University of Southern California Dissertations and Theses

Considering the effects of disfluent speech on children’s sentence processing capabilities and language development

(USC Thesis Other)

Considering the effects of disfluent speech on children’s sentence processing capabilities and language development

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Considering the effects of disfluent speech on children’s sentence processing
capabilities and language development
by
Cindy Chiang
A Dissertation Presented to the
Faculty of the USC Graduate School
University of Southern California
In Partial Fulfillment of the
Requirements for the Degree
Doctor of Philosophy
(PSYCHOLOGY)
August 2022
Copyright 2022 Cindy Chiang
To my family and family friends, at least three generations have shaped this dissertation. All of
your tall tales and generous words have made this possible!
ii
Acknowledgments
First and foremost, I would like to thank my advisor— Toby Mintz. This dissertation would not
have been possible without his mentorship. He has been incredibly supportive throughout my time
at USC. One notable example was when I mentioned that I would like to pursue an eye-tracking
project before I graduated, and he replied “We should make it happen.” As the the content of this
dissertation shows, it did happen.
Thank you to Elsi Kaiser, Frank Manis and Jason Zevin for having a hand in shaping the
content of this dissertation. The concepts driving this dissertation have been directly influenced by
the incredible classes and interactions I have had with you all. Elsi also kindly lent me the eye-
tracker that started the sequence of projects in this dissertation. I also owe an intellectual debt to
Jill de Villiers, who unintentionally introduced me to the body of literature that grounds this work.
A combination of the first BUCLD talks I attended with Jill and the work of a fellow Smithie who
did her honors thesis with Jill were the building blocks for Chapter 3.
This research would not have been possible without the hundreds children and parents who
participated my experiments throughout my time at USC, the kindness of the Twyla Ponton, and
the 17 research assistants who I have had the immense privilege of working with. I would also like
to acknowledge the support of a NSF Doctoral Dissertation Research Improvement Grant (NSF
#BCS-2041372), which funded the work described in this dissertation.
Thank you to my friends in psychology, linguistics and outside of academia— especially to
Katie Kim, Crystal Wang, Peter Wang, Maddy Jalbert, Yijing Lu, Helen Shiyang Lu, my office
neighbors, the duck gang, and the members of the Psycholinguistics Lab. You all have been the
best friends, the best colleagues, and the best people to grab food with.
Finally, thank you to my family for supporting me even when I could not fully explain what
I was doing or why. Thank you to my nieces and nephews— my first pilot subjects. Most of all,
thank you mom, dad, Angela, Tim, Natalie, May, Jenkin, aunts, and cousins. Words cannot express
how much your support means to me.
iii
Table of Contents
Dedication ii
Acknowledgments iii
List of Tables vi
List of Figures vii
Abstract ix
1 Chapter 1: Introduction 1
1.1 Linguistic and contextual cues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Overview of dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Chapter 2: Disfluencies activate alternative argument structures in toddlers during
online sentence processing 7
2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 Chapter 3: Uh what did you say? Children’s parsing preferences are altered by expe-
rience with disfluent sentences 32
3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
iv
3.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4 Chapter 4: Initial steps to incorporate disfluencies into a modelling framework 55
4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Extending model to simulate how children process ditransitives . . . . . . . . . . . 60
4.3 Future extensions: Training the model on naturalistic data . . . . . . . . . . . . . . 63
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Chapter 5: General Discussion 67
Bibliography 70
6 Appendix 80
6.1 Items and Proportion of 30 month olds on WordBank reported to produce the item
label in Study 1 and 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2 Estimated verb bias and frequency of verbs used in the experimental stimuli in
study 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.3 Estimated verb bias of verbs used in the experimental stimuli in Study 2 . . . . . . 84
6.4 Items in the experimental stimuli and Proportion of 30 month olds on WordBank
reported to produce the item label . . . . . . . . . . . . . . . . . . . . . . . . . . 85
v
List of Tables
2.1 Demographic information for the children in each condition in Study 1 . . . . . . . 11
3.1 Number of children in each condition in Study 2 . . . . . . . . . . . . . . . . . . . 36
3.2 Median age of participants in months in Study 2 . . . . . . . . . . . . . . . . . . . 37
4.1 Examples of chunk configuration for lexical items in a prepositional dative in an
ACT-R model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
vi
List of Figures
2.1 Schematic of a sample trial in Study 1 . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Proportions of looks to the N1 in all trials in Study 1. Lined regions represent
± 1 SE. A proportion of .25 indicates chance performance as indicated by the
horizontal dotted line. Black vertical line indicates offset of noise and onset of N1 . 19
2.3 Proportions of looks to the N2 in all trials in Study 1. Lined regions represent
± 1 SE. A proportion of .25 indicates chance performance as indicated by the
horizontal dotted line. Black vertical line indicates offset of N1 and earliest onset
of N2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Proportions of looks to the depiction of the PD in all trials by block and verb fre-
quency in Study 1. Lined regions represent± 1 SE. A proportion of .25 indicates
chance performance as indicated by the horizontal dotted line . . . . . . . . . . . . 21
2.5 Proportions of looks to the depiction of the PD in all trials by block in Study 1.
Lined regions represent± 1 SE. A proportion of .25 indicates chance performance.
Shaded region denotes significant cluster. . . . . . . . . . . . . . . . . . . . . . . 22
2.6 Proportions of looks to the depiction of the N1 in all trials by block and verb fre-
quency in Study 1. Lined regions represent± 1 SE. A proportion of .25 indicates
chance performance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Proportions of looks to the depiction of the N1 in all trials by verb frequency in
Study 1. Lined regions represent± 1 SE. A proportion of .25 indicates chance
performance. Shaded region denotes significant cluster . . . . . . . . . . . . . . . 25
3.1 Schematic of a critical trial in Study 2 . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Proportions of looks to the prime target in all trials by age and speaker type in
Study 2. Lined regions represent± 1 SE. Shaded region denotes significant cluster. 43
vii
3.3 Proportions of looks to the prime target in the last 3 trials by age and speaker type
in Study 2. Lined regions represent ± 1 SE. Shaded region denotes significant
cluster. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.1 Example of trial presentation in the web based version of the 3dccs conducted on
Lookit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2 Children’s flexibility score by age (3DCSS scores from Lookit study) . . . . . . . 64
viii
Abstract
Disfluencies, such as uh’s and um’s, frequently occur in everyday speech. Prior work estimates that
there are at least six disfluencies that occur for every 100 words and demonstrates that disfluencies
can impact how well adults process sentences and remember their content. Yet, how disfluencies
affect children’s understanding of sentences in the moment and their language development has
not been investigated. In this dissertation, I address this gap through two experiments. In chapter
2, I examine how toddlers’ ability to process sentences in the moment is impacted by disfluent
speech. Using an in-person eye-tracking study, I demonstrate that toddlers anticipate the more
accommodating sentence structures as they hear disfluencies— such as Jill gave the cracker to the
duck rather than Jill gave the duck the cracker. The types of structures they consider also appear
to impact how toddlers subsequently process sentences. These effects are modulated, as a result,
by their verb knowledge (what argument structures are available) and their recent experience. In
chapter 3, I investigate how brief experiences with disfluent speech impact children’s, three-year-
old and five-year-olds, predictions about upcoming sentences. I find that children’s experiences
with disfluent speech reduce children’s preferences for the structures they have recently heard.
I discuss these findings’ impacts on current theories of children’s syntactic development, which
depend on children’s preference for recently heard structures. In chapter 4, I discuss how a com-
putational model can be implemented to conduct a rigorous and comprehensive examination of
children’s syntactic development. Specifically, I discuss how coded corpora, findings from an on-
line study of children’s cognitive development, and extensions of the studies described in chapters
2 and 3 can inform this model. To summarize, across these chapters, I discuss the implication of
disfluent speech on current theories of children’s sentence processing and children’s syntactic de-
velopment. I also discuss the implications this work has for intervention for language disabilities,
such as specific language impairments (SLI).
ix
Chapter 1: Introduction
When children are listening to a speaker, they make predictions about what a speaker will say
next. These predictions are guided by a number of factors, including the probability that a verb
is associated with a particular structure — verb bias (Snedeker & Trueswell, 2004; Trueswell,
Sekerina, Hill, & Logrip, 1999). For example, following the phrase “Jill gave. . . , ” there are several
possible verb argument structures. Among other structures, a speaker could choose to produce are
double object (1) and prepositional datives (2). Since prepositional dative (PD) frames are more
likely to be linked to the verb threw, if a child is only considering verb bias, (2) is the structure
they are likely to anticipate following the phrase “Jill gave. . . ”.
(1) Jill gave the duck the cracker
(2) Jill gave the cracker to the duck.
However, while children heavily rely on verb bias to inform their predictions, verb argument
structures do not always follow verb bias. For example, with ditransitive verbs that can alternate
between both the DO and PD frame (e.g. gave), discourse factors — givenness, animacy, heaviness
— and intended meaning can influence adults’ and consequently caregivers’ use of each frame
(K. M. Snyder, 2003; Tily et al., 2009). When these instances occur, children may need to revise
their initial predictions. This is a difficult task for children (Weighall, 2008), and this difficulty has
widespread implications.
In the moment, an inability to revise predictions can create unintended parses of the sentence
(Snedeker, Worek, & Shafto, 2009). Children can parse sentences in ways that result in an unin-
tended interpretation of a sentence’s meaning. This can be problematic with ditransitive structures.
DO and PD frames are proposed to entail different verb meanings (Hovav & Levin, 2008). Thus,
an unrevised prediction may lead children to arrive at an unintended meaning in this context as
well.
This difficulty with revising predictions also has cascading effects for children’s word learning
and syntactic development. Prior work indicates that children experience difficulty with learning
1
word meanings when they need to revise their predictions about a sentences structure (Y. T. Huang
& Arnold, 2016; Lidz, White, & Baier, 2017). Moreover, the need to revise predictions is impli-
cated in the children’s delayed acquisition of passives and difficulty with morphology that comes
online later in a sentence (Y. T. Huang, Zheng, Meng, & Snedeker, 2013; Trueswell, Kaufman,
Hafri, & Lidz, 2012).
Children’s errorful predictions is also proposed to have consequences for theories of language
learning, particularly their understanding of verb bias. In error-based learning proposals, error
signals or mismatches between children’s initial parses and the actual structure they hear leads
to learning (Chang, Dell, & Bock, 2006; Lin & Fisher, 2017; Peter, Chang, Pine, Blything, &
Rowland, 2015). In these proposals, stronger error signals lead to stronger learning effects. As
a result, children’s errorful initial parses may also have a significant role in shaping their future
parsing preferences.
In sum, children’s errorful predictions can lead to difficulty for parsing in the moment and have
lasting consequences— both with hindering children’s word learning and syntactic development as
well as influencing children’s understanding of verb bias. This dissertation aims to assess whether
a previously unexplored cue, disfluencies ( thee uh), can be used by children to circumvent errorful
initial predictions (Chapter 2), whether use of this cue influences children’s parsing preferences
beyond the sentences they are embedded in (Chapter 3), and how a cognitive model can be imple-
mented to better explore how use of this cue interacts with various cognitive mechanisms (Chapter
4). In the subsequent sections in this chapter, I will discuss the linguistic and para-linguistic cues
that have been proposed to facilitate children’s parses of ambiguous sentences that can require re-
vision of initial predictions. I will then discuss disfluencies and the studies used to explore their
utility.
1.1 Linguistic and contextual cues
Several studies have demonstrated that children are often unable to revise predictions even when
linguistic and contextual cues, such as referential cues, prosodic cues, and plausibility, make al-
ternative structures more likely (Y. T. Huang et al., 2013; E. Kidd, Stewart, & Serratrice, 2011;
2
Snedeker & Trueswell, 2004; Snedeker et al., 2009; Trueswell et al., 1999). These studies have
examined children’s parses through eye tracking procedures in which children’s eye gaze towards
2x2 visual arrays containing representations of possible verb arguments as they listen to sentences
(e.g. visual world paradigm) were tracked as well as act out tasks. Sentences with prepositional
phrases (PP) that allow both VP and NP attachment were examined, such as Tickle the frog with
the feather. In this example, the VP attachment or instrument interpretation is the act of tickling
using the feather, while NP attachment or modifier interpretations is the frog with the feather.
In one such study, Snedeker and Trueswell (2004), manipulated the referential context and
verb bias. The referential context was manipulated through the visual scene which contained
either one or two possible referents (one frog holding a feather vs. one frog holding a candle
and one frog holding a feather). Adult participants relied on both referential context and verb bias,
preferring modifier interpretations when seeing two possible referents and using verb bias to reach
final parses. In contrast, five-year-olds did not show different preferences based on the referential
context. Instead, children only relied on verb bias. Children’s inability to use referential context
were attributed to the lower reliability of the referential context relative to verb bias (Trueswell,
Papafragou, & Choi, 2011).
However, children also do not appear to reliably anticipate alternate verb argument structures
in the presence of more reliable cues, such as intonational phrase (IP) breaks and plausibility.
Using similar methods as Snedeker and Trueswell (2004), Snedeker et al. (2009) manipulated
the plausibility of PP as an instrument modifier (ex: Tickle the frog with the candle). Five-year-
olds almost exclusively used verb bias, even when parses led to highly unlikely interpretations
(e.g. using a candle to tickle a frog). Similarly, manipulations of prosodic cues do not reliably
shift children’s parsing preferences. Snedeker and Yuan (2008) examined the effects of IP breaks
favoring instrument or modifier biases on children’s ability to revise predictions ( You can pinch the
frog...with the barrette. vs. You can pinch....the frog with the barrette). Children, as young as four-
years-old, could use intonational phrase. However, use of IP breaks faltered when children heard
instrument biased IP breaks first and heard modifier biased IP breaks afterwards. The inefficacy
of these cues were attributed to top-down nature of plausibility, children’s difficulties with quickly
integrating top-down information, the fragility of IP breaks as a cue, and an inability to inhibit
3
priming effects of prior cues.
A cue that children could potentially use to systematically anticipate non-standard or unex-
pected sentence structures is disfluencies. Disfluencies, speech disruptions produced by speakers,
are highly frequent and reliable in adult spontaneous speech. Critically, filled pauses (thee uh), the
most frequent disfluency (Bortfeld, Leon, Bloom, Schober, & Brennan, 2001; Fox Tree, 2001) ori-
ent adults’ parses towards otherwise less likely words or structures (Arnold, Tanenhaus, Altmann,
& Fagnano, 2004; Bailey & Ferreira, 2003) when other noises, such as construction noises and
beeps, do not (Arnold, Kam, & Tanenhaus, 2007).
Moreover, toddlers are sensitive to disfluencies (Soderstrom & Morgan, 2007) and can capi-
talize on their presence to infer a speaker’s referential intents. Children, as young as two years
and four months old, anticipate labels for infrequent or novel objects after filled pauses (C. Kidd,
White, & Aslin, 2011; Thacker, Chambers, & Graham, 2018a). Toddlers’ use of this cue also indi-
cates that its use may not be subject to the same cognitive constraints that restrict older children’s
use of IP breaks and plausibility as cues for alternative parses. Children also hear disfluencies
at increasing rates as the complexity of their input increases and there are more opportunities to
encounter parses that go against children’s initial predictions (C. Kidd et al., 2011). Additionally,
when speakers produce a less expected ditransitive structure, the structures also tend to contain
disfluencies and longer word durations (Tily et al., 2009). Thus, disfluencies may also be a more
reliable and plentiful cue in children’s input than cues that have been investigated in prior research.
In summary, prior research has been unable to identify a cue that is reliably used to help chil-
dren revise initial parses. This poses a significant issue. Verb arguments do not always follow
verb bias and inaccurate parses can result in misinterpretations of a speaker’s intended message.
Children’s inability to use cues have been associated with reliability of the cue, problems with
successfully inhibiting priming effects, and an inability to quickly integrate top-down information
(Snedeker & Huang, 2015; Trueswell & Gleitman, 2007). However, an unexplored cue, disfluen-
cies, could alert children to errorful initial parses. The cue is highly reliable, and children’s early
use of this cue in other domains suggests that its use may not be restricted by slow integration
times of top-down information nor be dependent on more developed cognitive capabilities. If chil-
dren are able to capitalize on this cue, further investigations is needed to understand the long term
4
implications of integrating this cue and how this process affects children’s language development
as well as current language learning proposals.
1.2 Overview of dissertation
In this dissertation, I explore the effects that disfluencies can have on children’s parses in the
moment and on the types of structures they subsequently encounter. I also discuss a cognitive
model that can be used to map children’s language development and investigate interactions be-
tween language specific and cognitive mechanisms. Specifically, in chapter 2, I examine whether
toddlers, two and three-year-olds, anticipate alternative verb argument structures following disflu-
encies. I assess the types of ditransitive structures (DO or PDs) that toddlers anticipate following
disfluencies and an everyday noise (control noise) by tracking toddlers’ looks to depictions of the
post-verbal nouns that correspond to DO and PD frames. I find that toddlers use disfluencies as a
cue to anticipate alternative argument structures. This use of disfluencies appeared to vary based
several factors including their familiarity with the verb and recent experience with ditransitive sen-
tences. These findings have implications for theories of sentence processing, as disfluencies may
be influencing children’s interpretations and parses of sentences in the moment.
I then investigate how this use of disfluencies impacts children’s parsing preferences in the
sentences they subsequently encounter in chapter 3. I examine the types of ditransitive structures
three and five-year-olds prefer after brief experiences with the same ditransitive structures embed-
ded with either a disfluency or an everyday noise. I assess the types of structures that these children
prefer by tracking their looks to depictions of the post-verbal nouns that correspond to DO and PD
frames after they have had experience with sentences with disfluencies and an everyday noise.
Overall, I find that children’s parsing preferences differ based on the type of noise embedded in the
sentences they have recently heard. The nature of these preferences differed by children’s age and
the types of speakers they heard indicating that children’s use of disfluencies may be modulated
by children’s general cognitive and social development. While these findings have implications
for theories of children’s syntactic development, further work is needed to understand how these
developmental factors interact with children’s use of disfluencies.
5
In chapter 4, I discuss the steps that can be taken to implement an extension of a cognitive
architecture model. This cognitive model can be used to investigate fine-grained interactions be-
tween language specific and general cognitive mechanisms. I, specifically discuss how this model
can be used to map children’s developing understanding of ditransitives and validated against data
from extensions of the studies in chapters 2 and 3, as well as the findings of a cognitive task that
I carried out with three and five-year-old children. I will also provide an overview of a corpus
that I have coded. This corpus will allow for a large scale implementation of the model, once the
cognitive model has been validated.
Lastly, in chapter 5, I present an overall discussion, and I also discuss possible future directions
for research on disfluencies.
6
Chapter 2: Disfluencies activate alternative argument structures in toddlers during
online sentence processing
2.1 Abstract
Children rely on a number of factors—such as recent experience and verb knowledge—to make
predictions about how a sentence will progress. However, speakers do not always produce sen-
tences that match children’s expectations. When this occurs, children need to revise their predic-
tions to match a speaker’s intended message. Prior studies have shown that this is difficult for
children under eight. I conducted an eye-tracking study to investigate whether the presence of a
disfluency— thee uhh—in an utterance can be used by 2 and 3-year-olds to pre-emptively anticipate
alternative structures and circumvent the need to revise predictions. Participants listened to ditran-
sitive sentences—Jill gave the cracker to the duck—with either an everyday noise or a disfluency
following the verb. I measured looks towards possible verb arguments during the everyday noises
and disfluencies, as well as after the onset of the first post-verbal noun. The looks of those who
heard construction noises differed from those who heard disfluencies. These looking preferences
appeared to shift based on children’s recent experience and verb knowledge. These findings indi-
cate that children as young as 2 and 3 years-old treat disfluencies differently from everyday noises
and disfluencies influence the argument structures they consider. I discuss possible implications
for the acquisition of verb argument structure.
2.2 Introduction
To arrive at a sentence’s intended meaning, children incrementally evaluate input and predict possi-
ble parses as the sentence progresses. This process requires children to keep in mind several options
and consider verb bias, the likelihood a verb occurs with various sentence structures (Snedeker &
Trueswell, 2004; Trueswell et al., 1999). For example, an utterance following the phrase ”the po-
liceman threw, ” could be continued with one of there are several alternatives arguments structures,
including a double object (DO) frame, The policeman threw the duck the cracker, and prepositional
dative (PD) frames— The policeman threw the cracker to the duck. Since threw is more likely to
7
occur in the PD frame over the DO frame, a listener would predict the PD frame if only considering
verb bias.
However, verb argument structures do not always follow verb biases. Speakers may choose to
use structures that go against verb bias to better convey meaning. For example, it has been proposed
that threw entails different event schemas in DO and PD frames. Threw entails successful transfer
of possession in the DO frame and ballistic motion in the PD frame (Green, 1981; Hovav & Levin,
2008; Levin, 2008; Oehrle, 1976). Thus, speakers may choose to produce the DO frame to better
suit a message’s intended meaning even when PD frames are generally preferred.
These instances can create immense difficulty for children under eight (Weighall, 2008). Chil-
dren are often unable to revise predictions that adhere to a verb’s bias, even when other factors,
such as referential context, make alternative structures more likely (E. Kidd et al., 2011; Snedeker
& Trueswell, 2004; Trueswell et al., 1999). This difficulty with revising predictions can negatively
impact syntactic development and how word meanings are learned (Y. T. Huang & Arnold, 2016;
Y. T. Huang et al., 2013; Lidz et al., 2017; Pozzan & Trueswell, 2015). The presence of cues, such
as intonational phrase breaks and plausibility, can alleviate some of the difficulties associated with
revising initial predictions (Snedeker et al., 2009; Snedeker & Yuan, 2008). However, children’s
ability to capitalize on these cues is limited by their difficulties with inhibiting initial predictions
and with rapidly integrating these cues. The reliability and the frequency of the cues also limit
their use (Snedeker & Huang, 2015).
Evidence suggests that a different cue — the disfluency thee uh— can be used by a younger age
group to circumvent some of the difficulties associated with revising initial predictions. This cue
is highly frequent in adult spontaneous speech (Bortfeld et al., 2001). Additionally, corpus studies
and experimental work suggest that it provides a reliable and meaningful indication of speech plan-
ning difficulties in adults. Speakers tend to produce the disfluency more frequently with infrequent
words, discourse new information, and low probability structures (Beattie & Butterworth, 1979;
Tily et al., 2009). Correspondingly, listeners can anticipate new information and infrequent words
after hearing a disfluency (Arnold et al., 2004; Barr & Seyfeddinipur, 2010; Corley, MacGregor,
& Donaldson, 2007). These effects do not hold when listeners hear other noises, such as construc-
tion noises and beeps, instead of disfluencies (Arnold et al., 2007). That is, typical predictions
8
are observed after the course of construction noises, while alternative predictions are observed af-
ter the course of a disfluency. Listeners will also shift parsing strategies following disruptions to
the speech stream, including after the disfluency thee uh, anticipating different syntactic structures.
Use of this cue, as a result, has been hypothesized to provide a way to resolve syntactic ambiguities
(Bailey & Ferreira, 2003).
Children as young as twenty-two months are also sensitive to this cue (Soderstrom & Morgan,
2007). Moreover, children as young as twenty-eight-month-olds can use the disfluency thee uh as
an indication of speakers’ referential intentions. Using an eye-tracking procedure, C. Kidd et al.
(2011) demonstrated that toddlers can use disfluencies as a cue for novel labels. At the start of
experiment, toddlers were visually presented with both a novel and familiar object while auditory
stimuli introduced and described the familiar object. Then, toddlers were instructed to look towards
an object. Half were fluent instructions, while the other half contained a two second disfluency.
Children as young as two years and four months old tended to look towards the unfamiliar and
previously unmentioned object during the disfluency, in anticipation that it would be labeled. This
early use of disfluencies indicates that its use may not be subject to the same cognitive constraints
that restrict older children’s use of other cues. This, in turn, raises the possibility that toddlers may
be able to use disfluencies as a cue to otherwise unexpected argument structures, providing a way
for children to identify verbs that can take on alternative argument structures and avoid revising
their initial predictions.
2.2.1 Disfluencies As a Cue For Alternative Ditransitive Argument Structures: The Current
Study
Despite the potential utility of disfluencies, whether children can use disfluencies as an indicator for
non-standard or otherwise low probability verb arguments has not yet been empirically tested. In
the current study, I investigate whether the cue can be used by two and three-year-olds to preemp-
tively anticipate alternative verb argument structures following ditransitive verbs. To investigate
this question, participants’ looks towards the referents of possible verb arguments were measured
as they heard either a baseline noise (construction noise) or the disfluency, thee uh, following the
verbs. I also investigate how this differs as a product of children’s verb knowledge.
9
I focus on the verb argument structure of ditransitive verbs for several reasons. First, many
ditransitive verbs in English can undergo dative alternation, occurring in both DO and PD frames.
While the elements involved are identical, the ordering of the object arguments varies on the surface
level after the verb. The recipient of the action or the indirect object is the first post-verbal argument
in DO structures and the second post-verbal argument in PD structures. The theme of the verb or
the direct object is the second post-verbal argument in DO structures and first post-verbal argument
in PD structures. Thus, parsing preferences for each of the structures can be measured following
the verb while keeping visual stimuli consistent.
Second, whether an adult speaker produces a DO or PD frame is highly probabilistic. Factors
including verb bias as well as the complexity, discourse status, and animacy of the verb arguments
influence the probability each frame is produced (Bresnan, Cueni, Nikitina, & Baayen, 2007).
Third, when adult speakers do produce low probability ditransitive frames, they are more likely
to produce disfluencies surrounding the verb or lengthen the pronunciation of the surrounding
words. Using the same fine-grained measures as Bresnan et al. (2007), Tily et al. (2009) found that
disfluencies are likely to occur in low probability DO and PD frames. The duration of words in
these types of structures are also more likely to be lengthened.
Fourth, evidence suggests that children are influenced by the same factors that guide adult’s
productions of ditransitive sentences (De Marneffe, Grimm, Arnon, Kirby, & Bresnan, 2012;
K. M. Snyder, 2003). Adults are influenced by the verb bias of ditransitive verbs and discourse
factors, such as prior mentions of the verb arguments and the length of the verb arguments (Tily
et al., 2009). However, children’s preferences appear to depend on their verb knowledge. With
newly learned verbs, children generally prefer PD frames. DO frames are more restrictive than PD
frames with the verbs that it takes on. PD frames are, thus, the more conservative choice. This
preference is consistently observed in both comprehension and production tasks carried out with
children under the age of four (Arunachalam, 2017; Conwell & Demuth, 2007; Rowland, Noble,
& Chan, 2014; Rowland & Noble, 2010).
Finally and critically, two and three-year-olds have knowledge of ditransitive frames (Gropen,
Pinker, Hollander, Goldberg, & Wilson, 1989; W. Snyder & Stromswold, 1997) but have varying
states of verb knowledge. This corresponds to the time period when children are starting to capi-
10
talize on the disfluencies in other domains. Thus, ditransitive frames can provide a way to test how
disfluencies impact children’s parsing preferences both when their knowledge of verb argument
structures are more stable and when it is still growing. Ditransitive frames also allow us to test the
effect of disfluencies on children’s parsing preferences at the earliest time point possible.
2.3 Methods
2.3.1 Participants
Forty-eight typically developing two and three-year-olds, who received primarily English input at
home, were recruited for the study (2;4-4;0, mean 36.8 months). A total of fifteen participants
were excluded from the final data analysis for squirminess (13), experimenter error (1), and high
number of trackloss points during the critical point of the trial (1).
Thirty-three children were included in the final analysis (mean age in months = 37.45; range
= 28-48 months). Sixteen participants heard disfluencies following verbs in dative constructions
(Disfluency Condition), and seventeen heard a construction noise following verbs in dative con-
structions (Control Condition). The demographics of the participants are shown in (Table 2.1).
Table 2.1: Demographic information for the children in each condition in Study 1
Condition
Age in Months Sex (N)
Mean SD F M
Construction Noise 37.23 4.87 10 7
Disfluency Noise 37.68 5.59 11 5
2.3.2 Stimuli and Design
The experiment consisted of twelve trials. Each trial had an introduction and test phase that
centered around the four different cartoon depictions. In total, children saw forty-eight images
11
throughout the trials. All but four images were depictions of nouns taken from the MCDI II (Fen-
son, Marchman, Thal, Dale, & Reznick, 2007). By-word summary data from Wordbank (Frank,
Mika, Daniel, & Marchman Virginia, 2016) indicated that the nouns taken from the MCDI II were
produced by a high proportion of American English speaking 30 month-olds. All nouns were also
highly frequent in child directed speech (see Appendix Table 6.1). All the audio children heard in
this experiment were audio recordings from a female native English speaker.
The critical sentences in these trials were ditransitive sentences. Throughout the experiment,
the critical sentence alternated between DO and PD frames to reduce potential structural priming
effects, the increased tendency to anticipate a sentence structure after recent experience with the
structure. The verbs in the critical sentences were always followed by a disfluency or construction
noise (e.g. The policeman threw thee uhh duck the cracker vs. The policeman threw the [construc-
tion noise] duck the cracker.) and varied trial to trial. Both the disfluency and construction noises
were two seconds long. This was the same length of the disfluency that children heard in a prior
study investigating children’s use of disfluencies in novel noun labeling task (C. Kidd et al., 2011).
Ditransitive verbs in this study varied by the frequency of the verb in a DO frame compared to
a PD frame (verb bias) and the general frequency of the verb. Verb biases and general frequencies
were drawn from a corpus study conducted by Gropen et al. (1989) as well as from the Brown
(1973) and MacWhinney (1991) corpora in the CHILDES database (MacWhinney, 2000). For my
manipulation of verb bias, I included verbs with high bias towards DO frames, verbs with equi-bias
towards both frames, and verbs with low bias towards DO frames. For my manipulation of general
verb frequency, I included both high frequency and low frequency ditransitive verbs. The general
frequency served as a rough approximation of how familiar children may be with the verbs. The
children in this study were likely to be less familiar with low frequency verbs, such as paid (see
Appendix Table 6.2 for the full list of verbs).
The presence of a disfluency or construction noise after verbs in a dative frame was a between-
subjects factor. Roughly half the participants heard construction noises following the ditransitive
verbs, while the other participants heard disfluencies. The manipulations of verb bias and the
general frequency of the verb were within subjects factors. The experiment was presented in
blocks, such that subjects saw all variations of the item types in the first half (Block 1) and in the
12
second half of the experiment (Block 2).
2.3.3 Procedure
Parents were asked to provide consent for their child’s participation prior to the start of the study.
Verbal assent from the child was also obtained at this time. During the main experiment, partic-
ipants sat in a sound attenuating booth on their parent’s lap approximately 70 centimeters away
from a Tobii T120 eye-tracker. Parents wore headphones to mask the audio recording and dark-
ened sunglasses to prevent viewing of the visual stimuli. Stimuli were presented on the eye-tracker
through Pygaze, an open source software package (Dalmaijer, Mathˆ ot, & Van der Stigchel, 2014).
Participants’ eye movements were sampled approximately every 17 milliseconds throughout the
experiment.
Before the main experiment began, the Pygaze software was used to calibrate the eyetracker.
Children were asked to look at a cartoon depiction of a happy face, which appeared in the four
corners and the center of the screen. At the start of each experimental trial, a happy face was
displayed in the center of the screen for 500 milliseconds, to orient children’s gaze towards the
neutral center position. Children were then presented with a 2x2 visual array. The array included
cartoon depictions of the agent of a ditransitive sentence (e.g. policeman), the theme of a ditransi-
tive sentence (e.g. cracker), the patient of the ditransitive sentence (e.g. the duck), and a distractor
object (e.g. tree). All possible verb arguments were limited to involve an inanimate indirect object
or theme (e.g. cracker) and an animate direct object or recipient (e.g. duck) through the visual dis-
play. Each of the images were introduced with a spoken introduction (e.g. This is a policeman.), as
each image appeared on the screen. The location of the images on the screen were counterbalanced
across participants and across trials, ensuring that looks towards targets were not confounded with
a particular location of the screen.
After the four objects were introduced, the center fixation happy face appeared again for 500
milliseconds to orient looks to the center, neutral position, while the four images remained on the
screen. Participants subsequently heard two filler sentences, containing transitive or intransitive
verbs, and the critical sentence. These filler sentences preceded or followed the critical sentence
13
(See Figure 2.1).
Figure 2.1: Schematic of a sample trial in Study 1
After the main experiment, parents with children from twenty-eight to thirty months were
asked to fill out the MacArthur-Bates Communicative Development Inventory, Level 2 Words and
Sentences form, while parents with children from thirty to thirty-seven months were asked to fill
out the MacArthur-Bates Communicative Development Inventory, Level 3 form (Dale, Reznick, &
Thal, 1998; Fenson et al., 2000). Both forms have vocabulary checklists. The Level 2 Words and
Sentences form also contains additional questions about children’s morphosyntactic development,
while the Level 3 form contains additional questions about the types of sentences children are
producing.
2.3.4 Window and measures of interest
In the analysis described in later sections, I focused on a shifted time window. Children in the
study’s age range have been estimated to need 270 milliseconds react and look towards stimuli in
visual world paradigm (Canfield et al., 1997). As a result, I shifted my window of analysis by 250
milliseconds. This follows the length of time that C. Kidd et al. (2011) shifted their window of
analysis in a study examining the effect of disfluencies on novel noun labeling with a similar age
14
group. I further shifted my time frame by 50 milliseconds to allow children to process any initial
phonological cues (i.e the start of the disfluencies and the initial segments of the N1), following
Fernald (2004).
To ensure that the participants were generally able to process the sentences included in the
experiment, I first examined overall looks towards the depictions of the post-verbal nouns (i.e.
N1 and N2) as the sentences unfolded. If participants were able to process the sentence, I would
expect above chance looks towards N1 and N2 after the speaker named them. The onset of N1
immediately followed the offset of the construction noise and disfluency. The duration of the N1
ranged from 331 milliseconds to 996 milliseconds. The onset and the offset of N2 varied. When
children heard DO frames, the onset of N2 immediately followed the offset of N1. When children
heard PD frames, the onset of N2 followed the offset of the preposition. The duration of the N2
similarly varied, ranging from 515 milliseconds to 1079 milliseconds. Since the duration of the
N1 and N2 varied, I examined children’s looks towards the N1 and N2 for 1000 milliseconds after
their respective onsets.
To examine whether children’s processing of the datives were differently impacted by the con-
struction noise or disfluency, I measured the proportion of looks towards all possible verb argu-
ments during the duration of the disfluencies and construction noises, lasting 2000 milliseconds.
This measure allowed us investigate whether disfluencies and construction noises impacted chil-
dren’s anticipation for different argument structures as they heard the sounds. Since the visual
display and the verb semantics limited the direct object (i.e. the inanimate target) to one cartoon
image, I examined anticipatory looks towards these images over the course of the disfluency and
construction noises in the shifted time frame. During this time, looks towards the inanimate target
reflected a parsing preference for the PD frame.
To examine whether this anticipation for different argument structures subsequently impacted
children’s sentence processing, I analyzed children’s looks towards the depiction of the N1 follow-
ing the disfluency and construction noise. In the study, the target sentence alternated between DO
and PD frames. As a result, the N1 was the inanimate target in half of the trials and the depiction
of the indirect object (i.e. the animate target) in the other half of the trials.
15
2.4 Predictions
Since prior work has indicated that children begin to produce DO and PD frames before the age of
three, I predicted that children in the current study would be able to process the sentences used here,
even as they heard the 2 seconds of sounds following the verb (Gropen et al., 1989). This would
be reflected in above chance looks towards the N1 and N2 shortly after their onset. I predicted that
the sounds would result in some variation both in the types of structures children anticipated as
they heard the sounds and how children processed the sentences.
Specifically, I hypothesized that participants’ anticipatory looks towards possible referents in
the current study would differ over the course of the construction noise and disfluency. Because
construction noises do not provide meaningful signals, I predicted that children’s looks in this
condition would provide a baseline of how children’s looks would vary when hearing a two second
disruption. With disfluencies, I predicted that children would treat disfluencies as a meaningful
signal and make predictions about the alternative types of structures that the speaker would produce
following the findings of C. Kidd et al. (2011). Thus, I should observe differences in children’s
looking preferences over the course of the 2 seconds of disfluencies and construction noises, as a
result.
Since prior work has indicated that disfluencies can also impact the types of words and sen-
tence structures that adults subsequently identify (Arnold et al., 2007; Bailey & Ferreira, 2003), I
also hypothesized that children’s parsing biases would differ following the sounds they heard. Fol-
lowing the construction noise, I predicted that children’s accurate identification of the N1 will be
high. This is the behavior I would expect, if children accurately process and understand the ditran-
sitives. In contrast, following the disfluency, I predicted that children’s accurate identification of
the N1 will be high but occur at a reduced rate. This is the behavior I would expect, if children are
generally able to accurately process and understand the ditransitives but are considering alternative
structures that can impact how they process the sentences.
16
Effects of verb knowledge and experience
I also expected that any preferences would vary by children’s verb knowledge and shift over the
course of the experiment. Children’s verb knowledge has been demonstrated to influence the types
of structures that children consider (Arunachalam, 2017; Fisher, 2002). Brief experiences with DO
and PD frames can shift children’s productions of ditransitive frames. This shift is especially prone
to occur when the brief experience differs from children’s initial expectations (Lin & Fisher, 2017;
Peter et al., 2015; Rowland, Chang, Ambridge, Pine, & Lieven, 2012). As a result, children’s
anticipation should vary in the first and half of the study (Block 1 and Block 2). The frames
that children anticipate in Block 2 should depend on how their strongly expectations at the start
of the experiment differed from the test sentences they heard in the first block of the study. The
influences of the first block could be quite complex, and hence, it is unclear what children’s precise
expectations for the second block will be.
Verb bias
I also initially expected that any preferences would vary by verb bias, because prior work indicates
that older children are influenced by verb bias in sentence processing tasks (Qi, Yuan, & Fisher,
2011; Snedeker & Trueswell, 2004; Trueswell et al., 1999). This is reflected in the stimuli set in
(Appendix 6.2)
However, since children’s recent experience (block number) and the verb knowledge were
hypothesized to impact both the types of predictions as well as their subsequent identification of
verb arguments, this left few trials of each verb type when accounting for the block and children’s
verb knowledge. Thus, I focused the analyses by block and verb knowledge in the subsequent
sections.
2.5 Results
To ensure that children were generally able to process the sentences. I first measured participants’
looks towards the N1 and N2 as the sentences unfolded. Participants identified both the N1 and
17
N2 above chance. Since there were four objects on the screen during each trial, a proportion of
25% would reflect looks at chance. Overall, children’s proportion of looks to the N1 was 51.24%
and the N2 was 44.26%. Figure 2.2 shows children’s looks towards the N1 following the offset
of the construction noise and disfluency, while Figure 2.3 shows children’s looks towards the N2
following the offset of the N1.
To examine the effects of disfluencies and construction noises over the course of their two sec-
ond duration, I analyzed children’s looks towards the inanimate target during this time window. I
then examined children’s identifications of the N1 over the course of the N1 duration to investi-
gate the influence of disfluencies and construction noise on how children processed the sentence.
During these two time windows, I analyzed the proportion of looks towards the target images (i.e.
inanimate target and N1) through cluster permutation analyses using the eyetrackingr package
(Dink & Ferguson, 2015) on R (R Core Team, 2020). This provided a conservative test of whether
looks towards the targets diverged for children in the control condition and the disfluency condition
over the window of interest (Y . Huang & Snedeker, 2020).
Participants with a high number of trackloss points during the critical trial were participants
with less than 25% of the trials. Valid trials with less than 25% of valid look during the window
of interest. Invalid looks were trackloss points, which included blinks and looks towards areas
outside the bounds of the eye tracking monitor. These points were marked by the eye tracker in the
data cleaning process. (32 trials were excluded for insufficient number of trials and 1 participant
was removed from the final analyses).
18
Figure 2.2: Proportions of looks to the N1 in all trials in Study 1. Lined regions represent± 1 SE.
A proportion of .25 indicates chance performance as indicated by the horizontal dotted line. Black
vertical line indicates offset of noise and onset of N1
Figure 2.3: Proportions of looks to the N2 in all trials in Study 1. Lined regions represent± 1 SE.
A proportion of .25 indicates chance performance as indicated by the horizontal dotted line. Black
vertical line indicates offset of N1 and earliest onset of N2
19
2.5.1 Anticipatory Looks
Visual inspection revealed that there were differences in children’s looking preferences in the first
and second half of the experiment. These preferences visually did not appear to vary by children’s
verb knowledge (Figure 2.4). In the first block, participants appeared to look towards the inanimate
target at chance rates as they heard the construction noise. Approximately, 25% of their looks
were towards the inanimate target. This differed when they heard disfluencies. Participants in this
condition looked towards the inanimate target at higher than chance rates. In the second block this
trend shifted. Participants appeared to look towards the inanimate target at chance, both when they
heard construction noises and disfluencies. To statistically assess these differences, I conducted
cluster permutation analyses on the data from Block 1 and Block 2.
Cluster permutation analyses consisted of two steps (see Maris and Oostenveld (2007) for de-
tails). In the first step, individual clusters were identified by grouping adjacent 50 millisecond time
bins with t-values greater than a pre-defined threshold, and the sum-statistics were obtained for
each cluster by summing the t-values associated with each time bin within the cluster. Follow-
ing prior work, I set the initial threshold to 1.5 (de Carvalho, Babineau, Trueswell, Waxman, &
Christophe, 2019; de Carvalho, Dautriche, Lin, & Christophe, 2017; Hahn, Snedeker, & Rabagliati,
2015).
In the second step, I examined the likelihood that these clusters occurred by chance. To do so, I
obtained a distribution of sum-statistics that I would expect to get by chance. This was achieved by
first randomly shuffling the category assignment—disfluency and construction noise–of each data
point, then running the same procedure as in the previous paragraph to identify the cluster with
the highest sum-statistic. This reshuffling and analysis procedure was carried out 1,000 times to
obtain a distribution of sum-statistics representing the chance distribution. I used that distribution
to obtain a p-value of my actual cluster sum-statistics.
20
Figure 2.4: Proportions of looks to the depiction of the PD in all trials by block and verb frequency
in Study 1. Lined regions represent± 1 SE. A proportion of .25 indicates chance performance as
indicated by the horizontal dotted line
21
Block 1
A cluster permutation analysis on Block 1 trials revealed that there was one significant cluster
between 600-1200 milliseconds (p = 0.05) after onset of the two seconds of sound proceeding the
verb and preceding the onset of the N1 (Figure 2.5). Children who heard the construction noise
appeared to look towards the inanimate target at chance throughout the duration of the noise, while
the children who heard disfluencies appeared to look towards the inanimate target at higher rates
during this time period.
Figure 2.5: Proportions of looks to the depiction of the PD in all trials by block in Study 1. Lined
regions represent± 1 SE. A proportion of .25 indicates chance performance. Shaded region de-
notes significant cluster.
Visual inspection revealed that these effects may be present both when children heard High
Frequency verbs and Low Frequency verbs (Figure 2.4). Separate cluster permutation analyses
with these trials did not support this observation. There was one marginally significant between
650 - 1100 milliseconds (p = 0.097) after onset of the two seconds of sound proceeding the verb and
preceding the onset of the N1 for children who heard High Frequency verbs. Children who heard
the construction noises following High frequency verbs appeared to look towards the inanimate
target at chance throughout the duration of the noise, while the children who heard disfluencies
appeared to look towards the inanimate target at higher rates during this time period. There were
22
no significant clusters for children who heard Low Frequency verbs. However, this difference
may be coming from the number of data points in this cross section of data and consequently
the statistical power associated with it. These results suggest that disfluencies may be leading
children to anticipate the PD frame at higher rates in the first block of the study. This is the more
accommodating ditransitive frame.
Block 2
A cluster permutation analysis on Block 2 trials did not reveal any significant clusters. Instead, the
proportion of children’s looks towards the inanimate target did not significantly differ based on the
type of sounds children heard following ditransitive verbs during the second half of the experiment.
Children in both conditions appeared to look towards the inanimate target at chance in the second
block (Figure 2.5).
2.5.2 N1 Region
Visual inspection revealed that, while children looked towards N1 above chance, there were gen-
erally differences in children’s looking preferences towards the N1 after they heard construction
noises and disfluencies (Figure 2.2). A cluster permutation analysis on all trials revealed that there
was one significant cluster between 300-1000 milliseconds (p = 0.004) after onset of the N1. Com-
pared to children who heard construction noise, children who heard disfluencies looked towards
the N1 at reduced rates.
Further visual inspection of the N1 region by verb frequency and block indicated that differ-
ences between the children’s looking preferences towards the N1 were particularly pronounced
following high frequency verbs (Figure 2.6). To statistically assess this difference, I conducted
cluster permutation analyses on trials with high frequency verbs and trials with low frequency
verbs.
23
Figure 2.6: Proportions of looks to the depiction of the N1 in all trials by block and verb frequency
in Study 1. Lined regions represent± 1 SE. A proportion of .25 indicates chance performance.
24
High Frequency Verbs
While visual inspection revealed that children who heard high frequency verbs looked towards
the N1 above chance during the window of interest, children who heard the construction noise
appeared to look more towards the N1 than the children who heard disfluencies (Figure 2.7). A
cluster permutation analysis on trials with high frequency verbs supported this observation. It
revealed that there was one significant cluster between 350-1000 milliseconds (p = 0.002) after
onset of the onset of the N1. This indicated that children who heard construction noises were
better able to identify the N1, following it’s onset, and that hearing disfluencies may have impacted
children’s sentence processing.
Figure 2.7: Proportions of looks to the depiction of the N1 in all trials by verb frequency in Study
1. Lined regions represent± 1 SE. A proportion of .25 indicates chance performance. Shaded
region denotes significant cluster
Visual inspection revealed that these effects were present in both blocks of the experiment
(Figure 2.6). Separate cluster permutation analyses on the cross sections of Block 1 and Block 2
data supported this observation. There was one significant cluster between 350-1000 milliseconds
(p = 0.006) after onset of the onset of the N1 in Block 1 and one marginally significant cluster
between 400-750 milliseconds (p = 0.074) after onset of the onset of the N1 in Block 2. This
indicated that children who heard construction noises looked towards the N1 more than the children
25
who disfluencies in both blocks of the experiment.
Low Frequency Verbs
Visual inspection revealed that children who heard low frequency verbs looked towards the N1
above chance during the window of interest. However, there did not appear to be a difference
between the looking preferences of children who heard construction noises and children who heard
disfluencies (Figure 2.7).
A cluster permutation analysis on trials with low frequency verbs supported this observation.
The analysis did not reveal any significant clusters. Children’s looks towards the N1 did not sig-
nificantly differ based on the type of sounds children heard following low frequency ditransitive
verbs.
2.6 Discussion
The findings suggest that two and three-year-olds make use of speakers’ disfluencies as a cue to
anticipate alternative argument structures, which could help them to circumvent some of the dif-
ficulties linked to revising predictions. There are overarching differences in what frames children
prefer when they hear a disfluency. Toddler’s use of disfluencies in this context brings up larger
questions about the types of alternative structures children anticipate when hearing disfluencies,
the effect of this cue on children’s language development, and the mechanisms driving the use of
this cue.
2.6.1 Considering alternatives
Prior work suggests that children’s parses of fluent ditransitive sentences are guided by their verb
knowledge and recent—even brief—experiences with ditransitive frames. With unfamiliar ditran-
sitive verbs, children prefer the more accommodating PD frame, as there are restrictions on the
type of verbs that can occur in DO frames (Gropen et al., 1989). It is unmarked and is less restric-
tive with the ditransitive verbs it can be linked to (Fisher, 2002; Naigles, 2002). When children
26
have brief experiences with ditransitives, these preferences can shift (Arunachalam, 2017; Lin &
Fisher, 2017; Peter et al., 2015; Rowland et al., 2012).
The findings from the current work suggest that the types of structures toddlers consider when
hearing disfluencies are similarly modulated by their recent experience and verb knowledge. This
is highlighted both by the types of structures toddlers appeared to anticipate in the first block of
the study when they heard disfluencies and in toddlers’ subsequent identification of the N1.
In the first block of the study, when toddlers heard disfluencies, they appeared to anticipate
the more conservative frame — the PD frame. This indicated that upon hearing the disfluency
toddlers may have defaulted to the more accommodating frame. An alternative factor that could
have contributed to this finding was verb bias. Toddlers’ preferences after hearing disfluencies
could be going against verb bias. In the current study, children heard an equal number of DO
biased verbs, PD biased verbs, and equi-biased verbs. Typically, children’s productions depend on
verb bias and an animacy preference (K. M. Snyder, 2003). Following, DO biased verbs and PD
biased verbs children tend to prefer DO and PD frames respectively. With equi-biased verbs, when
verb bias is not clear, children prefer animate verb arguments closer to the verb. In the current
study, this should create a preference for the DO frame, as two-thirds of the verbs in the study
skew children’s preferences towards the DO frame. Thus, one possibility for toddlers’ anticipation
of PD frames after hearing disfluencies in the current study could be a preference for frames that go
against the verb bias. However, a study with a larger sample size would be needed to clarify what
factors are at play. From the current analysis, it is not clear whether children generally prefer the
more accommodating frame after hearing a disfluency or if they prefer a frame that goes against
verb bias. Further analysis of the items by verb bias would not be informative without a larger
sample.
Toddler’s preferences for the PD frame as they heard disfluencies shifted over the course of the
experiment. In the second block, participants in the study no longer preferred different structures in
the disfluency and the control noise. This suggested that the types of frames that toddler’s preferred
as they heard disfluencies similarly shifted with brief experiences. Specifically, the strength of the
cue may be shifting over the course of the experiment.
Evidence to support this comes from toddlers’ looks towards the N1 following this sound. The
27
sounds that toddlers heard following the ditransitive verb impacted how they subsequently pro-
cessed DO and PD frames in both the first and second block of the study. Compared to toddlers
who heard construction noises, toddlers who heard disfluencies looked towards the N1 at lower
rates throughout the experiment. This difference appeared to hinge on the verb frequency— my
approximation for participants’ verb knowledge. Specifically, the effect was dependent on tod-
dlers’ looking preferences in trials with high frequency verbs or familiar verbs.
One possibility for this difference was toddlers’ ability to consider the different structures and
the processing cost that this consideration can incur. That is, if disfluencies cue children to consider
a number of different structures (e.g. PD and DO’s following ditransitive verbs), this consideration
may depend on their knowledge of what verb structure mappings are available. Without this verb
knowledge, children may be reluctant to make predictions about alternative ditransitive structures.
This possibility is supported by prior experimental work, which has demonstrated that without
sufficient experience children are reluctant to link DO structures to unfamiliar verbs (Arunachalam,
2017).
Collectively, these findings suggest that the structures children anticipate following disfluencies
depend on their verb knowledge and their recent experience. Further work is needed to understand
what effects disfluencies have on children’s sentence processing in a natural context where dis-
fluencies can provide a more reliably signal for otherwise low probability structures (Tily et al.,
2009). In the current study, disfluencies occurred with all of the sentences that children in the
disfluency condition heard. Thus, it is not clear if disfluencies create further processing difficulties
or can facilitate children’s sentence processing by providing an early cue.
2.6.2 Impact on language acquisition and syntactic development mechanisms
The current work also brings up several implications about the impact of disfluent speech on chil-
dren’s language development and on mechanisms of syntactic development. Most directly, dis-
fluencies may be able to serve as a broader signal that a verb can take on alternative argument
structures. This is supported by toddlers’ initial anticipation for PD frames as they heard disfluen-
cies in the first block of the study following both high and low frequency verbs. In the absence of
28
verb knowledge or when children have limited verb knowledge, this could be highly beneficial in
forming children’s knowledge of verb argument structure. This cue could be especially useful in
cases where verbs are highly biased in the input and when children have limited experience with a
verb.
Secondly, if children can use disfluencies to circumvent the need to revise predictions, this
cue could also have cascading effects in infant and young children’s ability to learn the mean-
ing of words and on children’s syntactic development. Revision has been found to limit infants’
and kindergarteners’ ability to map the meaning of novel nouns and to recall novel noun mean-
ings (Y. T. Huang & Arnold, 2016; Lidz et al., 2017). With adults, a need to revise predictions
can lead to delays in grammar learning (Pozzan & Trueswell, 2015). Although the current study
demonstrates that toddlers’ sentence processing can be disrupted by disfluencies, it is possible that
sentence processing can be facilitated in more natural contexts when disfluencies serve as a reliable
indicator of the types of structures speakers will produce.
The current results also hold implications for theories of syntactic development, in particular
with respect to theories positing that syntactic priming plays a role in developing verb argument
knowledge (Chang et al., 2006; Lin & Fisher, 2017; Peter et al., 2015). The changes in response
to disfluencies and construction noises that I observed between Blocks 1 and 2 may have been the
result of children being primed by the sentences in the first block of the experiment. Although this
interpretation of the block effect is only a speculation, results from Ziegler and Snedeker (2019),
using a similar paradigm and stimuli support this possibility. In that study, participants listened
to either PD or DO frames (e.g. priming sentences) before hearing test sentences. The priming
sentences increased the likelihood that children anticipated the structure of the priming sentence
when processing a test sentence. The observed block effects in the current study may have come
from similar priming effects. Moreover, it may be the case that disfluent sentences do not have
the same priming effect as sentences with construction noises. This possibility is supported by
ongoing work from priming studies investigating the effects of priming sentences with disfluencies
or construction noises on both children and adult listeners (Chiang & Mintz, 2020).
If children are also primed differently by disfluent sentences, this would have implications
for current theories of implicit learning, which only consider fluent speech. In theories of implicit
29
learning, the level of mismatch (error) between children’s initial expectations and the actual input is
used to shape children’s knowledge of verb argument structure. For example, if a child encounters
a low probability structure, such as a structure that does not adhere to verb bias, the child will
experience high error signals and high likelihood of verb bias learning. Low probability structures
are highly influential, as a result.
However, disfluencies usually occur in the same low probability structures. If children are sen-
sitive to this cue across sentences, this use would bring into question how often children actually
encounter sentence structures with high error signals. This would mean that low probability struc-
tures would be less influential in changing children’s verb biases and create some questions for the
proposed syntactic development mechanisms that depend on implicit learning.
Such an effect of disfluencies on priming could have widespread implications. Children are not
only primed by DO and PD frames, but are also primed by the use of many different structures,
including passives, prepositional phrases, and relative clauses (Chang et al., 2006; Havron et al.,
2020; Qi et al., 2011; Ryskin, Qi, Duff, & Brown-Schmidt, 2017; Savage, Lieven, Theakston, &
Tomasello, 2006). Future work is needed to more directly probe whether the effects I observe in
the current study are coming from priming effects, to clarify the extent to which there are long
term impacts, and to explore the how extensive the effects of disfluencies are.
2.6.3 Factors underlying children’s use of disfluencies in sentence processing
Another open question is what factors underlie children’s use of disfluencies as they are processing
sentences. One possibility is that disfluencies affect children’s speaker related attributions. Disflu-
encies often signal speech planning difficulties for adults (Bortfeld et al., 2001), and children may
make the inference that speakers are planning to use unusual sentence structures when they are
disfluent. Evidence from word learning studies points to this possibility. In word learning tasks,
young children will generally anticipate that speakers will refer to a novel object or a discourse new
object following a disfluency (C. Kidd et al., 2011; Owens & Graham, 2016). However, three and
four-year-olds vary how they interpret disfluencies, depending on qualities of the speaker. When
speakers’ are presented as forgetful, children will not anticipate discourse new objects follow-
30
ing a disfluency (Orena & White, 2015). Three and four-year-olds are also less likely to endorse
speakers who are consistently producing disfluencies while referring to familiar objects (White,
Nilsen, Deglint, & Silva, 2020). These studies suggest that children’s use of disfluencies is tied to
attributions made about the speaker.
Another factor that may be driving children’s use of disfluencies is increased attention to what
immediately follows the disfluency. Heightened attention to what may be following ditransitive
verbs could be leading children to consider alternative structures. Prior work on adult listeners’
use of disfluencies support this possibility. Adults have been found to more quickly identify words
following disfluencies (Fox Tree, 2001). Adult listeners have also been found to have better recall
of the words following disfluencies and the content of story passages told by disfluent speakers
(Corley et al., 2007; Fraundorf & Watson, 2011). Increased attention to the words following dis-
fluencies and the content of the story have been hypothesized to drive these findings. Further work
is needed to examine how these two factors impact how children use disfluencies as they process
sentences and whether there are potentially lasting impacts from making attributions of the speaker
and in the having heightened attention for disfluent sentences.
2.7 Conclusion
Children’s difficulty with revising initial predictions has been widely reported and has been im-
plicated in the delayed acquisition of various structures (Y. T. Huang et al., 2013; E. Kidd et al.,
2011; Snedeker & Trueswell, 2004; Trueswell et al., 1999; Weighall, 2008). The findings from the
current study indicate that two and three-year-olds use disfluencies as a signal to seek alternative
verb argument structures. Several factors impact how this cue is used, including familiarity with
the verb and recent experience with the sentences. A number of questions still remain about the
mechanisms driving this use of disfluencies and whether considering disfluencies will impact chil-
dren’s syntactic development. Nevertheless, the findings provide promising evidence that children
are using disfluencies as they process sentences, which may have lasting consequences. To my
knowledge, this is the first cue that children at this age are reliably using to anticipate alternative
structures.
31
Chapter 3: Uh what did you say? Children’s parsing preferences are altered by
experience with disfluent sentences
3.1 Abstract
Children can be primed to anticipate or produce a recently heard structure. This is more likely to
occur when the structures they hear, primes, are surprising. The disfluency (uh) is more likely to
occur in these surprising sentences. Moreover, recent work suggests that children may anticipate
alternative sentence structures when encountering this type of disfluency. This anticipation could
lead to differences in how well expectations align with the input and, consequently, differences in
how likely children are primed. However, whether disfluencies modulate priming effects has not
been explored. I examined this possibility with three and five-year-olds using a comprehension
priming task. I found that children’s parsing preferences were altered by disfluent primes. When
children heard disfluent primes, priming effects were reduced or children were primed by the
structure that was associated with the prime’s verb bias rather than the actual heard structure. The
strength and the onset of these effects appeared to differ by the reliability of the speaker and the
age of the children.
3.2 Introduction
Brief experience with sentence structures can have lasting impacts on the types of linguistic struc-
tures adults and children produce and anticipate in the future. Listeners have a tendency to produce
and anticipate the same sentence structures that they just heard (Bock, 1986). For example, when
a listener hears a prepositional dative (PD) (e.g. The grandma wrapped the present for the boy.),
they are more likely to produce and anticipate a PD rather than double object dative (DO) in the
future (e.g. The boy threw the ball to the dog. vs. The boy threw the dog the ball)
This tendency, structural priming, appears to have long lasting effects. Structural priming
can persist not only within conversations in spontaneous speech (Gries, 2005) but even over a
week after initial experiences with the prime structure in experimental contexts (Kaschak, Kutta,
32
& Schatschneider, 2011; Savage et al., 2006). Consequently, some theories propose that struc-
tural priming effects have long term consequences for the types of structures people produce and
anticipate over time.
Sentences with high levels of surprisal, mismatch between what listeners anticipate and what
speakers actually say, are the most likely to prime listeners (Bernolet & Hartsuiker, 2010). For ex-
ample, following the ditransitive verb wrapped, individuals can anticipate DOs and PDs (e.g. The
grandma wrapped the boy the present. vs. The grandma wrapped the present for the boy). Because
the verb wrapped has a PD verb bias (more likely to occur in a PD compared to a DO), individuals
are more likely to anticipate PDs with the verb wrapped and will experience low surprisal levels
when hearing a PD. This is the sentence structure that is more likely to occur with this verb. In
contrast, individuals are less likely anticipate DOs with the verb wrapped and will experience high
surprisal levels when hearing a DO. As a result, with the verb wrapped, individuals are more likely
to be primed by a DO, a low probability structure, than a PD.
3.2.1 Error-based learning
As a result of the lengthy effects of structural priming and the influence of surprisal, error-based-
learning mechanisms— learning from errorful predictions— is theorized to drive the development
and continued shaping of verb bias as well as children’s knowledge of verb argument structure (Lin
& Fisher, 2017; Qi et al., 2011; Ryskin et al., 2017). In error-based learning, the level of mismatch
or error between children’s predicted parse and the actual input is used to update verb bias (Chang
et al., 2006; Peter et al., 2015).
Support for error-based-learning is evident in experimental work using production priming
paradigms, in which children hear several prime sentences before they are asked to describe pic-
tures or videos of events that can elicit productions of the prime structure. Children’s verb biases
can shift within an experiment after structural priming manipulations (Lin & Fisher, 2017; Peter
et al., 2015; Qi et al., 2011; Rowland et al., 2012). There are also developmental differences for
these effects. With surprising sentences, three-year-olds are more prone to priming effects when
compared to five-year-olds (Peter et al., 2015).
33
However, research supporting the error-based-learning account and its effects focuses on fluent
speech. Disfluencies, such as uh’s and um’s, are frequent in adult speech and systematically pre-
cede less probable words and structures (Bortfeld et al., 2001; Fox Tree, 2001). Disfluent speech
can also impact the types of words and sentence structures that children anticipate. Children an-
ticipate otherwise low probability words after hearing disfluencies but not after hearing everyday
noises. Children similarly appear to consider alternative sentences structures (Chiang & Mintz, in
prep; Orena & White, 2015). Thus, if children hear sentences with disfluencies embedded in low
probability sentences, they may subsequently anticipate low probability sentences. This could lead
to lower levels surprisal and a reduced the likelihood of priming in the same sentences that should
be the most influential in error-based-learning accounts.
3.2.2 Current study
In the present study, I examined this possibility. Specifically, I asked how experience with low
probability ditransitives with disfluencies could affect the types of ditransitive structures 3 and
5-year-olds anticipated afterwards. I also asked how speaker specific traits—the speaker’s con-
sistency with producing the prime structure—impacts the types of structures these children an-
ticipate. Specifically, I asked whether children’s anticipation differed based on how reliably the
speaker produced the same structure within each trial. To address these question, I used a com-
prehension priming paradigm (conceptually following the procedures of Thothathiri and Snedeker
(2008b) and Ziegler and Snedeker (2019)) and manipulated the type of sentences children heard
within each trial.
Participants listened to several low probability sentences (primes) with either control noises
(construction noise) or disfluencies, while they looked at depictions of possible verb arguments.
Primes with control noises that were the same length as the disfluencies were included in this
study, because prior work has suggested that having extra time to process information could be an
important contributor to the effects of disfluencies (Corley & Stewart, 2008). Construction noises,
an everyday noise, was used to mirror the extra processing time that participants could get when
hearing disfluencies without providing the meaningful signal that disfluencies provide (Arnold et
al., 2007; Chiang & Mintz, in prep).
34
After participants listened to the primes, I measured whether or not participants were primed by
these recent experiences by measuring their looking preferences in a subsequent sentence. High
proportions of looks towards depictions of the verb arguments that corresponded to the prime
structure indicated priming. To examine the effect of the speaker, I also included an additional
variable. Participants either heard a speaker that always produced prime target matches or prime
target mismatches. That is, participants always either heard primes with that matched the target
sentences’ structure or primes that did not match the target sentences structure.
This design allowed us to examine any time course differences that can arise. Comprehension
priming paradigms have been demonstrated to be sensitive to subtle effects and have provided
insight into how these effects change over time. For example, Ziegler and Snedeker (2019) not
only detected priming effects through comprehension priming with adults, but also observed dif-
ferences in how preferences for the prime structures shifted with changing situational demands,
such as with speaker specific productions. In their experiment, participants heard audio recordings
of two prime sentences, ditransitives, while looking at four cartoon images. Adult’s preference
for the prime structure was measured during a subsequent target sentence. This preference was
measured through participants’ looks towards images corresponding likely continuation of either
DO or PDs during the duration of the verb and the onset of the noun following the verb— lasting a
total of approximately 600 milliseconds. When the participants in the study heard speakers who al-
ways produced prime structures that matched later utterances within a trial, participants integrated
the information more quickly. The participants switched their looks to the depiction of the noun
mentioned in the audio recording more quickly.
Moreover, looking time measures have been found to be sensitive to the effects of disfluencies.
Prior work has indicated that shifts in attention towards alternative words and structures following
disfluencies can be measured through looking preferences (Arnold et al., 2007, 2004; Chiang &
Mintz, in prep; C. Kidd et al., 2011; Orena & White, 2015; Thacker, Chambers, & Graham, 2018b).
As a result, a comprehension priming paradigm allowed us to investigate the effects of disfluencies
on both children’s general preferences for recently heard structures and how these preferences shift
over time.
The study was carried out on Lookit, an online experimental platform that collects event log
35
data and webcam recordings. This platform has conceptually replicated findings of several in per-
son eye-tracking studies and in principle can support the data collection needs for a comprehension
priming paradigm (Scott, Chu, & Schulz, 2017; Scott & Schulz, 2017).
3.3 Methods
3.3.1 Participants
Two hundred thirty-four typically developing 3- and 5-year-olds who primarily received English at
home were recruited for the study. A total of 134 participants were excluded from the final analysis
for irregular video lengths (23), audio buffering delays that exceeded 200ms (15), or for having
an insufficient number of valid trials (96). See data processing section for details on exclusion
criteria.
Thus, 100 children were included in the final analysis–– 56 three-year-olds and 44 five-year-
olds. The demographics and the conditions for these children are listed in Tables 3.1 and 3.2.
Table 3.1: Number of children in each condition in Study 2
Condition
Three-Year-Olds Five-Year-Olds)
PTM PTMM PTM PTMM
Primes with Construction Noise 16 13 15 9
Primes with Disfluencies 16 11 9 11
36
Table 3.2: Median age of participants in months in Study 2
Condition
Three-Year-Olds Five-Year-Olds)
PTM PTMM PTM PTMM
Primes with Construction Noise 38.5 42 65 64
Primes with Disfluencies 40 37 62 66
3.3.2 Stimuli and Design
The experiment consisted of eight trial— four critical trials and four distractor trials. Each of these
trials were associated with two images. In critical trials, these images were cartoon depictions of
possible patient (e.g nurse) and theme (e.g. money) for ditransitive sentences. Depictions of the
theme were always inanimate, while the patient were always animate. Each image was individually
presented on the screen as it was introduced with a verbal introduction. To ensure that any looking
preferences were not confounded with a particular location of the screen, the location of the images
were counterbalanced across participants and across trials.
The target sentences that participants heard were always ditransitive sentences with an equi-
biased verb followed by a two second construction noise. The target sentence structure alternated
between PD and DO frames. Half of the participants heard target sentences that matched the prime
structure (PTM), and half heard target sentences that mismatched the prime structure (PTMM)
The images that children saw throughout the trial were associated with one of the sentences
in the distractor trials and the target sentence in the critical trials . In total, children saw sixteen
unique images throughout the experiment. All but one image were depictions of highly frequent
nouns from the MCDI II (Fenson et al., 2007). By-word summary data from Wordbank (Frank et
al., 2016) indicated that the nouns taken from the MCDI II were produced by a high proportion of
American English speaking 30 month-olds (Appendix 6.1). Thus, children in this study should be
familiar with the images and labels that they were presented with.
37
3.3.3 Procedure
Participants accessed the experiment through the Lookit platform.
1
Before the start of the study,
parents were asked to provide verbal consent for their child’s participation and given the option
of viewing the experimental stimuli independently. Afterwards, to prevent viewing of the visual
stimuli during the main experiment, parents were asked to either have their eyes closed as their
child sat in their lap or have their child comfortably seated in front of a webcam alone. Participant
behavior was subsequently recorded for the remainder of the experiment through their webcams.
Children saw four critical trials and four distractor trials. These trials were interleaved to reduce
any spillover effects. Before the start of each trial, participants saw an attention-getter, a colorful
rotating ball, to focus their attention to the center of the screen for 3750 milliseconds. This was
accompanied by an audio recording alerting children to the start of a new trial. Two images were
presented on the far left and right side of the screen and were individually labelled by the speaker
(e.g. This is a nurse.)
Participants subsequently heard three sentences from the same speaker, a female native English
speaker. In distractor trials, participants heard three sentences that did not contain ditransitive
verbs. In critical trials, participants heard two prime sentences (ditransitives) before hearing a
target sentence. The prime sentences were always ditransitives going against verb bias (e.g. PD
frames with DO biased verbs or DO frames with PD biased verbs) and always contained either
a two second construction noise or disfluency following the verb. Half of the participants heard
prime sentences with construction noises, and half heard prime sentences with disfluencies. Figure
3.1 illustrates an example of a critical trial.
After the main experiment, parents were asked to fill out three sets of surveys regarding their
child’s mood and activities before the experiment, language input, average screen time, and vision,
as well as the browser used during the experiment. Finally, parents were asked to indicate the
privacy level for the webcam recordings. These levels included private (restricting access to the
research team working on the project), scientific and educational (allowing sharing to scientific
1
This experiment ran on the exp-lookit-images-audio frame on experiment runner version
7f470174f486396529154ee4ef2766b49c953a64
38
communities), and publicity (open to sharing to the general public for non-commercial purposes).
Recordings with privacy levels of scientific and educational or publicity have been made available
on Datavyu (Datavyu, 2014).
Figure 3.1: Schematic of a critical trial in Study 2
3.3.4 Window and measures of interest
Since children have been estimated to need 270 milliseconds to react and look towards stimuli in
visual world paradigm (Canfield et al., 1997), I focused on a shifted window of time. The window
of interest described in this section and later sections have been shifted by 250 milliseconds. This
shifted time window matches that of C. Kidd et al. (2011), a study investigating toddlers’ use of
disfluencies in novel noun labeling task.
The primary variable of interest was children’s likelihood to be primed. This was measured
through proportion of looks to the prime target during the 2,000 milliseconds of construction noise
in the target sentence and the 500 milliseconds prior to the onset of the noise. I examined whether
there were any differences in the 500 millisecond range prior to the onset of the noise to investigate
if there were any differences in looking preferences as participants heard the verb. The duration
of the verb ranged from 646 to 723 millisecond, and the average duration of the verb was 688
milliseconds. While the duration of the verb varied, examining this 500 millisecond range allowed
us to investigate the types of inferences children were making after they had some time to pro-
39
cess the initial phonological information of the verbs. In prior work, children’s anticipatory looks
towards possible verb referents have been examined with shifted time frames in order to give chil-
dren sufficient time to process initial speech segments (Fernald, 2004; Fernald, Zangl, Portillo, &
Marchman, 2008).
These 2,500 milliseconds occurred before children heard the first verb argument following
the verb, allowing us to measure any predictive looks towards the prime target. Comprehension
priming studies in the past have reported effects in similar regions — 500 milliseconds prior to the
onset of the first post-verbal during the duration of a verb (Ziegler & Snedeker, 2019).
The prime target was the depiction of the verb argument that would immediately follow the
verb in the prime structures. If children heard PD primes, the prime target was the depiction of the
theme (e.g. money). If children heard DO primes, the prime target was the was the depiction of
the patient (e.g. nurse). Looks towards prime target reflected a preference for the prime structure.
3.3.5 Data Processing
Webcam video recordings and log data, information reflecting when children saw and heard stim-
uli, were collected through Lookit. All videos were processed through the FFmpeg Python package
(Tomar, 2006) to have 20 frames per second or a frame every 50 milliseconds. The video of the
critical trials were coded by trained coders for looking directions every 50ms during the window of
interest using Datavyu (Datavyu, 2014). The looking directions corresponded to the prime target
and its alternative.
3.3.6 Predictions
Based on the prior literature, I hypothesized that participants’ looking preferences would differ by
three dimensions: the types of primes children heard, the age of the children, and the consistency of
the speaker. Specifically, I hypothesized that children who heard construction noise primes (control
condition) will have looking preferences following what has been reported in prior priming studies
(Lin & Fisher, 2017; Peter et al., 2015; Rowland et al., 2012; Ziegler & Snedeker, 2019). Since
40
children appear to parse sentences with construction noises like fluent sentences (Chiang & Mintz,
in prep), children should find these sentences surprising and be prone to structural priming effects.
I hypothesized that the type of structures that children anticipate after hearing disfluencies
would differ from structures anticipated after construction noise. Since toddlers consider different
ditransitive structures when hearing disfluencies following a ditransitive verb, children should find
these primes less surprising as they consider alternative structures (Chiang & Mintz, in prep).
Thus, these sentences should be less surprising and should have a reduced looking preference for
the prime structure. Alternatively, if children are considering alternative ditransitive structures,
they may even anticipate the structure that matches the verb bias of the prime structures rather than
the structures they hear.
I also hypothesized that these priming effects will differ by age. Prior work in production
priming paradigms indicated that five-year-olds exhibited reduced priming effects compared to
three-year-olds when the primes were surprising. However, three-year-olds likelihood to be primed
were more variable (Peter et al., 2015; Rowland et al., 2012). One reason for this variability could
be coming from task demands. Three-year-olds have been demonstrated to be more sensitive to
task demands in production priming studies (Shimpi, G´ amez, Huttenlocher, & Vasilyeva, 2007).
Thus, I not only expect three-year-olds to exhibit overall stronger priming effects in the control
condition—a stronger preference for the prime target structure—but also more variability. I also
predict that there will be age related differences when children hear disfluent primes. Specifically,
since older children have more developed verb knowledge, their representations of different verb
argument structures that can be linked to the ditransitive verbs may be more salient. This could
lead to stronger consideration of the actual verb bias.
Finally, I also hypothesized that the time-course of these looking preferences will differ by
the type of speaker children hear: speakers who always produced PTMs or who always produced
PTMMs. In comprehension priming studies with adults, participants who heard PTMs were better
able to process the target sentences than those who heard PTMMs (Ziegler & Snedeker, 2019).
Specifically, adults switched their looks from a distractor to target item more rapidly after hearing
the post-verbal arguments when they heard PTMs compared to PTMMs. Similarly, looking pref-
erences in the present study should differ based on the type of speaker children hear, if the type of
41
speaker impacts children’s consideration of primes.
3.4 Results
I analyzed the proportion of looks toward the prime target following the onset of the verb in the
target sentence through t-test cluster permutation analyses using the eyetrackingr package (Dink &
Ferguson, 2015). This method tested whether looks towards the prime target diverged for children
in the control condition compared to the disfluency condition. Several analyses, on cross sections
of the data, were conducted to examine whether the nature of this divergence varied by speaker
type and age group.
The analyses consisted of two steps (see Maris & Oostenveld, 2007, for details). In the first
step, individual clusters were identified by grouping adjacent 100 millisecond time bins with t-
values greater than a pre-defined threshold, and sum statistics were obtained for each cluster by
summing the t-values associated with each time bin within the cluster. Following prior work, I set
the initial threshold to 1.5 (de Carvalho et al., 2019, 2017; Hahn et al., 2015).
In the second step, I examined the likelihood that these clusters occurred by chance. To do
so, I obtained a distribution of sum statistics that I would expect to get by chance and used this
distribution to get a p-value for the cluster identified in the first step. I obtained a distribution of
sum statistics by conducting 1000 simulations. In each simulation, I shuffled the existing data. I
then identified any individual clusters on this shuffled dataset, following the procedure in the first
step. I grouped adjacent 100 millisecond time bins with the t-values greater than 1.5 and obtained
the corresponding sum statistics.
Cluster permutation analyses were conducted on the window of interest described in previous
sections. The analysis focused on 4 cross sections of the data: three-year-olds who heard PTMs,
three-year-olds who heard PTMMs, five-year-olds who heard PTMs, and five-year-olds who heard
PTMMs. The analyses described first include children’s looking preferences across all trials (Fig-
ure 3.2).
Next, I examined how children’s looking preference varied in the last three trials, when chil-
dren had some knowledge of the types of target sentences the speaker produced (i.e., targets that
42
matched or mismatched the structure of the primes). Measurements taken in the first trial are not
informative regarding whether knowledge of the speaker modulated priming effects (Figure 3.3).
Figure 3.2: Proportions of looks to the prime target in all trials by age and speaker type in Study
2. Lined regions represent± 1 SE. Shaded region denotes significant cluster.
Children with an insufficient number of trials were excluded from the final analyses. These
children had less than 2 valid trials. Trials were marked as invalid, if children did not look at the
computer screen for over 25% of the time during the window of interest. This often happened
when children closed their eyes during portions of the trial or were distracted by events or objects
in the home environment.
Data quality verification
A review of online methods used by developmental psychologists and validation studies of Lookit
indicated that participants can encounter technical difficulties surrounding stimuli presentation and
session recordings (Chuey et al., 2021; Scott et al., 2017). To minimize the amount of noise and
the effects of experimental manipulation confounds, I excluded participants with technical dif-
ficulties. I detected participants who had technical difficulties by examining the event log data
43
Figure 3.3: Proportions of looks to the prime target in the last 3 trials by age and speaker type in
Study 2. Lined regions represent± 1 SE. Shaded region denotes significant cluster.
collected through Lookit. The event log data included timestamps of when the audio started to
play, when the audio stopped playing, when the web cam recordings started, and when the web-
cam recordings stopped. Using these time stamps, I detected audio buffering times, sessions with
unintended interruptions caused by other computer based applications, and improperly uploaded
session recordings.
To detect audio buffering time, I compared the length of audio play time in each trial against
the intended amount of time each audio file was set to play. If the audio play time exceeded
the intended play time, this was indicative of audio buffering time. When the experiment was
first conducted, Lookit documentation listed 200 milliseconds as the maximum amount of timing
discrepancies that experimenters should expect. Thus, trials with audio buffering delays over 200
milliseconds were excluded from the final data analysis. Typically, a number of factors including
internet speed and computer performance can contribute to these audio buffering delays.
I detected irregular session recordings through similar methods. Using the event log times, I
detected video recordings with irregular lengths. This included trials with session recording video
44
lengths that did not encompass the time points of interest and video lengths that well exceeded
the lengths of usual video exports for this study on Lookit. These irregular times were often the
result of slow internet speeds cutting off session recordings during the upload process and phone or
messaging applications running in the background that interrupted the experiment’s audio stream
and extended the session recording.
The rate of exclusion for technical problems was similar to that reported by the original vali-
dation study for Lookit. This validation study also analyzed whether these exclusion criteria were
likely to affect disadvantaged populations disproportionately. These analyses indicated that these
populations were unlikely to be disproportionately affected (Scott et al., 2017).
Three-year-olds who heard Prime Target Mismatches (PTMMs)
A cluster permutation analysis on all trials revealed that when three-year-olds heard PTMMs there
was 1 significant cluster between 0–1000ms (p = 0.049). There were differences between the
looking preferences of the three-year-olds in the control condition and the three-year-olds in the
disfluency condition. The three-year-olds in the control condition looked more towards the prime
target than the three-year-olds in the disfluency condition. The three year-olds appeared to look less
towards the prime target during this time period when they heard disfluent primes from speakers
who produced PTMMs. When examining the last three trials, these trends extended and appeared
to strengthen, where there was 1 significant cluster between 0–1400ms (p = 0.029). These findings
indicated that three-year-old’s preference for the prime target diverged based on the the types of
sounds in the primes when they heard speakers who produce PTMMs.
Three-year-olds who heard Prime Target Matches (PTMs)
The pattern of results differed for three-year-olds who heard PTMs, compared to the PTMM re-
sults. Cluster permutation analyses revealed that when three-year-olds heard PTMs there was not
a significant divergence in looking preference for children in the control and disfluency condi-
tion. The three-year-olds appeared to look towards the prime target at similar rates throughout
the window of interest. Similarly, there was not a significant divergence when examining the last
45
three trials. These findings indicated that children’s preference for the prime target did not di-
verge based on the types of sounds in the primes—construction noises or disfluencies—when they
heard speakers who produced PTMs. Taken together with 3-year-olds’ behavior for inconsistent
speakers, these results provided support that three-year-olds factor in information about the speaker
when considering the primes they heard.
Five-year-olds who heard Prime Target Mismatches (PTMMs)
A cluster permutation analysis on all trials with five-year-olds who heard PTMMs did not reveal
any significant clusters. There were no significant differences in five-year-olds’ looking prefer-
ences in the control and the disfluency condition. However, when examining the last three trials, I
observe trends similar to those of the three-year-olds who heard PTMMs. There was 1 significant
cluster between 0 - 900 milliseconds (p = 0.033). The five-year-olds in the control condition look
more towards the prime target during this time period, while the five-year-olds in the disfluency
condition appeared look more towards the alternative. From visual inspection, this divergence
appeared to shift over the window of interest towards looking at the prime target at an at chance
level 900 milliseconds after the onset of the verb. These findings indicated that five-year-old’s
preference for the prime target diverge based on the sounds in the primes, construction noises and
disfluencies, when they hear speakers who produce PTMMs.
Five-year-olds who heard Prime Target Matches (PTMs)
A cluster permutation analysis on all trials with five-year-olds who heard PTMs revealed one sig-
nificant cluster between 1600 - 2600 milliseconds (p = 0.046). Unlike the three-year-olds in this
condition, there were differences between the looking preferences of the five-year-olds in the con-
trol condition and the five-year-olds in the disfluency condition. Five-year-olds in the control
condition looked more towards the prime target during this time period than the five-year-olds in
the disfluency condition. The five-year-olds in the disfluency condition appeared to look more to-
wards the alternative throughout this time period, specifically, the structure following the primes’
verb bias rather than the heard structure. When examining the last three trials, these trends were
46
extended and appeared to be stronger. There was 1 significant cluster between 1600–2600 millisec-
onds (p = 0.03). This indicated that, unlike the three-year-olds, five-year-old’s preference for the
prime target diverged based on the the sounds in the primes—construction noises or disfluencies—
when they hear speakers who produce PTMs.
Taken together with the results from the prior sections, these results provided support that that
five-year-olds, like three-year-olds, factor in information about the speaker when considering the
primes they heard.
3.5 Discussion
In an eye-tracking experiment, I examined whether primes with disfluencies and control noises
impacted the types of sentence structures children would subsequently anticipate. This effect was
modulated by children’s age and the speakers’ reliability in producing the same structure types. For
the most part, I found that when children heard primes with the control noise, visually, they had a
preference for the prime structure. When children heard primes with disfluencies, their preferences
for the prime structure was largely reduced or even followed the prime sentence’s verb bias rather
than the structures they actually heard.
Together, these results suggested that experience with disfluent sentences impacted children’s
parsing biases, but the nature of its effect was highly contextual and depended on many of the same
factors that influenced priming effects of fluent sentences as reported by previous studies (Peter et
al., 2015; Ziegler & Snedeker, 2019). Like studies with fluent speech, the current study’s priming
effects varied by children’s ages and the consistency of the speaker they heard.
These findings have implications for error-based learning proposals, which depend on surprisal
levels of sentences to predict the type of structures that children learn and the types of verb biases
they develop. The results indicated that the most influential sentences in error-based learning
proposals, low probability sentences, may not have the effect they have been proposed to have.
When low probability sentences were paired with disfluencies, they did not appear to have the
same priming effects as sentences with an everyday noise. These are the same sentences that are
more likely to contain disfluencies in naturalistic contexts. Combined, these findings suggest that
47
surprising sentences may not be as influential in shaping children’s verb bias as prior work has
proposed (Chang et al., 2006; Lin & Fisher, 2017; Rowland et al., 2012).
3.5.1 Speaker Reliability and Time course differences
In the present study, I found that the results from the disfluency condition mainly aligned with my
initial predictions. That is, disfluencies mostly led children to have a reduced preference for the
prime structures that they recently heard, compared to the baseline control condition. This was
evident in both three and five-year-old’s looks over the course of the window of interest (Figures
3.2 and 3.3). The strength and the onset of the disfluencies’ effects appeared to differ by the
reliability of the speaker and the age of the children.
The effect of speaker reliability was evident when first examining looking preferences across
all trials and cross sections of the data and comparing these trends to children’s looking preferences
in the last three trials. When looking at all trials, there were visually time course differences by
speaker type and age (Figure 3.2). Many of these differences were even more apparent in the
last three trials (Figure 3.3), providing support that some experience with a speaker could affect
children’s parsing preferences.
Trials with PTMs
There was one exception to my hypothesis: the three-year-olds who heard PTMs. For them, look-
ing preferences did not significantly diverge from those in the control condition (Figures 3.2 and
3.3). Visually, it appears that three-year-olds preference did not differ when they heard disflu-
ent primes. This trend notably diverged for five-year-olds who heard PTMs. Relative to those who
heard construction noise primes, those who heard disfluent primes had a reduced preference for the
prime target. These children, instead, had a strong looking preference for the image that matched
the primes’ actual bias in the second half of the window of interest. To inform a discussion about
why listeners might anticipate the structure that matches the verb bias, I will address why the older
children show this effect and the younger ones do not.
48
Linguistic and cognitive development may account for the differences observed in these two
age groups. One factor contributing to this difference could be children’s growing knowledge of
verb bias, which could impact children’s propensity to be primed by the prime structure. Five-year-
olds may have a stronger understanding of the verb bias, as they have had more language input and
experience with verbs. This may allow them to better access the verb arguments associated with
the verb bias of the verb. This could, consequently, lead the five-year-olds to more heavily consider
the actual verb bias and to a stronger preference for the structures that match these biases.
A related contributing factor could be children’s cognitive flexibility, the ability to switch be-
tween tasks and rules (De´ ak & Wiseheart, 2015). This ability has been implicated in the better
integration of cues that facilitate sentence processing (Snedeker et al., 2009; Snedeker & Yuan,
2008). Moreover, cognitive development is associated with better source monitoring capabilities
(Earhart & Roberts, 2014), and reductions in cognitive load can facilitate children’s use of disflu-
encies to adapt to speaker related preferences (Thacker et al., 2018b). Thus, an increased cognitive
flexibility could allow the five-year-olds to better integrate disfluencies as a cue, to consider the
alternatives when hearing a disfluency, switch their attention to the actual verb bias of the primes
rather than the sentences structures they have just heard, and make use of any inferences made
about a speaker.
Another important factor that could modulate children’s use of disfluencies is the type of attri-
butions children are making about speakers. It seemed clear that the type of speaker that children
heard in the disfluency condition influenced the types of sentences they subsequently anticipated.
However, it is unclear what specific attributions children were making.
One possibility is that children were making attributions about the trustworthiness of the
speaker. This type of inference could explain some of the developmental differences I observed
when children heard PTMs. Three-year-olds have been found to have difficulty with assessing the
trustworthiness of a speaker, while five-year-olds have not (Clegg, Kurkul, & Corriveau, 2019).
This suggests that when five-year-olds heard PTMs, they could have made the inference that a
disfluent speaker who produced PTMs were uncertain about the types of sentences they were pro-
ducing, and thus, not a trustworthy source of information. In this context, speakers expressed some
level of hesitancy and did not adhere to children’s expectations. This could have lead five-year-
49
olds to use the actual verb bias of the ditransitive verbs to inform their preferences rather than the
structures they heard.
Another possibility is that the five-year-olds who heard disfluencies were making attributions
related to production difficulties. Since the prime structures were always surprising and disfluen-
cies are associated with production difficulties, the five-year-olds could be making the attribution
that it was difficult for the speaker to produce atypical structures and expected the speaker to pro-
duce a typical structure in the absence of a disfluency. Three-year-olds may not have been able
to make or make use of these attributions for developmental reasons. Further work is needed to
disentangle these effects or to understand how these pieces interact, as speaker reliability can shape
language development and there are developmental shifts in how children consider the input they
hear.
Trials with PTMMs
For the children who heard PTMMs there were also developmental differences. However, unlike
those who heard PTMs, both the three and five-year-olds who heard disfluent primes and PT-
MMs had a decreased preference for the prime structure relative to those who heard primes with
construction noise and PTMMs. For three-year-olds who heard PTMMs, I visually observed a
sustained level of interest in the prime target during the first half of the window of interest in the
construction noise condition. This interest was reduced in the first half of the window of interest
for those in the disfluency condition. This difference was more pronounced when examining the
last three trials of the experiment.
In the last three trials, five-year-olds who heard PTMMs in the construction noise condition
similarly initially had a stronger level of interest in the prime target in the construction noise
condition compared to those in the disfluency condition. Similarly, the reduced preference for the
prime target occurred at the beginning of the window of interest.
Speakers who produce PTMMs and disfluent primes could be adhering to children’s expecta-
tions more closely. In this context, speakers are only producing disfluencies when they are pro-
ducing surprising structures. When there is not a disfluency, they are producing a structure that
50
is typically more accessible. This is the behavior that likely mirrors speakers’ use of disfluencies
in natural contexts, as disfluencies are more likely to occur with unexpected structures (Bortfeld
et al., 2001). This behavior could both adhere to expectations about a trustworthy speaker and to
expectations that disfluencies are an accurate indicator of production difficulties. Thus, it may be
easier for children to predict what the speaker will be doing when the speaker always produces PT-
MMs and why I observe differences in both three and five-year-old children’s looking preferences
when they heard primes with construction noises and disfluencies.
Time course differences
One interesting difference between children who heard PTMs and PTMMs were the observed
time course differences between the two conditions. For five-year-olds who heard PTMs looking
preferences in the construction noise and disfluency condition diverged towards the second half of
the window of interest. In contrast, for children who heard PTMMs, looking preferences in the
construction noise and disfluency condition diverged earlier in the window of interest.
Time course differences were also reflected in a previous comprehension priming study carried
out by Ziegler and Snedeker (2019). In this study, adult participants who heard PTMs switched
their looks towards the depictions of the target sentences post-verbal arguments more quickly than
participants who heard PTMMs after the offset of the ditransitive verbs. However, my findings did
not reflect the trends reported by Ziegler and Snedeker (2019). In the present study, participants in
the control condition who heard PTMs did not appear to identify the prime structure more quickly
after hearing the ditransitive verb. Visually, instead, it appears that participants who heard PTMMs
looked towards the prime structure at the beginning of the window of interest.
There may be several factors that contributed to this difference. Crucially, comprehension
priming paradigms with ditransitive sentences are sensitive to variations in experimental designs.
Ziegler and Snedeker (2019) also reported reported some differences in results from two other
studies (Arai, Van Gompel, & Scheepers, 2007; Thothathiri & Snedeker, 2008a) and attributed
these discrepancies to the sensitivity of comprehension priming studies to design changes. Three
key design differences may be contributing to the observed results of the current study.
51
First, the present study measured preference for the prime target across a longer window of
time and did not explicitly measure how quickly participants were able to process and identify
subsequent verb arguments in the target sentence (Ziegler & Snedeker, 2019). One possibility
is that children’s tendency to anticipate the prime structure towards the end of the noise when
listening to PTMs in the current study could similarly facilitate children’s ability to look towards
the depictions of the verb arguments in the target sentence more quickly. This would be in line
with Ziegler and Snedeker’s findings (2019).
Second, the populations of the two studies differed. Ziegler and Snedeker (2019) examined
priming effects in adults while the current study examined priming effects in three-year-olds and
five-year-olds. Prior literature indicated that there are developmental differences in how children
process sentences (Trueswell et al., 1999) and differences in the magnitude of structural priming
effects in children and adults, as measured through production priming paradigms (Peter et al.,
2015). Developmental changes could similarly have contributed to the different trends I observed.
As noted in the prior sections, I found developmental differences that impact how children process
prime sentences. Thus, how children consider PTMs and PTMMs may be different from adults.
Third, the current study included sentences that maximized surprisal levels, while Ziegler and
Snedeker (2019) did not. Instead, the surprisal levels of the sentences that participants heard were
varied. The surprisal levels of primes could have impacted the types of inferences participants
made about the speaker. Speculatively, in the current study, children may have made some infer-
ences about both the speaker and the structures they were hearing, as described in the previous
sections. That is, when hearing PTMs, children may think a speaker is atypical, as they are always
hearing atypical structures. In order to make make accurate predictions about the target sentences’
structure in this condition, children may be required to make inferences about the speaker, which
may have been a more difficult task. Support for this possibility comes from the developmental
differences I observed when the children heard PTMs. For those who heard PTMMs, children
may have been more willing to anticipate the prime structure, as there were some variations in the
speakers behavior and the speakers may not have been viewed as atypical.
Nonetheless, it seems clear that the type of speakers children heard and the age of children
in the present study contributed to the observed time course differences. While further work is
52
needed to investigate what factors contributed to these differences, these overlapping trends with
prior literature indicated that the results of the control condition can serve as an approximation
for measuring priming effects when children hear fluent sentences. Comparing children’s looking
preferences in the disfluency condition can, thus, provide some insight into the effects of disfluent
primes. At the same time, this comparison can help shed some insight into what factors and
mechanisms contribute to these effects, as well as the scope of these effects.
3.5.2 Future research
The current results provide support that children are influenced the prime sentences that they heard.
Under most circumstances, they anticipate . It also provides support for the hypothesis that chil-
dren’s anticipations following disfluencies are shaped not just by extra processing time afforded
by the disfluency noise but are influenced by their use of the disfluencies as a cue. These results
have implications for implicit learning theories, which have not considered the effect of disfluent
speech. Open questions remain about how children are using this cue. Specifically, what other
factors may be modulating the types of expectations children have as they hear disfluent primes
and what factors impact how their expectations.
Future work should address the types of attributions that children are making about speak-
ers and how this contributes to their interpretation of disfluencies. This could be critical, as
speaker attributions has been found to impact children’s language acquisition in other domains.
In word learning tasks, both toddlers and kindergartners consider the trustworthiness and reliabil-
ity of a speaker in word learning tasks (Dautriche, Goupil, Smith, & Rabagliati, 2021). However,
preschoolers and kindergartners’ attributions appear to be more flexible. These populations update
their beliefs about a speaker based on the types of labels and utterances a speaker produces across
an experiment (Sch¨ utte, Mani, & Behne, 2020; Scofield & Behrend, 2008).
This distinction may be particularly important for sentences with disfluencies. Children’s use of
disfluencies is modulated by the types of attributions they make about a speaker. In word learning
studies with disfluencies, young children have been reported to prefer discourse new labels when
a novel word is fronted by a disfluency (C. Kidd et al., 2011). This varies when children are led
53
to believe that a speaker is forgetful (Orena & White, 2015) and distracted (Yoon & Fisher, 2020).
Understanding the types of attributions that children make can, therefore, clarify the contexts under
which disfluencies can impact children’s future parsing preferences, word learning, and how social
development can impact these processes.
Future work should also work to understand the developmental trends for children’s use of
disfluencies. Findings from the current study suggest that some developmental factors contributed
to the current findings. While it is unclear whether these are linked to children’s cognitive or social
development, understanding developmental trends are critical. The incidences of disfluencies in
child directed speech increases with children’s age at the same time that they begin to hear more
complex sentence structures (C. Kidd et al., 2011). Disfluencies are also more likely to occur in
these same complex sentence structures (Rispoli & Hadley, 2001).
3.6 Conclusion
To summarize, the current study provides the first evidence that disfluencies can impact the types of
sentences that children anticipate in the future. Specifically, in many cases, experience with disflu-
ent primes reduced priming effects of surprising sentences or even led children to be primed by the
alternative structures that matched the bias of the verbs in the prime sentences. These effects varied
by children’s age and the types of speakers they heard. Further work needs to be conducted to elu-
cidate what cognitive and social factors contributed to these effects. Nevertheless, these findings
could have widespread implications for proposed language development mechanisms (error-based
learning) that depend on surprisal levels to predict learning. It would be critical to examine details
of these proposed mechanisms and assess the impacts of disfluent speech on these proposals.
54
Chapter 4: Initial steps to incorporate disfluencies into a modelling framework
The previous chapters provided experimental insight into how disfluencies impact children’s pars-
ing preferences in the moment and how experiences with disfluencies impact future parsing prefer-
ences. The current chapter will map out a framework for an extension of two cognitive models —
Lewis and Vasishth (2005) ACT-R model implemented through LISP as well as Engelmann, J¨ ager,
and Vasishth’s (2019) ACT-R model implemented on R (R Core Team, 2020). Implementation
of the model could offer an understanding of how the use of disfluencies interact with cognitive
mechanisms over time and with different types of input. These interactions could be subtle and
could vary by age and input.
I initially set out to implement and validate the model. However, the modifications of model
parameters and its validation are highly tied to the results of experimental data, particularly those
described in the previous chapter. While the data from the previous chapter provided broad insight
into the effects of disfluencies, the data required for a validation include precise latency measures
(how quickly participants process information). This type of data can only be obtained through
in-person experiments. As noted in the prior chapter, measures from web-based experiments can
be noisy, and in-person data could not be collected as a result of the COVID-19 pandemic. In the
subsequent sections, I lay out the planned framework for the model, the planned validation using
the larger scale experimental in-person data of the experiments discussed in the previous chapters
as well as a the results of a cognitive task, and completed corpus work that can facilitate large scale
training of the model for future computational research using this framework.
4.1 Background
One influential cognitive model that has been successful in the modelling sentence comprehension
and processing phenomena has been Lewis and Vasishth’s (2005) model. This model relied on
many of the concepts of the ACT-R framework and can implement parsing preference changes
with experience. This model has been used to examine a variety of experimental results related
to adults’ sentences processing (Dillon, Mishler, Sloggett, & Phillips, 2013; J¨ ager, Engelmann, &
Vasishth, 2015)
55
In the Lewis and Vasishth (2005) model, linguistic knowledge was represented through the
ACT-R framework’s declarative and procedural memory. Declarative memory contained the lexical
items and their associated features. These features include the functional category of each lexical
item, and the items that can proceed the lexical item. This information was represented in the form
of chunks, which could be modified and retrieved in the form of a buffer— a temporary chunk
held in short term memory. In contrast, the procedural memory acted as production rules that
were used to incrementally build syntactic trees as the lexical items became available to the model.
The manner in which the declarative and procedural memory pieces were retrieved depended on
the many of the parameters independently established from behavioral experiments and ACT-R
models (Anderson et al., 2004).
Parsing preferences in this model were developed through successful sentence comprehension
measures. In this system, comprehension was considered successful when the model’s syntactic
trees match the intended structure. How these structures were built largely depended on base
components of the ACT-R framework — how activation of each item in the declarative memory
were computed, how many cues are associated and activated for each item, and the retrieval latency
of each component following activation.
Broadly, parsing preferences were influenced by the frequency, reliability, and speed at which
the production rules were recalled, reinforced, and subsequently used to build these syntactic trees.
For example, if rules were successfully deployed, the efficiency of retrieval rules in future contexts
will be increased. Rules that were frequently and successfully used were more accessible. Con-
versely, if a rule failed to produce acceptable structures, it became harder for the system to deploy
it. If rules overlapped, in the case of ambiguous sentence structures or sentences requiring revision,
the system was influenced by prior instances of successful use when computing what kinds of trees
to construct.
More specifically, these preferences depended on the activation of each chunk containing a
lexical item and its associated features. The item with the highest activation was retrieved. An
item’s activation level depended on the item’s past usage (base activation as represented byB
i
), its
competitors’ activation levels (spreading activation as represented byS
i
), and noise involved with
each retrieval (ϵ i
). The values for each of these components were summed for items matching the
56
retrieval rule (Equation 4.1)
A
i
=B
i
+S
i
+ϵ i
(4.1)
The base activation was designed to account for the all previous past usages of the item (β i
), the
number of uses (n), the time since its last use (t
j
), and a constant decay factor accounting for the
decay in memory of this item since its last use (− d), as described in Equation 4.2.
B
i
=ln(
n
X
j=1
t
− d
j
)+β i
(4.2)
The spreading activation (Equation 4.3) accounted for the associative strength between an item
and cue(S
j
i
), which are weighted against the number of cues that matches the retrieval rules (W
j
).
In this model, cues were programmed in the declarative memory as features of each lexical item.
These cues are equally weighted in the Lewis and Vasishth (2005) model, but subsequent imple-
mentations of sentences processing models in ACT-R have made changes to the weights (see (see
J¨ ager, Mertzen, Van Dyke, & Vasishth, 2020; Vasishth, Nicenboim, Engelmann, & Burchert, 2021,
for details about the implications of weight changes))
S
i
=
X
j
W
j
S
j
i
(4.3)
The associative strength (Equation 4.4) of each cue was influenced by the maximum associated
strength (MAS) and the competition engendered by each matching cue (ln(fan
j
)). The MAS was
an ACT-R defaults setting that was validated from independent studies. Here, the fan effect equally
weighed the usefulness of each cue.
S
j
i
=MAS− ln(fan
j
) (4.4)
4.1.1 Tuning parameters
In recent extensions of the Lewis and Vasishth (2005) model, researchers have begun to tune the
parameter values and the features stored in the declarative memory to better fit a wider range of
57
experimental findings (Engelmann et al., 2019). In this section, I discuss this model in relation
to that of Lewis and Vasishth (2005), and discuss how the same architecture and differently tuned
parameters can be used to implement a model simulating children’s parses of ditransitives as well
as how their understanding of ditransitives develops.
The Engelmann, J¨ ager, and Vasishth’s (2019) model made three major changes to the Lewis
and Vasishth (2005) model to better account for the contextual cues that impact sentence parsing, to
better simulate retrieval and interference effects, and to better simulate the gradual learning process
that comes from experience with different sentence structures. Notably, the authors have speculated
that the changes to the architecture of the model will better simulate how children’s language
develops. This learning process is simulated by running a high number of simulations. Trained
models of adults parses have included 5,000 iterations (Engelmann, 2016,?; J¨ ager, Engelmann, &
Vasishth, 2017; Lewis & Vasishth, 2005).
These changes allow the model to better account for the findings in 77 cross-linguistic ex-
periments on the speed of adult’s processing of reflexives (ex: The farmer that grew the carrots
hurt himself with the pitchfork on the ground.). Specifically, they largely account for the effects
described in the J¨ ager et al. (2017) meta-analysis. The parameters and changes to the model corre-
spond to empirical findings that indicate which cues facilitate and interfere with processing.
One of the major changes that Engelmann et al. (2019) made was the addition of discourse
prominence to the base activation of each lexical item. In addition to taking into the prior usage,
the base activation of each item also included prominence (p
i
) (Equation 4.5). This was a constant
value that provided an additional boost to the base activation of an item every time it was suc-
cessfully retrieved. The model also accounted for discourse prominence in more contextual ways
through changes to the spreading activation and associative strength formulas. Changes to these
terms impacted how likely an item was retrieved and consequently the base activation of each item.
B
i
=ln(
n
X
j=1
t
− d
j
)+β i
+p
i
(4.5)
Another shift is in spreading activation formula (S
i
). This change incorporated the use of mul-
tiple discourse cues and a more graded definition of associative strength, as shown below (Equation
58
4.6).
S
i
=
X
jϵCues
W
j
S
j
i
(4.6)
The last change was to the associative strength of formula (S
j
i
). This formula incorporated
the use of a graded cross-association cue matching, the architecture of the model was modified to
include probabilistic components and that helped simulate a more graded learning process com-
pared to the fan effect used in the Lewis and Vasishth (2005) model. This change is shown in
Equation 4.7. Broadly, these probabilistic components (ln[P(i|j)]) weigh the likelihood that the
item is needed for the parse, occurrence of contextual cues, and the match quality which impacted
the saliency of an item— how well the cue and features corresponds to a procedural rule (see (see
Engelmann et al., 2019, for details and corresponding equations for each component)).
S
j
i
=MAS− ln[P(i|j)] (4.7)
From a practical perspective, this implementation impacts three key features in the R scripts:
chunk configuration, model formulas that determine the final activation rate, as well as parameters
associated with the base activation, cue matching, and retrieval cross-association cue matching.
First, the chunk configuration include more features than the Lewis and Vasishth (2005). These
are the prominence features. Along with the syntactic features, each lexical item in this model also
includes information about the cues associated with each item, such as animacy. Since one of the
primary goals of the simulations for this model is replicating latency measures, the approximate
time adults took to process each word in the words preceding the reflexive was also specified,
allowing the model to consider a baseline for retrieval measures. Second, the model formulas that
determine final activation rate are written and executed by calling functions within an R file in an
open source repository associated with the Engelmann et al. (2019) model. Third, the parameters
of the model are fit against the results of the J ¨ ager et al. (2017) meta-analysis. These include the
cross-association cue matching parameters.
Crucially, the implementation also resulted in a corresponding shiny application RStudio, Inc
(2013) that can be used to understand how relaxing and restricting parameters of the model would
59
impact the simulations of adults performance on parsing sentences with reflexives ( https://
engelmann.shinyapps.io/inter-act). Along with how well the prominence features
are weighted, these parameters also include the constant values associated with activation, such as
maximum association strength (MAS), noise (ϵ ), and penalties incurred when simulating memory
interference. This information is also available when running simulations and interpreting the
simulation output, but the shiny application offers a quickly interpretable set of summary graphs
through an easy to use GUI.
4.2 Extending model to simulate how children process ditransitives
Using the Engelmann, J¨ ager, and Vasishth’s (2019) framework to implement a model of how chil-
dren process ditransitives would be critical. Children similarly consider many discourse factors
that adults consider when processing and producing ditransitive sentences. Adults also consider
the animacy, the newness, and the grammatical complexity of the verb arguments Tily et al. (2009).
These discourse factors can be incorporated in the model.
Children are sensitive to these factors as well. Corpus work from De Marneffe et al. (2012)
demonstrates that the same discourse factors that adults consider can also influence children’s
productions of ditransitives. However, the findings from the study also indicated that children
weigh these factors differently. How these factors are weighed appear to depend on their perceived
reliability. This is supported by experimental work looking at how children process other syntactic
structures and their reliance on more reliable cues (Trueswell et al., 2012). Moreover, children’s
willingness to depend on these cues grow with their age and cognitive development (Chrysikou,
Novick, Trueswell, & Thompson-Schill, 2011; Snedeker et al., 2009; Snedeker & Yuan, 2008).
This body of work indicates that a successful cognitive model will, therefore, need to integrate
consideration of discourse factors. Thus, using an extension of the Engelmann et al. (2019) could
provide a good way to incorporate discourse factors. This could be accomplished by integrating
the cue based information that impacts children’s parses of ditransitives in the chunks for each
lexical item. Children’s graded consideration of the discourse cues can be simulated through the
cross-association cue matching parameters, which consider the match quality. These parameters,
60
cross-association cue matching parameters and match quality will need to be tuned based on the
experimental work.
These cue matching parameters may also need to be adjusted based on children’s linguistic and
cognitive development. Cognitive factors simulated in the model, memory decay and successful
retrieval, similarly may also need to be adjusted. Prior work has similarly found that children’s
memory capacity shift with age (Haden et al., 2011). This can be accomplished by the values of
salience and retrieval interference through the spreading activation values feeding into the calcula-
tions.
A first step that can be taken to accomplish this is to fit and validate a simplified model off of
five-year-old’s performance in a large scale in person version of the study presented in Chapter 3
along with a their performance on a cognitive task, which will be further discussed in the next sec-
tion. This implementation can be carried out by including a simplified set of discourse cues in the
experiment — animacy of the verb argument and the presence of a disfluency. Like the Engelmann
et al. (2019) model, the chunk configuration should include whether the cues are present for each
lexical item. The parameters can then be adjusted and fit five-year-olds results in the experimen-
tal task. This can occur through simulations with different settings. A similar shiny application
can also be developed by extending the existing shiny application from Engelmann et al. (2019)
to easily view how various parameter settings will impact the simulated data. This tuning should
also be relative to children’s performance on the cognitive task, which will be discussed in the next
section.
There are several benefits to tuning the parameters of the model with this population and task.
First, since five-year-olds’ verb knowledge and use of ditransitives are well established during
this time, fitting the model with this population will allow us to focus on fine tuning the parame-
ters related to children’s cognitive constraints. Second, the experiment can serve as a test of the
graded learning parameters, investigating whether the proposed architecture for gradual learning
will be appropriate to simulate children’s language development. Lastly, tuning the model against
experiment results can serve to test whether the application of the chunk configuration used by En-
gelmann et al. (2019) can successfully generalize to a structure with different cues within a more
controlled and predictable setting. This would be crucial before attempts are made to scale up the
61
model to accommodate naturalistic data.
4.2.1 Cognitive task
Considering experimental results of a cognitive task will be crucial. While fine tuning the param-
eters to match five-year-olds performance on that task will be informative about how cognitive
mechanisms are interacting as children processing sentences, examining how variability in cog-
nitive development through experimental measures will be crucial for future extensions of this
model. One option would be to adjust the cognitive parameters of the model relative to children’s
performance on a cognitive task. A task that can be used is the Three Dimension Changes Card
Sorting Task, a well validated cognitive flexibility task (De ´ ak & Wiseheart, 2015; Legare, Dale,
Kim, & De´ ak, 2018).
In this task, children are asked to sort the same set of cards by shape, color, and size. For
example, children could be presented with a card of a small yellow fish as shown in Figure 4.1.
They are then asked to select the box with an image that matches the image on the card’s shape—
the box with an image of a blue fish. Then, in a separate trial, they are asked to select the box with
an image that matches the image on the card’s color— the box with an image of a yellow dog.
Finally, they are asked to select the box with an image that matches the image on the card’s size—
the box with an image of a small bird.
Scores are determined by using children’s accuracy in switching to a new rule and is weighted
by the number of times they are able to switch between rules. For example, if a child successfully
sorts a card by shape first and then by color, they would get 1 point (1 accurate switch/ 1 switch
opportunity). However, if a child consistently sorts the card in the same category without regard
for the rule, they would get get 0 points (0 accurate switches). This task requires children to inhibit
the use of prior rules and flexibly switch to a new set of rules when sorting the cards.
I initially planned to examine children’s behavior in the task described in chapter 3 with this
task. Many of the participants who participated in the study described in Chapter 3 also participated
in a web-based version of the 3DCCS through Lookit (Scott et al., 2017; Scott & Schulz, 2017).
Children were initially recruited for this study shortly after their completion of the task described in
62
Figure 4.1: Example of trial presentation in the web based version of the 3dccs conducted on
Lookit
Chapter 3. However, there was a high attrition rate. Approximately 30% of the children recruited
for the study did not participate, and it was not possible to report variations in performance on
the main task as a function of flexibility score as a result. Nevertheless, the results of the task
indicated that there was some variability in the three, four, and five-year-olds’ performance on the
task. Visually, a scatter plot with children’s flexibility score (Figure 4.2) indicates that five-year-
olds did not reach ceiling performance. Thus, it may be a good starting point to examine how well
the children’s scores on this task and their performance in the main task described in Chapter 3 can
fit the model.
4.3 Future extensions: Training the model on naturalistic data
After understanding how model parameters should be tuned to best fit children’s behavior, an
extended model can be developed to simulate how children’s performance varies with verb knowl-
edge. As chapter 2 and prior work indicates, how children’s process and produce ditransitives
varies as a function of their verb knowledge (Arunachalam, 2017; Chiang & Mintz, in prep;
K. M. Snyder, 2003).
With the identified parameters, the settings can be further tuned to fit children’s developing
63
Figure 4.2: Children’s flexibility score by age (3DCSS scores from Lookit study)
understanding of ditransitive structures as it is trained through the ACT-R framework using ut-
terances from the CHILDES corpora (MacWhinney, 1991, 2000). Such a model could be used to
reinforce a knowledge base of verb bias, discourse cues, and syntactic constraints. Utterances from
CHILDES corpora have already been extracted and tagged. This tagged corpora provide some in-
formation about the frequency of lexical items and structure types in child directed speech over
time for sentences containing the verbs from the studies described in Chapters 2 and 3.
4.3.1 Extracted and tagged utterances
I extracted utterances from the CHILDES corpora (MacWhinney, 1991, 2000) using the childesr
package on R developed by Sanchez et al. (2019). This package allows researchers to extract infor-
mation from the CHILDES corpora in a reproducible manner. I used this R package to extract all
utterances from the North American database containing the following 12 words (stemmed) tagged
as verbs: show, read, find, bring, send, throw, pour, pay, save, kick, draw, and wrap. This yielded
a total of 42,013 utterances from 49 corpora, and included the utterances of parents, caretakers,
siblings, investigator, and the target child. Corresponding information about the child and the ut-
terance were also extracted at the same time. This information included the transcribed utterance,
64
the stemmed words in each utterance, the parts of speech for each stemmed word in the utterance,
the point in time the speakers produced the argument, and the age of the target child at the time of
the utterance.
I subsequently coded these utterances by sentence type. Since the primary sentence structures
of interest were ditransitive structures, the sentences were tagged as DO, PD, or Other (all sentence
types that were not DO or PDs). Arguments following the verbs in the ditransitive sentences were
then tagged by feature types using the same criteria as De Marneffe et al. (2012)— animacy,
pronominality, argument length, and prior argument structure. For prior argument structure, the
10 utterances preceding each sentence tagged as a DO or PD were also coded by sentence type. I
also extracted these utterances through childesr by using each utterances corresponding corpus and
utterance number tags. This yielded 67,344 more utterances. This number may appear to be low,
but many of the 10 utterances contained one of the twelve verbs or were part of the 10 utterances
preceding another sentence tagged as a PD or DO. These utterances were already coded and tagged
as a result.
Of the 67,344 utterances, there were 19,848 unique utterances. The corresponding files for
these utterances are available on an open source repository. In addition to these pieces, disfluencies
can be incorporated as a probabilistic cue. The regularity and reliability of the cue can be drawn
from work conducted by Tily et al. (2009).
Table 4.1 shows an example of how animacy and pronominality could be incorporated in the
chunks for the following sentence: The mom tossed it to the girl. In this table, cat refers to the
chunk’s current position on a simplified syntactic tree, while mom-cat refers to the part of the tree
where cat is projected from. In cases where the discourse features are not possible for a given
category (i.e. “tossed” in the example), the feature is marked as nil.
These naming conventions and the chunk configuration follows the chunks in the Engelmann
et al. (2019) model.
The effects of some of the other features that have been tagged in the corpora, such as prior
argument structure, can also be inputted as features of the chunk. However, since the model already
considers the last time a chunk has been recalled in it’s activation formula (t
− d
j
in Equation 4.5),
65
Table 4.1: Examples of chunk configuration for lexical items in a prepositional dative in an ACT-R
model
Chunk for
“the mom”
Chunk for
“tossed”
Chunk for
“it”
Chunk for
“to the girl”
name NP1 VP1 NP2 NP3
created 0 250 500 600
cat NP VP NP PP
mom-cat IP I-bar VP VP
role spec comp spec comp
animate yes nil no yes
pronominality no nil yes no
doing so should not be needed.
4.4 Conclusion
A successful implementation of the model would make several important theoretical and practical
contributions. Most directly, it would provide a comprehensive account of how children’s un-
derstanding of ditransitive sentences develops over time. Secondly, the results would reveal how
language specific and general cognitive mechanisms involved in language interact over time. The
model can provide subtle, fine-grained information about the interactions between each mecha-
nism allowing researchers to evaluate theories more rigorously. Because ACT-R was established
through behavioral research on general cognitive mechanisms, researchers will also be able to
adapt the model to test specific predictions about a broad range of topics including developmen-
tal language disabilities, such as specific language impairments (SLI). Using an adapted model
to study developmental language disabilities could allow researchers to pinpoint the underlying
mechanisms actively contributing to the difficulties present. This information could then be used
to develop effective and targeted interventions.
66
Chapter 5: General Discussion
The overarching goal of the dissertation is to provide an understanding of disfluencies’ impact on
children’s language development. The dissertation demonstrates that understanding how children
consider disfluencies is crucial to getting a fuller picture of how children understand and learn
from their input. In two eye-tracking studies, disfluencies influenced children’s initial predictions
(Chapter 2) and had downstream effects on how they subsequently parsed similar sentence struc-
tures (Chapter 3). There are direct and theoretical implications to these finding. The results also
identified open questions that future research should consider. I discuss a few of these implications
and open questions below.
Chapter 2 presents evidence that upon hearing disfluencies, even children as young as two-
years-old will anticipate a more conservative structure at a higher rate than when hearing an every-
day noise. Within a sentence, this anticipation also influences how children process the sentence.
The magnitude of these effects vary based on children’s verb knowledge and their recent experi-
ences with various sentence structures. These results have several implications.
Most directly, these findings provide the first evidence that disfluencies could act as a signal for
children to seek alternative verb argument structures. This may be a cue that children can capitalize
on to avoid errorful initial predictions, facilitating language processing and ensuring that children
arrive at the intended meaning of the sentence. This is significant as prior work has not found a
cue that children can robustly use. Even five-year-olds have difficulty using other cues, such as
prosodic, pragmatic, and referential information, that signal a need to revise predictions (Snedeker
& Huang, 2015; Trueswell & Gleitman, 2007).
These findings also have implications for research on word learning and children’s syntactic de-
velopment. Inaccurate parses are implicated in difficulties with learning word meanings (Havron,
de Carvalho, Fi´ evet, & Christophe, 2019; Lidz et al., 2017) and in the delayed acquisition of com-
plex structures like the passives (Y. T. Huang et al., 2013). Moreover, these results also have im-
plications for theories of language development, particularly those of error-driven-learning. These
theories stipulate that children learn the most from sentences that are surprising. That is, there is a
higher likelihood that children will anticipate and produce the low probability structures that they
67
have recently heard when compared to high probability structures (Lin & Fisher, 2017; Peter et al.,
2015). These are the same sentences that are the most likely to contain disfluencies (Bortfeld et
al., 2001). When these sentences contain disfluencies, they may no longer be surprising.
Chapter 3 addressed whether experiences with otherwise low probability sentences contain-
ing disfluencies would affect future children’s parsing preferences. I found that experiences with
disfluent sentences did influence three- and five-year-olds’ parsing preferences. Children mainly
exhibited a reduced preference for the sentence structures that they just heard when it contained a
disfluency rather than an everyday noise. Children’s preferences for the heard sentence structure
also varied by the types of speakers they heard and the children’s age. Children’s cognitive and
social development may be contributing to these variations.
These findings demonstrate that children consider low probability sentences with and without
disfluencies differently. This suggests that the most influential sentences for children’s language
learning error-based learning proposals may actually not have the proposed effect. This challenges
the body of work supporting error-based learning proposals, which only consider fluent speech.
Thus, the findings indicate that future work needs to consider children’s input in a more holistic
manner. The number of disfluencies that children hear are not trivial, and experiences with disfluent
sentences influences the types of parses that children anticipate in the future. C. Kidd et al. (2011)
estimated the rate of disfluencies when the children were two years old in the Soderstrom corpus
(2008) the rate of disfluencies was 1 in every 1000 words. The rate of disfluencies increased with
children’s age, and the authors acknowledged that this is likely an underestimation.
Speculatively, the rate of disfluencies in child-directed speech may be even higher in multi-
generational households where older adults participate in child rearing. Healthy older adults pro-
duce disfluencies at higher rates than healthy young adults (Arslan & G ¨ oksun, 2022). At the same
time, linguistic knowledge increases with age (Verhaeghen, Steitz, Sliwinski, & Cerella, 2003). As
such, older adults should be rich sources of information.
This population is important to consider for several reasons. First, the number of multi-
generational households has increased with time. Using decades of Census Bureau’s Current Pop-
ulation Survey records, the Pew Research Center estimated that the number of multi-generational
68
families has been increasing since the 1970s. In 2019, the percentage of the US population living in
multi-generational families reached 18%. This number is higher in BIPOC (Black, Indigenous, and
people of color) and economically disadvantaged families (Cohn, Horowitz, Minkin, Fry, & Hurst,
2022). Second, the current research indicates that the perceived reliability of the speaker may play
an integral role in how disfluent sentences are perceived. Healthy older adults are good sources of
information, and their use of disfluencies may be highly informative of how widely structures are
generally used. Thus, understanding how aging impacts the rate of disfluencies in child directed
speech and children’s syntactic development through experimental methods is crucial.
Another way of investigating this question would be to simulate children’s language develop-
ment with different sources of naturalistic input. Implementing the model discussed in Chapter
4 could be a good way to address the question. The discussed model could not only provide a
comprehensive account of how children’s understanding of ditransitive sentences develops over
time, it could also assess how different types of input could impact this process. The results could
reveal how language specific and general cognitive mechanisms involved in language interact over
time, as the model can provide fine-grained information about the interactions between each mech-
anism. This would allow researchers to evaluate theories more rigorously and develop language
interventions.
In sum, this dissertation presents on novel research on the effect of disfluencies on children’s
sentence processing and language development. It also suggests that future research needs to
consider the effects of disfluencies. Doing so, could better inform language development research,
especially among demographics that have been under represented in developmental psychology
research (i.e. BIPOC and economically disadvantaged populations).
69
Bibliography
Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Qin, Y . (2004). An
integrated theory of the mind. Psychological Review, 111(4), 1036.
Arai, M., Van Gompel, R. P., & Scheepers, C. (2007). Priming ditransitive structures in compre-
hension. Cognitive Psychology, 54(3), 218–250.
Arnold, J. E., Kam, C. L. H., & Tanenhaus, M. K. (2007). If you say thee uh you are describ-
ing something hard: The on-line attribution of disfluency during reference comprehension.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 33(5), 914.
Arnold, J. E., Tanenhaus, M. K., Altmann, R. J., & Fagnano, M. (2004). The old and thee, uh,
new: Disfluency and reference resolution. Psychological Science, 15(9), 578–582.
Arslan, B., & G¨ oksun, T. (2022). Aging, gesture production, and disfluency in speech: A compar-
ison of younger and older adults. Cognitive Science, 46(2), e13098.
Arunachalam, S. (2017). Preschoolers’ acquisition of novel verbs in the double object dative.
Cognitive Science, 41, 831–854.
Bailey, K. G., & Ferreira, F. (2003). Disfluencies affect the parsing of garden-path sentences.
Journal of Memory and Language, 49(2), 183–200.
Barr, D. J., & Seyfeddinipur, M. (2010). The role of fillers in listener attributions for speaker
disfluency. Language and Cognitive Processes, 25(4), 441–455.
Beattie, G. W., & Butterworth, B. L. (1979). Contextual probability and word frequency as
determinants of pauses and errors in spontaneous speech. Language and Speech, 22(3),
201–211.
Bernolet, S., & Hartsuiker, R. J. (2010). Does verb bias modulate syntactic priming? Cognition,
114(3), 455–461.
Bock, J. K. (1986). Syntactic persistence in language production. Cognitive Psychology, 18(3),
355–387.
Bortfeld, H., Leon, S. D., Bloom, J. E., Schober, M. F., & Brennan, S. E. (2001). Disfluency rates
in conversation: Effects of age, relationship, topic, role, and gender. Language and Speech,
44(2), 123–147.
70
Bresnan, J., Cueni, A., Nikitina, T., & Baayen, R. H. (2007). Predicting the dative alternation. In
Cognitive foundations of interpretation (pp. 69–94). KNAW.
Brown, R. (1973). A first language: The early stages . Cambridge, MA: Harvard University Press.
Canfield, R. L., Smith, E. G., Brezsnyak, M. P., Snow, K. L., Aslin, R. N., Haith, M. M., . . . Adler,
S. A. (1997). Information processing through the first year of life: A longitudinal study
using the visual expectation paradigm. Monographs of the Society for Research in Child
Development, i–160.
Chang, F., Dell, G. S., & Bock, K. (2006). Becoming syntactic. Psychological Review, 113(2),
234.
Chiang, C., & Mintz, T. H. (2020, September). Examining the mechanisms of structural prim-
ing: considering the effects of disfluent sentences on the implicit learning account. Poster
presented at the 26th Architectures and Mechanisms of Language Processing Conference,
Potsdam, Germany.
Chiang, C., & Mintz, T. H. (in prep). Disfluencies activate alternative argument structures in
toddlers during online sentence processing.
Chrysikou, E. G., Novick, J. M., Trueswell, J. C., & Thompson-Schill, S. L. (2011). The other side
of cognitive control: can a lack of cognitive control benefit language and cognition? Topics
in Cognitive Science, 3(2), 253–256.
Chuey, A., Asaba, M., Bridgers, S., Carrillo, B., Dietz, G., Garcia, T., . . . others (2021). Moderated
online data-collection for developmental research: methods and replications. Frontiers in
Psychology, 4968.
Clegg, J. M., Kurkul, K. E., & Corriveau, K. H. (2019). Trust me, i’ma competent expert: devel-
opmental differences in children’s use of an expert’s explanation quality to infer trustworthi-
ness. Journal of Experimental Child Psychology, 188, 104670.
Cohn, D., Horowitz, J. M., Minkin, R., Fry, R., & Hurst, K. (2022, Mar). The demographics of
multigenerational households. https://www.pewresearch.org/social-trends/2022/03/24/the-
demographics-of-multigenerational-households/. Pew Research Center.
Conwell, E., & Demuth, K. (2007). Early syntactic productivity: Evidence from dative shift.
Cognition, 103(2), 163–179.
71
Corley, M., MacGregor, L. J., & Donaldson, D. I. (2007). It’s the way that you, er, say it: Hesita-
tions in speech affect language comprehension. Cognition, 105(3), 658–668.
Corley, M., & Stewart, O. W. (2008). Hesitation disfluencies in spontaneous speech: The meaning
of um. Language and Linguistics Compass, 2(4), 589–602.
Dale, P. S., Reznick, J. S., & Thal, D. J. (1998). A parent report measure of language development
for three-year-olds. Infant Behavior and Development, 21, 370.
Dalmaijer, E. S., Mathˆ ot, S., & Van der Stigchel, S. (2014). Pygaze: An open-source, cross-
platform toolbox for minimal-effort programming of eyetracking experiments. Behavior
Research Methods, 46(4), 913–921.
Datavyu, T. (2014). Datavyu: A video coding tool. Databrary Project, New York University.
http://datavyu.org.
Dautriche, I., Goupil, L., Smith, K., & Rabagliati, H. (2021). Knowing how you know: Toddlers
reevaluate words learned from an unreliable speaker. Open Mind, 5, 1–19.
De´ ak, G. O., & Wiseheart, M. (2015). Cognitive flexibility in young children: General or task-
specific capacity? Journal of Experimental Child Psychology, 138, 31–53.
de Carvalho, A., Babineau, M., Trueswell, J. C., Waxman, S. R., & Christophe, A. (2019). Study-
ing the real-time interpretation of novel noun and verb meanings in young children. Frontiers
in Psychology, 274.
de Carvalho, A., Dautriche, I., Lin, I., & Christophe, A. (2017). Phrasal prosody constrains
syntactic analysis in toddlers. Cognition, 163, 67–79.
De Marneffe, M.-C., Grimm, S., Arnon, I., Kirby, S., & Bresnan, J. (2012). A statistical model of
the grammatical choices in child production of dative sentences. Language and Cognitive
Processes, 27(1), 25–61.
Dillon, B., Mishler, A., Sloggett, S., & Phillips, C. (2013). Contrasting intrusion profiles for
agreement and anaphora: Experimental and modeling evidence. Journal of Memory and
Language, 69(2), 85–103.
Dink, J. W., & Ferguson, B. (2015). eyetrackingr: An r library for eye-tracking data analysis.
Earhart, B., & Roberts, K. P. (2014). The role of executive function in children’s source monitoring
with varying retrieval strategies. Frontiers in Psychology, 5, 405.
72
Engelmann, F. (2016). Toward an integrated model of sentence processing in reading (Unpublished
doctoral dissertation). Universit¨ at Potsdam.
Engelmann, F., J¨ ager, L. A., & Vasishth, S. (2019). The effect of prominence and cue association
on retrieval processes: A computational account. Cognitive Science, 43(12), e12800.
Fenson, L., Marchman, V ., Thal, D., Dale, P., & Reznick, J. (2007). Macarthur-Bates Commu-
nicative Development Inventories. Baltimore, MD: Brookes Publishing Co.
Fenson, L., Pethick, S., Renda, C., Cox, J. L., Dale, P. S., & Reznick, J. S. (2000). Short-form ver-
sions of the Macarthur Communicative Development Inventories. Applied Psycholinguistics,
21(1), 95–116.
Fernald, A. (2004). The search for the object begins at the verb. In 29th annual boston university
conference on language development, november (pp. 4–7).
Fernald, A., Zangl, R., Portillo, A. L., & Marchman, V. A. (2008). Looking while listening:
Using eye movements to monitor spoken language. Developmental psycholinguistics: On-
line methods in children’s language processing, 44, 97.
Fisher, C. (2002). Structural limits on verb mapping: the role of abstract structure in 2.5-year-olds’
interpretations of novel verbs. Developmental Science, 5(1), 55–64.
Fox Tree, J. E. (2001). Listeners’ uses of um and uh in speech comprehension. Memory &
Cognition, 29(2), 320–326.
Frank, C., Michael, Mika, B., Daniel, Y ., & Marchman Virginia, A. (2016). Wordbank: An open
repository for developmental vocabulary data. Journal of Child Language, 1–18.
Fraundorf, S. H., & Watson, D. G. (2011). The disfluent discourse: Effects of filled pauses on
recall. Journal of Memory and Language, 65(2), 161–175.
Green, G. M. (1981). Pragmatics and syntactic description. Studies in the Linguistic Sciences,
11(1), 27–37.
Gries, S. T. (2005). Syntactic priming: A corpus-based approach. Journal of Psycholinguistic
Research, 34(4), 365–399.
Gropen, J., Pinker, S., Hollander, M., Goldberg, R., & Wilson, R. (1989). The learnability and
acquisition of the dative alternation in English. Language, 203–257.
Haden, C. A., Ornstein, P. A., O’Brien, B. S., Elischberger, H. B., Tyler, C. S., & Burchinal, M. J.
73
(2011). The development of children’s early memory skills. Journal of Experimental Child
Psychology, 108(1), 44–60.
Hahn, N., Snedeker, J., & Rabagliati, H. (2015). Rapid linguistic ambiguity resolution in young
children with autism spectrum disorder: Eye tracking evidence for the limits of weak central
coherence. Autism Research, 8(6), 717–726.
Havron, N., de Carvalho, A., Fi´ evet, A.-C., & Christophe, A. (2019). Three-to four-year-old
children rapidly adapt their predictions and use them to learn novel word meanings. Child
Development, 90(1), 82–90.
Havron, N., Scaff, C., Carbajal, M. J., Linzen, T., Barrault, A., & Christophe, A. (2020). Priming
syntactic ambiguity resolution in children and adults. Language, Cognition and Neuro-
science, 35(10), 1445–1455.
Hovav, M. R., & Levin, B. (2008). The english dative alternation: The case for verb sensitivity1.
Journal of Linguistics, 44(1), 129–167.
Huang, Y ., & Snedeker, J. (2020). Evidence from the visual world paradigm raises questions about
unaccusativity and growth curve analyses. Cognition, 200, 104251.
Huang, Y. T., & Arnold, A. R. (2016). Word learning in linguistic context: Processing and memory
effects. Cognition, 156, 71–87.
Huang, Y. T., Zheng, X., Meng, X., & Snedeker, J. (2013). Children’s assignment of grammat-
ical roles in the online processing of mandarin passive sentences. Journal of Memory and
Language, 69(4), 589–606.
J¨ ager, L. A., Engelmann, F., & Vasishth, S. (2015). Retrieval interference in reflexive processing:
experimental evidence from mandarin, and computational modeling. Frontiers in Psychol-
ogy, 6, 617.
J¨ ager, L. A., Engelmann, F., & Vasishth, S. (2017). Similarity-based interference in sentence
comprehension: Literature review and bayesian meta-analysis. Journal of Memory and Lan-
guage, 94, 316–339.
J¨ ager, L. A., Mertzen, D., Van Dyke, J. A., & Vasishth, S. (2020). Interference patterns in subject-
verb agreement and reflexives revisited: A large-sample study. Journal of Memory and
Language, 111, 104063.
74
Kaschak, M. P., Kutta, T. J., & Schatschneider, C. (2011). Long-term cumulative structural priming
persists for (at least) one week. Memory & Cognition, 39(3), 381–388.
Kidd, C., White, K. S., & Aslin, R. N. (2011). Toddlers use speech disfluencies to predict speakers’
referential intentions. Developmental Science, 14(4), 925–934.
Kidd, E., Stewart, A. J., & Serratrice, L. (2011). Children do not overcome lexical biases where
adults do: The role of the referential scene in garden-path recovery. Journal of Child Lan-
guage, 38(1), 222–234.
Legare, C. H., Dale, M. T., Kim, S. Y ., & De´ ak, G. O. (2018). Cultural variation in cognitive
flexibility reveals diversity in the development of executive functions. Scientific Reports ,
8(1), 1–14.
Levin, B. (2008). Dative verbs: A crosslinguistic perspective. Lingvisticæ Investigationes, 31(2),
285–312.
Lewis, R. L., & Vasishth, S. (2005). An activation-based model of sentence processing as skilled
memory retrieval. Cognitive Science, 29(3), 375–419.
Lidz, J., White, A. S., & Baier, R. (2017). The role of incremental parsing in syntactically
conditioned word learning. Cognitive Psychology, 97, 62–78.
Lin, Y ., & Fisher, C. (2017). Error-based learning: A mechanism for linking verbs to syntax. In
Proceedings of the 39th Annual Meeting of the Cognitve Science Society (pp. 754–759).
MacWhinney, B. (1991). The CHILDES project: Tools for analyzing talk. Hillsdale, NJ, US:
Lawrence Erlbaum Associates, Inc.
MacWhinney, B. (2000). The CHILDES project: Tools for analyzing talk. (3rd ed.). Mahwah, NJ:
Lawrence Erlbaum Associates.
Maris, E., & Oostenveld, R. (2007). Nonparametric statistical testing of eeg-and meg-data. Journal
of Neuroscience Methods, 164(1), 177–190.
Naigles, L. R. (2002). Form is easy, meaning is hard: Resolving a paradox in early child language.
Cognition, 86(2), 157–199.
Oehrle, R. T. (1976). The grammatical status of the english dative alternation (Unpublished
doctoral dissertation). Massachusetts Institute of Technology.
Orena, A. J., & White, K. S. (2015). I forget what that’s called! Children’s online processing of
75
disfluencies depends on speaker knowledge. Child Development, 86(6), 1701–1709.
Owens, S. J., & Graham, S. A. (2016). Thee, uhh disfluency effect in preschoolers: A cue to
discourse status. British Journal of Developmental Psychology, 34(3), 388–401.
Peter, M., Chang, F., Pine, J. M., Blything, R., & Rowland, C. F. (2015). When and how do
children develop knowledge of verb argument structure? Evidence from verb bias effects in
a structural priming task. Journal of Memory and Language, 81, 1–15.
Pozzan, L., & Trueswell, J. C. (2015). Revise and resubmit: How real-time parsing limitations
influence grammar acquisition. Cognitive Psychology, 80, 73–108.
Qi, Z., Yuan, S., & Fisher, C. (2011). Where does verb bias come from? experience with par-
ticular verbs affects on-line sentence processing. In Proceedings of 35th Boston University
Conference on Language Development (pp. 500–512).
R Core Team. (2020). R: A language and environment for statistical computing [Computer soft-
ware manual]. Vienna, Austria. Retrieved fromhttps://www.R-project.org/
Rispoli, M., & Hadley, P. (2001). The leading-edge: the significance of sentence disruptions in the
development of grammar. Journal of Speech, Language & Hearing Research, 44(5).
Rowland, C. F., Chang, F., Ambridge, B., Pine, J. M., & Lieven, E. V . (2012). The development of
abstract syntax: Evidence from structural priming and the lexical boost. Cognition, 125(1),
49–63.
Rowland, C. F., Noble, C. H., & Chan, A. (2014). Competition all the way down: How children
learn word order cues to sentence meaning. In Competing motivations in grammar and usage
(pp. 125–143). Oxford University Press.
Rowland, C. F., & Noble, C. L. (2010). The role of syntactic structure in children’s sentence
comprehension: Evidence from the dative. Language Learning and Development, 7(1), 55–
75.
RStudio, Inc. (2013). Easy web applications in r. [Computer software manual]. (URL:http://
www.rstudio.com/shiny/)
Ryskin, R. A., Qi, Z., Duff, M. C., & Brown-Schmidt, S. (2017). Verb biases are shaped through
lifelong learning. Journal of Experimental Psychology: Learning, Memory, and Cognition,
43(5), 781.
76
Sanchez, A., Meylan, S. C., Braginsky, M., MacDonald, K. E., Yurovsky, D., & Frank, M. C.
(2019). childes-db: A flexible and reproducible interface to the child language data exchange
system. Behavior Research Methods, 51(4), 1928–1941.
Savage, C., Lieven, E., Theakston, A., & Tomasello, M. (2006). Structural priming as implicit
learning in language acquisition: The persistence of lexical and structural priming in 4-year-
olds. Language Learning and Development, 2(1), 27–49.
Sch¨ utte, F., Mani, N., & Behne, T. (2020). Retrospective inferences in selective trust. Royal
Society Open Science, 7(2), 191451.
Scofield, J., & Behrend, D. A. (2008). Learning words from reliable and unreliable speakers.
Cognitive Development, 23(2), 278–290.
Scott, K., Chu, J., & Schulz, L. (2017). Lookit (part 2): Assessing the viability of online develop-
mental research, results from three case studies. Open Mind, 1(1), 15–29.
Scott, K., & Schulz, L. (2017). Lookit (part 1): A new online platform for developmental research.
Open Mind, 1(1), 4–14.
Shimpi, P. M., G´ amez, P. B., Huttenlocher, J., & Vasilyeva, M. (2007). Syntactic priming in 3-
and 4-year-old children: Evidence for abstract representations of transitive and dative forms.
Developmental Psychology, 43(6), 1334.
Snedeker, J., & Huang, Y. T. (2015). Sentence processing. In E. L. Bavin & L. R. Naigles
(Eds.), The Cambridge handbook of child language (2nd ed., p. 409-437). Cambridge, UK:
Cambridge University Press.
Snedeker, J., & Trueswell, J. C. (2004). The developing constraints on parsing decisions: The role
of lexical-biases and referential scenes in child and adult sentence processing. Cognitive
Psychology, 49(3), 238–299.
Snedeker, J., Worek, A., & Shafto, C. (2009). The role of lexical bias and global plausibility in
children’s online parsing: A developmental shift from bottom-up to top-down cues. Talk pre-
sented at the 45th annual Boston University Conference on Language Development, Boston.
Boston, MA.
Snedeker, J., & Yuan, S. (2008). Effects of prosodic and lexical constraints on parsing in young
children (and adults). Journal of Memory and Language, 58(2), 574–608.
77
Snyder, K. M. (2003). The relationship between* form and* function in ditransitive constructions
(Unpublished doctoral dissertation). Princeton University.
Snyder, W., & Stromswold, K. (1997). The structure and acquisition of English dative construc-
tions. Linguistic Inquiry, 281–317.
Soderstrom, M., Blossom, M., Foygel, R., & Morgan, J. L. (2008). Acoustical cues and grammat-
ical units in speech to two preverbal infants. Journal of Child Language, 35(4), 869–902.
Soderstrom, M., & Morgan, J. L. (2007). Twenty-two-month-olds discriminate fluent from disflu-
ent adult-directed speech. Developmental Science, 10(5), 641–653.
Thacker, J. M., Chambers, C. G., & Graham, S. A. (2018a). Five-year-olds’ and adults’ use of
paralinguistic cues to overcome referential uncertainty. Frontiers in Psychology, 9, 143.
Thacker, J. M., Chambers, C. G., & Graham, S. A. (2018b). When it is apt to adapt: Flexible rea-
soning guides children’s use of talker identity and disfluency cues. Journal of Experimental
Child Psychology, 167, 314–327.
Thothathiri, M., & Snedeker, J. (2008a). Give and take: Syntactic priming during spoken language
comprehension. Cognition, 108(1), 51–68.
Thothathiri, M., & Snedeker, J. (2008b). Syntactic priming during language comprehension in
three-and four-year-old children. Journal of Memory and Language, 58(2), 188–213.
Tily, H., Gahl, S., Arnon, I., Snider, N., Kothari, A., & Bresnan, J. (2009). Syntactic probabilities
affect pronunciation variation in spontaneous speech. Language and Cognition, 1(2), 147–
165.
Tomar, S. (2006). Converting video formats with ffmpeg. Linux Journal(146), 10.
Trueswell, J. C., & Gleitman, L. R. (2007). Learning to parse and its implications for language
acquisition. In The oxford handbook of psycholinguistics (pp. 635–656). Oxford University
Press, England.
Trueswell, J. C., Kaufman, D., Hafri, A., & Lidz, J. (2012). Development of parsing abilities
interacts with grammar learning: Evidence from Tagalog and Kannada. In Proceedings of
the 36th annual boston university conference on language development (pp. 620–632).
Trueswell, J. C., Papafragou, A., & Choi, Y . (2011). Referential and syntactic processes: what
develops? The Processing and Acquisition of Reference, 65–108.
78
Trueswell, J. C., Sekerina, I., Hill, N. M., & Logrip, M. L. (1999). The kindergarten-path effect:
Studying on-line sentence processing in young children. Cognition, 73(2), 89–134.
Vasishth, S., Nicenboim, B., Engelmann, F., & Burchert, F. (2021). Computational models of
retrieval processes in sentence processing:(trends in cognitive sciences 23, 968–982; 2019).
Trends in Cognitive Sciences, 25(10), 911.
Verhaeghen, P., Steitz, D. W., Sliwinski, M. J., & Cerella, J. (2003). Aging and dual-task perfor-
mance: a meta-analysis. Psychology and Aging, 18(3), 443.
Weighall, A. R. (2008). The kindergarten path effect revisited: Children’s use of context in
processing structural ambiguities. Journal of Experimental Child Psychology, 99(2), 75–
95.
White, K. S., Nilsen, E. S., Deglint, T., & Silva, J. (2020). That’s thee, uuh blicket! How does
disfluency affect children’s word learning? First Language, 40(1), 3–20.
Yoon, S. O., & Fisher, C. (2020). Children’s attribution of disfluency to different sources. In
Cogsci.
Ziegler, J., & Snedeker, J. (2019). The use of syntax and information structure during language
comprehension: Evidence from structural priming. Language, Cognition and Neuroscience,
34(3), 365–384.
79
Appendix
6.1 Items and Proportion of 30 month olds on WordBank reported to produce the item
label in Study 1 and 2
Item Proportion of 30 month olds
bowl 0.8
book 0.95
carrot 0.91
box 0.85
flowers 0.97
cracker 0.86
milk 0.96
money 0.79
cookie 0.94
ball 0.98
star 0.82
present 0.7
shoe 0.97
hat 0.93
sun 0.83
grass 0.77
table 0.82
tree 0.88
truck 0.93
plant 0.6
bubbles 0.92
Continued on next page
80
Continued from previous page
Item Proportion of 30 month olds
basket 0.67
cheese 0.92
door 0.91
grandpa 0.94
mom 0.99
cowboy 0.36
mailman 0.47
man 0.67
policeman 0.48
dad 0.99
doctor 0.76
teacher 0.56
girl 0.75
kid NA
grandma 0.98
dog 0.97
baby 0.96
horse 0.91
lady 0.46
woman NA
duck 0.91
cat 0.96
nurse 0.26
student NA
bunny 0.85
Continued on next page
81
Continued from previous page
Item Proportion of 30 month olds
nanny NA
boy 0.8
82
6.2 Estimated verb bias and frequency of verbs used in the experimental stimuli in study 1
Verb Verb Bias Frequency in Child Directed Speech
show DO Bias High
read Equi-Bias High
find PD Bias High
bring DO Bias High
send Equi-Bias High
throw PD Bias High
pour DO Bias Low
pay Equi-Bias Low
save PD Bias Low
kick DO Bias Low
draw Equi-Bias Low
wrap PD Bias Low
83
6.3 Estimated verb bias of verbs used in the experimental stimuli in Study 2
Verb Verb Bias Sentence type containing the verb
bring DO Bias Prime Sentence
kick DO Bias Prime Sentence
pour DO Bias Prime Sentence
show DO Bias Prime Sentence
find PD Bias Prime Sentence
save PD Bias Prime Sentence
throw PD Bias Prime Sentence
wrap PD Bias Prime Sentence
draw Equi-Bias Target Sentence
pay Equi-Bias Target Sentence
read Equi-Bias Target Sentence
send Equi-Bias Target Sentence
84
6.4 Items in the experimental stimuli and Proportion of 30 month olds on WordBank re-
ported to produce the item label
Item Proportion of 30 month olds
baby 0.96
ball 0.98
basket 0.67
book 0.95
cheese 0.92
flowers 0.97
grass 0.77
money 0.79
mom 0.99
nurse 0.26
plant 0.6
star 0.82
sun 0.83
table 0.82
tree 0.88
woman NA
85

Abstract (if available)

Abstract Disfluencies, such as uh’s and um’s, frequently occur in everyday speech. Prior work estimates that there are at least six disfluencies that occur for every 100 words and demonstrates that disfluencies can impact how well adults process sentences and remember their content. Yet, how disfluencies affect children’s understanding of sentences in the moment and their language development has not been investigated. In this dissertation, I address this gap through two experiments. In chapter 2, I examine how toddlers’ ability to process sentences in the moment is impacted by disfluent speech. Using an in-person eye-tracking study, I demonstrate that toddlers anticipate the more accommodating sentence structures as they hear disfluencies— such as Jill gave the cracker to the duck rather than Jill gave the duck the cracker. The types of structures they consider also appear to impact how toddlers subsequently process sentences. These effects are modulated, as a result, by their verb knowledge (what argument structures are available) and their recent experience. In chapter 3, I investigate how brief experiences with disfluent speech impact children’s, three-year-old and five-year-olds, predictions about upcoming sentences. I find that children’s experiences with disfluent speech reduces children’s preferences for the structures they have recently heard. I discuss these findings’ impacts on current theories of children’s syntactic development, which depend on children’s preference for recently heard structures. In chapter 4, I discuss how a computational model can be implemented to conduct a rigorous and comprehensive examination of children’s syntactic development. Specifically, I discuss how coded corpora, findings from an online study of children’s cognitive development, and extensions of the studies described in chapters 2 and 3 can inform this model. To summarize, across these chapters, I discuss the implication of disfluent speech on current theories of children’s sentence processing and children’s syntactic development. I also discuss the implications this work has for intervention for language disabilities, such as specific language impairments (SLI).

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

When things are left unsaid: existential and anaphoric implicit objects in discourse

PDF

Processing the dynamicity of events in language

PDF

Syntactic and non-syntactic factors in reflexive pronoun resolution in Mandarin Chinese

PDF

Discourse-level processing of nominal possessive constructions

PDF

Statistical learning is a two-step process

PDF

Dynamics of multiple pronoun resolution

PDF

Learning repetition rules and non-adjacent dependencies from human actions in nine-month-old infants

PDF

Can infants discriminate between declaratives and interrogatives?

PDF

Examining the neuroanatomical correlates of reading in developmental dyslexia

PDF

Schema architecture for language-vision interactions: a computational cognitive neuroscience model of language use

PDF

Syntax-aware natural language processing techniques and their applications

PDF

Exploring the effects of Korean subject marking and action verbs’ repetition frequency: how they influence the discourse and the memory representations of entities and events

PDF

Effects of language familiarity on talker discrimination from syllables

PDF

Investigating coordination of lexical and structural information cross-linguistically

PDF

Comparative iIlusions at the syntax-semantics interface

PDF

Young children’s demand for reciprocal engagement: evidence from typically-developing and autistic preschoolers

PDF

Building blocks of action and goal comprehension in a newborn visual system

PDF

Generalized surface correspondence in reduplicative opacity

PDF

The development of object recognition in the newborn brain

PDF

When and how infants discriminate between declaratives and interrogatives

Asset Metadata

Creator Chiang, Cindy (author)

Core Title Considering the effects of disfluent speech on children’s sentence processing capabilities and language development

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Psychology

Degree Conferral Date 2022-08

Publication Date 07/22/2022

Defense Date 06/27/2022

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

Tag disfluencies,language development,OAI-PMH Harvest,sentence processing,surprisal

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Mintz, Toben Herbert (committee chair), Kaiser, Elsi Heilala (committee member), Manis, Franklin (committee member), Zevin, Jason (committee member)

Creator Email cindyc@usc.edu,cindyychiang@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC111373898

Unique identifier UC111373898

Legacy Identifier etd-ChiangCind-10923

Document Type Dissertation

Format application/pdf (imt)

Rights Chiang, Cindy

Type texts

Source 20220722-usctheses-batch-961 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

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

Repository Email cisadmin@lib.usc.edu