Chapter 1. Introduction
1.3 The role of verb bias in sentence processing
Previous studies have investigated the role of verb bias in sentence processing
(Trueswell et al.,1993; Garnsey et al, 1997; Wilson & Garnsey, 2009), and attempted to
validate the two-stage model and the constraint-based model. Trueswell et al. (1993)
in their series of experiments suggested that verb subcategorization information was
accessed rapidly in sentence processing. They conducted a norming study to classify the
verbs that are strongly biased to certain subcategorization. With the result of sentence
completion test in which participants had to complete the sentence fragment “Subject
Verb ___________” (e.g. John insisted _________.), eight verbs were classified as
Noun Phrase (NP) biased verbs and eight verbs as Sentential complement (S) biased
verbs. Each of the verbs was adopted to construct the auditory sentence fragment “the
noun phrase + either verb + (that)” (e.g. The old man insisted (S-bias verb) /observe
(NP-bias verb). Each fragment was then paired with two targets: the nominative
pronoun he and the accusative pronoun him (e.g. The old man insisted him/he; observed
him/he). After the target was shown on the screen, the participant had to name the target word as soon as possible. The result showed that the naming times were fastest on him
following NP-bias verb and slowest on him following S-bias verb, suggesting verb bias
subcategorization information was immediately accessed. In their subsequent
experiment which investigated how subcategorization information was used in resolving
the ambiguous noun phrase (e.g. (NP-bias: The student forgot (that) the solution was in
the back of the book.; S-bias: The student hoped that the solution was in the back of the
book.), they proved that subcategorization information was used to determine whether
the attachment of a noun phrase was the NP complement or the subject of a sentence
complement. On the be (e.g. was) verb, in sentence with NP-bias verb, larger reading
times was shown on the sentence which did not contain that than the sentence
containing that. However, no such difference was found in sentence with S-bias verb.
The result demonstrated that subcategorization information was used to determine the
role of the noun phrase – either the NP complement or the subject of the sentence
complement.
However, Garnsey et al. (1997) suggested that the design of Truewell and his
colleagues (2003) was problematic. In Truewell et al. (2003)’s study, in each sentence
set, the plausibility of the post-verbal noun phrase (e.g. the directions) to the main verb
were not balanced (8). For instance, nouns that were plausible as direct object of the
NP-bias verb (e.g. remembered) were sometimes less plausible for S-bias verbs (e.g.
claimed).
(8) Examples of Truewell et al. (2003)’s study a. Mr. Smith remembered the directions b. Mr. Smith claimed the directions…
Hence, Garnsey et al. (1997) conducted a self-pace reading task which
concerned the plausibility of the post-verbal noun phrase and increased the amount of
the selected verbs. In their studies, forty-eight verbs were selected and each of them
were constructed four sentence versions (9). In 9(a), the “decision” was plausible as
the direct object of “regretted”, but the “reporter” was not.
(9) Examples of the sentence structure in Garnsey et al. (1997)’s study a. The senior senator regretted (that) the decision had ever been made public.
b. The senior senator regretted (that) the reporter had ever seen the report.
The result presented that verb bias plays an important role in initial sentence
processing. For instance, in sentence with direct-object (DO) bias verbs, first-pass
times on the temporarily ambiguous noun phrase (e.g. “the reporter) were slower than
the same noun phrase in an unambiguous sentence (e.g. “that the reporter”) and the
plausible noun phrase in an ambiguous sentence. (e.g. the decision). This reflected that
the noun phrase was interpreted as the direct object after these verbs. Therefore, this
study also supported that verb bias can rapidly resolve the temporarily ambiguity and
supported the constraint-based model.
In Wilson and Garnsey (2009)’s study, self-paced reading task and
eye-movement experiment were also conducted to investigate the verb bias. Three types of
sentence were constructed for each verb type (DO-and SC-bias verb): a. verb followed
by direct object continuation b. verb followed by sentential complement continuation. c.
verb followed by the complementizer that and sentential complement continuation (10).
Self-paced reading task revealed the longer reading time on the mismatch between the
verb bias and sentence continuation. The direct object continuation was read
significantly slower after SC-bias verbs than DO-bias verbs. In contrast, sentence
complement continuation following DO-bias verbs was read significantly slower than
SC-bias verbs. This result provided evidence that verb bias guides the early stage of
sentence comprehension. The same paradigm was also used in the eye-tracking
experiment.
The result showed that when readers encountered the sentential complement
following the DO-bias verb, they slowed down on the disambiguation region (the
underlined region) and re-read earlier sentence regions; however, when they read the
direct object following the SC- bias verb, they directly went back and re-read earlier
regions. The findings supported the constraint-based model. If the simplest structure is
expected at the initial stage, there is no need of re-reading in the condition of direct
object continuation. Yet, the readers slowed down when they encountered the sentential
complement followed by DO-bias verb. This supported that verb bias led the readers to
expect the potential continuations following DO-bias verb. Hence, in parsing sentences,
verb bias immediately comes into play. This
An ERPs study conducted by Osterhout and Holcomb (1992) found that both SC-bias verb
a. The ticket agent admitted the mistake because she had been caught.
b. The ticket agent admitted the mistake might not have been caught.
c. The ticket agent admitted that the mistake might not have been caught.
DO-bias verb
a. The talented photographer accepted the money because he was asked twice.
b. The talented photographer accepted the money might not be legally obtained.
c. The talented photographer accepted that the money might not be legally obtained.
(10) Examples of Wilson & Garnsey’s (2009) experiment
P600 effect. In the ungrammatical condition, the auxiliary verb in the sentence
containing transitive verb (e.g., “He forced the patient was lying.”) elicited a more
positive-going waveform within 500-800ms window than that in the sentence
containing intransitive verb (e.g., “He hoped the patient was lying.”). As for the
condition of less-preferred but grammatical structure, the auxiliary verb in the sentence
containing transitively biased verb (e.g., “He charged the patient was lying.”) elicited a
more positive-going waveform than that in the intransitively biased verb condition (e.g.,
“He believed the patient was lying.”). The finding indicated that the parsers can use the
information about verb subcategorization during sentence processing.
So far, only a few studies have discussed the verb bias effect in Mandarin RC
processing. Lin and Garnsey (2011) have found a verb bias effect on complex sentence
processing in Mandarin. Verbs were chosen and categorized based on the corpus study
done by Lu and Garnsey (2008,2009) in which fifty sentences for each verb were
selected from Chinese Gigaword and then hand-coded. Two types of verb bias (DO-bias
and SC-bias verbs) were manipulated as the main verb in the sentence with Mandarin
objective relative clauses (11).
The result showed longer reading time on the second verb (e.g. 痛罵/scold)
after the DO-bias verb. This difficulty lasts on the subsequent words, relative clause
marker DE, and RC head noun (e.g.學生/student). When reading DO-bias verb in the
main clause, readers did not expect another verb and thus slowed down when another
verb appeared. Therefore, the finding supported that verb bias influence readers’
expectation on the complex sentence processing.
1.3 Current study
In view of the issues mentioned above, the goal of this study is to investigate the
role of verb bias which carries both semantic and syntactic information in real-time
Mandarin RCs processing. This study will apply the Event-Related Potentials (ERPs)
measurement to achieve this goal due to its high temporal resolution. Several ERPs (11) Examples of sentence structure
P600, and frontal negativity. N400, a negative-going waveform peaking around 400ms,
was characterized as reflecting the semantic processing (Kutas & Hillyard,1980a,
1980b; Kutas & Federmeier, 2000). N400 enhances for words that do not fit to the
previous semantic context. In a classic study, larger amplitude of N400 was found on
“cry” that was semantically incongruent to the sentence (e.g. “The pizza was too hot to
cry”) than “eat” that was congruent (e.g. “The pizza was too hot to eat” ) (Kutas &
Hillyard,1980). The P600, positive-going waveform peaking around 500ms, indexes the
syntactic processing (Osterhout and Holcomb,1992; Hagoort et al. 1993). Syntactic
violation elicited the P600. For instance, the ungrammatical sentence elicited P600s
than the grammatical sentence. “to” following the transitive verb which leads to
ungrammaticality (e.g. “The woman persuaded to answer the door.”) elicited P600 than
that following the intransitive verb “The woman struggled to prepare the meal”. As for
frontal negativity, it was shown to reflect thematic-role ambiguity and (King & Kutas,
1995), the process of establishing reference (Barkley et al., 2015; Van Berkum et al,
2007; Nieuwland et al., 2006). Given the knowledge on the ERPs components
regarding to language processing, two ERPs experiments would be conducted. The
possible effects observed on those ERPs components might be able to address this issue.
The purpose of the first experiment is to examine the verb bias effect on online ORCs
processing. The second experiment focuses on how verb bias influences online SRCs
processing. With the use of ERP technique and the understanding about the verb bias
effect on ORCs and SRCs, this study attempts to discuss how verb bias incrementally
affects the real-time Mandarin RCs processing. Furthermore, “word order” might also
be an intervening (/or crucial??) role in this issue. The importance of word order in
Mandarin has been underlined by researchers (Chao, 1968; Chen 1995; Ho, 1993).
Mandarin relies heavily on word order as an underlying marking feature for meaning
(Ho, 1993) since Mandarin lacks case and agreement markings (Chao, 1965; Chen
1995; Ho, 1993). Moreover, word order plays an important role in information
structuring (Chen, 1995). Given that ORCs and SRCs are different in their word order;
that is, ORCs follow Mandarin word order, but SRCs do not, the role of “word order” in
RCs processing should be also taken into account.
Chapter 2. Event-Related Potential Studies: The influence of verb bias on ORC processing
Current study aims to investigate the verb bias effect on real-time ORCs processing.
The norming study would be conducted to categorize types of verb bias. Each type of
verb bias would be followed by an object-modifying Object Relative Clause (ORC).
ERP analysis for four regions of ORC would be carried out to examine the incremental
influence of verb bias.
2.1 Materials and Methods
2.1.1 Participants
Thirty-four right-handed undergraduate and graduate students between the age of 18
and 28 participated in this study. Participants were all native speakers of Mandarin
Chinese in Taiwan with no history of neurological or psychiatric disorders. Each
participant signed the consent before the experiment and were paid $500 NT of their
participation. All participants were evaluated for their working memory and verbal
memory capacities with working memory test of the Wechsler Adult Intelligence
Scales-fourth edition (WAIS-IV; Weschler, 2008) and reading span test.
(1) Working memory test (WAIS-IV): This test measure participants’ digit span.
Participants were required to memorize and recall the numbers in forward,
backward and ascending order. The total score of working memory is 48.
(2) Reading span: In this test, participants were required to read aloud a series of
sentences (series of 2, 3 4, 5, 6 sentences), and at the end of that series, they had to
recall the final word of each sentence. After recalling the final words, they had to
answer one comprehension question. The total score of reading span ranges from 1
to 6 (0.5 point as a score unit). This test has been shown to correlate with
participants’ reading ability (Baddeley, 1979; Daeneman & Carpenter, 1980; King
&Just, 1991). Table 2.1 presents the results of the behavioral tests.
Table 2.1. The result of behavioral tests
Behavioral tests Mean score Range
WAIS-IV (working memory capacity) 35.68 (3.97) 28-43
Reading span test 2.51 (0.77) 1.5-4.5
Note. Standard deviations are in parentheses.
2.1.2 Experimental Stimuli
This study aimed to manipulate three types of verb bias, Direct Object (DO),
Sentential Complement (SC), and Equilibrium Balanced (EQ) bias verb in the ORC
sentences. All the target sentences were constructed with the following syntactic
structure of the target sentences: Subject + DO/SC/EQ-bias main verb + ORC structure
(embedded RC noun + RC verb + RC marker DE + head noun). To achieve this goal,
we first conducted a norming procedure to quantify the verb bias.
Norming study of verb bias
44 high frequency verbs that can take both direct object and sentential complement
and have a frequency of 40 or greater per million words were selected from Academia
Sinica Balanced Corpus of Modern Chinese, version 4.0 (Sinica Corpus). The reason
for using the high-frequency verb is to ensure that there would be sufficient sentences
using that using that particular verb in the corpus for the analysis on the classification of
verb bias. The mean frequency of the high frequency verbs is 133 (± 153). The total of 300 sentences using that particular verb were extracted and the continuation following
the verb was then hand coded by two raters: DO for direct object, SC for taking
sentential complement, and Other for not taking either DO or SC or for not easily to be
reconstructed. If the counts of the given verb taking Other category exceeded the
threshold of 25%, this verb was excluded. These verbs were then classified into three
types of biases: DO-bias, SC-bias, and EQ-bias verb, based on the following definitions.
Verbs were classified as SC-bias if they occurred at least twice as often with a sentential
complement than with a direct object and categorized as DO-bias if they appeared at
least twice as often with a direct object than with a sentential complement. Verbs were
classified as EQ-bias if they were approximately equally followed by direct objects and
by sentential complements. Examples of the annotation process and the result of the
classification are shown in Table 2.2.
Table 2.2. Examples of the classification of the verb bias Types of
verb bias
Examples of annotation process Percentage of taking either types of
With the criteria mentioned above, three verbs were excluded, two for taking over
25% of the Other syntactic patterns, and one for inconsistent coding results between two
raters. Thus, 14 DO-bias, 13 SC-bias, and 14 EQ-bias verbs with medium to high
interrater agreement (DO-bias verb: kappa = 0.8, p >.05 ; SC-bias verb : kappa = 0.9;
DO-bias verb: kappa = 0.8, p >.05 and EQ-bias verb: kappa = 0.3, p >.05) and no
differences on the word frequency were thus chosen to construct target sentences (F=
0.9, p >.05) (Table 2.3).
Table 2.3. The characteristics of verb bias
Percentage of taking either continuations
Frequency of the verb (per million words) DO-bias verb 76% (taking direct objects) 113±111
SC- bias verb 75% (taking sentential complements) 167±190 EQ- bias verb 45%,45% (taking direct objects;
sentential complements)
96±96
The construction of target sentences and fillers
The present experiment consisted of 41 target sentences with the manipulation of
14 DO-bias, 13 SC-bias, and 14 EQ-bias verbs as well as 27 fillers (Table 2.4). Each
critical region of the ORC, were matched for word length, word frequency and the associations between critical words with no differences between conditions. The
embedded RC noun, RC verb, and head noun were all two-character words. The word
frequency of embedded RC noun and head noun in each condition was low-medium
word frequency, the frequency of 58 or greater per million words, with no differences
between conditions (embedded RC noun: F = 0.165, p > .05; head noun: F= 0.38, p
> .05). To further control the word predictability, the association between words were
computed using word2vec (Mikolov et al., 2013). The values of the association between
words, as shown in Table 2.5, including association between main verb and embedded
RC noun (F = 0.019, p > .05), as well as association between head noun and main verb
(F= 2.267, p >.05), RC verb (F= 1.87, p >.05), and embedded RC noun (F=0.621, p>
0.5) has to be lower than the value 0.3 (more than 0.3: high associations between words,
0.2-0.1: medium-low associations, lower than 0.1: low associations) with no differences
between conditions.
Additionally, 27 filler sentences were created for preventing the participants from
developing strategies, including three syntactic structures—a. simple SVO structure:
subject + main verb + (adjectives) object b. subject + main verb + sentential
complement with embedded ORC/SRC, and c. subject + main verb + sentential
complement (Table 2.4).
Table 2.4. Examples of target and filler sentences Target sentences
Main Clause Objective Relative Clause (ORC)
Subject Main verb embedded RC verb RC Head
verb biases)
Types of filler sentences Examples subject + main verb + (adjectives) object 他質疑長官。
subject + main verb + sentential
complement with embedded ORC/SRC
他否認黨團提名立委為候選人。
subject + main verb + sentential complement
他擔心孩子想不開。
Table 2.5. The characteristics of the stimuli Frequency (per million words) SC-bias 22.10±16.50 94.52±71.85 17.30±10.90 EQ-bias 25.80±19.98 73.45±74.16 18.40±14.76
Association values One-third of the experimental sentences were followed by one true/false
comprehension question to ensure that participants payed attention on reading
comprehension during the experiment.
2.1.3 Predictions
The first experiment aims to shed light on how verb bias influences the ORC
processing. It was hypothesized that the sentences with DO-bias verbs that were
expected to be followed by a direct object, would be more difficult to process than the
sentences with SC-bias verbs that were expected to be followed by a sentential
complement. This processing difficulty may be shown on RC verb and head noun.
Predictions on RC verb and head noun were listed as follows:
I. RC verb
The difficulty may be first shown on the RC verb which may be indexed by
P600 or frontal positivity. DO-bias verbs were expected to be followed by a direct
object, so the appearance of the RC verb may be a non-preferred structure that may
lead to larger P600 or frontal positivity as compared to that followed by SC-bias
verbs.
II. Head noun
When parsing the sentences with DO-bias verbs, readers might engage in the
process of referential binding. That is, when they found out that the embedded RC
noun they read was not the direct object of the main verb, they had to look for the
actual direct object for the main verb. direct object for the main verb. This process
difficulty of searching for the direct object and establishing the referential binding
might reflect on N400 and frontal negativity, respectively.
Aside from DO-biased and SC-biased verbs, some verbs do not have clear
tendency of taking either more direct objects or sentential complements. Those
verbs were classified as EQ-bias verbs. The sentences with EQ-bias verbs may
serve as a baseline, to be compared with DO-biased and SC-biased verbs
conditions. Or alternatively, they may exhibit a similar processing pattern to either
DO-bias or SC-bias condition. In order to delve into the role that each verb bias
plays in the ORC processing, contrasts between conditions, including – SC-DO,
SC- EQ, and DO-EQ bias contrasts, were be performed on N400 and frontal
positivity.
2.1.4 Procedure
Each participant first underwent the ERPs experiment and then three consecutive
behavioral assessments. For the ERPs experiment, participants were given 13 trials for
practice session, 67 randomized experimental trials in two sessions. Participants were
seated 90cm from the computer screen in a quiet room, with visual sentences presented
centrally. Each trial began with the fixation ”+” for 200ms. Sentences were presented
word by word. Each word displayed for 600ms with 400ms inter-stimulus interval (ISI).
The sentence ended with a period. If the participants were ready for the next trial, they
pressed the “enter” button; otherwise, the period marker “。” lasted for 400ms.
One-third of the sentences were followed by a comprehension question. When the
One-third of the sentences were followed by a comprehension question. When the