnRnH1I/wQLCQEWysdP/US1n54ax2OukVS27FnQEtEy8jJchQ7xW/un1JkogKQzjWuuO5sfFTrAwjnE4K3UTTGJMRHtCOpQJHVPvp7OIJOrFKH4VS2RIGzdTfEymOtB5Hge2MsBnqRW8q/ud1EhNe+SkTcWKoIPNFYcKRkWj6PuozRYnhY0swUczeisgQK0yMDalgQ/AWX14mzfOK51a824tS9TqLIw9HcAxl8OASqnADdWgAAQHP8ApvjnZenHfnY96ac7KZQ/gD5/MHr7SQQQ== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahHiy7Iuix2IvHCvYD26Vk02wbm02WJCuWpf/CiwdFvPpvvPlvTNs9aOuDgcd7M8zMC2LOtHHdbye3srq2vpHfLGxt7+zuFfcPmlomitAGkVyqdoA15UzQhmGG03asKI4CTlvBqDb1W49UaSbFnRnH1I/wQLCQEWysdP/US1n54ax2OukVS27FnQEtEy8jJchQ7xW/un1JkogKQzjWuuO5sfFTrAwjnE4K3UTTGJMRHtCOpQJHVPvp7OIJOrFKH4VS2RIGzdTfEymOtB5Hge2MsBnqRW8q/ud1EhNe+SkTcWKoIPNFYcKRkWj6PuozRYnhY0swUczeisgQK0yMDalgQ/AWX14mzfOK51a824tS9TqLIw9HcAxl8OASqnADdWgAAQHP8ApvjnZenHfnY96ac7KZQ/gD5/MHr7SQQQ==
x
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpNJ5rHMzIphyT948aCIV//Hm3/jJNmDJhY0FFXddHdFCWfG+v63V1hb39jcKm6Xdnb39g/Kh0dNo1JNaIMornQ7woZyJmnDMstpO9EUi4jTVjS+nfmtR6oNU/LBThIaCjyQLGYEWyc1n3oZG0175Ypf9edAqyTISQVy1Hvlr25fkVRQaQnHxnQCP7FhhrVlhNNpqZsammAyxgPacVRiQU2Yza+dojOn9FGstCtp0Vz9PZFhYcxERK5TYDs0y95M/M/rpDa+DjMmk9RSSRaL4pQjq9DsddRnmhLLJ45gopm7FZEh1phYF1DJhRAsv7xKmhfVwK8G95eV2k0eRxFO4BTOIYArqMEd1KEBBEbwDK/w5invxXv3PhatBS+fOYY/8D5/AO0mj1g= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpNJ5rHMzIphyT948aCIV//Hm3/jJNmDJhY0FFXddHdFCWfG+v63V1hb39jcKm6Xdnb39g/Kh0dNo1JNaIMornQ7woZyJmnDMstpO9EUi4jTVjS+nfmtR6oNU/LBThIaCjyQLGYEWyc1n3oZG0175Ypf9edAqyTISQVy1Hvlr25fkVRQaQnHxnQCP7FhhrVlhNNpqZsammAyxgPacVRiQU2Yza+dojOn9FGstCtp0Vz9PZFhYcxERK5TYDs0y95M/M/rpDa+DjMmk9RSSRaL4pQjq9DsddRnmhLLJ45gopm7FZEh1phYF1DJhRAsv7xKmhfVwK8G95eV2k0eRxFO4BTOIYArqMEd1KEBBEbwDK/w5invxXv3PhatBS+fOYY/8D5/AO0mj1g= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpNJ5rHMzIphyT948aCIV//Hm3/jJNmDJhY0FFXddHdFCWfG+v63V1hb39jcKm6Xdnb39g/Kh0dNo1JNaIMornQ7woZyJmnDMstpO9EUi4jTVjS+nfmtR6oNU/LBThIaCjyQLGYEWyc1n3oZG0175Ypf9edAqyTISQVy1Hvlr25fkVRQaQnHxnQCP7FhhrVlhNNpqZsammAyxgPacVRiQU2Yza+dojOn9FGstCtp0Vz9PZFhYcxERK5TYDs0y95M/M/rpDa+DjMmk9RSSRaL4pQjq9DsddRnmhLLJ45gopm7FZEh1phYF1DJhRAsv7xKmhfVwK8G95eV2k0eRxFO4BTOIYArqMEd1KEBBEbwDK/w5invxXv3PhatBS+fOYY/8D5/AO0mj1g= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpNJ5rHMzIphyT948aCIV//Hm3/jJNmDJhY0FFXddHdFCWfG+v63V1hb39jcKm6Xdnb39g/Kh0dNo1JNaIMornQ7woZyJmnDMstpO9EUi4jTVjS+nfmtR6oNU/LBThIaCjyQLGYEWyc1n3oZG0175Ypf9edAqyTISQVy1Hvlr25fkVRQaQnHxnQCP7FhhrVlhNNpqZsammAyxgPacVRiQU2Yza+dojOn9FGstCtp0Vz9PZFhYcxERK5TYDs0y95M/M/rpDa+DjMmk9RSSRaL4pQjq9DsddRnmhLLJ45gopm7FZEh1phYF1DJhRAsv7xKmhfVwK8G95eV2k0eRxFO4BTOIYArqMEd1KEBBEbwDK/w5invxXv3PhatBS+fOYY/8D5/AO0mj1g=
x
i(j+C)
AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahIpRdEfRY7MVjBfuB7VKyabaNzSZLkhXL0n/hxYMiXv033vw3pu0etPXBwOO9GWbmBTFn2rjut5NbWV1b38hvFra2d3b3ivsHTS0TRWiDSC5VO8CaciZowzDDaTtWFEcBp61gVJv6rUeqNJPizoxj6kd4IFjICDZWun/qpaz8cFY7nfSKJbfizoCWiZeREmSo94pf3b4kSUSFIRxr3fHc2PgpVoYRTieFbqJpjMkID2jHUoEjqv10dvEEnVilj0KpbAmDZurviRRHWo+jwHZG2Az1ojcV//M6iQmv/JSJODFUkPmiMOHISDR9H/WZosTwsSWYKGZvRWSIFSbGhlSwIXiLLy+T5nnFcyve7UWpep3FkYcjOIYyeHAJVbiBOjSAgIBneIU3RzsvzrvzMW/NOdnMIfyB8/kDrKaQPw== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahIpRdEfRY7MVjBfuB7VKyabaNzSZLkhXL0n/hxYMiXv033vw3pu0etPXBwOO9GWbmBTFn2rjut5NbWV1b38hvFra2d3b3ivsHTS0TRWiDSC5VO8CaciZowzDDaTtWFEcBp61gVJv6rUeqNJPizoxj6kd4IFjICDZWun/qpaz8cFY7nfSKJbfizoCWiZeREmSo94pf3b4kSUSFIRxr3fHc2PgpVoYRTieFbqJpjMkID2jHUoEjqv10dvEEnVilj0KpbAmDZurviRRHWo+jwHZG2Az1ojcV//M6iQmv/JSJODFUkPmiMOHISDR9H/WZosTwsSWYKGZvRWSIFSbGhlSwIXiLLy+T5nnFcyve7UWpep3FkYcjOIYyeHAJVbiBOjSAgIBneIU3RzsvzrvzMW/NOdnMIfyB8/kDrKaQPw== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahIpRdEfRY7MVjBfuB7VKyabaNzSZLkhXL0n/hxYMiXv033vw3pu0etPXBwOO9GWbmBTFn2rjut5NbWV1b38hvFra2d3b3ivsHTS0TRWiDSC5VO8CaciZowzDDaTtWFEcBp61gVJv6rUeqNJPizoxj6kd4IFjICDZWun/qpaz8cFY7nfSKJbfizoCWiZeREmSo94pf3b4kSUSFIRxr3fHc2PgpVoYRTieFbqJpjMkID2jHUoEjqv10dvEEnVilj0KpbAmDZurviRRHWo+jwHZG2Az1ojcV//M6iQmv/JSJODFUkPmiMOHISDR9H/WZosTwsSWYKGZvRWSIFSbGhlSwIXiLLy+T5nnFcyve7UWpep3FkYcjOIYyeHAJVbiBOjSAgIBneIU3RzsvzrvzMW/NOdnMIfyB8/kDrKaQPw== AAAB8XicbVBNSwMxEJ2tX7V+VT16CRahIpRdEfRY7MVjBfuB7VKyabaNzSZLkhXL0n/hxYMiXv033vw3pu0etPXBwOO9GWbmBTFn2rjut5NbWV1b38hvFra2d3b3ivsHTS0TRWiDSC5VO8CaciZowzDDaTtWFEcBp61gVJv6rUeqNJPizoxj6kd4IFjICDZWun/qpaz8cFY7nfSKJbfizoCWiZeREmSo94pf3b4kSUSFIRxr3fHc2PgpVoYRTieFbqJpjMkID2jHUoEjqv10dvEEnVilj0KpbAmDZurviRRHWo+jwHZG2Az1ojcV//M6iQmv/JSJODFUkPmiMOHISDR9H/WZosTwsSWYKGZvRWSIFSbGhlSwIXiLLy+T5nnFcyve7UWpep3FkYcjOIYyeHAJVbiBOjSAgIBneIU3RzsvzrvzMW/NOdnMIfyB8/kDrKaQPw==
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D
···
AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D AAAB7HicbVBNS8NAEJ34WetX1aOXYBE8lUQE9Vb04rGCaQttKJvNpl272Q27E6GU/gcvHlS8+oO8+W/ctjlo64OBx3szzMyLMsENet63s7K6tr6xWdoqb+/s7u1XDg6bRuWasoAqoXQ7IoYJLlmAHAVrZ5qRNBKsFQ1vp37riWnDlXzAUcbClPQlTzglaKVml8YKTa9S9WreDO4y8QtShQKNXuWrGyuap0wiFcSYju9lGI6JRk4Fm5S7uWEZoUPSZx1LJUmZCcezayfuqVViN1HalkR3pv6eGJPUmFEa2c6U4MAselPxP6+TY3IVjrnMcmSSzhcluXBRudPX3ZhrRlGMLCFUc3urSwdEE4o2oLINwV98eZkE57Xrmn9/Ua3fFGmU4BhO4Ax8uIQ63EEDAqDwCM/wCm+Ocl6cd+dj3rriFDNH8AfO5w8at48D g
i(j C)
AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQD4ZdEfQYzMVjBPPAZAmzk04yOju7zMwKYclfePGgiFf/xpt/4yTZgyYWNBRV3XR3BbHg2rjut5NbWV1b38hvFra2d3b3ivsHTR0limGDRSJS7YBqFFxiw3AjsB0rpGEgsBU81qZ+6wmV5pG8M+MY/ZAOJR9wRo2V7oe9lJcfzmqnk16x5FbcGcgy8TJSggz1XvGr249YEqI0TFCtO54bGz+lynAmcFLoJhpjyh7pEDuWShqi9tPZxRNyYpU+GUTKljRkpv6eSGmo9TgMbGdIzUgvelPxP6+TmMGVn3IZJwYlmy8aJIKYiEzfJ32ukBkxtoQyxe2thI2ooszYkAo2BG/x5WXSPK94bsW7vShVr7M48nAEx1AGDy6hCjdQhwYwkPAMr/DmaOfFeXc+5q05J5s5hD9wPn8AlVeQMA== AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQD4ZdEfQYzMVjBPPAZAmzk04yOju7zMwKYclfePGgiFf/xpt/4yTZgyYWNBRV3XR3BbHg2rjut5NbWV1b38hvFra2d3b3ivsHTR0limGDRSJS7YBqFFxiw3AjsB0rpGEgsBU81qZ+6wmV5pG8M+MY/ZAOJR9wRo2V7oe9lJcfzmqnk16x5FbcGcgy8TJSggz1XvGr249YEqI0TFCtO54bGz+lynAmcFLoJhpjyh7pEDuWShqi9tPZxRNyYpU+GUTKljRkpv6eSGmo9TgMbGdIzUgvelPxP6+TmMGVn3IZJwYlmy8aJIKYiEzfJ32ukBkxtoQyxe2thI2ooszYkAo2BG/x5WXSPK94bsW7vShVr7M48nAEx1AGDy6hCjdQhwYwkPAMr/DmaOfFeXc+5q05J5s5hD9wPn8AlVeQMA== AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQD4ZdEfQYzMVjBPPAZAmzk04yOju7zMwKYclfePGgiFf/xpt/4yTZgyYWNBRV3XR3BbHg2rjut5NbWV1b38hvFra2d3b3ivsHTR0limGDRSJS7YBqFFxiw3AjsB0rpGEgsBU81qZ+6wmV5pG8M+MY/ZAOJR9wRo2V7oe9lJcfzmqnk16x5FbcGcgy8TJSggz1XvGr249YEqI0TFCtO54bGz+lynAmcFLoJhpjyh7pEDuWShqi9tPZxRNyYpU+GUTKljRkpv6eSGmo9TgMbGdIzUgvelPxP6+TmMGVn3IZJwYlmy8aJIKYiEzfJ32ukBkxtoQyxe2thI2ooszYkAo2BG/x5WXSPK94bsW7vShVr7M48nAEx1AGDy6hCjdQhwYwkPAMr/DmaOfFeXc+5q05J5s5hD9wPn8AlVeQMA== AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQD4ZdEfQYzMVjBPPAZAmzk04yOju7zMwKYclfePGgiFf/xpt/4yTZgyYWNBRV3XR3BbHg2rjut5NbWV1b38hvFra2d3b3ivsHTR0limGDRSJS7YBqFFxiw3AjsB0rpGEgsBU81qZ+6wmV5pG8M+MY/ZAOJR9wRo2V7oe9lJcfzmqnk16x5FbcGcgy8TJSggz1XvGr249YEqI0TFCtO54bGz+lynAmcFLoJhpjyh7pEDuWShqi9tPZxRNyYpU+GUTKljRkpv6eSGmo9TgMbGdIzUgvelPxP6+TmMGVn3IZJwYlmy8aJIKYiEzfJ32ukBkxtoQyxe2thI2ooszYkAo2BG/x5WXSPK94bsW7vShVr7M48nAEx1AGDy6hCjdQhwYwkPAMr/DmaOfFeXc+5q05J5s5hD9wPn8AlVeQMA==
g
ij
AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJJJlkdmaZ6RXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEiks+v63t7a+sbm1Xdgp7u7tHxyWjo4bVqeG8TrTUptWRC2XQvE6CpS8lRhO40jyZjS+m/nNJ26s0OoRJwkPYzpQoi8YRSc1Bt1MjKbdUtmv+HOQVRLkpAw5at3SV6enWRpzhUxSa9uBn2CYUYOCST4tdlLLE8rGdMDbjioacxtm82un5NwpPdLXxpVCMld/T2Q0tnYSR64zpji0y95M/M9rp9i/CTOhkhS5YotF/VQS1GT2OukJwxnKiSOUGeFuJWxIDWXoAiq6EILll1dJ47IS+JXg4apcvc3jKMApnMEFBHANVbiHGtSBwQie4RXePO29eO/ex6J1zctnTuAPvM8f0w2PRw== AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJJJlkdmaZ6RXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEiks+v63t7a+sbm1Xdgp7u7tHxyWjo4bVqeG8TrTUptWRC2XQvE6CpS8lRhO40jyZjS+m/nNJ26s0OoRJwkPYzpQoi8YRSc1Bt1MjKbdUtmv+HOQVRLkpAw5at3SV6enWRpzhUxSa9uBn2CYUYOCST4tdlLLE8rGdMDbjioacxtm82un5NwpPdLXxpVCMld/T2Q0tnYSR64zpji0y95M/M9rp9i/CTOhkhS5YotF/VQS1GT2OukJwxnKiSOUGeFuJWxIDWXoAiq6EILll1dJ47IS+JXg4apcvc3jKMApnMEFBHANVbiHGtSBwQie4RXePO29eO/ex6J1zctnTuAPvM8f0w2PRw== AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJJJlkdmaZ6RXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEiks+v63t7a+sbm1Xdgp7u7tHxyWjo4bVqeG8TrTUptWRC2XQvE6CpS8lRhO40jyZjS+m/nNJ26s0OoRJwkPYzpQoi8YRSc1Bt1MjKbdUtmv+HOQVRLkpAw5at3SV6enWRpzhUxSa9uBn2CYUYOCST4tdlLLE8rGdMDbjioacxtm82un5NwpPdLXxpVCMld/T2Q0tnYSR64zpji0y95M/M9rp9i/CTOhkhS5YotF/VQS1GT2OukJwxnKiSOUGeFuJWxIDWXoAiq6EILll1dJ47IS+JXg4apcvc3jKMApnMEFBHANVbiHGtSBwQie4RXePO29eO/ex6J1zctnTuAPvM8f0w2PRw== AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0R9Bj04jGCeUCyhNnJJJlkdmaZ6RXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEiks+v63t7a+sbm1Xdgp7u7tHxyWjo4bVqeG8TrTUptWRC2XQvE6CpS8lRhO40jyZjS+m/nNJ26s0OoRJwkPYzpQoi8YRSc1Bt1MjKbdUtmv+HOQVRLkpAw5at3SV6enWRpzhUxSa9uBn2CYUYOCST4tdlLLE8rGdMDbjioacxtm82un5NwpPdLXxpVCMld/T2Q0tnYSR64zpji0y95M/M9rp9i/CTOhkhS5YotF/VQS1GT2OukJwxnKiSOUGeFuJWxIDWXoAiq6EILll1dJ47IS+JXg4apcvc3jKMApnMEFBHANVbiHGtSBwQie4RXePO29eO/ex6J1zctnTuAPvM8f0w2PRw==
g
i(j+C)
AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQEcKuCHoM5uIxgnlgsoTZSScZnZ1dZmaFsOQvvHhQxKt/482/cZLsQRMLGoqqbrq7glhwbVz328mtrK6tb+Q3C1vbO7t7xf2Dpo4SxbDBIhGpdkA1Ci6xYbgR2I4V0jAQ2Aoea1O/9YRK80jemXGMfkiHkg84o8ZK98NeyssPZ7XTSa9YcivuDGSZeBkpQYZ6r/jV7UcsCVEaJqjWHc+NjZ9SZTgTOCl0E40xZY90iB1LJQ1R++ns4gk5sUqfDCJlSxoyU39PpDTUehwGtjOkZqQXvan4n9dJzODKT7mME4OSzRcNEkFMRKbvkz5XyIwYW0KZ4vZWwkZUUWZsSAUbgrf48jJpnlc8t+LdXpSq11kceTiCYyiDB5dQhRuoQwMYSHiGV3hztPPivDsf89ack80cwh84nz+SSZAu AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQEcKuCHoM5uIxgnlgsoTZSScZnZ1dZmaFsOQvvHhQxKt/482/cZLsQRMLGoqqbrq7glhwbVz328mtrK6tb+Q3C1vbO7t7xf2Dpo4SxbDBIhGpdkA1Ci6xYbgR2I4V0jAQ2Aoea1O/9YRK80jemXGMfkiHkg84o8ZK98NeyssPZ7XTSa9YcivuDGSZeBkpQYZ6r/jV7UcsCVEaJqjWHc+NjZ9SZTgTOCl0E40xZY90iB1LJQ1R++ns4gk5sUqfDCJlSxoyU39PpDTUehwGtjOkZqQXvan4n9dJzODKT7mME4OSzRcNEkFMRKbvkz5XyIwYW0KZ4vZWwkZUUWZsSAUbgrf48jJpnlc8t+LdXpSq11kceTiCYyiDB5dQhRuoQwMYSHiGV3hztPPivDsf89ack80cwh84nz+SSZAu AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQEcKuCHoM5uIxgnlgsoTZSScZnZ1dZmaFsOQvvHhQxKt/482/cZLsQRMLGoqqbrq7glhwbVz328mtrK6tb+Q3C1vbO7t7xf2Dpo4SxbDBIhGpdkA1Ci6xYbgR2I4V0jAQ2Aoea1O/9YRK80jemXGMfkiHkg84o8ZK98NeyssPZ7XTSa9YcivuDGSZeBkpQYZ6r/jV7UcsCVEaJqjWHc+NjZ9SZTgTOCl0E40xZY90iB1LJQ1R++ns4gk5sUqfDCJlSxoyU39PpDTUehwGtjOkZqQXvan4n9dJzODKT7mME4OSzRcNEkFMRKbvkz5XyIwYW0KZ4vZWwkZUUWZsSAUbgrf48jJpnlc8t+LdXpSq11kceTiCYyiDB5dQhRuoQwMYSHiGV3hztPPivDsf89ack80cwh84nz+SSZAu AAAB8XicbVDLSgNBEOyNrxhfUY9eBoMQEcKuCHoM5uIxgnlgsoTZSScZnZ1dZmaFsOQvvHhQxKt/482/cZLsQRMLGoqqbrq7glhwbVz328mtrK6tb+Q3C1vbO7t7xf2Dpo4SxbDBIhGpdkA1Ci6xYbgR2I4V0jAQ2Aoea1O/9YRK80jemXGMfkiHkg84o8ZK98NeyssPZ7XTSa9YcivuDGSZeBkpQYZ6r/jV7UcsCVEaJqjWHc+NjZ9SZTgTOCl0E40xZY90iB1LJQ1R++ns4gk5sUqfDCJlSxoyU39PpDTUehwGtjOkZqQXvan4n9dJzODKT7mME4OSzRcNEkFMRKbvkz5XyIwYW0KZ4vZWwkZUUWZsSAUbgrf48jJpnlc8t+LdXpSq11kceTiCYyiDB5dQhRuoQwMYSHiGV3hztPPivDsf89ack80cwh84nz+SSZAu
1
2C+1
P
c
g
j+c
AAACBXicbVDLSsNAFL2pr1pfUZciDBZBKNSkG3VX7MaNUMHYQlPiZDppx04ezEyEErJy46+4caHi1n9w5984fSy0euDC4Zx7ufceP+FMKsv6MgoLi0vLK8XV0tr6xuaWub1zI+NUEOqQmMei7WNJOYuoo5jitJ0IikOf05Y/bIz91j0VksXRtRoltBvifsQCRrDSkmfuu4HAJLPzrNao2DlyZRp6BPW97K5Ccs8sW1VrAvSX2DNSrh9bl7cA0PTMT7cXkzSkkSIcS9mxrUR1MywUI5zmJTeVNMFkiPu0o2mEQyq72eSNHB1qpYeCWOiKFJqoPycyHEo5Cn3dGWI1kPPeWPzP66QqOO1mLEpSRSMyXRSkHKkYjTNBPSYoUXykCSaC6VsRGWCdi9LJlXQI9vzLf4lTq55V7Su7XD+HKYqwBwdwBDacQB0uoAkOEHiAJ3iBV+PReDbejPdpa8GYzezCLxgf3w+pmT0= AAACBXicbVC7SgNBFJ2Nrxhfq5YSGAyCEIg7adQumMZGiOCaQHZZZiezyZjZBzOzQli2svFXbCxUbC3t7fwbJ49CEw9cOJxzL/fe4yecSWVZ30ZhaXllda24XtrY3NreMXf3bmWcCkJtEvNYdHwsKWcRtRVTnHYSQXHoc9r2h82x376nQrI4ulGjhLoh7kcsYAQrLXlm2QkEJhnKs3qzinLoyDT0COx72V2V5J5ZsWrWBHCRoBmpNE6sK7s8+Gx55pfTi0ka0kgRjqXsIitRboaFYoTTvOSkkiaYDHGfdjWNcEilm03eyOGRVnowiIWuSMGJ+nsiw6GUo9DXnSFWAznvjcX/vG6qgjM3Y1GSKhqR6aIg5VDFcJwJ7DFBieIjTTARTN8KyQDrXJROrqRDQPMvLxK7XjuvoWtUaVyAKYrgAByCY4DAKWiAS9ACNiDgATyBF/BqPBrPxpvxPm0tGLOZffAHxscPuo2afg== AAACBXicbVC7SgNBFJ2Nrxhfq5YSGAyCEIg7adQumMZGiOCaQHZZZiezyZjZBzOzQli2svFXbCxUbC3t7fwbJ49CEw9cOJxzL/fe4yecSWVZ30ZhaXllda24XtrY3NreMXf3bmWcCkJtEvNYdHwsKWcRtRVTnHYSQXHoc9r2h82x376nQrI4ulGjhLoh7kcsYAQrLXlm2QkEJhnKs3qzinLoyDT0COx72V2V5J5ZsWrWBHCRoBmpNE6sK7s8+Gx55pfTi0ka0kgRjqXsIitRboaFYoTTvOSkkiaYDHGfdjWNcEilm03eyOGRVnowiIWuSMGJ+nsiw6GUo9DXnSFWAznvjcX/vG6qgjM3Y1GSKhqR6aIg5VDFcJwJ7DFBieIjTTARTN8KyQDrXJROrqRDQPMvLxK7XjuvoWtUaVyAKYrgAByCY4DAKWiAS9ACNiDgATyBF/BqPBrPxpvxPm0tGLOZffAHxscPuo2afg== AAACBXicbVBNS8NAEN3Ur1q/oh5FWCyCUChJL+qt2IvHCsYW2hA22027dncTdjdCCTl58a948aDi1f/gzX/jts1BWx8MPN6bYWZemDCqtON8W6WV1bX1jfJmZWt7Z3fP3j+4U3EqMfFwzGLZDZEijAriaaoZ6SaSIB4y0gnHranfeSBS0Vjc6klCfI6GgkYUI22kwD7uRxLhzM2zRqvm5rCvUh5gOAyy+xrOA7vq1J0Z4DJxC1IFBdqB/dUfxDjlRGjMkFI910m0nyGpKWYkr/RTRRKEx2hIeoYKxInys9kbOTw1ygBGsTQlNJypvycyxJWa8NB0cqRHatGbiv95vVRHF35GRZJqIvB8UZQyqGM4zQQOqCRYs4khCEtqboV4hEwu2iRXMSG4iy8vE69Rv6y7N261eVWkUQZH4AScARecgya4Bm3gAQwewTN4BW/Wk/VivVsf89aSVcwcgj+wPn8AUH2X8Q==
G
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc=
(b) Context encoder
ˆ
Y
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG
G
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc=
(W
Y
G
ij
+b
Y
)
AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvYh7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkD/piXlg== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvYh7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkD/piXlg== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvYh7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkD/piXlg== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvYh7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkD/piXlg==
(c) Turn predictor
G
ij
AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc= AAAB7XicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegBz1GMA9IljA7mU0mmccyMyuEJf/gxYMiXv0fb/6Nk2QPmljQUFR1090VJZwZ6/vfXmFtfWNzq7hd2tnd2z8oHx41jUo1oQ2iuNLtCBvKmaQNyyyn7URTLCJOW9H4dua3nqg2TMlHO0loKPBAspgRbJ3UvOtlbDTtlSt+1Z8DrZIgJxXIUe+Vv7p9RVJBpSUcG9MJ/MSGGdaWEU6npW5qaILJGA9ox1GJBTVhNr92is6c0kex0q6kRXP190SGhTETEblOge3QLHsz8T+vk9r4OsyYTFJLJVksilOOrEKz11GfaUosnziCiWbuVkSGWGNiXUAlF0Kw/PIqaV5UA78aPFxWajd5HEU4gVM4hwCuoAb3UIcGEBjBM7zCm6e8F+/d+1i0Frx85hj+wPv8AaHtjyc=
(W
Z
G
ij
+b
Z
)
AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvaB7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkDAcKXmA== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvaB7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkDAcKXmA== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvaB7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkDAcKXmA== AAACBXicbVDLSgMxFM3UV62vUZe6CBahIpQZEXRZdKHLCvaB7TBk0rSNTTJDkhHKMBs3/oobF4q49R/c+Tdm2llo64ELJ+fcS+49QcSo0o7zbRUWFpeWV4qrpbX1jc0te3unqcJYYtLAIQtlO0CKMCpIQ1PNSDuSBPGAkVYwusz81gORiobiVo8j4nE0ELRPMdJG8u39rqIDjiotP7lL4ZWf0PsUHsMgex75dtmpOhPAeeLmpAxy1H37q9sLccyJ0JghpTquE2kvQVJTzEha6saKRAiP0IB0DBWIE+UlkytSeGiUHuyH0pTQcKL+nkgQV2rMA9PJkR6qWS8T//M6se6fewkVUayJwNOP+jGDOoRZJLBHJcGajQ1BWFKzK8RDJBHWJriSCcGdPXmeNE+qrlN1b07LtYs8jiLYAwegAlxwBmrgGtRBA2DwCJ7BK3iznqwX6936mLYWrHxmF/yB9fkDAcKXmA==
ˆ
Z
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH
(d) Session predictor
Figure 5.1: Diagram of encoding and decoding networks
Behavioral observation and coding can be applied at a variety of temporal
granularities, including at the utterance, turn, and session levels. The behavioral
codes, or labels, annotated in a particular segment are often co-occurring and
related. Thus, viewing these individual labels as a set of relevant labels, allows for
casting the behavioral coding problem as a multi-label learning problem. In this
multi-label learning scenario, a sample, e.g., w
ij
, has an associate set of labels,Y
ij
,
where Y
ij
(l) = 1 if the l
th
label is true for that sample and Y
ij
(l) = 0, otherwise.
38
For session level labels, the posterior of estimated labels from the predictor is
averaged across turns in each session:
^
Z
i
=
1
M
i
M
i
X
j=1
^
Z
ij
; (5.1)
for the session level prediction.
5.2.1 Multi-label Learning with Deep Neural Networks
Deep neural networks provide a
exible architecture for multi-input and multi-
output learning paradigms. A multi-output network can be interpreted as a multi-
label network when the network weights are fully shared by the multi-label outputs.
For the multi-label loss we use binary cross entropy loss, summed across the
multi-label outputs. This loss does not explicitly take into account correlations
between the labels but rather relies on the shared network weights to encode this
information. The multi-label binary cross entropy loss is given by:
E
bce
=
N
X
i=1
M
i
X
j=1
L
X
l=1
Y
ij
(l) log

^
Y
ij
(l)

+ (1Y
ij
(l)) log

1
^
Y
ij
(l)

:
(5.2)
5.2.2 Multi-label Sample Weights
Class imbalance is a common problem in machine learning that can drastically
impact model training and generalization. A common approach is to weigh the
loss function so losses incurred by samples of less frequent classes are weighted
more heavily, so as to increase the impact of those samples in the model. In multi-
label problems class imbalance is no longer clearly dened as labels are no longer
individual but parts of a multi-label set. One option would be to weigh the loss
39
function according to the power set of the multi-label set, 2
Y
. However, because the
power set grows exponentially with the number of labels the label co-occurrence
distribution becomes sparse which can lead to overtting. In [66], the authors
propose an algorithm that seeks to jointly learn binary class learning for each
label and multi-class learners for rst order pairs of labels to create a predictive
multi-label model. While, this approach has shown promise for addressing class
imbalance in multi-label datasets, it does require learning additional parameters
which is undesirable in deep learning settings where the number of parameters to
be learned is typically already high. Instead, we propose a heuristic approach that
weighs the loss function according to the average frequency of the labels appearing
in a given label set. Each label contributes a weight according to the inverse
frequency of that individual label, i.e.,

ij
(l) =
8
>
>
<
>
>
:
P
ij
1Y
ij
(l)
P
ij
Y
ij
(l)
; ifY
ij
(l) = 1:
1; ifY
ij
(l) = 0:
(5.3)
The mean of these weights is taken as the multi-label sample weight:

ij
=
1
L
L
X
l=1

ij
(l): (5.4)
Thus when a given loss is computed for that sample it is weighted according to:
E =
N
X
i=1
M
i
X
j=1

ij
Loss(Y
ij
;
^
Y
ij
): (5.5)
40
5.3 Multi-task Learning for Behavioral Coding
We show an overview of the proposed single-task and multi-task learning systems
in gure 5.2. Our multi-task system follows an adversarial approach proposed
by [34]. This system consists of encoders for each individual task as well as a shared
encoder. The output of these encoders is then concatenated and fed to a predictor
for each task as well as a shared predictor that attempts to discriminate between the
tasks. The gradient from the task discriminator is reversed to the shared encoder
in order to make the shared encoder task invariant. Additionally, orthogonality
constraints are placed on the encoder outputs between the task specic and shared
encoders in order to ensure that they are not encoding redundant representations.
The total multi-task loss is computed as:
E
total
=
X
m
E
m
+E
task
+
E
di
; (5.6)
whereE
m
is the loss of them
th
task,E
task
is the loss of the task discriminator,E
di
is the loss of the orthogonality constraint, and and
are hyper-parameters for
weighting the respective losses. The task discriminator attempts to predict which
domain a particular sample belongs to. It is a single feed-forward layer with a
sigmoid activation, i.e.,
^
T
ij
=(W
T
G
shared
ij
+b
T
): (5.7)
The task discriminator loss is the binary cross entropy between the reference task
label, T
ij
, and the prediction from the task discriminator
^
T
ij
, i.e.,
E
task
=
N
X
i=1
M
i
X
j=1
L
X
l=1
T
ij
(l) log

^
T
ij
(l)

+ (1T
ij
(l)) log

1
^
T
ij
(l)

:
(5.8)
41
The gradient from the task discriminator is reversed to the shared encoder making
it increasingly dicult to predict the task as training proceeds thus resulting in a
task invariant representation in the shared encoder. The orthogonality constraint
loss is given by:
E
di
=
X
m
jj(G
shared
)
T
G
m
jj
2
; (5.9)
whereG
shared
is the output of the shared encoder,G
m
is the result of them
th
task
encoder, andjjjj
2
is the squared Frobenius norm.
Predictor AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKezmosegF48RzEOSJczOziZD5rHMzAphyVd48aCIVz/Hm3/jJNmDJhY0FFXddHdFKWfG+v63V9rY3NreKe9W9vYPDo+qxycdozJNaJsornQvwoZyJmnbMstpL9UUi4jTbjS5nfvdJ6oNU/LBTlMaCjySLGEEWyc9tjSNGbFKD6s1v+4vgNZJUJAaFGgNq1+DWJFMUGkJx8b0Az+1YY61ZYTTWWWQGZpiMsEj2ndUYkFNmC8OnqELp8QoUdqVtGih/p7IsTBmKiLXKbAdm1VvLv7n9TObXIc5k2lmqSTLRUnGkVVo/j2KmabE8qkjmGjmbkVkjDUm1mVUcSEEqy+vk06jHvj14L5Ra94UcZThDM7hEgK4gibcQQvaQEDAM7zCm6e9F+/d+1i2lrxi5hT+wPv8AfYvkH4= AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKezmosegF48RzEOSJczOziZD5rHMzAphyVd48aCIVz/Hm3/jJNmDJhY0FFXddHdFKWfG+v63V9rY3NreKe9W9vYPDo+qxycdozJNaJsornQvwoZyJmnbMstpL9UUi4jTbjS5nfvdJ6oNU/LBTlMaCjySLGEEWyc9tjSNGbFKD6s1v+4vgNZJUJAaFGgNq1+DWJFMUGkJx8b0Az+1YY61ZYTTWWWQGZpiMsEj2ndUYkFNmC8OnqELp8QoUdqVtGih/p7IsTBmKiLXKbAdm1VvLv7n9TObXIc5k2lmqSTLRUnGkVVo/j2KmabE8qkjmGjmbkVkjDUm1mVUcSEEqy+vk06jHvj14L5Ra94UcZThDM7hEgK4gibcQQvaQEDAM7zCm6e9F+/d+1i2lrxi5hT+wPv8AfYvkH4= AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKezmosegF48RzEOSJczOziZD5rHMzAphyVd48aCIVz/Hm3/jJNmDJhY0FFXddHdFKWfG+v63V9rY3NreKe9W9vYPDo+qxycdozJNaJsornQvwoZyJmnbMstpL9UUi4jTbjS5nfvdJ6oNU/LBTlMaCjySLGEEWyc9tjSNGbFKD6s1v+4vgNZJUJAaFGgNq1+DWJFMUGkJx8b0Az+1YY61ZYTTWWWQGZpiMsEj2ndUYkFNmC8OnqELp8QoUdqVtGih/p7IsTBmKiLXKbAdm1VvLv7n9TObXIc5k2lmqSTLRUnGkVVo/j2KmabE8qkjmGjmbkVkjDUm1mVUcSEEqy+vk06jHvj14L5Ra94UcZThDM7hEgK4gibcQQvaQEDAM7zCm6e9F+/d+1i2lrxi5hT+wPv8AfYvkH4= AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKezmosegF48RzEOSJczOziZD5rHMzAphyVd48aCIVz/Hm3/jJNmDJhY0FFXddHdFKWfG+v63V9rY3NreKe9W9vYPDo+qxycdozJNaJsornQvwoZyJmnbMstpL9UUi4jTbjS5nfvdJ6oNU/LBTlMaCjySLGEEWyc9tjSNGbFKD6s1v+4vgNZJUJAaFGgNq1+DWJFMUGkJx8b0Az+1YY61ZYTTWWWQGZpiMsEj2ndUYkFNmC8OnqELp8QoUdqVtGih/p7IsTBmKiLXKbAdm1VvLv7n9TObXIc5k2lmqSTLRUnGkVVo/j2KmabE8qkjmGjmbkVkjDUm1mVUcSEEqy+vk06jHvj14L5Ra94UcZThDM7hEgK4gibcQQvaQEDAM7zCm6e9F+/d+1i2lrxi5hT+wPv8AfYvkH4=
ˆ
Y
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG
Encoder AAAB7nicbVBNS8NAEJ3Ur1q/qh69BIvgqSS96LEogscK9gPaUDabSbt0sxt2N0IJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZemHKmjed9O6WNza3tnfJuZW//4PCoenzS0TJTFNtUcql6IdHImcC2YYZjL1VIkpBjN5zczv3uEyrNpHg00xSDhIwEixklxkrdO0FlhGpYrXl1bwF3nfgFqUGB1rD6NYgkzRIUhnKidd/3UhPkRBlGOc4qg0xjSuiEjLBvqSAJ6iBfnDtzL6wSubFUtoRxF+rviZwkWk+T0HYmxIz1qjcX//P6mYmvg5yJNDMo6HJRnHHXSHf+uxsxhdTwqSWEKmZvdemYKEKNTahiQ/BXX14nnUbd9+r+Q6PWvCniKMMZnMMl+HAFTbiHFrSBwgSe4RXenNR5cd6dj2VrySlmTuEPnM8fQESPfg== AAAB7nicbVBNS8NAEJ3Ur1q/qh69BIvgqSS96LEogscK9gPaUDabSbt0sxt2N0IJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZemHKmjed9O6WNza3tnfJuZW//4PCoenzS0TJTFNtUcql6IdHImcC2YYZjL1VIkpBjN5zczv3uEyrNpHg00xSDhIwEixklxkrdO0FlhGpYrXl1bwF3nfgFqUGB1rD6NYgkzRIUhnKidd/3UhPkRBlGOc4qg0xjSuiEjLBvqSAJ6iBfnDtzL6wSubFUtoRxF+rviZwkWk+T0HYmxIz1qjcX//P6mYmvg5yJNDMo6HJRnHHXSHf+uxsxhdTwqSWEKmZvdemYKEKNTahiQ/BXX14nnUbd9+r+Q6PWvCniKMMZnMMl+HAFTbiHFrSBwgSe4RXenNR5cd6dj2VrySlmTuEPnM8fQESPfg== AAAB7nicbVBNS8NAEJ3Ur1q/qh69BIvgqSS96LEogscK9gPaUDabSbt0sxt2N0IJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZemHKmjed9O6WNza3tnfJuZW//4PCoenzS0TJTFNtUcql6IdHImcC2YYZjL1VIkpBjN5zczv3uEyrNpHg00xSDhIwEixklxkrdO0FlhGpYrXl1bwF3nfgFqUGB1rD6NYgkzRIUhnKidd/3UhPkRBlGOc4qg0xjSuiEjLBvqSAJ6iBfnDtzL6wSubFUtoRxF+rviZwkWk+T0HYmxIz1qjcX//P6mYmvg5yJNDMo6HJRnHHXSHf+uxsxhdTwqSWEKmZvdemYKEKNTahiQ/BXX14nnUbd9+r+Q6PWvCniKMMZnMMl+HAFTbiHFrSBwgSe4RXenNR5cd6dj2VrySlmTuEPnM8fQESPfg== AAAB7nicbVBNS8NAEJ3Ur1q/qh69BIvgqSS96LEogscK9gPaUDabSbt0sxt2N0IJ/RFePCji1d/jzX/jts1BWx8MPN6bYWZemHKmjed9O6WNza3tnfJuZW//4PCoenzS0TJTFNtUcql6IdHImcC2YYZjL1VIkpBjN5zczv3uEyrNpHg00xSDhIwEixklxkrdO0FlhGpYrXl1bwF3nfgFqUGB1rD6NYgkzRIUhnKidd/3UhPkRBlGOc4qg0xjSuiEjLBvqSAJ6iBfnDtzL6wSubFUtoRxF+rviZwkWk+T0HYmxIz1qjcX//P6mYmvg5yJNDMo6HJRnHHXSHf+uxsxhdTwqSWEKmZvdemYKEKNTahiQ/BXX14nnUbd9+r+Q6PWvCniKMMZnMMl+HAFTbiHFrSBwgSe4RXenNR5cd6dj2VrySlmTuEPnM8fQESPfg==
w
ij
AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY=
(a) Single Task
Task Predictor AAAB9XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SaBm0sYyQL0jOsLe3SZbs7R27c0o48j9sLBSx9b/Y+W/cJFdo4oOBx3szzMwLEikMuu63U9jY3NreKe6W9vYPDo/KxydtE6ea8RaLZay7ATVcCsVbKFDybqI5jQLJO8Hkdu53Hrk2IlZNnCbcj+hIiaFgFK300KRmQhqah4JhrAflilt1FyDrxMtJBXI0BuWvfhizNOIKmaTG9Dw3QT+jGgWTfFbqp4YnlE3oiPcsVTTixs8WV8/IhVVCMoy1LYVkof6eyGhkzDQKbGdEcWxWvbn4n9dLcXjtZ0IlKXLFlouGqSQYk3kEJBSaM5RTSyjTwt5K2JhqytAGVbIheKsvr5N2req5Ve++Vqnf5HEU4QzO4RI8uII63EEDWsBAwzO8wpvz5Lw4787HsrXg5DOn8AfO5w9VnpJj AAAB9XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SaBm0sYyQL0jOsLe3SZbs7R27c0o48j9sLBSx9b/Y+W/cJFdo4oOBx3szzMwLEikMuu63U9jY3NreKe6W9vYPDo/KxydtE6ea8RaLZay7ATVcCsVbKFDybqI5jQLJO8Hkdu53Hrk2IlZNnCbcj+hIiaFgFK300KRmQhqah4JhrAflilt1FyDrxMtJBXI0BuWvfhizNOIKmaTG9Dw3QT+jGgWTfFbqp4YnlE3oiPcsVTTixs8WV8/IhVVCMoy1LYVkof6eyGhkzDQKbGdEcWxWvbn4n9dLcXjtZ0IlKXLFlouGqSQYk3kEJBSaM5RTSyjTwt5K2JhqytAGVbIheKsvr5N2req5Ve++Vqnf5HEU4QzO4RI8uII63EEDWsBAwzO8wpvz5Lw4787HsrXg5DOn8AfO5w9VnpJj AAAB9XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SaBm0sYyQL0jOsLe3SZbs7R27c0o48j9sLBSx9b/Y+W/cJFdo4oOBx3szzMwLEikMuu63U9jY3NreKe6W9vYPDo/KxydtE6ea8RaLZay7ATVcCsVbKFDybqI5jQLJO8Hkdu53Hrk2IlZNnCbcj+hIiaFgFK300KRmQhqah4JhrAflilt1FyDrxMtJBXI0BuWvfhizNOIKmaTG9Dw3QT+jGgWTfFbqp4YnlE3oiPcsVTTixs8WV8/IhVVCMoy1LYVkof6eyGhkzDQKbGdEcWxWvbn4n9dLcXjtZ0IlKXLFlouGqSQYk3kEJBSaM5RTSyjTwt5K2JhqytAGVbIheKsvr5N2req5Ve++Vqnf5HEU4QzO4RI8uII63EEDWsBAwzO8wpvz5Lw4787HsrXg5DOn8AfO5w9VnpJj AAAB9XicbVA9SwNBEJ2LXzF+RS1tFoNgFe7SaBm0sYyQL0jOsLe3SZbs7R27c0o48j9sLBSx9b/Y+W/cJFdo4oOBx3szzMwLEikMuu63U9jY3NreKe6W9vYPDo/KxydtE6ea8RaLZay7ATVcCsVbKFDybqI5jQLJO8Hkdu53Hrk2IlZNnCbcj+hIiaFgFK300KRmQhqah4JhrAflilt1FyDrxMtJBXI0BuWvfhizNOIKmaTG9Dw3QT+jGgWTfFbqp4YnlE3oiPcsVTTixs8WV8/IhVVCMoy1LYVkof6eyGhkzDQKbGdEcWxWvbn4n9dLcXjtZ0IlKXLFlouGqSQYk3kEJBSaM5RTSyjTwt5K2JhqytAGVbIheKsvr5N2req5Ve++Vqnf5HEU4QzO4RI8uII63EEDWsBAwzO8wpvz5Lw4787HsrXg5DOn8AfO5w9VnpJj MISC Predictor AAAB9XicbVA9SwNBEN3zM8avqKXNYhCswl0aLYNptBAimg9IzrC3N5cs2ds9dveUcOR/2FgoYut/sfPfuEmu0MQHA4/3ZpiZFyScaeO6387K6tr6xmZhq7i9s7u3Xzo4bGmZKgpNKrlUnYBo4ExA0zDDoZMoIHHAoR2M6lO//QhKMynuzTgBPyYDwSJGibHSw831XR03FISMGqn6pbJbcWfAy8TLSRnlaPRLX71Q0jQGYSgnWnc9NzF+RpRhlMOk2Es1JISOyAC6lgoSg/az2dUTfGqVEEdS2RIGz9TfExmJtR7Hge2MiRnqRW8q/ud1UxNd+BkTSWpA0PmiKOXYSDyNAIdMATV8bAmhitlbMR0SRaixQRVtCN7iy8ukVa14bsW7rZZrl3kcBXSMTtAZ8tA5qqEr1EBNRJFCz+gVvTlPzovz7nzMW1ecfOYI/YHz+QO1CJH8 AAAB9XicbVA9SwNBEN3zM8avqKXNYhCswl0aLYNptBAimg9IzrC3N5cs2ds9dveUcOR/2FgoYut/sfPfuEmu0MQHA4/3ZpiZFyScaeO6387K6tr6xmZhq7i9s7u3Xzo4bGmZKgpNKrlUnYBo4ExA0zDDoZMoIHHAoR2M6lO//QhKMynuzTgBPyYDwSJGibHSw831XR03FISMGqn6pbJbcWfAy8TLSRnlaPRLX71Q0jQGYSgnWnc9NzF+RpRhlMOk2Es1JISOyAC6lgoSg/az2dUTfGqVEEdS2RIGz9TfExmJtR7Hge2MiRnqRW8q/ud1UxNd+BkTSWpA0PmiKOXYSDyNAIdMATV8bAmhitlbMR0SRaixQRVtCN7iy8ukVa14bsW7rZZrl3kcBXSMTtAZ8tA5qqEr1EBNRJFCz+gVvTlPzovz7nzMW1ecfOYI/YHz+QO1CJH8 AAAB9XicbVA9SwNBEN3zM8avqKXNYhCswl0aLYNptBAimg9IzrC3N5cs2ds9dveUcOR/2FgoYut/sfPfuEmu0MQHA4/3ZpiZFyScaeO6387K6tr6xmZhq7i9s7u3Xzo4bGmZKgpNKrlUnYBo4ExA0zDDoZMoIHHAoR2M6lO//QhKMynuzTgBPyYDwSJGibHSw831XR03FISMGqn6pbJbcWfAy8TLSRnlaPRLX71Q0jQGYSgnWnc9NzF+RpRhlMOk2Es1JISOyAC6lgoSg/az2dUTfGqVEEdS2RIGz9TfExmJtR7Hge2MiRnqRW8q/ud1UxNd+BkTSWpA0PmiKOXYSDyNAIdMATV8bAmhitlbMR0SRaixQRVtCN7iy8ukVa14bsW7rZZrl3kcBXSMTtAZ8tA5qqEr1EBNRJFCz+gVvTlPzovz7nzMW1ecfOYI/YHz+QO1CJH8 AAAB9XicbVA9SwNBEN3zM8avqKXNYhCswl0aLYNptBAimg9IzrC3N5cs2ds9dveUcOR/2FgoYut/sfPfuEmu0MQHA4/3ZpiZFyScaeO6387K6tr6xmZhq7i9s7u3Xzo4bGmZKgpNKrlUnYBo4ExA0zDDoZMoIHHAoR2M6lO//QhKMynuzTgBPyYDwSJGibHSw831XR03FISMGqn6pbJbcWfAy8TLSRnlaPRLX71Q0jQGYSgnWnc9NzF+RpRhlMOk2Es1JISOyAC6lgoSg/az2dUTfGqVEEdS2RIGz9TfExmJtR7Hge2MiRnqRW8q/ud1UxNd+BkTSWpA0PmiKOXYSDyNAIdMATV8bAmhitlbMR0SRaixQRVtCN7iy8ukVa14bsW7rZZrl3kcBXSMTtAZ8tA5qqEr1EBNRJFCz+gVvTlPzovz7nzMW1ecfOYI/YHz+QO1CJH8
MISC Encoder AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiEfQgVLQf0C4lm51tQ7PJkmSFsvRvePGgiFf/jDf/jWm7B219MPB4b4aZeUHCmTau++0U1tY3NreK26Wd3b39g/LhUVvLVFFoUcml6gZEA2cCWoYZDt1EAYkDDp1g3Jj5nSdQmknxaCYJ+DEZChYxSoyV+ne3Dw18LagMQQ3KFbfqzoFXiZeTCsrRHJS/+qGkaQzCUE607nluYvyMKMMoh2mpn2pICB2TIfQsFSQG7Wfzm6f4zCohjqSyJQyeq78nMhJrPYkD2xkTM9LL3kz8z+ulJrr0MyaS1ICgi0VRyrGReBYADpkCavjEEkIVs7diOiKKUGNjKtkQvOWXV0m7VvXcqndfq9Sv8jiK6ASdonPkoQtURzeoiVqIogQ9o1f05qTOi/PufCxaC04+c4z+wPn8AfxikPw= AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiEfQgVLQf0C4lm51tQ7PJkmSFsvRvePGgiFf/jDf/jWm7B219MPB4b4aZeUHCmTau++0U1tY3NreK26Wd3b39g/LhUVvLVFFoUcml6gZEA2cCWoYZDt1EAYkDDp1g3Jj5nSdQmknxaCYJ+DEZChYxSoyV+ne3Dw18LagMQQ3KFbfqzoFXiZeTCsrRHJS/+qGkaQzCUE607nluYvyMKMMoh2mpn2pICB2TIfQsFSQG7Wfzm6f4zCohjqSyJQyeq78nMhJrPYkD2xkTM9LL3kz8z+ulJrr0MyaS1ICgi0VRyrGReBYADpkCavjEEkIVs7diOiKKUGNjKtkQvOWXV0m7VvXcqndfq9Sv8jiK6ASdonPkoQtURzeoiVqIogQ9o1f05qTOi/PufCxaC04+c4z+wPn8AfxikPw= AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiEfQgVLQf0C4lm51tQ7PJkmSFsvRvePGgiFf/jDf/jWm7B219MPB4b4aZeUHCmTau++0U1tY3NreK26Wd3b39g/LhUVvLVFFoUcml6gZEA2cCWoYZDt1EAYkDDp1g3Jj5nSdQmknxaCYJ+DEZChYxSoyV+ne3Dw18LagMQQ3KFbfqzoFXiZeTCsrRHJS/+qGkaQzCUE607nluYvyMKMMoh2mpn2pICB2TIfQsFSQG7Wfzm6f4zCohjqSyJQyeq78nMhJrPYkD2xkTM9LL3kz8z+ulJrr0MyaS1ICgi0VRyrGReBYADpkCavjEEkIVs7diOiKKUGNjKtkQvOWXV0m7VvXcqndfq9Sv8jiK6ASdonPkoQtURzeoiVqIogQ9o1f05qTOi/PufCxaC04+c4z+wPn8AfxikPw= AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiEfQgVLQf0C4lm51tQ7PJkmSFsvRvePGgiFf/jDf/jWm7B219MPB4b4aZeUHCmTau++0U1tY3NreK26Wd3b39g/LhUVvLVFFoUcml6gZEA2cCWoYZDt1EAYkDDp1g3Jj5nSdQmknxaCYJ+DEZChYxSoyV+ne3Dw18LagMQQ3KFbfqzoFXiZeTCsrRHJS/+qGkaQzCUE607nluYvyMKMMoh2mpn2pICB2TIfQsFSQG7Wfzm6f4zCohjqSyJQyeq78nMhJrPYkD2xkTM9LL3kz8z+ulJrr0MyaS1ICgi0VRyrGReBYADpkCavjEEkIVs7diOiKKUGNjKtkQvOWXV0m7VvXcqndfq9Sv8jiK6ASdonPkoQtURzeoiVqIogQ9o1f05qTOi/PufCxaC04+c4z+wPn8AfxikPw= CTRS Encoder AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiETxW7Re0S8lmZ9vQbLIkWaEs/RtePCji1T/jzX9j2u5BWx8MPN6bYWZekHCmjet+O4WNza3tneJuaW//4PCofHzS0TJVFNpUcql6AdHAmYC2YYZDL1FA4oBDN5g05n73CZRmUrTMNAE/JiPBIkaJsdKg0Xp4xLeCyhDUsFxxq+4CeJ14OamgHM1h+WsQSprGIAzlROu+5ybGz4gyjHKYlQaphoTQCRlB31JBYtB+trh5hi+sEuJIKlvC4IX6eyIjsdbTOLCdMTFjverNxf+8fmqiaz9jIkkNCLpcFKUcG4nnAeCQKaCGTy0hVDF7K6Zjogg1NqaSDcFbfXmddGpVz61697VK/SaPo4jO0Dm6RB66QnV0h5qojShK0DN6RW9O6rw4787HsrXg5DOn6A+czx8VKJEM AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiETxW7Re0S8lmZ9vQbLIkWaEs/RtePCji1T/jzX9j2u5BWx8MPN6bYWZekHCmjet+O4WNza3tneJuaW//4PCofHzS0TJVFNpUcql6AdHAmYC2YYZDL1FA4oBDN5g05n73CZRmUrTMNAE/JiPBIkaJsdKg0Xp4xLeCyhDUsFxxq+4CeJ14OamgHM1h+WsQSprGIAzlROu+5ybGz4gyjHKYlQaphoTQCRlB31JBYtB+trh5hi+sEuJIKlvC4IX6eyIjsdbTOLCdMTFjverNxf+8fmqiaz9jIkkNCLpcFKUcG4nnAeCQKaCGTy0hVDF7K6Zjogg1NqaSDcFbfXmddGpVz61697VK/SaPo4jO0Dm6RB66QnV0h5qojShK0DN6RW9O6rw4787HsrXg5DOn6A+czx8VKJEM AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiETxW7Re0S8lmZ9vQbLIkWaEs/RtePCji1T/jzX9j2u5BWx8MPN6bYWZekHCmjet+O4WNza3tneJuaW//4PCofHzS0TJVFNpUcql6AdHAmYC2YYZDL1FA4oBDN5g05n73CZRmUrTMNAE/JiPBIkaJsdKg0Xp4xLeCyhDUsFxxq+4CeJ14OamgHM1h+WsQSprGIAzlROu+5ybGz4gyjHKYlQaphoTQCRlB31JBYtB+trh5hi+sEuJIKlvC4IX6eyIjsdbTOLCdMTFjverNxf+8fmqiaz9jIkkNCLpcFKUcG4nnAeCQKaCGTy0hVDF7K6Zjogg1NqaSDcFbfXmddGpVz61697VK/SaPo4jO0Dm6RB66QnV0h5qojShK0DN6RW9O6rw4787HsrXg5DOn6A+czx8VKJEM AAAB83icbVBNSwMxEM3Wr1q/qh69BIvgqez2osdiETxW7Re0S8lmZ9vQbLIkWaEs/RtePCji1T/jzX9j2u5BWx8MPN6bYWZekHCmjet+O4WNza3tneJuaW//4PCofHzS0TJVFNpUcql6AdHAmYC2YYZDL1FA4oBDN5g05n73CZRmUrTMNAE/JiPBIkaJsdKg0Xp4xLeCyhDUsFxxq+4CeJ14OamgHM1h+WsQSprGIAzlROu+5ybGz4gyjHKYlQaphoTQCRlB31JBYtB+trh5hi+sEuJIKlvC4IX6eyIjsdbTOLCdMTFjverNxf+8fmqiaz9jIkkNCLpcFKUcG4nnAeCQKaCGTy0hVDF7K6Zjogg1NqaSDcFbfXmddGpVz61697VK/SaPo4jO0Dm6RB66QnV0h5qojShK0DN6RW9O6rw4787HsrXg5DOn6A+czx8VKJEM
CTRS Predictor AAAB9XicbVA9T8MwEHX4LOWrwMhiUSExVUkXGCu6MBbol9SGynGc1qpjR/YFVEX9HywMIMTKf2Hj3+C2GaDlSSc9vXenu3tBIrgB1/121tY3Nre2CzvF3b39g8PS0XHbqFRT1qJKKN0NiGGCS9YCDoJ1E81IHAjWCcb1md95ZNpwJZswSZgfk6HkEacErPRQb97d44ZmIaeg9KBUdivuHHiVeDkpoxyNQemrHyqaxkwCFcSYnucm4GdEA6eCTYv91LCE0DEZsp6lksTM+Nn86ik+t0qII6VtScBz9fdERmJjJnFgO2MCI7PszcT/vF4K0ZWfcZmkwCRdLIpSgUHhWQQ45JpREBNLCNXc3orpiGhCwQZVtCF4yy+vkna14rkV77Zarl3ncRTQKTpDF8hDl6iGblADtRBFGj2jV/TmPDkvzrvzsWhdc/KZE/QHzucPzd+SDA== AAAB9XicbVA9T8MwEHX4LOWrwMhiUSExVUkXGCu6MBbol9SGynGc1qpjR/YFVEX9HywMIMTKf2Hj3+C2GaDlSSc9vXenu3tBIrgB1/121tY3Nre2CzvF3b39g8PS0XHbqFRT1qJKKN0NiGGCS9YCDoJ1E81IHAjWCcb1md95ZNpwJZswSZgfk6HkEacErPRQb97d44ZmIaeg9KBUdivuHHiVeDkpoxyNQemrHyqaxkwCFcSYnucm4GdEA6eCTYv91LCE0DEZsp6lksTM+Nn86ik+t0qII6VtScBz9fdERmJjJnFgO2MCI7PszcT/vF4K0ZWfcZmkwCRdLIpSgUHhWQQ45JpREBNLCNXc3orpiGhCwQZVtCF4yy+vkna14rkV77Zarl3ncRTQKTpDF8hDl6iGblADtRBFGj2jV/TmPDkvzrvzsWhdc/KZE/QHzucPzd+SDA== AAAB9XicbVA9T8MwEHX4LOWrwMhiUSExVUkXGCu6MBbol9SGynGc1qpjR/YFVEX9HywMIMTKf2Hj3+C2GaDlSSc9vXenu3tBIrgB1/121tY3Nre2CzvF3b39g8PS0XHbqFRT1qJKKN0NiGGCS9YCDoJ1E81IHAjWCcb1md95ZNpwJZswSZgfk6HkEacErPRQb97d44ZmIaeg9KBUdivuHHiVeDkpoxyNQemrHyqaxkwCFcSYnucm4GdEA6eCTYv91LCE0DEZsp6lksTM+Nn86ik+t0qII6VtScBz9fdERmJjJnFgO2MCI7PszcT/vF4K0ZWfcZmkwCRdLIpSgUHhWQQ45JpREBNLCNXc3orpiGhCwQZVtCF4yy+vkna14rkV77Zarl3ncRTQKTpDF8hDl6iGblADtRBFGj2jV/TmPDkvzrvzsWhdc/KZE/QHzucPzd+SDA== AAAB9XicbVA9T8MwEHX4LOWrwMhiUSExVUkXGCu6MBbol9SGynGc1qpjR/YFVEX9HywMIMTKf2Hj3+C2GaDlSSc9vXenu3tBIrgB1/121tY3Nre2CzvF3b39g8PS0XHbqFRT1qJKKN0NiGGCS9YCDoJ1E81IHAjWCcb1md95ZNpwJZswSZgfk6HkEacErPRQb97d44ZmIaeg9KBUdivuHHiVeDkpoxyNQemrHyqaxkwCFcSYnucm4GdEA6eCTYv91LCE0DEZsp6lksTM+Nn86ik+t0qII6VtScBz9fdERmJjJnFgO2MCI7PszcT/vF4K0ZWfcZmkwCRdLIpSgUHhWQQ45JpREBNLCNXc3orpiGhCwQZVtCF4yy+vkna14rkV77Zarl3ncRTQKTpDF8hDl6iGblADtRBFGj2jV/TmPDkvzrvzsWhdc/KZE/QHzucPzd+SDA==
Shared Encoder AAAB9XicbVBNS8NAEJ34WetX1aOXxSJ4KkkveiyK4LGi/YA2ls1m0i7dbMLuRiml/8OLB0W8+l+8+W/ctjlo64OBx3szzMwLUsG1cd1vZ2V1bX1js7BV3N7Z3dsvHRw2dZIphg2WiES1A6pRcIkNw43AdqqQxoHAVjC8mvqtR1SaJ/LejFL0Y9qXPOKMGis93A2owpBcS5aEqHqlsltxZyDLxMtJGXLUe6WvbpiwLEZpmKBadzw3Nf6YKsOZwEmxm2lMKRvSPnYslTRG7Y9nV0/IqVVCEiXKljRkpv6eGNNY61Ec2M6YmoFe9Kbif14nM9GFP+YyzQxKNl8UZYKYhEwjICFXyIwYWUKZ4vZWwmwOlBkbVNGG4C2+vEya1YrnVrzbarl2mcdRgGM4gTPw4BxqcAN1aAADBc/wCm/Ok/PivDsf89YVJ585gj9wPn8AGOiSOw== AAAB9XicbVBNS8NAEJ34WetX1aOXxSJ4KkkveiyK4LGi/YA2ls1m0i7dbMLuRiml/8OLB0W8+l+8+W/ctjlo64OBx3szzMwLUsG1cd1vZ2V1bX1js7BV3N7Z3dsvHRw2dZIphg2WiES1A6pRcIkNw43AdqqQxoHAVjC8mvqtR1SaJ/LejFL0Y9qXPOKMGis93A2owpBcS5aEqHqlsltxZyDLxMtJGXLUe6WvbpiwLEZpmKBadzw3Nf6YKsOZwEmxm2lMKRvSPnYslTRG7Y9nV0/IqVVCEiXKljRkpv6eGNNY61Ec2M6YmoFe9Kbif14nM9GFP+YyzQxKNl8UZYKYhEwjICFXyIwYWUKZ4vZWwmwOlBkbVNGG4C2+vEya1YrnVrzbarl2mcdRgGM4gTPw4BxqcAN1aAADBc/wCm/Ok/PivDsf89YVJ585gj9wPn8AGOiSOw== AAAB9XicbVBNS8NAEJ34WetX1aOXxSJ4KkkveiyK4LGi/YA2ls1m0i7dbMLuRiml/8OLB0W8+l+8+W/ctjlo64OBx3szzMwLUsG1cd1vZ2V1bX1js7BV3N7Z3dsvHRw2dZIphg2WiES1A6pRcIkNw43AdqqQxoHAVjC8mvqtR1SaJ/LejFL0Y9qXPOKMGis93A2owpBcS5aEqHqlsltxZyDLxMtJGXLUe6WvbpiwLEZpmKBadzw3Nf6YKsOZwEmxm2lMKRvSPnYslTRG7Y9nV0/IqVVCEiXKljRkpv6eGNNY61Ec2M6YmoFe9Kbif14nM9GFP+YyzQxKNl8UZYKYhEwjICFXyIwYWUKZ4vZWwmwOlBkbVNGG4C2+vEya1YrnVrzbarl2mcdRgGM4gTPw4BxqcAN1aAADBc/wCm/Ok/PivDsf89YVJ585gj9wPn8AGOiSOw== AAAB9XicbVBNS8NAEJ34WetX1aOXxSJ4KkkveiyK4LGi/YA2ls1m0i7dbMLuRiml/8OLB0W8+l+8+W/ctjlo64OBx3szzMwLUsG1cd1vZ2V1bX1js7BV3N7Z3dsvHRw2dZIphg2WiES1A6pRcIkNw43AdqqQxoHAVjC8mvqtR1SaJ/LejFL0Y9qXPOKMGis93A2owpBcS5aEqHqlsltxZyDLxMtJGXLUe6WvbpiwLEZpmKBadzw3Nf6YKsOZwEmxm2lMKRvSPnYslTRG7Y9nV0/IqVVCEiXKljRkpv6eGNNY61Ec2M6YmoFe9Kbif14nM9GFP+YyzQxKNl8UZYKYhEwjICFXyIwYWUKZ4vZWwmwOlBkbVNGG4C2+vEya1YrnVrzbarl2mcdRgGM4gTPw4BxqcAN1aAADBc/wCm/Ok/PivDsf89YVJ585gj9wPn8AGOiSOw==
ˆ
T
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VihX9CEstlu2rWbTdidCCX0b3jxoIhX/4w3/43bNgdtfTDweG+GmXlhKoVB1/121tY3Nre2Szvl3b39g8PK0XHbJJlmvMUSmehuSA2XQvEWCpS8m2pO41DyTji+m/mdJ66NSFQTJykPYjpUIhKMopV8f0Qxb077uXic9itVt+bOQVaJV5AqFGj0K1/+IGFZzBUySY3peW6KQU41Cib5tOxnhqeUjemQ9yxVNOYmyOc3T8m5VQYkSrQthWSu/p7IaWzMJA5tZ0xxZJa9mfif18swuglyodIMuWKLRVEmCSZkFgAZCM0ZyokllGlhbyVsRDVlaGMq2xC85ZdXSfuy5rk17+GqWr8t4ijBKZzBBXhwDXW4hwa0gEEKz/AKb07mvDjvzseidc0pZk7gD5zPH4zSkgE= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VihX9CEstlu2rWbTdidCCX0b3jxoIhX/4w3/43bNgdtfTDweG+GmXlhKoVB1/121tY3Nre2Szvl3b39g8PK0XHbJJlmvMUSmehuSA2XQvEWCpS8m2pO41DyTji+m/mdJ66NSFQTJykPYjpUIhKMopV8f0Qxb077uXic9itVt+bOQVaJV5AqFGj0K1/+IGFZzBUySY3peW6KQU41Cib5tOxnhqeUjemQ9yxVNOYmyOc3T8m5VQYkSrQthWSu/p7IaWzMJA5tZ0xxZJa9mfif18swuglyodIMuWKLRVEmCSZkFgAZCM0ZyokllGlhbyVsRDVlaGMq2xC85ZdXSfuy5rk17+GqWr8t4ijBKZzBBXhwDXW4hwa0gEEKz/AKb07mvDjvzseidc0pZk7gD5zPH4zSkgE= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VihX9CEstlu2rWbTdidCCX0b3jxoIhX/4w3/43bNgdtfTDweG+GmXlhKoVB1/121tY3Nre2Szvl3b39g8PK0XHbJJlmvMUSmehuSA2XQvEWCpS8m2pO41DyTji+m/mdJ66NSFQTJykPYjpUIhKMopV8f0Qxb077uXic9itVt+bOQVaJV5AqFGj0K1/+IGFZzBUySY3peW6KQU41Cib5tOxnhqeUjemQ9yxVNOYmyOc3T8m5VQYkSrQthWSu/p7IaWzMJA5tZ0xxZJa9mfif18swuglyodIMuWKLRVEmCSZkFgAZCM0ZyokllGlhbyVsRDVlaGMq2xC85ZdXSfuy5rk17+GqWr8t4ijBKZzBBXhwDXW4hwa0gEEKz/AKb07mvDjvzseidc0pZk7gD5zPH4zSkgE= AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VihX9CEstlu2rWbTdidCCX0b3jxoIhX/4w3/43bNgdtfTDweG+GmXlhKoVB1/121tY3Nre2Szvl3b39g8PK0XHbJJlmvMUSmehuSA2XQvEWCpS8m2pO41DyTji+m/mdJ66NSFQTJykPYjpUIhKMopV8f0Qxb077uXic9itVt+bOQVaJV5AqFGj0K1/+IGFZzBUySY3peW6KQU41Cib5tOxnhqeUjemQ9yxVNOYmyOc3T8m5VQYkSrQthWSu/p7IaWzMJA5tZ0xxZJa9mfif18swuglyodIMuWKLRVEmCSZkFgAZCM0ZyokllGlhbyVsRDVlaGMq2xC85ZdXSfuy5rk17+GqWr8t4ijBKZzBBXhwDXW4hwa0gEEKz/AKb07mvDjvzseidc0pZk7gD5zPH4zSkgE=
ˆ
Y
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfkgTyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGUhJIG
ˆ
Z
ij
AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH AAAB83icbVBNS8NAEJ34WetX1aOXxSJ4KokIeix68VjBfmATyma7bdduNmF3IpSQv+HFgyJe/TPe/Ddu2xy09cHA470ZZuaFiRQGXffbWVldW9/YLG2Vt3d29/YrB4ctE6ea8SaLZaw7ITVcCsWbKFDyTqI5jULJ2+H4Zuq3n7g2Ilb3OEl4ENGhEgPBKFrJ90cUs4e8l4nHvFepujV3BrJMvIJUoUCjV/ny+zFLI66QSWpM13MTDDKqUTDJ87KfGp5QNqZD3rVU0YibIJvdnJNTq/TJINa2FJKZ+nsio5Exkyi0nRHFkVn0puJ/XjfFwVWQCZWkyBWbLxqkkmBMpgGQvtCcoZxYQpkW9lbCRlRThjamsg3BW3x5mbTOa55b8+4uqvXrIo4SHMMJnIEHl1CHW2hAExgk8Ayv8Oakzovz7nzMW1ecYuYI/sD5/AGWDpIH
w
ij
AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY= AAAB9HicbVBNS8NAEN3Ur1q/qh69BIvgqSQi6LHoxWMF+wFtKJvtpF272cTdSbWE/A4vHhTx6o/x5r9x2+agrQ8GHu/NMDPPjwXX6DjfVmFldW19o7hZ2tre2d0r7x80dZQoBg0WiUi1fapBcAkN5CigHSugoS+g5Y+up35rDErzSN7hJAYvpAPJA84oGsnrIjxh+pj1Un6f9coVp+rMYC8TNycVkqPeK391+xFLQpDIBNW64zoxeilVyJmArNRNNMSUjegAOoZKGoL20tnRmX1ilL4dRMqURHum/p5Iaaj1JPRNZ0hxqBe9qfif10kwuPRSLuMEQbL5oiARNkb2NAG7zxUwFBNDKFPc3GqzIVWUocmpZEJwF19eJs2zqutU3dvzSu0qj6NIjsgxOSUuuSA1ckPqpEEYeSDP5JW8WWPrxXq3PuatBSufOSR/YH3+ALvikrY=
(b) Multitask
Figure 5.2: Overview of single task and multitask systems
42
5.4 Learning with Context
Because the context of a turn, i.e., the surrounding turns, can provide important
discriminative information for decoding a particular turn, the word encoder is
followed by a context encoder. The context encoder is provided both the vector,
H
ij
, the result of the word encoder, as well as the role of the speaker of the turn,
r
ij
(i.e., an indicator variable of whether the speaker is the therapist or client).
The speaker role undergoes a single linear transformation layer whose output is of
equal size to the number of labels L. These two variable are then concatenated
into a vector representation of the turn, X
ij
, i.e.,
X
ij
= [H
ij
;W
X
r
ij
+b
X
]: (5.10)
The current turn representation, X
ij
, is then concatenated with the C preced-
ing and following turns, X
C
ij
=fX
i(jC)
;X
i(jC+1)
;:::;X
ij
;:::;X
i(j+C1)
;X
i(j+C)
g
and input a BiLSTM which will encode the turn context representations. The
hidden states of the BiLSTM are averaged resulting in a turn context vector rep-
resentation, G
ij
. A visualization of the context encoder is given in gure 5.1b.
5.5 Model Implementation and Training
All models are implemented in Keras [59] with Theano as the backend [60]. An
early stopping procedure is used in which training is terminated if loss on the
validation set does not improve after consecutive epochs. The validation set is
10% of the data from the training set which is chosen randomly. All models
are optimized using the adam algorithm with a learning rate of 10
3
[62]. As a
metric of comparison we use f1-score (macro average across labels). The reported
43
results are the mean f1-score of the network being trained from 10 dierent random
initializations.
Word embeddings vectors (300 dimensional) are pre-trained on the training
data using word2vec [55]. All hidden layers are of the same dimension as the word
vectors (initialized with Glorot uniform). The data is separated into batches of 32
samples and shued randomly between epochs.
As described in section 5.3 we use an adversarial approach, with the hyper-
parameters = 0:05 and
= 0:01 as recommended in [34] for multi-task learning.
The unshared encoding layers are initialized with the weights from the multi-label
systems trained in the previous experiments. The learning rate of the optimizer is
reduced to 10
4
to allow for ne tuning of these layers.
5.6 Experiments and Results
5.6.1 Multi-label learning
In this section we compare single-label (SL) and multi-label (ML) approaches with
and without sample weighting (sw) for predicting behavioral codes in our exposi-
tory domains. In table 5.1, we show prediction results (f1-score) for the MISC and
CTRS behavioral code prediction tasks. As a point of reference, we include the
f1-score for each behavioral code if that behavior is considered present in every
turn referred to as `baseline' in the table. Due to the imbalance in the labels
in both datasets, sample weights improve performance for both single-label and
multi-label prediction.
For MISC code prediction, the multi-label approaches outperform single-label
approaches both with and without sample weighting. The highest per-code results
are for multi-label without sample weights for the codes FA, GI, QUC, QUO,
44
REC, RES, and FN. The occurrence of these codes is more balanced thus they
do not benet from the sample weighting scheme. However, the more unbalanced
codes (MIA, MIN, POS, and NEG) are predicted best by the multi-label system
with sample weights, due to the infrequency in which they occur in the data. On
average, the ML-sw approach resulted in the highest performance for the MISC
prediction task.
With respect to the CTRS behavioral code prediction task, the multi-label
system with sample weighting yielded the highest (or tied) f1-score performance
for all CTRS codes. In this case, however, the mutli-label system without sample
weighting did not outperform the single-label approach using sample weights.
5.6.2 Multi-task Learning
In this section we evaluate the performance of a multi-task model that aims to learn
both MISC and CTRS behavioral codes. We show the results for the multi-task
model in table 5.1. The average performance of both tasks is improved compared
to their single-task counterparts.
In the MISC task, prediction of MIN (MI non-adherent) had the largest rel-
ative improvement (18.1%) with respect to the best result from the single-task
approaches. MI non-adherent behaviors include confrontation, direction, warning,
advising and raising concern without permission. These are considered negative
therapist behaviors in motivation interviewing and are to be avoided by therapists
practicing MI due to its non-confrontational and non-adversarial nature. For this
reason, their occurrence is very rare in the MI corpus. The combined occurrences
of these ve behaviors combined account for only 2.3% of all therapist turns in
the dataset. Cognitive behavior therapy has a distinctly dierent approach with
respect to directive statements. This is evidenced by the behavior of assigning
45
homework (HW), an essential element of CBT. Thus statements that would be
considered non-adherent in MI counseling are much more likely to occur in CBT
sessions. In this way the multi-task system allows for a better representation of
these behaviors even though the turns in the CBT dataset are not labeled for these
behaviors.
With respect to the CTRS task, guided discovery (GD) had the largest relative
improvements (8.5%) for the multi-task system with respect to the single-task sys-
tem. Guided discovery is a behavior in which the therapist \uses exploration and
questioning to help patients see new perspectives." In the MI data, questions are
explicitly labeled and therefore encoded by the system. Because of the important
relation between questioning and GD, the multi-task system enables a better rep-
resentation for decoding this behavior, despite not having turns manually labeled
as questions.
5.6.3 Turn Context Learning
In table 5.2, we present results comparing prediction performance when adding
context to the multi-label prediction task. Adding contextual information pro-
vides increased performance for almost every behavioral code. In the MISC task,
the codes REC and RES (complex and simple re
ections) had the largest relative
improvement (17.9% and 17.5% respectively). Re
ections are when a therapist
restates information provided by the client and are either a slight rephrasing (sim-
ple) or add signicant meaning or emphasis (complex) to the client's statements.
Thus, it is intuitive that prediction of these behaviors would benet from knowing
the surrounding utterances.
In the CTRS task, the codes guided discovery (GD), strategy for change (SC),
and pacing and timing (PT), had the most relative performance improvement with
46
context (12.6%, 7.3%, and 6.9% respectively). Guided discover and strategy for
change are both behaviors which re
ect the therapists' conceptualization of the
client's concern and their approach for addressing them. Therefore, these are
behaviors that unfold and occur throughout the session not in isolation. Pacing
and timing (PT) re
ects the therapist's ability to manage the pace of the session
over the course of the session and thus turn context will provide useful information
about the therapist's skill in this regard.
5.6.4 Multi-label Multi-task Learning with Context
We show the results for the a system combining the multi-label and multi-task
paradigms with context in table 5.2. The mutli-label, multi-task system with
context achieved the highest combined performance for the two tasks. The aver-
age performance for the CTRS task did not improve in this setting versus the
single-task multi-label system with context. The CTRS prediction loss typically
converges more quickly than that of the MISC, likely due in part to the amount of
available data as well as the level of supervision (session labels versus turn labels).
The multi-label multi-task system achieved higher performance than the single task
system for 8 of 11 MISC behavioral codes and 4 of 11 CTRS behavioral codes. The
CTRS behavioral codes interpersonal eective (IP) and understanding (UN) had
the most extreme label imbalance and thus only UN achieved performance above
the baseline in the case of multi-label single-task with context of 4 turns.
5.7 Conclusions
In this chapter we proposed multi-label and multi-task approaches for behavioral
coding of psychotherapy interactions. We demonstrated that by incorporating
47
these paradigms, which help re
ect the complexities of these data, better prediction
of behaviors in these sessions is achieved. The multi-label multi-task system with
turn context achieved the highest combined prediction for the behavioral coding
tasks.
48
Table 5.1: Comparison of single-label, multi-label,and multi-label multi-task sys-
tems
code baseline SL SL-sw ML ML-sw ML-MT
MISC
FA 0.289 0.909 0.887 0.919 0.903 0.911
GI 0.264 0.757 0.709 0.771 0.743 0.760
QUC 0.156 0.672 0.586 0.625 0.598 0.659
QUO 0.122 0.802 0.639 0.798 0.787 0.801
REC 0.143 0.498 0.484 0.522 0.504 0.564
RES 0.185 0.476 0.516 0.491 0.429 0.486
MIA 0.118 0.556 0.442 0.517 0.548 0.576
MIN 0.018 0.001 0.112 0.066 0.199 0.235
FN 0.637 0.960 0.963 0.964 0.949 0.958
POS 0.117 0.286 0.363 0.316 0.379 0.381
NEG 0.094 0.185 0.322 0.252 0.339 0.354
AVG 0.195 0.555 0.548 0.567 0.580 0.608
CTRS
AG 0.667 0.718 0.718 0.716 0.784 0.790
AT 0.605 0.654 0.654 0.654 0.714 0.731
CO 0.776 0.327 0.776 0.776 0.778 0.776
FB 0.636 0.686 0.686 0.687 0.751 0.772
GD 0.636 0.672 0.672 0.667 0.693 0.752
HW 0.605 0.671 0.672 0.661 0.743 0.654
IP 0.929 0.000 0.929 0.929 0.929 0.929
KC 0.667 0.692 0.692 0.687 0.717 0.753
PT 0.696 0.705 0.705 0.720 0.741 0.798
SC 0.605 0.644 0.644 0.642 0.695 0.744
UN 0.800 0.071 0.800 0.800 0.800 0.800
AVG 0.688 0.531 0.723 0.722 0.758 0.773
49
Table 5.2: Multi-label Learning with Context
C 0 1 2 3 4
ML
MISC
FA 0.903 0.918 0.912 0.918 0.917
GI 0.743 0.762 0.756 0.770 0.764
QUC 0.598 0.648 0.634 0.653 0.659
QUO 0.787 0.803 0.801 0.809 0.809
REC 0.504 0.549 0.560 0.558 0.594
RES 0.429 0.463 0.461 0.495 0.504
MIA 0.548 0.565 0.532 0.570 0.558
MIN 0.199 0.213 0.191 0.224 0.220
FN 0.949 0.956 0.956 0.960 0.954
POS 0.379 0.405 0.371 0.408 0.401
NEG 0.339 0.372 0.361 0.365 0.384
AVG 0.580 0.605 0.594 0.612 0.615
CTRS
AG 0.784 0.771 0.787 0.732 0.766
AT 0.714 0.733 0.749 0.691 0.739
CO 0.778 0.787 0.792 0.790 0.789
FB 0.751 0.750 0.770 0.716 0.754
GD 0.693 0.741 0.772 0.758 0.780
HW 0.743 0.705 0.731 0.637 0.737
IP 0.929 0.929 0.929 0.929 0.929
KC 0.717 0.736 0.765 0.722 0.743
PT 0.741 0.767 0.779 0.780 0.792
SC 0.695 0.726 0.746 0.695 0.743
UN 0.800 0.800 0.800 0.800 0.803
AVG 0.758 0.768 0.784 0.750 0.780
50
Table 5.3: Multi-task Learning with Context
C 0 1 2 3 4
MISC
FA 0.911 0.918 0.918 0.917 0.919
GI 0.760 0.775 0.774 0.761 0.776
QUC 0.659 0.656 0.667 0.638 0.686
QUO 0.801 0.809 0.812 0.812 0.806
REC 0.564 0.592 0.576 0.572 0.570
RES 0.486 0.519 0.504 0.499 0.516
MIA 0.576 0.580 0.587 0.551 0.581
MIN 0.235 0.223 0.221 0.229 0.208
FN 0.958 0.959 0.956 0.960 0.960
POS 0.381 0.396 0.416 0.332 0.397
NEG 0.354 0.377 0.372 0.383 0.391
AVG 0.608 0.619 0.619 0.605 0.619
CTRS
AG 0.790 0.739 0.772 0.771 0.741
AT 0.731 0.712 0.739 0.742 0.707
CO 0.776 0.775 0.777 0.783 0.774
FB 0.772 0.753 0.778 0.753 0.712
GD 0.752 0.771 0.770 0.746 0.764
HW 0.654 0.643 0.723 0.735 0.703
IP 0.929 0.929 0.929 0.929 0.929
KC 0.753 0.753 0.757 0.726 0.726
PT 0.798 0.797 0.828 0.800 0.794
SC 0.744 0.737 0.752 0.702 0.718
UN 0.800 0.800 0.800 0.800 0.800
AVG 0.773 0.765 0.784 0.772 0.761
51
Chapter 6
Conclusions and Future Work
6.1 Conclusions
In this disseration, we proposed and experimentally investigated several machine
learning paradigms for the task of human behavioral coding. In chapter 3, we pro-
posed a deep learning system which maps from turn level feature representations to
turn level behavioral codes, then from turn level behavioral codes to session level
empathy. We demonstrated the proposed network, when trained using a two stage
approach, outperforms a shallow network based only on turn feature representa-
tions and performs comparably to a system predicting empathy from the manually
annotated turn level behavioral codes. In chapter 4, we proposed a word level
LSTM network with an attention mechanism for predicting turn level behavioral
codes. We demonstrated that the attention-based network improves prediction of
several behavioral codes which also adding increased model interpretability. In
chapter 5, we proposed a system that integrated multi-label and multi-task learn-
ing approaches for predicting behavioral codes in two psychotherapy domains. We
demonstrated that these approaches oer increased accuracy for both prediction
tasks.
52
6.2 Potential Applications of Automatic Behav-
ioral Coding
Providing automated methods for coding behaviors which occur in psychotherapy
interactions has many potential applications. One of the rst proposed applica-
tions is the task of evaluating therapist ecacy from therapy audio recordings
using a speech pipeline system which performs audio segmentation, automatic
transcription, and behavioral code prediction [67]. Such a framework could enable
patients to choose their therapist based on empirically derived quality metrics
rather than word of mouth and online reviews. Additionally, this could enable for
monitoring patient progress and tracking of behavioral changes and symptoms over
time. Automatic behavioral coding (ABC) has the potential to augment therapist
understanding of their clients and the quality of the therapy they are providing
by allowing rapid monitoring and feedback of their therapy sessions. Furthermore,
lessons learned from ABC developments within the psychotherapy domain may
provide insights to automatic understanding and modeling of human behaviors in
other human-human and human-computer interaction domains.
53
6.3 Future Directions
6.3.1 Automatic Behavioral Coding in the `Wild'
There are many potential avenues for extending the proposed work. One key
direction is to investigate how these learning paradigms are aected by imperfect
word, speaker, and turn boundary information that would be derived from a speech
pipeline system [10]. Such an investigation is necessary to determine the feasibil-
ity of incorporating complex learning paradigms in a truly automatic behavioral
coding system.
While the present work did not incorporate multi-modal feature representa-
tions, it is an important line of inquiry. As discussed in section 1.3, there are
many behavioral cues which are important for the behavioral coding task. One
such eort [68], proposes fusing lexical and prosodic information in an attentional
LSTM to predict behaviors in MI therapy sessions. The promising results of this
initial step encourage further exploration of this area.
6.3.2 Transfer Learning and Domain Adaptation
There are many machine learning paradigms which are not discussed in this work
that could potentially be of interest for behavioral coding domains. Domain adap-
tation networks learn representations from data of one domain and then adapt the
representation to data of a target domain [69]. This approach is a type of inductive
transfer learning, where data from domains that is readily available can be used
to augment learning for domains where data is harder to collect. This could be of
special interest in psychotherapy and behavioral health domains where data are
often of an extremely private and sensitive nature.
54
6.3.3 Beyond Behavioral Coding
There are many ways that advances in signal processing and machine learning can
aid mental and behavioral health domains beyond automating behavioral coding.
One important direction is tracking of client symptoms and progress longitudi-
nally. Furthermore statistical modeling of therapeutic conversations may aid in
the development data driven therapy quality metrics.
55
Reference List
[1] R. Bakeman, \Behavioral observation and coding," Handbook of research
methods in social and personality psychology, pp. 138{159, 2000.
[2] Z. E. Imel, B. E. Wampold, S. D. Miller, and R. R. Fleming, \Distinctions
without a dierence: direct comparisons of psychotherapies for alcohol use
disorders." Psychology of Addictive Behaviors, vol. 22, no. 4, p. 533, 2008.
[3] C. Lord, M. Rutter, P. DiLavore, S. Risi, K. Gotham, and S. Bishop, Autism
diagnostic observation schedule: ADOS-2. Western Psychological Services
Torrance, 2012.
[4] G. Margolin, P. Oliver, E. Gordis, H. O'Hearn, A. Medina, C. Ghosh, and
L. Morland, \The nuts and bolts of behavioral observation of marital and
family interaction," Clinical Child and Family Psychology Review, vol. 1, no. 4,
pp. 195{213, 1998.
[5] A. Christensen, D. Atkins, S. Berns, J. Wheeler, D. Baucom, and L. Simpson,
\Traditional versus integrative behavioral couple therapy for signicantly and
chronically distressed married couples," Journal of Consulting and Clinical
Psychology, vol. 72, no. 2, pp. 176{191, 2004.
[6] W. R. Miller and G. S. Rose, \Toward a theory of motivational interviewing."
American Psychologist, vol. 64, no. 6, p. 527, 2009.
[7] T. A. Creed, S. A. Frankel, R. E. German, K. L. Green, S. Jager-Hyman,
K. P. Taylor, A. D. Adler, C. B. Wolk, S. W. Stirman, S. H. Waltman et al.,
\Implementation of transdiagnostic cognitive therapy in community behav-
ioral health: The beck community initiative." Journal of consulting and clin-
ical psychology, vol. 84, no. 12, p. 1116, 2016.
[8] S. Narayanan and P. G. Georgiou, \Behavioral signal processing: Deriving
human behavioral informatics from speech and language," Proceedings of the
IEEE, 2012.
56
[9] M. Black, A. Katsamanis, B. Baucom, C. Lee, A. Lammert, A. Christensen,
P. Georgiou, and S. Narayanan, \Toward automating a human behavioral
coding system for married couples interactions using speech acoustic features,"
Speech Communication, 2011.
[10] B. Xiao, C. Huang, Z. E. Imel, D. C. Atkins, P. Georgiou, and S. S. Narayanan,
\A technology prototype system for rating therapist empathy from audio
recordings in addiction counseling," PeerJ Computer Science, vol. 2, p. e59,
2016.
[11] M. Tanana, K. A. Hallgren, Z. E. Imel, D. C. Atkins, and V. Srikumar, \A
comparison of natural language processing methods for automated coding of
motivational interviewing," Journal of substance abuse treatment, 2016.
[12] D. Bone, S. L. Bishop, M. P. Black, M. S. Goodwin, C. Lord, and S. S.
Narayanan, \Use of machine learning to improve autism screening and diag-
nostic instruments: eectiveness, eciency, and multi-instrument fusion,"
Journal of Child Psychology and Psychiatry, vol. 57, no. 8, pp. 927{937, 2016.
[13] B. Xiao, P. G. Georgiou, Z. E. Imel, D. C. Atkins, and S. Narayanan, \Mod-
eling therapist empathy and vocal entrainment in drug addiction counseling."
in Proceedings of Interspeech, 2013, pp. 2861{2865.
[14] B. Xiao, Z. E. Imel, D. C. Atkins, P. G. Georgiou, and S. S. Narayanan, \Ana-
lyzing speech rate entrainment and its relation to therapist empathy in drug
addiction counseling," in Sixteenth Annual Conference of the International
Speech Communication Association, 2015.
[15] J. Gibson, N. Malandrakis, F. Romero, D. C. Atkins, and S. S. Narayanan,
\Predicting therapist empathy in motivational interviews using language fea-
tures inspired by psycholinguistic norms," in Sixteenth Annual Conference of
the International Speech Communication Association, 2015.
[16] V. P erez-Rosas, R. Mihalcea, K. Resnicow, S. Singh, L. Ann, K. J. Goggin,
and D. Catley, \Predicting counselor behaviors in motivational interviewing
encounters," in Proceedings of the 15th Conference of the European Chapter
of the Association for Computational Linguistics: Volume 1, Long Papers,
vol. 1, 2017, pp. 1128{1137.
[17] P. G. Georgiou, M. Black, A. Lammert, B. Baucom, and S. Narayanan,
\That's aggravating, very aggravating: Is it possible to classify behaviors in
couple interactions using automatically derived lexical features?" in Aective
Computing and Intelligent Interaction (ACII), 2011.
57
[18] B. Xiao, P. G. Georgiou, B. Baucom, and S. S. Narayanan, \Power-spectral
analysis of head motion signal for behavioral modeling in human interaction,"
in IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), May 2014, pp. 4593{4597.
[19] Y. LeCun, Y. Bengio, and G. Hinton, \Deep learning," nature, vol. 521, no.
7553, p. 436, 2015.
[20] D. Bahdanau, K. Cho, and Y. Bengio, \Neural machine translation by jointly
learning to align and translate," arXiv preprint arXiv:1409.0473, 2014.
[21] V. Mnih, N. Heess, A. Graves et al., \Recurrent models of visual attention,"
in Advances in neural information processing systems, 2014, pp. 2204{2212.
[22] K. Cho, A. Courville, and Y. Bengio, \Describing multimedia content using
attention-based encoder-decoder networks," IEEE Transactions on Multime-
dia, vol. 17, no. 11, pp. 1875{1886, 2015.
[23] Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, and E. Hovy, \Hierarchical
attention networks for document classication," in Proceedings of NAACL-
HLT, 2016, pp. 1480{1489.
[24] M.-L. Zhang and Z.-H. Zhou, \A review on multi-label learning algorithms,"
IEEE transactions on knowledge and data engineering, vol. 26, no. 8, pp.
1819{1837, 2014.
[25] M. S. Sorower, \A literature survey on algorithms for multi-label learning,"
Oregon State University, Corvallis, vol. 18, 2010.
[26] M.-L. Zhang and Z.-H. Zhou, \Ml-knn: A lazy learning approach to multi-
label learning," Pattern recognition, vol. 40, no. 7, pp. 2038{2048, 2007.
[27] A. Elissee and J. Weston, \A kernel method for multi-labelled classication,"
in Advances in neural information processing systems, 2002, pp. 681{687.
[28] M.-L. Zhang and Z.-H. Zhou, \Multilabel neural networks with applications to
functional genomics and text categorization," IEEE Transactions on Knowl-
edge and Data Engineering, vol. 18, no. 10, pp. 1338{1351, 2006.
[29] R. Caruana, \Multitask learning," in Learning to learn. Springer, 1998, pp.
95{133.
[30] M. Lan, J. Wang, Y. Wu, Z.-Y. Niu, and H. Wang, \Multi-task attention-
based neural networks for implicit discourse relationship representation and
identication," in Proceedings of the 2017 Conference on Empirical Methods
in Natural Language Processing, 2017, pp. 1299{1308.
58
[31] Y. Yan, E. Ricci, R. Subramanian, G. Liu, O. Lanz, and N. Sebe, \A multi-
task learning framework for head pose estimation under target motion," IEEE
transactions on pattern analysis and machine intelligence, vol. 38, no. 6, pp.
1070{1083, 2016.
[32] G. Pons and D. Masip, \Multi-task, multi-label and multi-domain learning
with residual convolutional networks for emotion recognition," arXiv preprint
arXiv:1802.06664, 2018.
[33] Y. Zhang, P. Zhang, and Y. Yan, \Attention-based lstm with multi-task learn-
ing for distant speech recognition," in Proceedings of Interspeech, 2017, pp.
3857{3861.
[34] P. Liu, X. Qiu, and X. Huang, \Adversarial multi-task learning for text clas-
sication," arXiv preprint arXiv:1704.05742, 2017.
[35] D. Tang, B. Qin, and T. Liu, \Document modeling with gated recurrent neural
network for sentiment classication," in Proceedings of the 2015 conference on
empirical methods in natural language processing, 2015, pp. 1422{1432.
[36] J. Li, M.-T. Luong, and D. Jurafsky, \A hierarchical neural autoencoder for
paragraphs and documents," arXiv preprint arXiv:1506.01057, 2015.
[37] A. L. Katole, K. P. Yellapragada, A. K. Bedi, S. S. Kalra, and M. S. Chaitanya,
\Hierarchical deep learning architecture for 10k objects classication," arXiv
preprint arXiv:1509.01951, 2015.
[38] H. Lakkaraju, R. Socher, and C. Manning, \Aspect specic sentiment analysis
using hierarchical deep learning," in NIPS Workshop on Deep Learning and
Representation Learning, 2014.
[39] J. Gibson, A. Katsamanis, F. Romero, B. Xiao, P. Georgiou, and
S. Narayanan, \Multiple instance learning for behavioral coding," IEEE
Transactions on Aective Computing, vol. 8, no. 1, pp. 81{94, 2017.
[40] D. Can, D. C. Atkins, and S. S. Narayanan, \A dialog act tagging approach
to behavioral coding: A case study of addiction counseling conversations,"
in Sixteenth Annual Conference of the International Speech Communication
Association, 2015.
[41] M. Tanana, K. Hallgren, Z. Imel, D. Atkins, P. Smyth, and V. Srikumar,
\Recursive neural networks for coding therapist and patient behavior in moti-
vational interviewing," in Proceedings of the 2nd Workshop on Computational
Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Real-
ity, 2015, pp. 71{79.
59
[42] B. Xiao, D. Can, J. Gibson, Z. E. Imel, D. C. Atkins, P. Georgiou, and
S. Narayanan, \Behavioral coding of therapist language in addiction counsel-
ing using recurrent neural networks," in Proceedings of Interspeech, 2016, pp.
908{912.
[43] J. Gibson, D. Can, B. Xiao, Z. E. Imel, D. C. Atkins, P. Georgiou, and S. S.
Narayanan, \A deep learning approach to modeling empathy in addiction
counseling," in Proceedings of Interspeech, 2016.
[44] K. Singla, Z. Chen, N. Flemotomos, J. Gibson, D. Can, D. C. Atkins, and
S. Narayanan, \Using prosodic and lexical information for learning utterance-
level behaviors in psychotherapy," in Proceedings of Interspeech, 2018.
[45] N. Flemotomos, V. R. Martinez, J. Gibson, D. C. Atkins, T. A. Creed,
and S. Narayanan, \Language features for automated evaluation of cognitive
behavior psychotherapy sessions," in Proceedings of Interspeech, 2018.
[46] J. Gaume, G. Gmel, M. Faouzi, and J.-B. Daeppen, \Counselor skill in
uences
outcomes of brief motivational interventions," Journal of Substance Abuse
Treatment, vol. 37, no. 2, pp. 151{159, 2009.
[47] J. McCambridge, M. Day, B. A. Thomas, and J. Strang, \Fidelity to moti-
vational interviewing and subsequent cannabis cessation among adolescents,"
Addictive Behaviors, vol. 36, no. 7, pp. 749{754, 2011.
[48] D. C. Atkins, M. Steyvers, Z. E. Imel, and P. Smyth, \Scaling up the evalu-
ation of psychotherapy: evaluating motivational interviewing delity via sta-
tistical text classication," Implementation Science, vol. 9, no. 1, p. 49, 2014.
[49] S. P. Lord, E. Sheng, Z. E. Imel, J. Baer, and D. C. Atkins, \More than
re
ections: Empathy in motivational interviewing includes language style syn-
chrony between therapist and client," Behavior Therapy, 2014.
[50] J. S. Baer, E. A. Wells, D. B. Rosengren, B. Hartzler, B. Beadnell, and
C. Dunn, \Agency context and tailored training in technology transfer: A
pilot evaluation of motivational interviewing training for community coun-
selors," Journal of Substance Abuse Treatment, vol. 37, no. 2, pp. 191{202,
2009.
[51] W. R. Miller, T. B. Moyers, D. Ernst, and P. Amrhein, \Manual for the moti-
vational interviewing skill code (MISC)," Center on Alcoholism, Substance
Abuse and Addictions, University of New Mexico, 2003.
[52] R. Elliott, A. C. Bohart, J. C. Watson, and L. S. Greenberg, \Empathy."
Psychotherapy, vol. 48, no. 1, p. 43, 2011.
60
[53] J. S. Beck, Cognitive behavior therapy: Basics and beyond. Guilford press,
2011.
[54] J. Young and A. T. Beck, \Cognitive therapy scale: Rating manual," Unpub-
lished manuscript, p. 36th, 1980.
[55] T. Mikolov, K. Chen, G. Corrado, and J. Dean, \Ecient estimation of word
representations in vector space," arXiv preprint arXiv:1301.3781, 2013.
[56] S. Hochreiter and J. Schmidhuber, \Long short-term memory," Neural com-
putation, vol. 9, no. 8, pp. 1735{1780, 1997.
[57] Y. Bengio, P. Simard, and P. Frasconi, \Learning long-term dependencies with
gradient descent is dicult," Neural Networks, IEEE Transactions on, vol. 5,
no. 2, pp. 157{166, 1994.
[58] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, \Distributed
representations of words and phrases and their compositionality," in Advances
in neural information processing systems, 2013, pp. 3111{3119.
[59] F. Chollet, \Keras," https://github.com/fchollet/keras, 2015.
[60] F. Bastien, P. Lamblin, R. Pascanu, J. Bergstra, I. J. Goodfellow, A. Berg-
eron, N. Bouchard, and Y. Bengio, \Theano: new features and speed improve-
ments," Deep Learning and Unsupervised Feature Learning NIPS 2012 Work-
shop, 2012.
[61] X. Glorot and Y. Bengio, \Understanding the diculty of training deep feed-
forward neural networks," in International conference on articial intelligence
and statistics, 2010, pp. 249{256.
[62] D. Kingma and J. Ba, \Adam: A method for stochastic optimization," arXiv
preprint arXiv:1412.6980, 2014.
[63] B. Xiao, P. G. Georgiou, and S. Narayanan, \Analyzing the language of thera-
pist empathy in motivational interview based psychotherapy," in Asia-Pacic
Signal and Information Processing Association, 2012, pp. 1{4.
[64] D. Can, R. A. Mar n, P. G. Georgiou, Z. E. Imel, D. C. Atkins, and S. S.
Narayanan, \\it sounds like...": A natural language processing approach to
detecting counselor re
ections in motivational interviewing." Journal of coun-
seling psychology, vol. 63, no. 3, p. 343, 2016.
[65] J. Gibson, D. Can, P. Georgiou, D. C. Atkins, and S. S. Narayanan, \Attention
networks for modeling behaviors in addiction counseling," in Proceedings of
Interspeech, 2017.
61
[66] M.-L. Zhang, Y.-K. Li, and X.-Y. Liu, \Towards class-imbalance aware multi-
label learning." in IJCAI, 2015, pp. 4041{4047.
[67] B. Xiao, Z. E. Imel, P. G. Georgiou, D. C. Atkins, and S. S. Narayanan,
\`rate my therapist': Automated detection of empathy in drug and alcohol
counseling via speech and language processing," PloS one, vol. 10, no. 12, p.
e0143055, 2015.
[68] K. Singla, Z. Chen, N. Flemotomos, J. Gibson, D. Can, D. Atkins, and
S. Narayanan, \Using prosodic and lexical information for learning utterance-
level behaviors in psychotherapy," in Proceedings of Interspeech, 2018.
[69] X. Glorot, A. Bordes, and Y. Bengio, \Domain adaptation for large-scale
sentiment classication: A deep learning approach," in Proceedings of the 28th
international conference on machine learning (ICML-11), 2011, pp. 513{520.
62

Abstract (if available)

Abstract Modeling human behavior in conversational interactions is a complex and challenging task. This dissertation posits that by employing modeling and learning techniques which better reflect the complexities of human interaction, behaviors can be better predicted and understood. We propose several machine learning approaches for modeling these complex and important phenomena. Specifically, we investigate machine learning paradigms including deep, attentional, multi-label, multi-task, and hierarchical learning models to achieve this aim. We discuss how these learning techniques reflect important aspects of understanding human interactions. ❧ Using behavioral coding in psychotherapy as an expository domain, we experimentally validate the proposed approaches. For this purpose, we describe two datasets of therapist-client therapy interactions, from the domains of motivational interviewing and cognitive behavior therapy, and the specified behavioral constructs of interest in each. We demonstrate that by incorporating and merging the proposed learning approaches we achieve increased performance in the behavioral coding task and discuss the insights gained from these approaches. We discuss the potential implications and applications of this work for automating and augmenting behavioral coding and understanding.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Human behavior understanding from language through unsupervised modeling

PDF

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

PDF

Automatic quantification and prediction of human subjective judgments in behavioral signal processing

PDF

Modeling expert assessment of empathy through multimodal signal cues

PDF

Interaction dynamics and coordination for behavioral analysis in dyadic conversations

PDF

Computational modeling of behavioral attributes in conversational dyadic interactions

PDF

Behavioral signal processing: computational approaches for modeling and quantifying interaction dynamics in dyadic human interactions

PDF

Statistical inference for dynamical, interacting multi-object systems with emphasis on human small group interactions

PDF

Spoken language processing in low resource scenarios with applications in automated behavioral coding

PDF

Multimodal and self-guided clustering approaches toward context aware speaker diarization

PDF

Toward robust affective learning from speech signals based on deep learning techniques

PDF

Speech recognition error modeling for robust speech processing and natural language understanding applications

PDF

Computational modeling of mental health therapy sessions

PDF

Fair Machine Learning for Human Behavior Understanding

PDF

Multimodal representation learning of affective behavior

PDF

A computational framework for diversity in ensembles of humans and machine systems

PDF

Learning multi-annotator subjective label embeddings

PDF

Learning shared subspaces across multiple views and modalities

PDF

Towards social virtual listeners: computational models of human nonverbal behaviors

PDF

Improving modeling of human experience and behavior: methodologies for enhancing the quality of human-produced data and annotations of subjective constructs

Asset Metadata

Creator Gibson, James Joseph (author)

Core Title Machine learning paradigms for behavioral coding

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 08/01/2018

Defense Date 06/12/2018

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag human behavioral signal processing,machine learning,natural language processing,OAI-PMH Harvest,psychotherapy,speech processing

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Narayanan, Shrikanth (committee chair), Georgiou, Panayiotis (committee member), Margolin, Gayla (committee member)

Creator Email jimmygibson3@gmail.com,jjgibson@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-45118

Unique identifier UC11671396

Identifier etd-GibsonJame-6595.pdf (filename),usctheses-c89-45118 (legacy record id)

Legacy Identifier etd-GibsonJame-6595.pdf

Dmrecord 45118

Document Type Dissertation

Format application/pdf (imt)

Rights Gibson, James Joseph

Type texts

Source 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 a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA