UNDERSTANDING ACCEPTABILITY JUDGEMENTS: GRAMMATICAL

KNOWLEDGE VS. LEXICAL SEARCH

By

Darby Grachek

A THESIS

Submitted to

Michigan State University

in partial fulfillment of the requirements

for the degree of

Linguistics – Master of Arts

2021

ABSTRACT

UNDERSTANDING ACCEPTABILITY JUDGEMENTS: GRAMMATICAL

KNOWLEDGE VS. LEXICAL SEARCH

By

Darby Grachek

In this thesis, the source of gradience in acceptability judgments is discussed (Scholes

1966) and a set of experiments is performed which attempt to attribute gradience more con-

cretely to either phonotactic knowledge or lexical knowledge. Two phonotactic acceptability

judgment tasks are implemented to better understand whether reaction time can lessen the

influence of lexical information on phonotactic acceptability judgments. Following results

from Fox (1984) which show weaker influence from lexical information when less response

time is allowed, I hypothesize that phonotactic information should be immediately acces-

sible for participants, but that a lexical search takes more time to perform.

In turn, an

acceptability judgment task which allots less response time to participants should result in

less influence from lexical information in their responses. By comparing the resulting par-

ticipant judgments to gradient and categorical language models, I show that lexical access

is still present at early reaction times, meaning reaction time was not useful in removing

the influence of lexical information from phonotactic acceptability judgments in this set of

experiments. This prompts a discussion of other possible models which can feasibly be used

to understand these judgments and the source of their gradience.

ACKNOWLEDGEMENTS

I owe a huge thank you to all of the faculty in the linguistics department at MSU, but I

first want to extend an extra huge thank you to my advisor, Dr. Karthik Durvasula. Karthik

has not only supported my ideas and interests, but also challenged me to do more than I

ever imagined I could on my own. He has helped me to become a more competent linguist,

and a more confident person. I feel extremely lucky that I was able learn so much from him

during these past two years.

I’m also extremely grateful for the other members of my committee, Dr. Yen-Hwei Lin

and Dr. Betsy Sneller, for their insightful comments and helpful criticisms of the work that

went into this thesis. Thank you to Dr. Cristina Schmitt and Dr. Alan Munn for giving

me my first ever opportunity to do linguistics research and inspiring me to apply to the

MA program at MSU in the first place. I also want to acknowledge Dr. Deo Ngonyani for

providing me with my first teaching opportunity, and Dr. Suzanne Wagner her continuous

support during my time at MSU.

Thank you to the members of the Phonogroup and Child Language Acquisition Lab for

listening to my ideas, and also sharing their own. I’m also indebted to the members of the

Writing Accountability Group for their incredible feedback and encouragement during the

writing process for this project.

Having great friends to lean on made the biggest difference when I felt overwhelmed

and discouraged, and I’m forever grateful to the great friends that struggled through class

projects and deadlines with me, especially Sarah Sirna, Shannon Cousins, Yongqing Ye,

Daniel Greeson, Rachel Stacey, and Michaela Smith.

I’m also always grateful to my parents, Paul and Shara, and my brother, Brooks, for

supporting my dreams and ideas even when they have no idea what I’m talking about.

iii

v

vi

1

3
4
7

10
10
12
13

14
14
18
19
20
20
23
29

31

37

39

41

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 1

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 2 BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Source of Gradient Judgements . . . . . . . . . . . . . . . . . . . . . . .
Evidence from Speech Perception . . . . . . . . . . . . . . . . . . . . . . . .

1
2

CHAPTER 3 EXPERIMENT . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Participants and Procedure
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Hypotheses

3
4
5

6

7

CHAPTER 4 RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1
Comparison to Scholes (1966) . . . . . . . . . . . . . . . . . . . . . .
6.2
Analysis of Reaction Time . . . . . . . . . . . . . . . . . . . . . . . .
Primary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introducing the Models . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1
7.2
Modeling Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Post-Hoc Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8

CHAPTER 5 DISCUSSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 6 CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iv

LIST OF TABLES

Table 1:

Table 2:

Table 3:

Table of R-Squared Values for Neighborhood Density model, Gross Phono-
tactic model, and MaxEnt model of mean participant responses per stim-
ulus.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

R-sqaured values and confidence interval ranges for Neighborhood Den-
sity model, Gross Phonotactic model, and MaxEnt model of mean par-
ticipant responses per stimulus for the non-speeded condition only.

.

28

List of all stimuli used in this set of experiments along with their corre-
sponding CMU Glyphs, neighborhood density scores, gross status, and
previous rating provided by participants in Scholes (1966). . . . . . . . . .

40

v

LIST OF FIGURES

Figure 1:

Figure 2:

Figure 3:

Counts of ‘1’ responses per stimulus across all participants. It is sep-
arated the between speeded and non-speeded condition. The y-axis
shows each stimulus in IPA, and the x-axis shows the number of times
all participants responded with a ‘1’, meaning the stimulus was judged
to be a ‘good’ possible word of English. . . . . . . . . . . . . . . . . . . .

A scatter plot showing a comparison of the correlation between the
speeded and non-speeded conditions and a perfect correlation. The red
line represents a perfect correlation (a correlation between non-speeded
responses and themselves) while the blue line represents the actual cor-
relation between the two experimental conditions.

. . . . . . . . . . . . .

A scatter plot showing the correlation between the total 1-responses for
the speeded and non-speeded conditions and the original ratings (which
are also in the form of counts) from the Scholes (1966). The y-axis
presents the mean ratings for the speeded trials and non-speeded trials
for the present study, and the x-axis presents the ratings from Scholes
(1966). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 4: Mean reaction time was calculated for each stimulus across participants
and separated by condition. The y-axis presents each stimulus in IPA,
and the x-axis measures the reaction time in seconds.

. . . . . . . . . . .

Figure 5:

Figure 6:

Figure 7:

A scatter plot showing the correlation between the participant mean
responses and MaxEnt harmony scores between conditions. The x-axis
contains the portrays the responses means and the y-axis portrays the
MaxEnt harmony scores. . . . . . . . . . . . . . . . . . . . . . . . . . . .

A scatter plot showing the correlation between the participant mean
responses and neighborhood density scores between conditions. The x-
axis contains the portrays the responses means and the y-axis portrays
the neighborhood density scores.

. . . . . . . . . . . . . . . . . . . . . .

A violin plot with overlaying box plots showing a comparison between
the proportion of 1-responses responses and gross status for the speeded
and non-speeded conditions. The ‘valid’ gross status label means that
the initial cluster in the stimulus is present in English, and ‘invalid’ label
means that the initial cluster in the stimulus is not present in English.
The x-axis contains the gross status information and the y-axis portrays
the proportion of 1-responses.

. . . . . . . . . . . . . . . . . . . . . . . .

15

17

19

20

24

25

26

vi

Figure 8:

A plot showing the R-squared values for the speeded and non-speeded
conditions in each of the different models, along with the confidence
intervals for each R-squared value. This shows that all the R-squared
values are captured within the confidence intervals for the opposing con-
dition, meaning that there are no significant differences between condi-
tions for any model.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 9:

A ridge plot showing the mean ratings for each individual stimulus.
These ratings are aggregated across all participants and plotted next
to one another in order to discern how similarly individual participants
behaved to one another between conditions for each stimulus item.

. . .

27

30

vii

CHAPTER 1

INTRODUCTION

Native speakers have intuitions about which combinations of sounds are considered to be

‘good’ or ‘bad’ in their language. These are referred to as phonotactic judgments. Concerning

these judgments, I aim to address two main questions in my thesis: (1) Where do those

judgments come from; (2) More specifically, what kinds of factors are influential in making

those judgments?

It has been well-established that speaker’s phonotactic judgments are gradient (Scholes

1966; Bailey and Hahn 2001), although the source of that gradience has been hotly debated.

The two main factors that have been proposed in the literature as potential sources of

information for phonotactic acceptability judgments are probabilistic phonotactic knowledge

and lexical knowledge. However, it is not clear which of these factors is responsible for the

gradience observed in previous studies.

Several studies investigating the source of this gradience have found that when modeling

these judgments, language models which incorporate gradience have been found to be good

predictors of participant data (Albright 2009; Hayes and Wilson 2008).

Conversely, more recent studies by Gorman (2013) and Sarver (2020) show that cate-

gorical models are just as successful at producing an accurate model of speaker judgments.

This raises the question, if gradient grammatical knowledge does not significantly improve a

model’s ability to replicate native speaker judgments, is gradience really an integral part of

the grammar? It could be possible that the gradience observed in experiments from Scholes

(1966) and others is the result of factors outside of the phonotactic grammar, such as lexical

knowledge or task effects.

The source of gradience in phonotactic judgments is of importance for understanding the

scope of what phonological theory needs to account for. In claiming that gradient phonotactic

judgments are possible, we necessitate that our theory of phonology has to be able to process

1

fine-grained information in a gradient fashion. This is a much more complex version of the

phonology compared to a the categorical view. In order to avoid this overly-complex view of

phonological theory, I seek to provide evidence for categorical phonotactic judgments (Berg

2018). This would mean that the phonology does not need to account for the gradient

judgments seen in previous phonotactic acceptability judgment tasks. Instead, it would be

possible to attribute that gradience to an extra-phonological factor like lexical information

or task effects.

In this thesis, I investigated whether the gradience observed in previous judgments from

Scholes (1966) and others is really the result of the phonotactic grammar, or if influence

from lexical knowledge is actually responsible. In order to tease apart these two factors, I

preformed a set of experiments inspired by Fox (1984), who studied the influence of lexical

knowledge on the Ganong effect by looking at the effect of reaction time on participant

responses. The intent of this set of experiments was to separate the influence of lexical

knowledge from probabilistic phonotactic knowledge in speaker’s judgments, in order to

attribute the gradience in speaker judgments clearly to either phonotactic knowledge or

lexical knowledge. Results show that differences in response time are not enough to separate

these two factors prompting a discussion of the Cohort Model and other ways to determine

the role the lexical and phonotactic knowledge in phonotactic acceptability judgements.

2

CHAPTER 2

BACKGROUND

An early theory of phonotactic knowledge and judgments was proposed by Chomsky

and Halle (1968). They observed that speakers have clear categorical preferences for certain

sound sequences over others, even in novel words. For example, there is a clear difference

in the acceptability of the two novel words [blIk] and [bnIk]. Even though both are nonce

words and are therefore unattested as words of English, [blIk] is judged to be acceptable

while [bnIk] is not.

Chomsky and Halle’s (1968) explanation for that preference is the existence of sequence

structure constraints. These are feature-based constraints on which types of segments can

occur next to each other in a sound sequence. Given the judgments of native speakers,

preferences for some sequences over others can be modeled through a constraint against

word initial stop and nasal combinations, with no such constraint against stop and liquid

combinations.

Contrasting with Chomsky and Halle’s categorical account, evidence of gradient phono-

tactic judgments has also been attested. Scholes (1966) investigated whether speakers could

assign different ‘levels’ of acceptability, showing that speaker judgments are gradient, not

categorical. In order to demonstrate this gradience, he conducted an experiment where novel

word stimuli were presented auditorily to 35 seventh-graders with the following prompt:

Suppose these are foreign words which English wishes to borrow; which ones will

be admitted in their present form and which ones will be changed?

These same participants were also told, falsely, that some of the words they were hearing

were real words of English that the participants simply had not heard of before. This was

done presumably to encourage speakers to treat the nonce words more like existing English

3

words and better ensure that nonce word judgments would be comparable to those of real

English words.

Stimuli were all in the form of CCVC(C), with the initial consonant cluster being used to

test speaker’s judgments on phonotactic grammaticality. The participants heard a stimulus,

and responded with either ‘yes’ (it could be borrowed into English) or ‘no’ (it could not be

borrowed into English) on a worksheet.

Results showed that participant responses varied greatly and indicated that there are in

fact different ‘levels’ of acceptability in participant’s judgments of phonotactic acceptability.

Scholes concluded from this that any model which seeks to replicate speaker phonotactic

acceptability judgments would have to take into account that participants seem to have a

gradient measure of phonotactic acceptability.

1 The Source of Gradient Judgements

While a number of works add empirical support to Scholes’ (1966) claims that phonotactic

judgments are gradient (Albright 2007; Albright and Hayes 2003; Bailey and Hahn 2001),

the source of that gradience is often disagreed upon. Some claim that lexical information

is responsible for the observed gradient judgments, while others claim that probabilistic

phonotactic information is actually the source. For those that claim lexical information

is the source, factors like neighborhood density are often cited as being correlated with

acceptability judgments. Neighborhood density is the number of words that are similar to

the target word, usually by a one phoneme difference. Vitevitch and Luce (1999) found

that neighborhood density has an effect on how quickly participants are able to respond

to word and non-word stimuli. Specifically, high neighborhood density words are perceived

slower and less accurately than low neighborhood density words, but novel words with high

neighborhood density are responded to more quickly than low density novel words.

Bailey and Hahn (2001) found that probabilistic phonotactic information and lexical

information (namely, neighborhood density and token frequency) both have a significant

4

effect on speaker judgments. They found that lexical information accounted for 23% of the

variance in speaker judgments (or 29% using their Generalized Neighborhood Model which

also takes frequency information into account), while the model of probabilistic phonotactic

information they tested accounted for 18% of the variance. While this shows that both factors

could affect speaker’s judgments about what is a possible word of English, it also shows that

there is a very large portion of the variance in speaker judgments that is unaccounted for by

both factors. This could either mean that there are factors besides probabilistic phonotactic

knowledge and lexical knowledge that affect the level of gradience in speaker judgments,

or that the current models in use lack the nuance to correctly identify the source of more

variance in the results.

Albright (2007) claims that a phonotactic grammar, defined as a grammar that includes

bigram probabilities, is the main source of gradient judgments. In order to show this, he

used phonotactic probabilities to create a model of gradient participant behavior. He claims

that using these probabilities to model participant judgments does a better job of replicating

gradient responses in phonotactic judgment tasks than those which use lexical information

like neighborhood density.

Shademan (2006) also suggests that probabilistic phonotactic information is a better

predictor of speaker judgments. She investigated the possibility that variance in speaker

judgments could be a task effect related to whether or not there are real words in the stim-

uli. In order to demonstrate this, she implemented two versions of a nonce word acceptability

judgment task. One condition contained only nonce words, and the other contained both

nonce words and real words. Stimuli varied in lexical similarity using neighborhood density

(low and high) and in phonotactic probability (high probability, low probability, and phono-

tactic violations were all present). For the acceptability judgment tasks, Shademan observed

phonotactic probability ratings were similar across both conditions. That is, participants

consistently rated the lower probability forms as being less acceptable than the higher prob-

ability forms (as expected) in both conditions. However, the effect of lexical density was

5

not consistent between the two conditions. Shademan found that participants who saw the

condition with both real and nonce words gave slightly lower ratings to the nonce words with

low neighborhood density than those who saw only the nonce word condition. This suggests

that the effects of lexical similarity may be more sensitive to task effects like whether or not

real words are present in the stimuli. This, Shademan claims, makes phonotactic probability

more stable and therefore, a better predictor of speaker judgments.

However, it might also be the case that phonotactic information is not simply less sensitive

to task effects (and more reliable for producing speaker judgments as a result), but it is also

not the source of the variability in speaker judgment. If there is more variability in speaker

judgments with regard to lexical information, this makes a better case for lexical information

being the source of gradient judgments, instead of phonotactic probabilities.

While several studies have found probabilistic phonotactic knowledge to be a more reliable

factor in capturing the gradient behavior of phonotactic acceptability judgments (Albright

2007; Bailey and Hahn 2001; Shademan 2006), there is also evidence that lexical knowledge

is equally accurate in replicating these judgments. One such study, conducted by Gorman

(2013), compares 4 different models to measure which model types most accurately account

for speaker judgments. The first model, the Gross Phonotactic Model, evaluated data by

categorizing each nonce word as either well-formed (containing no phonotactic violations),

or ill-formed (containing phonotactic violations). A token is considered well-formed when
the constituent sequences have a non-zero frequency in a representative sample.1 This is

evaluated at both the level of the onset and the rhyme.

The second model used neighborhood density as a measure to assess whether a nonce

word would be judged favorably by participants. Gorman adopts the one-phoneme metric

to define neighborhood density, meaning that a word which can be formed by making a

one-phoneme change to the nonce word in question is counted as a ‘neighbor’. The more

lexical neighbors a word has, the more likely it is to be rated as phonotactically acceptable.
1Refers to whether the sound sequence is present in the Carnegie Mellon University

dictionary (Weide 1994).

6

The third model measured word-likeliness by using bigram frequency, and the fourth

model is the MaxEnt model, which claims that phonotactic judgements can be modeled by

using assigned weights according to the principle of maximum entropy (Hayes and Wilson

2008). This model can supposedly capture both categorical and gradient phonotactic pat-

terns and does not need to be provided with constraints in advance. Instead, it only needs

to be trained on a sample set of words.

Gorman found that the Gross Phonotactic Model (a categorical model) and the Neigh-

borhood Density Model performed better than both the bigram frequency model and the

MaxEnt model. That is, they are able to more reliably predict participant acceptability

judgments. This shows that gradient models do not reliably predict intermediate ratings for

nonce words. Importantly, Gorman also found that the Gross Phonotactic Model (which

uses phonotactic knowledge) and the Neighborhood Density Model (which uses lexical infor-

mation) perform at about the same level.

2 Evidence from Speech Perception

The above discussion clearly shows that there are multiple different views on the source

of the gradience in speaker judgments. Adding to this is the fact that experimental results

will always look somewhat gradient when comparing the behavior of many independent

participant judgments (Armstrong, L. R. Gleitman, and H. Gleitman 1983). Clearly, it is

important to find some way to be objective in measuring gradience. One way to forge ahead

is to see how other related domains have dealt with the issue of multiple sources. In this

thesis, I will use a technique inspired by work from the speech perception literature that

attempts to tease apart lexical knowledge from early perception.

A specific case in speech perception where a similar question arose is the Ganong effect.

The Ganong effect is a phenomenon in which an ambiguous segment is more likely to be

identified as one sound over another if the sound forms a word with the surrounding segments.

For example, a sound that is ambiguous between [t] and [d] might be more often identified

7

as a [d] when it is heard in the context of the rhyme [-æS], since tash isn’t a word of English

(and dash is). However, if the ambiguous segment was heard in the context of the rhyme

[-Ekst], participants would be more likely to identify the first segment as a [t], since text

is a real English word and dext is not. By implementing speech perception research, Fox

(1984) developed a better understanding of the Ganong effect and its interaction with lexical

knowledge. His goal was to investigate the difference between mechanisms of the speech

perception system, and post-perceptual decision-making mechanisms, which can utilize such

factors as lexical information, using reaction time to mediate between them.

Previous accounts of the Ganong effect characterized its interaction with the speech per-

ception process by using the criterion-shift model (Ganong 1980). This model allows lexical

information to affect and bias the process of phonetic categorization, meaning that it oc-

curs simultaneously with phonetic perception. However, Fox claims that the difference in

responses is a product of a post-categorization process where possible phonetic characteriza-

tions for each of the non-word tokens would be changed to categorizations that formed real

words after initial phonetic categorizations were made. This would mean that phonetic cat-

egorization occurs first, and afterwards, lexical information influences those categorizations.

This model is referred to as the categorical model.

In order to provide evidence for the categorical model, Fox set up two experiments with

segments that were ambiguous between [b] and [d]. Some of the ambiguous segments were

onsets for real English words, and some were onsets of nonce words. Participants were then

asked to identify whether the first segment was a [b] or a [d]. The only difference between

the two experiments was that in one version, participants were told to respond as quickly as

possible, and in the other, they were not given any time limit on their responses.

Results showed that the more time participants had to provide their response, the more

they responded with the segment that formed a real English word. This suggests that when

participants are given less time to process a stimulus, there is less influence from lexical

knowledge. This also suggests that the categorical model provides a better description of

8

the speech perception process.

More recent work has questioned this result (Rysling et al. 2015; Kingston et al. 2016)

showing that lexical knowledge is not completely absent at shorter reaction times, but it is

instead lessened compared to longer reaction times. This is somewhat concerning for Fox’s

findings, but it is at least still true that lexical knowledge has less influence over participant

decisions at shorter reaction times.

Fox’s findings lead us to question whether the influence of lexical knowledge on phonotac-

tic acceptability judgments can be separated to some extent from probabilistic phonotactic

knowledge using reaction time as a way to mediate its presence.

In this thesis I will use methodology outlined by Fox (1984) in order to further our

understanding of phonotactic acceptability judgments. Specifically, how does timing affect

the influence of lexical knowledge on phonotactic acceptability judgments, and is there a

reduction in the gradience of said judgments when participants are given less time to make

a decision?

Taking Fox’s line of reasoning into account, it is possible that pushing participants to

answer quickly would allow less time for post-perceptual mechanisms, like lexical knowledge,

to influence their decision. If in fact the gradience in judgments comes from some sort of

lexical search or lexical comparison, that should be easier to conduct with additional time.

Therefore, the acceptability judgments should become more gradient (in line with lexical

statistics) as the participant has more time.

9

CHAPTER 3

EXPERIMENT

In order to examine the effect of reaction time on phonotactic acceptability judgments, I

performed two phonotactic acceptability judgment tasks - one that is speeded, and one that

is not.

If the gradience in the responses comes largely from a lexical comparison/search,

then the speeded task should not allow for as much gradience in participant responses as

the non-speeded task. This is because there is less time available for a search of the lexicon

to be conducted and therefore less lexical knowledge is available to influence acceptability

judgments.

3 Stimuli

Stimuli included 46 of the nonsense words from Scholes’ (1966) study, selected from the

original 61 so that there is still a range of licit to illicit phonotactic sequences, plus four

additional stimuli added in order to incorporate more possible clusters like [Tw-] (full list of

stimuli located in the appendix). All stimuli are monosyllabic nonce words and contain a

number of different bi-consonantal clusters. There were 50 total clusters ranging from those

that occur in English ([sm-], [dr-]), to clusters that never occur in English ([zf-], [bv-]). The

stimuli also incorporated a number of different rhymes so that participants won’t be clued in

to the fact that the onset clusters are what is being tested. Each stimulus was presented four

times in order to capture gradience at the individual level for each participant. An equal

ratio of fillers to stimuli were also included, which were also repeated four times each. This

was done in order to ensure that participants remain sufficiently ignorant to the task’s focus

on the consonant clusters in the stimuli. This means there were 50 stimuli and 50 fillers

which each repeated four times resulting in a total of 400 tokens for each participant. All

stimuli were recorded by a native American English speaker (the author) using a Samson

USB Studio Meteor Microphone.

10

Admittedly, such a large number of stimuli made the experiment quite long and poten-

tially exhausting for participants. It is true that it is harder to trust judgments from speakers

who are exhausted, especially when the experiment is administered online as this experiment

was. To combat this, I incorporated an optional two-minute break into the middle of the

experiment (after the first 200 tokens were presented) so that participants were able to rest

during the experiment if necessary. If a participant wanted to end their break early, they

could press the spacebar to skip the break and continue with the task.

As for fillers, careful thought was put into the nonce words used as fillers in this experi-

ment. This is because using clusters that are more complex than the clusters used in the test

stimuli (e.g. tri-consonantal clusters like [ftl-]), might have resulted in skewed judgments. In

other words, having very ‘bad’ clusters in the fillers might have pushed some of the judgments

of the test clusters, which are less common or not present in English, farther in the direction

of being ‘good’, simply because they are not as shockingly bad as the tri-consonantal cluster

in the example above.

In order to avoid skewing judgments, the fillers contained mostly

simplex onsets. If they did have complex consonant clusters, they were not in the onset of

the word (to contrast them from the stimuli), but in the coda instead. This was done to draw

the participant’s attention away from the clusters when making judgments about words.

Another aspect of the stimuli that required careful consideration was the process of

recording the clusters for each experimental token. Having unattested consonant clusters in

the stimuli means that there may be small articulatory differences in the way that a native

English speaker pronounces clusters that are present in English, and those that are not. For

example, an English speaker may inadvertently produce a schwa vowel between each segment

in the cluster [bd-]. This might lead participants to judge those tokens, not as clusters, but

as CVCVC sequences ([b@da] instead of [bda]).

In order to avoid this, I produced all of

the stimuli with a schwa vowel between the first and second segment of the bi-consonantal

clusters. Then, that schwa vowel was spliced out using Praat (Boersma and Weenink 2016)

so that there were no accidental articulatory cues in the speech signal. This was done by

11

splicing each token at zero-crossings in order to make the stimuli sound as close to natural

speech as possible. For example, to get the test stimulus [bna], I pronounced [b@na], then

spliced out the [@] between the [b] and [n] segments.

4 Participants and Procedure

One-hundred-thirteen participants were recruited online via Prolific (Palanab and Schit-

ter 2018), but thirty-four were disqualified as they did not finish the experiment. Seventy-

nine participants completed the study and were paid $10.11 per hour for their participation.

Most participants took around twenty minutes to complete the entire experiment, meaning

payment for each person was around $3.37. However, nineteen additional participants were

eliminated for either not responding to any stimuli, or having answers which were too uni-

form showing that the participant was just clicking through the task without paying proper

attention. Ultimately, sixty participants were used for data analysis. A between-subjects

design was implemented to run the two different versions of the experiment, meaning that

thirty participants received the speeded acceptability judgement task, and thirty received

the non-speeded task. This was done in order to make sure that participants did not have

any previous information about the stimuli used in each task (since the same stimuli will be

used in both). PsychoPy (Peirce et al. 2019) was used to design the experiment, and and

Pavlovia (Peirce et al. 2019; Palanab and Schitter 2018) was used to run the experiment

online.

During the experiment, participants were presented with each stimulus auditorily and

asked to respond whether they thought the nonce word was ‘good’ or ‘bad’ as an example of

a typical word of English. The crucial experiment design manipulation was that one version

allowed participants to respond at their own pace, and the other pushed participants to re-

spond as quickly as possible. This was implemented via a training session where participants

were encouraged to answer as quickly as possible during the speeded task. They were also

told that if they did not answer quickly enough, the experiment would move on without

12

them and their response would not be recorded. This was not present in the non-speeded

trial. Instead, the experiment would not continue until it received the participants answer.

The inter-trial interval (ITI) for both tasks was 1000 ms.

During each trial, participants were given the prompt “What do you think of the following

word as a possible word of English?" and each nonce word stimulus was presented one at a

time. Participants were then asked to provide their acceptability judgments for the stimulus,

which they selected by pressing either the ‘1’ key for good, or the ‘0’ key for bad.

I chose to utilize a binary choice component in this set of experiments for two reasons:

One is to replicate Scholes’ original 1966 experiment where he also only allowed a binary

response. The other is that I assumed that in a speeded task like the one used here, that a

Likert scale would not be fully explored by participants due to the speed with which they

were meant to respond to the stimuli. In order to balance the two experiments, I used a

binary forced choice design for both the speeded and non-speeded version of the task.

5 Hypotheses

Based on Fox (1984), I hypothesized that in the speeded trials, participant’s judgments

would be less gradient than in the non-speeded trials. Assuming that the gradience in a

speaker’s judgment comes from lexical information such as neighborhood density or fre-

quency, it should take more time for the speaker to perform a search of the lexicon than it

should for them to rely solely on phonotactic knowledge. In the speeded trials, I hypothe-

sized that participant judgments would become more dependent on phonotactic knowledge.

If Shademan (2006) is correct, then the responses using phonotactic knowledge alone should

be less varied than those using lexical knowledge. Whether or not lexical information is be-

ing used will be dictated by how much time the participant has to make their judgment. In

other words, the speeded trials should prevent lexical information from being used, prompt-

ing a less gradient response than the non-speeded trials (where lexical information is readily

available).

13

CHAPTER 4

RESULTS

All data analysis was conducted in R (R Core Team 2019) using the meta-library tidyverse

(Wickham et al. 2019). The results are split up into two sections: the preliminary analysis

and the primary analysis. The preliminary analysis was preformed in order to have a more

holistic view of the data. These preliminary results include a comparison of the number of ‘1’

responses (‘good’ responses) in each condition, a discussion of regression towards the mean,

correlations between conditions, an analysis of reaction time, and a comparison between the

present study and a previous phonotactic acceptability judgement task by Scholes (1966).

The primary results contain model comparisons which were used to evaluate whether the

patterns of responses between the speeded and non-speeded conditions were different from

one another in a more objective fashion, which is central to the assessment of my hypotheses.

6 Preliminary Results

In order to understand how participants responded to each stimulus, the total number

of ‘1’ responses (meaning the participants thought the nonce word sounded like English)

per stimulus were analyzed according to experimental condition (speeded and non-speeded).
Figure 1 below shows that participant responses between conditions were quite similar.1

Many of the highest rated stimuli in the speeded condition are also some of the highest rated

stimuli for the non-speeded condition and vice versa. Figure 1 also contains information

about the gross status of the stimuli, which in this case is a measure of whether the initial
1A note about the scale of the x-axis: This scale is in terms of raw counts - it goes up to
around 60-70 because each stimulus was repeated 4 times for 30 participants. This means
that for each stimulus, there are 120 total times that a stimulus can be responded to by
the participants for each condition. Looking at proportions of participant responses, they
responded with ‘0’, meaning they thought the nonce word would not be a good example of
a typical word of English, about half the time. This checks out when observing the totals
for the counts of ‘1’ responses because 60 is about half of 120.

14

cluster is observed in English. This is calculated by verifying whether the cluster is present

in a representative sample of English (here the Carnegie Mellon University Pronunciation

dictionary is used (Weide 1994)). If a stimulus is labeled as ‘valid’, that means that its initial

cluster is present in English, and if a stimulus is labeled as ‘invalid’, that means its initial

cluster is not present in English.

Figure 1: Counts of ‘1’ responses per stimulus across all participants. It is separated the
between speeded and non-speeded condition. The y-axis shows each stimulus in IPA, and
the x-axis shows the number of times all participants responded with a ‘1’, meaning the
stimulus was judged to be a ‘good’ possible word of English.

Figure 1 also clearly shows that the responses for the speeded condition tend to be closer

15

to the center of the x-axis than those in the non-speeded condition. This can be interpreted as

regression toward the mean. (This will be more salient in Figure 2). A regression toward the

mean indicates that participants in the speeded trial were answering with responses that were

closer to the mean of the data than in the non-speeded version. This could be interpreted in

two ways: In one case, regression toward the mean would be interpreted as there being less

variation in participant responses for this condition (as my hypothesis predicts). Another

interpretation is that participants in the speeded condition were just guessing more often

than the non-speeded participants. Of the two possible interpretations, the latter is more

detrimental to the interests of the present study because it could allude to there being a

less systematic response strategy for the participants in the speeded condition. Under that

assumption, it may be that since participants were under more pressure to answer quickly,

they were more likely to randomly guess at an answer rather than produce a judgement

rooted in phonotactic knowledge.

One way to challenge the notion that participants were guessing more often in the speeded

condition is to compare how similar the standard deviation is for responses to each stimulus

between the two conditions. The standard deviations of the responses for both conditions

are quite similar to one another (speeded = 14.18, non-speeded = 15.94; standard deviation

was calculated using the counts for each stimulus aggregated across participants).

It is

worth noting that the standard deviation is slightly lower for the speeded condition than the

non-speeded condition (this is the right direction for my prediction that the speeded version

would be less variable than the non-speeded due to less influence from the lexicon). However,

the similar standard deviation scores suggest that the answers given by participants for each

condition were similar enough to rule out the second interpretation of the speeded responses’

regression toward the mean - that the participants were just randomly guessing more often

for the speeded task. This can be further verified by considering the correlation between the

responses for each condition.

16

Figure 2: A scatter plot showing a comparison of the correlation between the speeded and
non-speeded conditions and a perfect correlation. The red line represents a perfect correlation
(a correlation between non-speeded responses and themselves) while the blue line represents
the actual correlation between the two experimental conditions.

The correlation value between the two conditions is 0.85 (quite high) which can also

be visualized in Figure 2 above. The red line portrays a perfect correlation (in this case,

it is a correlation between the non-speeded condition and itself), and the blue line is the

correlation between the speeded and non-speeded conditions. If the blue line is close to the

red line, it would suggest that the results of the two experiments are very similar. The

red and blue lines are in fact quite close, providing evidence for highly correlated responses

between conditions. This shows that the speeded and non-speeded responses are actually

highly correlated, which makes the possibility of the speeded responses being the result of

random guessing less probable.

However, does this mean that the responses for each condition are so similar that the

previously reported difference in the level of variance is not significant? One way to answer

this question is to look at the confidence intervals of the correlation value. The null hypothesis

predicts that there is no difference between the pattern of responses in each condition,

17

meaning that the correlation value should be 1. If the confidence interval contains the value

1, that would suggest my result as not significantly different from the prediction made by the

null hypothesis. The actual confidence interval for the correlation value is 0.7416 - 0.9097,

suggesting that there is some evidence pointing to participants having substantially different

behavior between the two experimental conditions. Although participant response patterns

are highly correlated, they are potentially different enough to provide evidence for a different

response pattern between the speeded and non-speeded tasks. However, a correlation may

not be the best way to look at a difference in the level of gradience between conditions.

In order to be objective about how gradience is measured, I also fitted the data of both

conditions to gradient and categorical models, which is outlined in section 4.2.2.

6.1 Comparison to Scholes (1966)

It is also a useful check to compare the results of this experiment to the results of Scholes

(1966), since Scholes had a similar procedure to the one used here, and a subset of the

stimuli used in Scholes (1966) were used in the present study. The correlation between

the participants responses for the non-speeded condition is 0.69, and 0.63 for the speeded

condition.

It makes sense that the non-speeded version has a slightly higher correlation

than the speeded one since Scholes’ task did not involve a time limit for responses. The

correlation value for the non-speeded version suggests a positive correlation between between

the behavior of participants in both studies. A scatter plot that shows evidence of the

correlation between the responses from Scholes (1966) and both the speeded and non-speeded

conditions is shown below in Figure 3.

18

Figure 3: A scatter plot showing the correlation between the total 1-responses for the speeded
and non-speeded conditions and the original ratings (which are also in the form of counts)
from the Scholes (1966). The y-axis presents the mean ratings for the speeded trials and
non-speeded trials for the present study, and the x-axis presents the ratings from Scholes
(1966).

6.2 Analysis of Reaction Time

An analysis of the reaction times between the speeded and non-speeded conditions will

inform us about how well the experimental designs for each condition were able to regulate

the speed at which participants responded. The speeded condition timed out and did not

record a participant’s answer if they were too slow with their response. This resulted in a

loss of 695 responses in the speeded condition (this is equal to .05% of the responses from

each participant). Ideally, the responses for the speeded condition will be consistently faster

than the responses in the non-speeded condition.

As expected, the reaction times in the non-speeded condition are much more varied than

in the speeded condition. The average reaction time for the speeded trials is much shorter

at 934 ms than the average reaction time in the non-speeded trials at 1388 ms (a difference

19

of about 450 ms). This can also be seen in Figure 4.

A correlation between the speeded and non-speeded reaction times shows that the two

are not highly correlated (0.24). A weak correlation between the reaction times for both

conditions shows that the mean of reaction time between the speeded and non-speeded

conditions was significantly different.

Figure 4: Mean reaction time was calculated for each stimulus across participants and sep-
arated by condition. The y-axis presents each stimulus in IPA, and the x-axis measures the
reaction time in seconds.

7 Primary Results

7.1

Introducing the Models

It is important to consider how gradience should be measured when observing the pat-

tern of participant responses between each condition. It is true that random variation is

20

always possible in experiments as a result of noise in participant behavior (Armstrong, L. R.

Gleitman, and H. Gleitman 1983). Because of this, it might be difficult to tease apart gen-

uine gradience in phonotactic judgments from inevitable non-systematic variation between

different participants. In order to separate systematic gradience from non-systematic vari-

ation, model comparisons were performed between the mean of participant responses for

each stimulus and several different types of models. The goal of these model comparisons

was to measure differences between the response patterns in the speeded and non-speeded

conditions more objectively (i.e. avoiding incorporating non-systematic variation into the

comparison).

In order to perform each model comparison, mean participant responses per stimulus item

were fitted to several models of phonotactic acceptability. My original hypothesis was that

there is a decrease in the influence of lexical knowledge at shorter reaction times. If lexical

knowledge is hypothesized to be responsible for the gradience in participant judgements,

then the non-speeded task should have more gradience in the responses than the speeded

task. This is because participants will have had more time to access lexical knowledge than

in the speeded task.

In order to find evidence for that claim, the results of the speeded

condition and non-speeded condition would need to show significantly different effect sizes

for each of the models.

If the model fits decrease between the non-speeded and speeded

conditions, it may be possible to reason that the relevant knowledge that the model is trying

to capture is used less in the speeded condition. If it is true that lexical knowledge is used

less in the speeded condition, and the gradience is largely coming from such knowledge, then

the fits of the gradient models should worsen in the speeded task. In contrast, the fit of the

categorical model should improve in the speeded condition.

The three specific models that were used in this analysis are the MaxEnt Model (Hayes

and Wilson 2008), the Neighborhood Density Model, and the Gross Phonotactic Model

(Gorman 2013).

The first model mentioned above, the MaxEnt Model (Hayes and Wilson 2008), is a

21

gradient model that uses weighted constraints pre-determined by a training data set to

establish the acceptability of a given word. The model calculates a harmony score for each

word input into the model using the following equation:

N(cid:88)

h(x) =

WiCi(x)

(4.1)

i=1

by(cid:80)N

Here, the weight of the constraint i is represented by Wi, while Ci represents the number
of violations of that constraint, and the summation over all of the constraints is represented
i=1. The harmony score of x (h(x)) is calculated by summing the products of the weight
of each constraint and the number of violations of the corresponding constraint. These scores

were obtained via the UCLA phonotactic learner (Hayes and Wilson 2008). The MaxEnt

model uses only phonological information to form the harmony scores that will be used to

estimate participant acceptability judgements. If the data fits best to this model, it would

show that a grammar which is able to use gradient phonotactic information when making

phonotactic judgements is the most accurate way to conceptualize the patterns in the data.

The second model implemented here is the Neighborhood Density Model, which uses

neighborhood density scores as a way to rate words as acceptable or unacceptable. These

scores were collected using the iPhod neighborhood density calculator (Vaden, Halpin, and

Hickok 2009). The higher the neighborhood density score, the more acceptable the word

should be for participants. This is an important model for us to compare participant re-

sponses to because it uses lexical information to make its predictions about phonotactic

acceptability. If the data fits better to this model than other models, that would mean that

a lexical factor is better at explaining the patterns in the data than a phonotactic one.

The third model, the Gross Phonotactic Model, is a categorical model that uses gross

phonotactic information to judge whether words are acceptable or not.

In this case, the

gross status of a cluster is considered valid if it is present in the Carnegie Mellon University

dictionary (Weide 1994), and invalid if is not present. Essentially, a cluster is considered

‘good’ if it occurs in English and ‘bad’ if it does not. If the data fits best to this model, it

22

would show that a categorical measure utilizing only phonotactic information does the best

job of explaining the patterns in participant responses.

In order to compare the two experimental conditions to neighborhood density and gross

phonotactic status, a linear regression was performed between the by-item average partici-

pant response (for each condition) and the neighborhood density scores for each stimulus. A

separate regression was performed between mean participant response and gross status for

each stimulus. A regression analysis was also performed between mean participant response

and the MaxEnt harmony scores for each stimulus. The central comparison for my hypothe-

sis is between the difference in the effect size of speeded and non-speeded conditions for each

model. A general analysis of which model has the best fit to the data is also performed in

order to compare these results to the results of previous studies, but remains secondary to

the analysis of the model fits between conditions.

7.2 Modeling Results

The resulting R-squared values from each of the models are recorded in Table 1 below.

This shows the effect size of the comparison between the model and the responses (basi-

cally how well the responses fit each particular model). The crucial comparison is between

the speeded and non-speeded effect sizes for each model, with a higher R-squared value in-

dicating a better fit to the model. Additionally, Figure 5 shows the correlations between

participant response means and the MaxEnt harmony scores, while Figure 6 shows the cor-

relation between the participant response means and neighborhood density. Figure 7 shows

a comparison in the proportion of 1-responses and gross status. These plots all show that in

each case, the distribution of responses between conditions is quite similar.

Speeded

Non-speeded

0.19
0.10

Neighborhood Density Gross Status MaxEnt

0.21
0.19

0.39
0.35

Table 1: Table of R-Squared Values for Neighborhood Density model, Gross Phonotactic
model, and MaxEnt model of mean participant responses per stimulus.

23

Figure 5: A scatter plot showing the correlation between the participant mean responses and
MaxEnt harmony scores between conditions. The x-axis contains the portrays the responses
means and the y-axis portrays the MaxEnt harmony scores.

24

Figure 6: A scatter plot showing the correlation between the participant mean responses
and neighborhood density scores between conditions. The x-axis contains the portrays the
responses means and the y-axis portrays the neighborhood density scores.

25

Figure 7: A violin plot with overlaying box plots showing a comparison between the propor-
tion of 1-responses responses and gross status for the speeded and non-speeded conditions.
The ‘valid’ gross status label means that the initial cluster in the stimulus is present in
English, and ‘invalid’ label means that the initial cluster in the stimulus is not present
in English. The x-axis contains the gross status information and the y-axis portrays the
proportion of 1-responses.

The next step for this model comparison is finding out whether there are significant

differences between the model effect sizes for the speeded and non-speeded conditions. Once

again, confidence intervals were calculated for each model to see if the R-squared value for one

condition is contained within the confidence interval of the other condition. If the confidence

intervals for the speeded conditions for each model do not contain the R-squared value for

the non-speeded conditions (and vice versa), one can conclude that the differences in the

R-squared values (and therefore the model effect sizes) are significantly different between

conditions. This would provide evidence for participant acceptability judgements being

significantly affected by the amount of time they were given to respond. The confidence

intervals for each model are visible in Figure 8 below as the bars surrounding the R-squared

values for each model:

26

Figure 8: A plot showing the R-squared values for the speeded and non-speeded conditions
in each of the different models, along with the confidence intervals for each R-squared value.
This shows that all the R-squared values are captured within the confidence intervals for the
opposing condition, meaning that there are no significant differences between conditions for
any model.

In all cases, the R-squared values for both the speeded and non-speeded conditions are

contained within the confidence intervals for both individual conditions, meaning that the

change between the R-squared values for each condition can be viewed as non-significant.

Also pictured is a null model, which only contains information about the means in the data.

It is shown to have the best fit out of all of the models utilized here. Further details on this

model will be discussed below.

Although it is not central my hypothesis, an investigation of which model fit best to the

experimental data was also performed in order to compare the present study to previous

ones. Table 2 below summarizes the R-squared values and confidence intervals for each of

the models for only the non-speeded condition. This is because the non-speeded condition

27

is more comparable to previous studies than the speeded condition.

Neighborhood Density Gross Status MaxEnt

R-squared Values
Confidence Intervals

0.10

-0.05 - 0.25

0.19

0.35

0.00 - 0.37

0.16 - 0.55

Table 2: R-sqaured values and confidence interval ranges for Neighborhood Density model,
Gross Phonotactic model, and MaxEnt model of mean participant responses per stimulus
for the non-speeded condition only.

The Neighborhood Density Model has the lowest R-squared values for both conditions

and is therefore the worst-performing model to fit to this data. This means that lexical

information does not contribute much explanatory power to the pattern of judgements here.

The Gross Phonotactic Model does a bit better, showing that gross phonotactic information,

which is a categorical phonotactic measure, improves our ability to explain the patterns in

this judgement data compared to just using lexical information. The highest R-squared

value obtained for both conditions is from the MaxEnt Model. Before checking levels of

significance, this seems to indicate that gradient phonotactic information can explain more

patterns in the judgements for this set of experiments compared to both of the other models.

It is also clear that the speeded condition fits better to every model. Though it is important

to point out that, according to Cohen’s Rule of Thumb, all these effect sizes are quite small

for behavioral experiment data (Cohen 1992). Furthermore, the confidence interval for the

MaxEnt Model contains the value for the Gross Status Model (and vice versa) suggesting

that the MaxEnt and Gross Status models are not significantly different from one another.

The confidence interval for the Neighborhood Density Model also contains the R-squared

value for the Gross Status Model (and again vice versa) meaning that the Neighborhood

Density and Gross Status models are also not significantly different from each other. The

only significant improvement is between the Neighborhood Density Model and the MaxEnt

Model. Overall these comparisons should be interpreted with caution, since there are not

very substantial differences between the fits for any of the models.

To examine this further, a null model which contains no predictors for either condition

28

was also performed. It produced an R-squared value of 0.82 for the speeded condition, and

0.80 for the non-speeded condition. This is a much larger effect size than any of the models

mentioned above. This would imply that there are no systematic patterns in the data since

a null model containing no predictors has a better fit to the data than any other model

utilized in the analysis. However, this is strange since Figure 2 shows that the responses

for both experimental conditions are highly correlated, suggesting that the responses from

participants were not random and instead have some systematic behavior associated with

them. This might imply that there is a model which has not yet been identified that would

better explain the patterns that exist in this acceptability judgment data. This will be more

thoroughly explored in the Discussion section.

From the investigation above, it can be concluded that differences between the model

fits for the speeded and non-speeded conditions are not significant, and it was therefore not

possible in my experiments to separate lexical knowledge from phonotactic knowledge using

reaction time. Additionally, a null model produced the largest R-squared value in relation to

this data, which along with the high correlation between the two experimental conditions,

shows that there is a model that has not yet been entertained which may better explain the

patterns in this data.

8 Post-Hoc Testing

In order to check how consistent responses were between individual participants, by-

participant mean ratings for each stimulus is plotted below in Figure 9. Ratings for each

stimulus seem to be more or less consistent, with the ratings for most stimuli having a uni-

modal distribution. However, there are items that have a bimodal distribution, meaning

that participant responses to that item were not as consistent as those with unimodal dis-

tributions. This might be evidence that there were a few stimuli which induced perceptual

illusions for some participants more than others. However, if there is an effect from these

stimuli, it is likely that the effect is small since the majority of participants behave similarly

29

in terms of their ratings for the majority of the stimuli. The idea of perceptual illusions in

the stimuli is further investigated in the Discussion section.

Figure 9: A ridge plot showing the mean ratings for each individual stimulus. These ratings
are aggregated across all participants and plotted next to one another in order to discern
how similarly individual participants behaved to one another between conditions for each
stimulus item.

30

CHAPTER 5

DISCUSSION

Given that the difference in model fits between the speeded and non-speeded conditions

are so small, I am not able to clearly confirm the hypothesis that I initially predicted.

Although, there may be a few possible contributing factors which can be investigated in

order to improve future iterations of this experiment. One potential issue could be that the

behavior of online participants is inherently noisier than in laboratory conditions. The R-

squared values for this data are smaller than those seen for in-person experiments. Without

being able to supervise participants online, I am unable to verify if they are paying sufficient

attention to the task, wearing proper headphones, running the experiment from a computer

(not a mobile device) etc. It would be more ideal to repeat this experiment with participants

in-person and under laboratory conditions post-COVID in order to replicate the results and

verify the effects reported in section 4.

Related to rerunning the experiment, there are a few possible problems that could be

improved upon in future versions of the experimental design that may reveal participant

behavior which more closely resembles what was originally predicted for these tasks. One

being that there is a design flaw in the speeded version of the experiment that did not give

participants negative feedback when they were too slow to respond to the stimuli. Because of

this, there are 695 total stimuli that do not have responses associated with them in the results

of the speeded condition. By using feedback incorporated into the experimental design, I

could enforce a quicker response time for all stimuli with less risk of losing out on responses

in cases where participants lost track of how little time they had to answer.

There are also some potential improvements to be made with the recording and splicing

of the stimuli. As outlined in section 3.1, I pronounced all the stimuli with a schwa vowel

[@] in between the two segments of the initial cluster. I then spliced it out in an attempt

to remove any unintentional articulatory cues that may have been present if I had instead

31

attempted to simply pronounce the illicit cluster. However, in attempting to minimize those

articulatory cues, it is possible that some sounds introduced in the stimuli were unnatural

for English speakers. For example, in a cluster like [dr-], the [d] is actually pronounced more

like an affricate than a stop. However, when I pronounced the segments [d] and [r] with the

schwa between them, the [d] was pronounced just like a regular plosive [d]. This resulted in

some clusters sounding different to how they would be heard in natural English speech. If

participants picked up on the differences in these particular clusters, their ratings may have

been influenced by the fact that the usually licit clusters sounded different than normal (and

potentially less licit).

Additionally, some participants may not have heard the stimuli as intended. Since many

of the stimuli contained illicit consonant clusters that are not usually present in English, it

is possible that these clusters may have produced auditory illusions (Dupoux et al. (1999),

Durvasula and Kahng (2016)). These illusions could have made the illicit clusters in the

stimuli sound more like the English clusters which participants are more familiar with. This

may have led to them rating some stimuli as ‘good’ because they were making judgments

based on the illusory sequences, not the illicit clusters.

In order to check whether perceptual illusions were prevalent in this set of tasks, it is

helpful to look at specific clusters that are prone to being misperceived by English speakers to

see how the participants in this task responded to them. Davidson and Shaw (2012) identify

3 common perceptual illusions in English and the environments that they are most likely

to occur. These include prothesis illusions (adding a sound to the beginning of o a word),

most often found in fricative–initial sequences, deletion or change of the first consonant,

often found in stop–nasal sequences, and vowel insertion, which is common in stop–stop

sequences. There are not perfect examples of all of these environments in the stimuli used

for this experiment, but close candidates can be observed to approximate these environments.

Many of the clusters occurring in these environments are not usually considered to be

licit in English, so it would make sense for participants to rate them as ‘bad’ more often. If

32

it is instead the case that the stimuli containing these clusters are highly rated, there would

be reason to believe that perceptual illusions interfered with participant’s judgments in these

tasks.

Fricative initial sequences like those in [fsEt] and [zkip] have a neighborhood density of 4

and 1 (respectively), invalid gross status, and low ratings from participants in Scholes (1966).

These factors would would predict that these stimuli be rated poorly by the participants in

this study. Referring back to Figure 1, we can see that both stimuli are in fact rated quite low

for both conditions. This tells us that prothesis illusions were probably not very prevalent

for speakers in this set of tasks.

There are no stop-nasal sequences present in the stimuli for this experiment, but there

are obstruent-nasal sequences, such as the fricative-nasal sequences in [vnEt] and [fmæt].

This environment should trigger consonant deletion or change perceptual illusions. Both

stimuli have low neighborhood density scores ([vnEt] is 2 and [fmæt] is 4), they both have

an invalid gross status, and were both rated poorly by participants in Scholes (1966). This

again predicts that they should have received low ratings from participants in the present

task. Ratings for [vn Et] are indeed quite low (around 20 total ‘good’ responses), however,

the ratings for [fmæt] are almost double that of [vnEt] at 41 total ‘good’ responses. The

higher ratings for [fmæt] may point to consonant deletion or change illusions being more

prevalent for participants in these tasks.

There are also no stop-stop clusters in the stimuli used in this experiments, but there are

obstruent-obstruent clusters like those in [ftIn] and [Sp eIl]. These kinds of clusters should

trigger vowel insertion perceptual illusions. Both stimuli have a neighborhood density score

of 2, an invalid gross status, and both are rated poorly by participants in Scholes (1966).

This should again mean that these stimuli are rated low by participants in this set of tasks

as well, and that turns out to be true for both conditions, meaning that illusory vowels do

not seem to be a common perceptual illusion for this set of data.

To summarize this discussion of perceptual illusions, it is possible that consonant deletion

33

or change illusions were present for participants here, but no real evidence of vowel insertion

or prothesis was found. Ultimately, it is difficult to verify whether these illusions are present

or not without making changes to the experimental design. One way to address this would

be to implement a stimulus transcription task into the experimental design after stimuli that

have consonant clusters that are prone to being misperceived. Such a task would involve par-

ticipants typing out exactly what they thought they heard after rating the the stimuli. This

would allow us to verify whether participants are hearing the stimulus as it was presented, or

it they are perceiving auditory illusions. Additionally, it could also function as way to check

if participants are paying attention as they go through the experiment. Recognizing whether

auditory illusions are present will help us to better understand the knowledge behind phono-

tactic judgements, since it will be clearer which sequences participants are actually using to

make judgments. Having an accurate view of this may also change the way that the models

fit to the data.

Considering the very small differences in the behavior of participants between the speeded

and non-speeded trials, it is clear that it was not possible in my experiments to separate

phonological and lexical knowledge using reaction time. However, the strong correlation be-

tween the two experimental conditions indicates that participant judgements are not random,

meaning that there is another model that I have not yet explored which could better explain

the patterns observed in this set of experiments. One possibility is the the Cohort Model

(W. D. Marslen-Wilson and Welsh 1978), which predicts that the lexicon is activated at an
extremely early stage in the process of speech recognition.1 The Cohort Model predicts that

the lexicon is almost immediately activated upon hearing an utterance, and all the possible

words that the utterance can map to are activated at once. As the speaker hears more input,

the possibilities for how many words that utterance might map to are narrowed. The main

claim of the Cohort Model is that the recognition system is able to identify possible words in

the lexicon so soon after the beginning of a word, that acoustic-phonetic input alone cannot

1Suggested by Louis Goldstein (personal communication).

34

be the only source of information in identifying sound sequences, even at early stages of

perception.

The evidence for the above statement can be attributed to speech shadowing tasks used

to measure the time it takes to recognize words in continuous speech contexts (W. Marslen-

Wilson 1973; W. D. Marslen-Wilson 1975). The results of these tasks show that words could

be accurately identified and responded to in 250-275 ms. Marslen-Wilson assigns about 75-

100 ms of this response time to processes involved in response integration and execution,

meaning that participants were initiating their responses between 150 to 200 ms after the

beginning of each word. Recall that the mean of the reaction times in the speeded version of

the experiment presented here is 934 ms, which is much later than the point that the lexicon

is hypothesized to be activated by the Cohort Model.

According to this model, even though there is a large difference between the mean reaction

times for the speeded and non-speeded conditions, the lexicon is activated too quickly in the

process of speech perception for its effects to be eliminated at earlier reaction times. The

Cohort Model might even lead us to make the opposite prediction from my initial one based

on Fox’s (1984) findings - that the more time participants have to make a decision about the

utterance the have heard, the more possible candidates which the utterance could map to

are ruled out. This would mean that participant behavior is less gradient in the non-speeded

condition than in the speeded one.

Although the Cohort Model’s predictions are based on real word recognition, there is no

reason believe that the lexicon would not also be activated at the same speed in a nonce

word context. If the activation of the lexicon is as integral to identifying possible words as

the Cohort Model claims, it should also be active in a phonotactic acceptability judgement

task which asks participants to evaluate how ‘good’ of an English-sounding word a nonce

word is.

In order to determine whether this is a more appropriate model for this set of

tasks, a computational implementation of the Cohort Model would be necessary to compare

with the models outlined in section 4.2.1, as well as the null model. Regardless, it is clear

35

that more research is necessary to understand the implications of this model for phonotactic

acceptability judgments.

36

CHAPTER 6

CONCLUSION

In an attempt to attribute gradience in phonotactic judgments to lexical information

rather than phonotactic information, I performed two variations of an acceptability judgment

task. One pushed participants to provide their judgments as quickly as possible, and one

allowed participants to provide their judgments at their own pace (inspired by Fox (1984)).

I predicted that the level of gradience in the speeded condition would be lower due to a

lessened availability of lexical information.

This appeared to be confirmed when observing the spread of the data for each condi-

tion. The speeded condition had a lower standard deviation (0.127) than the non-speeded

condition (0.133). However, as discussed in section 4.2.1, this is not a completely objec-

tive measurement of gradience. There could be many factors that contribute to the level

of gradience in participant responses that are not related to the source of their phonotactic

knowledge. To avoid mistaking gradience due to unrelated factors as significant, mean par-

ticipant responses for each stimulus were fitted to three different models which make various

predictions about the source and level of gradience in the phonotactic grammar. The result-

ing R-squared values from those models showed that there were no significant differences in

the level of variation between the speeded and non-speeded condition, indicating that lexical

and phonotactic knowledge cannot be separated via reaction time. This would suggest that

lexical access is still present at early reaction times, supporting claims about early Ganong

Effects (Rysling et al. 2015; Kingston et al. 2016).

The effects sizes for each of the models was also examined and revealed that the Max-

Ent Model (utilizes gradient phonotactic information), and the Gross Phonotactic Model

(utilizes categorical phonotactic information) performed similarly, while the Neighborhood

Density Model (utilizes lexical information), performed slightly worse. A null model with

no predictors had the largest effect size out of all the models tested, suggesting a lack of

37

systematic patterns in the results. However, a strong correlation value for response patterns

between conditions implies that the pattern of responses here is not random, but that there

may be an additional model that would better explain the patterns in participant judgments.

The Cohort Model does well in explaining the patterns observed in the present study, and it

is possible that more research which considers this model in relation to phonotactic accept-

ability judgments could be fruitful. It also confirms that varying participant response time

is not a productive means of separating the effects of phonotactic and lexical influence on

phonotactic acceptability judgments.

Regardless of the results of this set of experiments, the significance of this study is that

it explores a way to vary methodology in experimentation to contribute to theoretical views

of phonology. Specifically, it attempts to simplify the theory of phonology by finding ways

to show how certain effects (like gradience in judgments of phonotactics) may not need to

be captured and explained by the phonological grammar. They can instead be explained by

other sources, such as the lexicon, in order to lessen the scope of variation that the theory

of phonology needs to account for.

38

APPENDIX

39

APPENDIX

Neighborhood Density Gross Status Rating

valid
valid
valid
valid
valid
valid
valid
valid
valid
valid
valid
invalid
invalid
valid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
valid
invalid
invalid
invalid
valid
valid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
invalid
valid
invalid

33
33
32
32
31
31
31
29
28
28
27
27
22
20
19
14
14
13
13
13
11
11
10
10
8
8
7
7
6
6
6
6
5
4
3
3
3
3
2
2
1
1
1
0
0
0
NA
NA
NA
NA

CMUPD
# IPA
G R AH1 N
[gô@n]
1
S T IH1 N
[stIn]
2
S M AE1 T
[smæt]
3
P R AH1 N
[pô@n]
4
F L ER1 K
5
[߀k]
D R AH1 NG
[dô@N]
6
T R AH1 N
[tô@n]
7
F R AH1 N
[fô@n]
8
S N EH1 T
[snEt]
9
S P AH1 NG
10
[sp@N]
[gl@N]
11
G L AH1 NG
[mô@N] M R AH1 NG
12
SH L ER1 K
[SlÄk]
13
S K IY1 P
[skip]
14
15
[vô@n]
V R AH1 N
S R AH1 N
[sô@n]
16
[vlÄk]
17
V L ER1 K
[ml@N] M L AH1 NG
18
SH T IH1 N
[StIn]
19
20
[fpeIl]
F P EY1 L
ZH R AH1 N
[Zô@n]
21
F SH IH1 P
[fSIp]
22
SH N EH1 T
[SnEt]
23
F T IH1 N
[ftIn]
24
25
[zô@n]
Z R AH1 N
N R AH1 N
[nô@n]
26
[Smæt]
27
SH M AE1 T
[sfId]
28
[zlÄk]
29
30
[ztIn]
[fsEt]
31
[vz@t]
32
[SfId]
33
[znæt]
34
35
[fnEt]
[fkip]
36
[vtIn]
37
[zvip]
38
[fmæt]
39
40
[SpeIl]
[vnEt]
41
[Skip]
42
[ZpeIl]
43
[zkip]
44
45
[vpeIl]
[ZvIl]
46
[dwIl]
47
[Tw@p]
48
[twip]
49
50
[dwæt]

S F IY1 D
Z L ER1 K
Z T IH1 N
F S EH1 T
V Z AH1 T
SH F IY1 D
Z N AE1 T
F N EH1 T
F K IY1 P
V T IH1 N
Z V IY1 L
F M AE1 T
SH P EY1 L
V N EH1 T
SH K IY1 P
ZH P EY1 L
Z K IY1 P
V P EY1 L
ZH V IY1 L
D W IH1 L
TH W AH P
T W IY1 P
DH W AE T

18
18
13
11
11
7
12
12
7
11
9
1
4
15
4
9
2
4
3
4
4
2
2
2
5
5
1
7
2
1
4
1
1
1
2
1
2
2
4
2
2
2
1
1
2
1
7
0
5
1

Table 3: List of all stimuli used in this set of experiments along with their corresponding
CMU Glyphs, neighborhood density scores, gross status, and previous rating provided by
participants in Scholes (1966).

40

BIBLIOGRAPHY

41

BIBLIOGRAPHY

Albright, Adam (2009). “Feature-based generalization as a source of gradient acceptability.”

In: Phonology.

— (2007). “Natural classes are not enough: Biased generalization in novel onset clusters”.

In:

Albright, Adam and Bruce Hayes (2003). “Rules vs. analogy in English past tenses: A com-

putational/experimental study.” In: Cognition 90, pp. 119–161.

Armstrong, Sharon Lee, Lila R. Gleitman, and Henry Gleitman (1983). “What some concepts
might not be”. In: Cognition 13.3, pp. 263–308. issn: 0010-0277. doi: https://doi.org/
10.1016/0010-0277(83)90012-4. url: http://www.sciencedirect.com/science/
article/pii/0010027783900124.

Bailey, Todd M. and Ulrike Hahn (2001). “Determinants of Wordlikeness: Phonotactics or

Lexical Neighborhoods?” In: Journal of Memory and Language. 44, pp. 568–591.

Berg, Hugo A. Van Den (2018). “Occam’s Razor: From Ockham’s via Moderna to Modern
Data Science”. In: Science Progress 101.3. PMID: 30025552, pp. 261–272. doi: 10.3184/
003685018X15295002645082.

Boersma, Paul and David Weenink (2016). Praat: doing Phonetics by Computer [Computer

program]. Version 6.0.19, retrieved 13 June 2016 from http://www.praat.org/.

Chomsky, Noam and Morris Halle (1968). The Sound Pattern of English. New York, Evanston,

and London: Harper and Row.

Davidson, Lisa and Jason A. Shaw (2012). “Sources of illusion in consonant cluster percep-
tion”. In: Journal of Phonetics 40.2, pp. 234–248. issn: 0095-4470. doi: http://dx.doi.
org/10.1016/j.wocn.2011.11.005. url: http://www.sciencedirect.com/science/
article/pii/S0095447011001057.

Dupoux, Emmanuel et al. (1999). “Epenthetic vowels in Japanese: A perceptual illusion?” In:
Journal of Experimental Psychology: Human Perception and Performance 25.6, pp. 1568–
1578. issn: 1939-1277(ELECTRONIC);0096-1523(PRINT). doi: 10.1037/0096-1523.
25.6.1568.

Durvasula, Karthik and Jimin Kahng (2016). “The role of phrasal phonology in speech
perception: What perceptual epenthesis shows us”. In: Journal of Phonetics 54, pp. 15–
34. issn: 0095-4470. doi: http://dx.doi.org/10.1016/j.wocn.2015.08.002. url:
http://www.sciencedirect.com/science/article/pii/S0095447015000674.

42

Fox, Robert Allen (1984). ““Effect of Lexical Status on Phonetic Categorization””. In: Journal

of Experimental Psychology: Human Perception and Performance.

Ganong, William F (1980). “Phonetic categorization in auditory word perception.” In: Jour-

nal of experimental psychology: Human perception and performance 6.1, p. 110.

Gorman, Kyle (2013). “Generative Phonotactics”. PhD thesis. University of Pennsylvania.

Hayes, Bruce and Colin Wilson (2008). “A Maximum Entropy Model of Phonotactics and
Phonotactic Learning”. In: Linguistic Inquiry 39.3, pp. 379–440. doi: 10.1162/ling.
2008.39.3.379.

Kingston, J. et al. (2016). “Eye movement evidence for an immediate Ganong effect.” In:
Journal of experimental psychology. Human perception and performance 42 12, pp. 1969–
1988.

Marslen-Wilson, W. (1973). “Linguistic Structure and Speech Shadowing at Very Short La-

tencies”. In: Nature 244, pp. 522–523.

Marslen-Wilson, William D. (1975). “Sentence Perception as an Interactive Parallel Process”.
In: Science 189.4198, pp. 226–228. issn: 0036-8075. doi: 10.1126/science.189.4198.
226. eprint: https://science.sciencemag.org/content/189/4198/226.full.pdf.
url: https://science.sciencemag.org/content/189/4198/226.

Marslen-Wilson, William D. and Alan Welsh (1978). “Processing Interaction and Lexical
Access during Word Recognition in Continious Speech”. In: Cognitive Psychology 10,
pp. 29–63.

Palanab, Stefan and Christian Schitter (2018). “Prolific.ac—A subject pool for online exper-
iments”. In: Journal of Behavioral and Experimental Finance. doi: https://doi.org/
10.1016/j.jbef.2017.12.004.

Peirce, J. W. et al. (2019). “PsychoPy2: experiments in behavior made easy”. In: Behavior

Research Methods. doi: 10.3758/s13428-018-01193-y.

R Core Team (2019). R: A Language and Environment for Statistical Computing. R Foun-
dation for Statistical Computing. Vienna, Austria. url: https://www.R-project.org/.

Rysling, Amanda et al. (2015). “Early Ganong effects”. In: ICPhS.

Sarver, Isaac (2020). “A Systemic Evaluation of Computational Models of Phonotactics”.
MA. These. Michigan State University. doi: https://doi.org/10.25335/5mch-cz06.

Scholes, R. J. (1966). Phonotactic grammaticality. The Hague: Mouton.

43

Shademan, Shabnam (2006). ““Is Phonotactic Knowledge Grammatical Knowledge?””. In:

Proceedings of the 25th West Coast Conference on Formal Linguistics.

Vaden, K.I., H.R. Halpin, and G.S. Hickok (2009). Irvine Phonotactic Online Dictionary,

Version 2.0. [Data file]. Available from http://www.iphod.com.

Vitevitch, Michael S. and Paul A. Luce (1999). “Probabilistic Phonotactics and Neighbor-
hood Density in Spoken Word Recognition”. In: Journal of Memory and Language 40,
pp. 374–408.

Weide, Robert L. (1994). CMU Pronouncing Dictionary. http://www.speech.cs.cmu.edu/cgi-

bin/cmudict.

Wickham, Hadley et al. (2019). “Welcome to the tidyverse”. In: Journal of Open Source

Software 4.43, p. 1686. doi: 10.21105/joss.01686.

44