Semantic processing when syntactic requirement is not satisfied during reading Chinese A-not-A questions

(1)

during reading Chinese A-not-A questions

Yue Qin

A thesis submitted in partial fulfillment of the requirements for the degree of

Master of Science

(Clinical Linguistics)

at the Joint European Erasmus Mundus Master’s Programme in Clinical Linguistics (EMCL+)

UNIVERSITY OF GRONINGEN

(2)

during reading Chinese A-not-A questions

Yue Qin

Under the supervision of Dr. Srdjan Popov at the University of Groningen and Dr. Jinxing Yue at Harbin Institute of Technology

(3)

ACKNOWLEDGEMENTS

I want to first expresses my most sincere gratitude to my supervisors Dr. Srdjan Popov and Dr. Jinxing Yue. They gave timely and careful guidance throughout the process of thesis writing, from topic selection, data collection to the completion of this thesis. Dr.. Jinxing Yue invested a lot of effort and put forward many useful suggestions during the process of creating stimuli and analysing data. Dr. Srdjan Popov patiently listened to my progress reports in thesis classes and gave feedback in time to ensure the smooth progress of the project. I would like to thank my supervisors for their careful guidance on academic writing, from the structure of the article to the punctuation and spaces.

Thanks to all the people who spared their valuable time reading weird sentences in both pretests and the self-paced reading experiment. Specially, I would like to thank my friends and classmates during undergraduate studies (Lin Gao, Xuejie Bao, Fei Du, Zeyu Pan & Wenqian Li), current and previous EMCL students (Qingqing Guo, Dingxing Zhang & Jindan Yang), the new friends I met in Europe (Yanmiao Li & Pengxiang Cheng) for their generous help to spread the questionnaire and to recruit

(4)

participants for my experiment.

I am grateful to all the teachers who taught me and have given me endless enlightenment as well as all the EMCL students in the 2018-2020 cohort who accompanied me through this wonderful and unforgettable journey in life.

Finally, I would like to thank my family whose understanding and encouragement is the motivation for me over the two years of studying abroad. Without their support, this thesis would not have been successfully completed.

(5)

ABSTRACT

Existing theoretical models of language comprehension agree on the idea that various sources of information including syntax and semantics have to be integrated to guarantee successful understanding of sentences. However, the debates on the interaction between syntactic and semantic information is still ongoing. The current study aims to investigate how semantics of a constituent word is processed in grammatical constructions. Taking advantage of the A-not-A construction, a highly formal syntactic structure in Mandarin Chinese, a self-paced reading paradigm was conducted to compare participants’ performance on reading interrogative questions with grammatical or ungrammatical A-not-A constructions. The syntactic violation was constructed by replacing the second A with another verb B to build two ungrammatical A-not-B constructions in which the constituent word B was either semantically congruent (B1) or incongruent (B2) with the global context. Two hypotheses about semantic processing of this special construction were examined with reaction time data. the purely morphological hypothesis argues that once the first A is presented, the processing of the second A should be highly morphological. Therefore,

(6)

no semantic integration is expected for the second A within A-not-A constructions, which is in line with the syntax-first hypothesis which claims that unsuccessful syntactic analysis will block follow-up semantic processing. Alternatively, the integrative view of sentence comprehension insists that various language sources of information including the semantic context have to be integrated in time during language comprehension. According to this hypothesis, a graded semantic processing of the constituent word with the global context should be observed for the second A. Results in this study showed that mean reading times were significantly shorter in the grammatical condition but did not differ within the two ungrammatical conditions, one with semantic congruence and another being incongruous to the global context. This provides supporting evidence for the purely morphological hypothesis, which suggests an absence of semantic processing in constructions with highly formal constraints. Findings from the current study challenge the universality of the integrative view of language processing in psycholinguistics, calling for the verification of existing theoretical models and the establishment of more comprehensive theories of language comprehension, with a focus on highly formal constructions.

(7)

LIST OF TABLES

Table 1.1 Summary of theoretical models of language comprehension...18 Table 2.1 Examples of sentences for baseline and experimental conditions...39 Table 3.1 Descriptive statistics for reaction times at target word positions across conditions... ..48

(10)

LIST OF FIGURES

Figure 1.1 Three language related ERP components... 16 Figure 1.2 Syntactic structure of a declarative sentencewith interrogative features....30 Figure 1.3 Movement of [+Q] feature in A-not-A construction...31 Figure 3.1 Mean reaction times at verb positions across conditions...49 Figure 3.2 Mean reaction times at the post second-verb position across conditions... 50

(11)

Chapter 1 Introduction

Sentence comprehension requires the listener/reader to quickly integrate various sources of linguistic information including phonological, morphological, syntactic, and semantic information. At present, researchers have reached a wide consensus on the following two aspects. First, sentence processing is compositional and dynamic in that words’ meanings are accessed linearly and incrementally (Hagoort, 2005; Werning, Hinzen, & Machery, 2011). In linguistic analysis, comprehenders follow the principle of immediacy, rather than waiting until the complete sentence is read or heard. Second, in the process of sentence comprehension, the parser not only takes advantage of the grammatical structure of the sentence (e.g., Ferreira & Clifton, 1986; Frazier, 1979, 1987; Frazier & Rayner, 1982), but also comprehensively use multiple pieces of information such as syntactic features of the words (e.g., Trueswell & Tanaenhaus, 1994; Hagoort, Wassenaar & Brown, 2003; Friederici, Gunter, Hahne, & Mauth, 2004), semantic information of vocabulary and sentence (e.g., Taraban & McClelland, 1988; Kounios & Holcomb, 1992; Kutas & Hillyard, 1980), contextual

(12)

information (e.g., Glucksberg, Kreuz, & Rho, 1986; Spivey-Knowlton, & Tanenhaus, 1998; see Simpson, 1994 for a review), and statistical information about frequency of certain sentence structure (see Ellis, 2002 for a review). However, it remains a debated issue as to how these different sources of information are integrated during language comprehension. This chapter will first give a brief introduction of several representative theoretical models of sentence comprehension, and then provide supporting evidences of models based on behavioural and neuroimaging studies. Finally, we will introduce the current study on the semantic processing of A-not-A questions in Mandarin Chinese.

1.1 Theoretical models of sentence comprehension

Models of language comprehension all agree that syntactic and semantic information have to be integrated in time to facilitate understanding the whole sentence, but show controversy on how these processes interact with each other (Hagoort, 2003). In the field of syntactic analysis, the two types of most classical models are the two-stage serial model, as represented by the garden path model (e.g., Ferreira & Clifton, 1986; Frazier, 1979, 1987; Frazier & Rayner, 1982) and the parallel one-stage model (also called parallel processing model), as represented by the constraint-based model (e.g., MacDonald, 1994; Trueswell & Tanaenhaus, 1994). More recently, scholars have proposed the unrestricted competition model (Traxler, Pickering, & Clifton, 1998; Van Gompel, Pickering, & Traxler, 2001; Van Gompel, Pickering,, Pearson, & Liversedge, 2005) and the three-stage model (Friederici, 2002). We will briefly review these

(13)

models one by one.

1.1.1 Serial two-stage model

Serial two-stage models include garden path model (Frazier & Fodor, 1978; Frazier, 1979, 1987; Frazier & Rayner, 1982), a computational model proposed by Abney (1989), an incremental comprehension model proposed by Crocker (1996), Pritchett’s model (1992), etc. This type of models postulates that in the process of language comprehension, syntactic analysis needs to go through two stages. In the first stage, only limited information is used to quickly establish the initial analysis. Subsequently, various information needs to be integrated to process the sentence.

One of the most important and classical two-stage models of language comprehension is the garden path model (Frazier, 1979, 1987; Frazier & Rayner, 1982). The first stage of the garden path model involves purely structural processing. Only syntactic information (e.g., part of speech) is used to quickly establish the initial syntactic structure. At this stage, the process of syntactic analysis follows two principles, namely, the Minimum Attachment Principle and the Late Closure Principle (Frazier & Fodor, 1978). The second stage is the semantic processing stage, which uses non-structural information such as global contexts and argument structure of verbs for syntactic analysis.

1.1.2 Parallel one-stage model

Since the 1990s, some scholars have found that in addition to syntactic information, lexical information, frequency of certain sentence structures, and

(14)

contextual information also make contributions during the first stage of sentence processing (e.g., Trueswell & Tanaenhaus, 1994; Spivey-Knowlton, & Tanenhaus, 1998). Therefore, they proposed parallel one-stage models. One of the most well-known and influential models is the constraint-based model (also called constraint satisfaction model, MacDonald et al., 1994; McRae, Spivey-Knowlton, & Tanenhaus, 1998; Spivey & Tanenhaus, 1998; Tanenhaus & Trueswell, 1995; Taraban, & McClelland, 1988). This model regards all kinds of information as constraints to establish a syntactic structure. Therefore, the establishing of a syntactic structure is a process to satisfy these constraints.

In the process of analysing sentences, the parser simultaneously employs syntactic information and non-syntactic information to activate various analyses. The degree of activation of different analyses is decided by the strengths of multiple supporting streams of information. However, instead of immediately choosing one analysis and giving up the others, various analyses are activated at the same time, but the analysis with the highest activation intensity will be located in the most prominent place. An analysis is selected as the final analysis only if it reaches the specific degree of activation. In this model, various analyses will compete with each other for a long time (MacDonald et al., 1994). Therefore, the competition between multiple analyses with similar degrees of activation results in greatest processing difficulty of syntactic analysis in ambiguous sentences.

The experimental basis of the parallel one-stage model first came from studies examining the role that semantic information plays in the initial stage of syntactic

(15)

analysis. For example, Taraban and McClelland (1988) compared the reading times of sentences like in (1) through a self-paced reading experiment.

(1a) The thieves stole all the paintings in the museum while the guard slept.

(1b) The thieves stole all the paintings in the night while the guard slept.

The prepositional phrase in the museum in sentence (1a) should be linked to the noun phrase the paintings, and the prepositional phrase in the night in sentence (1b) should be linked to the verb stole. According to the garden path model, sentence (1b) should have a shorter reading time than sentence (1a) which violates the Minimal Attachment Principle. However, the results of the experiment happened to be contrary to expectations. Taraban and McClelland (1988) provided a possible explanation that the syntactic parser uses semantic information to build sentence structure from the first stage of syntactic analysis, thus avoiding reanalysis of the initial representation.

In addition to semantic information, the thematic role of verbs also affects the first stage of syntactic analysis, thereby eliminating the garden path effect. For example, Trueswell et al. (1994) conducted an experiment using sentences The fossil / archaeologist examined...as experimental materials. Measurements on eye movements and reading times of the subjects revealed that people tend to regard The fossil examined... as a reduced relative clause when they read the verb examined, and view The archaeologist examined… as a subject-verb structure. This showed that when subjects see the same verb, they immediately have a tendency to build different sentence structures based on thematic information of the verb.

(16)

For the above research results, proponents of the garden path model represented by Frazier have different interpretations (Frazier & Clifton, 1996). They believe that the above results were produced because the second stage of the serial model started very quickly, and that the experiments that aimed to investigate the first stage of sentence comprehension were actually examining the second stage. In sum, when will the semantic and syntactic information interact with each other during sentence comprehension remains a problem to be studied.

1.1.3 Unrestricted competition model

Apart from the above two types of classical models, some scholars (e.g., Traxler et al., 1998; van Gompel et al, 2001, 2005) proposed an unrestricted competition model. The model is proposed mainly based on studies investigating the process of disambiguation when subjects read ambiguous sentences. Van Compel et al. (2001) investigated the eye movements and reading times of their subjects and found that the reading times of ambiguous sentences were shorter than those of unambiguous sentences, causing less processing difficulty. These results were difficult to explain with the garden path model or the constraint-based model. According to the garden path model, the reading times of the unambiguous sentence should be the shortest. However, the experimental results did not match this prediction. According to the constraint-based model, ambiguous sentences should cause the greatest competition due to almost the same activation degree between the establishment of two sentence structures. The processing difficulties of sentences with ambiguity should be the

(17)

largest, reflected by the longest reaction times. This prediction was also inconsistent with the experimental results.

In order to explain the above results, the unrestricted competition model (Traxler et al, 1998; van Gompel et al, 2001, 2005) was proposed. Similar to the constraint-based model, various information including syntax and semantics are thought to interact with each other to establish multiple initial analyses in the unrestricted competition model. But only the one with highest consistency is selected. In other words, this model chooses only one analysis at a time, rather than keeping different analyses competing with each other, which is different from the constraint-based model. This model effectively integrates the garden path model and the constraint-based model. In the process of reading ambiguous sentences, no matter what kind of analysis is used in the initial stage of syntactic analysis, it is compatible with the following information. Therefore, the processing difficulties caused by ambiguous sentences should be minimal.

1.1.4 Three-stage model

The point of dispute between the garden path model or the constraint-based model lies in sentence processing behaviors in different stages. Friederici (2002) proposed a neural model called the three-stage model on the basis of event-related potential (ERP) research studying the neural mechanism of auditory sentence processing. Electrophysiological techniques (i.e., ERPs) can test the brain's response to specific materials with a millisecond accuracy, thereby helping solve the debates

(18)

between serial twp-stage or parallel one-stage models of language comprehension.

Figure 1.1

Three language related ERP components

Note. (a) N400, (b) very early left-anterior negativity (ELAN), and (c) P600. Solid lines represent the correct condition, and dotted lines the syntactic and semantic violations. Reprinted from ‘Towards a neural basis of auditory sentence processing’, by A. D. Friederici, 2002, Trends in cognitive sciences, 6(2), p.78-84. Copyright 2002 by Elsevier Science Ltd.

As is shown in Figure 1.1, syntactic and semantic violations produce different ERP responses. Traditionally, sentences with semantic violations can trigger a centro-posterior bilateral negativity observed 400 miles after the target word in the sentence, the so-called N400 effect (Kounios & Holcomb, 1992; Kutas & Hillyard, 1980). Syntactic violations cause a different response depending on the situation. If the violation is perceived in the early stage, that is, 100-500 milliseconds after the onset of the target word, the word category violation will trigger the Early Left

(19)

Anterior Negativity (ELAN) while morphosyntactic errors such as agreement violations will induce LAN effects (Friederici, Hahne, & Mecklinger, 1996; Münte & Heinze, 1994; Münte, Heinze, & Mangun, 1993). In the later stage of sentence analysis, that is, 600-1000 ms after the onset of the target word, repair and reanalysis takes place (e.g., garden path sentences that require syntactic revision) leading to the P600 effect (Friederici, 2002).

According to the three-stage model, the first stage is a structural analysis stage located between 100 to 300 ms (after stimulus onset). The initial sentence structure needs to be established based on the part of speech information (e.g., whether the target word is a noun or a verb). The second stage is between 300 and 500 ms when the processing of lexical-semantic and morphosyntactic information takes place in parallel without interacting with each other. In the third stage between 500 and 1000 milliseconds, the parser integrates various sources of information to establish the final sentence structure. Thus, in the first stage, syntactic processing of the sentence structure information precedes and is independent of processing of semantic features. However, in the third stage, syntactic and semantic information is finally integrated together on the sentence level.

To sum up, a central issue under debate in online language processing is the relationship between syntactic and semantic information. Existing models differ in terms of how syntactic and semantic processing interact with each other. Table 1.1 shows a summary of these models of language comprehension.

(20)

Table 1.1

Summary of theoretical models of language comprehension

Models Syntax Semantics Syntax-semantics interaction Serial two-stage

model (Garden path model)

Stage 1 Stage 2 Interact at a late stage

Stage 1: purely structural analysis Stage 2: semantic integration and it can affect initial syntactic analysis

Parallel one-stage model

(Constraint-based model)

Stage 1 Stage 1 Interact from an early stage

Initially employ syntactic and non-syntactic information to activate various analyses; select one analysis later

Unrestricted competition model

Stage 1 Stage 1 Interact at an early stage

initially employ syntactic and non-syntactic information to activate various analyses; select one analysis immediately

Three-stage model Stage 1, 3

Stage 2 Interact at a late stage Stage 1: syntactic structure

Stage 2: independent processing of semantic and morphosyntactic information

Stage 3: integration of different types of information and syntactic revision

(21)

The serial, as well as the syntax-first models (e.g., Garden path model & Three-staged model), propose that syntactic information is processed independently of semantic information at an early stage and that semantic information could have influence on the initial structural analysis at later stages of sentence comprehension (Frazier, 1987; Friederici, 2002). In contrast, interactive models (e.g., Constraint-based model) propose that various types of linguistic information, including syntax and semantics, are activated and processed in a parallel and interactive manner throughout comprehension processes (MacDonald et al., 1994; McClelland et al., 1989; Trueswell & Tanenhaus, 1994). As to how syntactic and semantic information interact and integrate into the global context, the answer remains unclear. More research is needed to verify different models on language comprehension.

1.2 Previous research on the syntax-semantics interaction

In the previous section, we discussed some representative and influential syntactic analysis models. Among these theoretical models of language comprehension, many argue for the syntax-first hypothesis in sentence comprehension. In the initial stage of language comprehension, the sentence structure information processing is prior to and independent of the processing of semantic information. This hypothesis states that (1) initial mental representation of sentence comprehension depends on intactness of syntactic information; and (2) failed processing of structural information can block follow-up semantic integration.

(22)

Several previous studies have been conducted on the syntax-first hypothesis to explore how does the language system integrate syntactic and semantic information during the processing of a constituent word which modulates many aspects such as the construction structure and the semantic context. These studies used online measurements of sentence processing such as the event-related potential (ERP) technique or behavioural reaction times in self-paced reading experiments.

1.2.1 Event-related potentials studies

Neurocognitive studies have confirmed that syntactic category plays an important role during sentence comprehension in European languages such as German and French (Hahne & Friederici, 2002; Friederici, Gunter, Hahne, & Mauth, 2004; Isel, Hahne, Maess & Friederici, 2007). The results of these ERP studies revealed that at early stages, syntactic processing indexed by the ELAN or LAN proceed independently of semantic processing indexed by the N400 while these two types of processes interact at later stages indexed by the P600 under certain conditions, supporting the serial, syntax-first models of sentence comprehension.

For example, Isel et al.(2007) conducted an ERP experiment on French sentences containing a subject-modifying relative clause (SRC). Sentences were controlled for their semantic congruence between subject nouns and critical verbs and/or the grammaticality realized by prepositional phrases. The ERP data collected under double violations revealed that syntactic and semantic information did not interact in the early time window. The similarity of the ERP pattern between the pure syntactic

(23)

and the double violation condition provided strong evidence for the syntax-first hypothesis. Similar results were found in other studies during language comprehension (Hahne & Friederici, 2002; Friederici, Gunter, Hahne, & Mauth, 2004). However, experimental materials of most current research were created in European languages such as German and French. Conclusions drawn in this way may not be suitable for other languages (Mitchell, 1984). Through the investigation of different languages, we may discover which syntactic analysis processes are universal to all languages, and which processes are unique to certain languages.

Therefore, the next question that researchers need to answer is whether the syntax-first hypothesis holds true for Chinese. Largely distinct from European languages, Chinese lacks inflectional morphology to mark syntactic category or other features. Some constructions in Chinese have unique linguistic properties that can help us to determine whether to accept or discard the syntax-first models. Several studies have been carried out to explore the interplay between semantic and syntactic processing in Mandarin Chinese (Ye, Luo, Friederici, & Zhou, 2006; Yu & Zhang, 2008; Liu, Li, Shu, Zhang, & Chen, 2010; Zhang, Yu, & Boland, 2010; Wang, Mo, Xiang, Xu, & Chen, 2013; Yang, Wu, & Zhou, 2015).

Ye et al. (2006) first examined the interaction between syntactic and semantic information represented in Chinese sentences with preposition-object phrases, as is shown in (2). In their study, sentences with the preposition ‘ba’ were auditorily presented in four conditions: correct, semantically incorrect, syntactically incorrect, or combinations of both semantic and syntactic violations. A semantic violations was

(24)

realized by a mismatch between the critical verb and its object noun while a syntactic violation was constructed by removing the object noun immediately after the preposition ‘ba’.

(2) 设计师制作新衣，把布料裁了

She4ji4shi1 zhi4zuo3 xin1yi1, ba3 bu4liao4 cai2 le

stylist make new dresses, [PREP] cloth cut [perfective marker]

'To make new dresses, the stylist cut cloth.’

The syntactic violation elicited an early anterior negativity and a broadly distributed negativity between 300-500ms. The semantic violation evoked an early starting N400 effect. In the combined violation condition, broad negativities which were more negative than the other incorrect conditions was observed within the N400 latency range (250-400ms). The early anterior negativity, however, did not differ between the syntactic and the combined condition in the 150–250ms time window.

The results suggested that syntactic processes precede semantic processes. They were independent from each other in the early time window (150–250 ms) but interacted in a later processing phase (250–400 ms). It seems that the comprehension of Chinese ‘ba’ sentence, like European languages, conforms to the syntax-first hypothesis. However, such conclusion is questionable because in this study, the critical word occurred at the end of the sentence where wrap-up effects in sentence-final positions may influence ERPs (Hagoort, 2003; Hagoort et al., 2003; Osterhout & Nicol, 1999). Another shortcoming of this experiment lies in the way

(25)

they displayed experimental stimuli. Test sentences were presented from the auditory modality in the study. Large number of homophones in Chinese may be a confounding factor misleading the interpretation for the results.

Later, other studies investigated syntactic and semantic violations with visually presented Chinese sentences. Two of the studies (Yu & Zhang, 2008; Wang et al. 2013) manipulated the transitivity of verbs in Chinese marked active/passive constructions while the other two studies (Zhang et al., 2010; Yang et al., 2015) controlled semantic consistency (consistent vs. inconsistent) and syntactic category (noun vs. verb) of the critical verb. Results from all the studies showed N400-P600 pattern for sentences with semantic anomaly, syntactic anomaly, or combined anomalies opposed to those under congruent condition. Overall, the similarity between semantics-only violated condition and sentences with both semantic and syntactic violations in the earlier time window (i.e., N400 window) suggested that syntactic processing in Chinese is not a prerequisite for semantic processing.

Liu et al. (2010) further examined the interaction between structural analysis and semantic processing in Chinese four-character idioms using ERPs. They made manipulations for the final words of the idioms, yielding the following four conditions: congruous condition (e.g., 笑里藏刀 xiao4li3cang2dao1 ‘hiding a dagger behind one’s smiles’), synonym condition (e.g., 笑里藏剑 xiao4li3cang2jian4 ‘hiding a sword behind one’s smiles’), semantic violation (e.g., 笑里藏房 xiao4li3cang2fang2 ‘hiding a room behind one’s smiles’) and combined violation (e.g., 笑里藏投 xiao4li3cang2tou2 ‘hiding throw behind one’s smiles’).

(26)

Results showed that the N400 effect found in this experiment reflected graded semantic distance in the three incongruent conditions involving synonyms, semantic violations, and combined violations. No difference was found between the semantic violation and the combined violation condition. Liu et al. (2010) therefore concluded that semantic processing of Chinese idioms is independent of structural information, that is, semantic integration could still proceed even when syntactic processing fails.

In addition, researchers have evidence for context-dependent semantic processing during idiom processing, which seemed to support the integrative view of language comprehension (Rommers, Dijkstra, & Bastiaansen, 2013; Zhang, 2019). Rommers et al. (2013) conducted an electrophysiological research on Dutch idiomatic expressions comparing participants’ performance on idiom comprehension with two equally predictable contexts: literal and idiomatic. Sentences in literal contexts elicited a reduced N400 effect for semantically related words over unrelated words which was not observed in idioms. This result suggested that the activation of literal integration process relies on contextual cues. Findings from Chinese revealed similar results. Zhang (2019) examined semantic processing of Chinese idioms in base forms as well as their variants within literally biased and figuratively biased contexts. ERP data showed a facilitation effect of different contexts on the semantic processing of both basic and variant forms for Chinese three-character idioms.

1.2.2 Self-paced reading studies

(27)

comprehension is the self-paced reading paradigm. In such experiments, participants’ reading times on each word are recorded automatically and the length of the reaction time indicates the processing difficulty encountered during sentence processing. Although behavioral studies did not provide unequivocal evidence for either serial or interactive models on language comprehension (Pickering, Clifton, & Crocker, 2000), sentence processing on a syntax-semantics interaction can still be tested using self-paced reading.

For example, a self-paced reading study (Serafini, Franzon, Arcara, Bertocci & Zanini, 2016) investigating number agreement to the Italian version of quantification expressions such as ‘a bag / bunch of lemons’. The verb following the expression could either agree with the quantifier forming a syntactic number agreement or agree with the noun to yield a semantic number agreement. Measurements of reaction time showed a lack of significant difference at the critical word between experimental conditions, suggesting that no extra semantic processing costs were required during the purely syntactic processing of number agreement.

A recent self-paced reading study (Bousquet, Swaab, & Long, 2020) examined three types of bias in English temporarily ambiguous direct object / sentential complement sentences: structural bias (frequency of a verb to co-occur with sentence structures), lexical bias (the co-occurrence of verbs with other lexical items) and global bias (probabilistic knowledge about combinations of syntactic structure and semantic context). Reading times in the critical word region only provided evidence for structural bias as a reliable predictor. At the post-target word, global bias, however,

(28)

outweighed structural and lexical bias. The results suggested that people have immediate access to structural information while context-specific, semantic information becomes increasingly informative as processing proceeds.

The effect observed after the manipulated word can be explained by spillover effects (Mitchell, 1984), which were frequently reported in self-paced reading experiments. Such effects refer to a phenomenon that the processing of the critical word may be incomplete and will spill over to the next region. Therefore, facilitation effects or processing difficulties represented by decreased or increased reaction times compared to that of the baseline condition were found at the word immediately after the target word rather than in the critical region.

Spillover effects can be seen more directly through the comparison between results from behavioural studies and those of neuroimaging research (e.g., ERPs). Bicknell, Elman, Hare, McRae, and Kutas (2010) tested participants’ event knowledge during online processing of sentences in which the combinations of the agent and the verb exhibit a preference for its patient nouns. For example, the mechanic checked had stronger expectations for an event that includes an automobile (such as checked the brakes) than does the journalist checked. They employed both self-paced reading and event-related brain potential (ERP) experiments. The reaction times were significantly shorter at the word after the patient under congruent conditions, but semantics-related N400s were observed directly at the patient noun. Taken altogether, previous literature on the syntax-first hypothesis showed opposite patterns for the semantic processing of a constituent word at a early stage of

(29)

language comprehension in European languages and Chinese. Regarding context-dependent semantic processing, cross-linguistic studies on idiom processing found that semantic processing relies on contextual information during language comprehension. However, quite few studies have investigated semantic processing of a constituent word within constructions with highly formal constraints, e.g., A-not-A constructions in Mandarin Chinese, in which semantic integration of the constituent word with the global context may be totally unnecessary. Taking into account that idiomatic expressions and A-not-A constructions share similarities that semantic computation of these structures is not theoretically necessary, conclusions drawn from idiom processing may shed light on the processing of A-not-A constructions. But it remains unknown whether context-dependent semantic integration still takes place in highly formal constructions like A-not-A constructions. Further research is needed to reveal the cognitive processing of syntactic and semantic information during language comprehension of A-not-A constructions and thus providing experimental evidence for the absence of semantic processing in these structures when the syntactic requirement is not satisfied.

1.3 The present study

The current experiment was designed to explore the interplay between syntactic structures and semantic context with a self-paced reading paradigm. We made manipulations for a special interrogative A-not-A construction in Mandarin Chinese, constructing ungrammatical conditions in which the critical verb is either semantically

(30)

congruent or incongruent with the sentential global context. In this section, linguistic features of this special grammatical construction will be introduced, followed by hypotheses and predictions.

1.3.1 A-not-A question

A-not-A question is a special type of interrogative question in Chinese. It has a form of [X-negation-X]. A predicate has to be followed by a negation marker plus a (partial) copy of the predicate itself, as is shown in (3).

(3) 吃不吃

chi1 bu4 chi1

eat not eat

'eat or not’

Generally speaking, there are three basic sub-types of A-not-A questions in Mandarin Chinese, namely, ‘A-not-A’, ‘A-not-AB’ and ‘AB-not-A’. The derivation of A-not-A constructions requires a reduplication of the first A after the negation marker ‘bu4’ in Chinese. For transitive verbs in Mandarin Chinese, only two sub-types: ‘A-not-A’ and ‘A-not-AB’ can be generated, as is illustrated in (4).

There may be two possible ways to construct an A-not-A question for double-syllable verbs. The first is to duplicate the whole word A, thus forming sentences such as (4a) and the other is to only replicate the first syllable of verb A, as is shown in (4b).

(31)

(4a) 你喜欢不喜欢这本书？

ni3 xi3huan1 bu4 xi3huan1 zhe4 ben3 shu1

you like not like this CL book

'Do you like this book or not?’

(4b) 你喜不喜欢这本书？

ni3 xi3 bu4 xi3huan1 zhe4 ben3 shu1

(4c) *你喜欢不喜这本书？

ni3 xi3huan1 bu4 xi3 zhe4 ben3 shu1

(4d) 你读不读这本书？

ni3 du2 bu4 du2 zhe4 ben3 shu1

you read not read this CL book

'Do you want to read this book or not?’

But for mono-syllable verbs, it is impossible to delete anything morphologically, as is shown in (4d). Sentences like (4c) in which the second syllable is repeated to

(32)

build the construction are ungrammatical. Only if A refers to a verb-noun phrase consisting of a monosyllabic or disyllabic verb with an object noun, the generation of the third sub-type ‘AB-not-A’ is valid in Chinese. One example for a verb-noun phrase (买书 mai3shu1 ‘buy books’) is provided in (5).

(5a) 你买书不买书？

ni3 mai3shu1 bu4 mai3shu1

you buy-book not buy-book

'Do you buy a book or not?’

(5b) 你买不买书？

ni3 mai3 bu4 mai3shu1

you buy not buy-book

(5c) 你买书不买？

ni3 mai3shu1 bu4 mai3

you buy-book not buy

Regarding the generation mechanism of A-not-A construction, Huang (1982) argued that A-not-A question is derived from a simple declarative sentence carrying a [+Q] feature at the INFL0_{node, as illustrated in Figure 1.2.}

(33)

Figure 1.2

Syntactic structure of a declarative sentence with interrogative features

Figure 1.3

(34)

In order to convert a declarative sentence into an A-not-A question, the interrogative feature [+Q] has to move from INFL0_{to [Spec, CP] at Logical Form.}

The syntactic structure after movement is shown in Figure 1.3.

According to Huang (1982), the generation mechanism of A-not-A questions is purely morphological or even phonological. If this construction is driven by syntax, sentences such as in (4b) where part of the first A (double-syllable verb) does not convey meaning would be totally unacceptable as purely syntactic processes cannot target just part of a word. Alternatively, he argued that the A-not-A structure is generated from a reduplication rule in a morphological or even phonological fashion. The structure generated in this way is independent of syntactic constraints, and (4b) should thus be acceptable. Based on these theories, the processing of A-not-A constructions could be highly formal in that comprehenders only mechanically check whether the second word is a morphological copy of the first A, totally ignoring word properties of the second verb. If it is the case, no semantic integration of global context will take place in A-not-A constructions. This would be also consistent with the prediction of syntax-first models in which the initial analysis requires syntactic information of the sentence structure but later semantic integration is blocked due to unsuccessful syntactic processing.

(35)

the first A. This construction rule modulates both semantic and syntactic expectations within the A-not-A structure. If either one of the two verbs within the construction is replaced with another transitive verb, sentences will become ungrammatical regardless of matched word properties (e.g., semantics). Taking advantage of the linguistic features of A-not-A construction, Yue (2012) experimentally studied this special construction. He employed an ERP study to investigate the predictive processes during reading Chinese interrogative questions. The first verb A was substituted by another verb which was semantically consistent with global context, written as B. Two conditions (A-not-A and B-not-A) were therefore constructed. Results showed a weak N400 between the waveform under these two conditions but could not be verified by statistics, suggesting that verbs that agree with the global context did not elicit semantic processing, but instead inducing formal and syntactic processing difficulties. This interesting finding raised another question: Does global context information constrain the semantic processing of the second verb within A-not-A constructions? However, as stimuli used in Yue’s study were all semantic congruent with the global context, his data cannot clarify whether the lack of semantic processing difficulty reflected by N400 effects in the A-not-A construction rule violated condition is due to the absence of semantic processing of constructions or because of an agreement with the global context. The former possibility is proposed based on Huang’s purely morphological view of A-not-A constructions. This view also sounds like syntax-first hypothesis, according to which semantic processing will be blocked if constructional requirement is not satisfied. But the latter

(36)

context-dependent view based on interactive models of language comprehension is more global in nature allowing all possible sources of information to be used to support sentence comprehension. The current study was therefore conducted to explore whether global context information will constrain the semantic processing of a constituent word within the A-not-A construction, a highly formal structure in Mandarin Chinese.

1.3.2 Hypotheses and predictions

Two theoretical hypotheses can be proposed to explain semantic processing of Chinese A-not-A questions. First, according to syntactic theories, A-not-A constructions are purely formal and driven by an abstract interrogative operator, Q [+A-not-A] located at the I-IP position (Huang, 1982). Comprehension of such a construction with highly formal constrains does not involve any semantic computation on the second A if the first A has been given. If this purely morphological hypothesis is true, we do not expect significant difference in the reaction times at or after the constituent word (the second verb in this study) between construction-rule violated sentences regardless of its semantic congruence or incongruence with the global context. But for sentences with correct A-not-A constructions, reaction times directly at or after the manipulated verb position will be expected to be shorter compared to sentences violating construction rules. The prediction of absent semantic processing is also in line with some theoretical models including serial two-stage models e.g., garden path model (Ferreira & Clifton, 1986; Frazier, 1979, 1987; Frazier & Rayner,

(37)

1982) and the three-stage model (Friederici, 2002). These models are supportive of syntax-first hypothesis in which failed processing of structural information will block follow-up semantic integration.

Alternatively, A-not-A questions can be explained from the integrative view of sentence comprehension supported by parallel one-stage models e.g., constraint-based model (MacDonald, 1994; Trueswell & Tanaenhaus, 1994) as well as the unrestricted competition model (van Gompel el al，2001，2005). In these models, various sources of information such as the sentence structure, semantic properties of words, the global context are supposed to be integrated together during language comprehension. If this is true, a graded effect on reaction time at or immediately after the critical verb under conditions should appear. To be more specific, sentences with syntactically correct structures will be better comprehended than incorrect ones, and also among ungrammatical conditions, participants will spend less time during the comprehension of sentences with semantically congruous verbs as opposed to those involving verbs inconsistent with the global context in the sentence.

(38)

Chapter 2 Method

2.1 Participants

Sixty-three native speakers of Mandarin Chinese (33 female; mean age 24.5, SD = 2.1, ranging from 20 to 30) born and raised in Mainland China participated in this study. They were equally distributed to one of three lists; each list had 10 male and 11 female participants. All the participants were receiving university education or had finished their bachelor’s degree studies. All of them were right-handed and had a normal or corrected-to-normal vision. None of them had neurological or psychiatric disorders. No one had participated in the pretests of materials described below. Participants gave written informed consent before the experiment. A participant was paid with about €4 upon completing the experiment. This study was carried out in accordance with the Declaration of Helsinki and had been approved by the Ethics Committee (CETO) of Faculty of Arts, University of Groningen.

(39)

2.2 Materials

The materials were adapted from a previous ERP study on Chinese A-not-A questions (Yue, 2012). In the original study, the first verb within an A-not-A construction was replaced with another mono-syllable transitive verb to construct two conditions, namely, A-not-A and B-not-A. Critical verbs in both conditions were semantically acceptable within the global context but only one was grammatically correct. Differently from the original experimental design, the second verbs instead of the first ones were substituted, as the purpose of this study was to investigate the effect that global context may have on semantic computation for A-not-A constructions. It was because with original stimuli we could not rule out the possibility that participants would make judgments on the grammaticality of the whole sentence as soon as they came across a semantically unacceptable verb and they might just skip reading the left part of the sentences. Should they adopt this strategy, construction constraints would exert no influence on sentence comprehension, which ran against our research goals.

In the current study, the congruence of semantics was manipulated on the basis of the interaction between object nouns and words at the second verb position within the constructions. Apart from a grammatical construction functioning as a baseline (A-not-A condition), two other ungrammatical conditions were added. Both of them violated the A-not-A construction rule but differed in semantic congruence with the global context: for sentences in one condition, the second verbs can be integrated into

(40)

preceding context with no difficulty (A-not-B1 condition); for sentences in the other condition, the second verbs are designed to cause difficulties in semantic integration with global context (A-not-B2 condition).

Thirty sets of sentences were selected on the basis of two pretests (see below). Examples of one set are shown in Table 2.1. Each set shared the same contextual information and sentence structure but only differed in the second verb of A-not-A construction. All the second verbs were monosyllabic transitive verbs matched for visual complexity (measured by the number of strokes) and word frequency. For effect of visual complexity, one-way ANOVA found no statistical differences for the means of number of strokes across all three conditions (mean of A-not-A condition = 9.38, SD = 3.30; mean of A-not-B1 condition = 9.49, SD = 3.24; mean of A-not-B2 condition = 9.41, SD = 2.59, F(2,114) = .012, p = .988). Furthermore, word frequency was controlled for the second verb according to Leiden Weibo Corpus (Van Esch, 2012) based on messages from Sina Weibo, a Chinese online social platform similar to Twitter. As participants in this study were recruited online and tests (pretests and self-paced reading) were completed online, this corpus could better point out which words were frequently used in the context of online reading. Further one-way ANOVA showed no significant difference between the means of word frequency across all three conditions (mean of A-not-A condition = 304.6, SD = 453.1; mean of A-not-B1 condition = 315.5, SD = 473.7; mean of A-not-B2 condition = 330.1, SD = 434.6, F(2,114) = .031, p = .97).

(41)

Table 2.1

Examples of sentences for baseline and experimental conditions Cond Contextual sentence Question sentence

A 在/餐厅，老师 /问/ 友人 /

说:

at canteen teacher ask friend say

At the canteen, the teacher asks friend:

老王，饺子，你 /吃/不/吃/ 一盘？

Lao Wang dumplings you eat not eat one plate Lao Wang, do you want to eat one plate of dumplings (or not)?

B1 在_{/餐厅，老师 / 问 / 友人}

/说:

*老王，饺子，你 /吃/ 不/买/ 一盘？ Lao Wang dumplings you eat not buy one plate

Lao Wang, do you want to eat or buy one plate of dumplings?

B2 在/餐厅，老师 / 问 / 友人

/说:

*老王，饺子，你 /吃/ 不 /踢/ 一盘？ Lao Wang dumplings you eat not kick one plate

Lao Wang, do you want to eat or kick one plate of dumplings?

Note. Cond = Condition. A = A-not-A construction rule obeyed and semantically acceptable (baseline), B1 = A-not-A construction rule violated but semantically acceptable, B2 = A-not-A construction rule violated and semantically unacceptable.

(42)

Critical words in the second verb position were underlined.

All the verbs under A-not-B2 conditions met the following criteria: (1) All the verbs are non-directional verbs. Directional verbs such as 来 lai2: ‘to come’ were not considered as test stimuli because they can function as complement of the first verb within A-not-A construction, which violates the construction rule but forming a grammatical phrase (e.g., 起不来 qi3bu4lai3 ‘cannot get up’). (2) All the single-character verbs were frequently used ones in spoken Mandarin Chinese according to CCL (Centre for Chinese Linguistics, Peking University) corpus (Zhan, Guo & Chen, 2003). Verbs with low occurrence in oral contexts (e.g., 绘 hui4 ‘to draw’) or those used mostly in Chinese dialects (e.g.,整 zheng3 ‘to do’) were excluded from test stimuli. (3) All the verbs should have semantic constraints for object nouns as their themes. For example, the verb 吃 chi1 ‘to eat’ requires solid food as its object while the verb 喝 he1 ‘to drink’ should be followed by a noun indicating liquid food. Verbs like 摆 bai3 ‘to place’ did not satisfy this criterion and thus they were not taken into account. (4) In order to make sure that verbs under A-not-A and A-not-B1 conditions are equally semantically acceptable in the given global context, the plausibility of Verb-Noun phrases was checked in terms of presence of phrases consisting of the second verb and the object noun in a large-scale Chinese multi-field BCC (BLCU Chinese Corpus) (Xun, Rao, Xiao & Zang, 2016), which has a word count of 15 billion words from newspapers, literature, technology, etc. No entry was found for verb-noun pairs under A-not-B2 condition. Additionally, The object nouns in the test sentences were pre-positioned at the subject position to provide more

(43)

contextual information before participants encountered the critical second verb within A-not-A construction.

For each set of sentences, two filler sentences were created. One of the two fillers was grammatical with the other one being ungrammatical determined by semantic congruence between the verb and object noun. There were 78 filler sentences in total. Filler sentences were ‘what’ questions. All the sets of test sentences were divided into three lists. An equal number of sentences belonging to each of the three conditions appeared in one list. Fillers remained the same for all three lists. One participant received only one list. Therefore, one participant saw only one sentence out of three sentences forming one stimulus set. In each list, sentences were randomized across participants.

Each reading trial was followed by a comprehension question. The information that was useful for successfully answering the single-choice question was set in three different regions: identity of speakers, places where conversations happened and prepositional object nouns. The sequences of options as well as the serial numbers of correct answers were randomized automatically by programming.

2.3 Pretests

In order to avoid potential influence that confounding factors may have on this experiment, several pretests were carried out, namely, Semantic Rating Test for semantic congruence between the critical verb under each condition and its object noun, along with Sentence Acceptability Test for grammaticality of sentences. The

(44)

two pretests would be first introduced, followed by the inclusion criteria for sentences involved in the actual experiment.

2.3.1 Semantic Rating Test

A semantic acceptability test was conducted to evaluate semantic relatedness between the transitive verbs and object nouns under all three conditions. This was done to make sure that participants have close to equal acceptance of verb-noun pairs under A-not-A and A-not-B1 conditions but would not accept combinations of verbs and object nouns in the third condition. A group of 30 native speakers of Mandarin Chinese were recruited for this test. Each participant saw a questionnaire consisting of 40 items randomly chosen from 117 verb-noun pairs. Participants were instructed to rate their acceptability on a 5-point scale for verb-noun phrases under all three conditions (1 = totally unacceptable, 5 = totally acceptable).

2.3.2 Sentence Acceptability Test

Another group of 42 native Chinese speakers who did not participate in the reading experiment was recruited to finish a sentence acceptability test by rating on a 2-point scale (1 = unacceptable, 2 = acceptable) for both test sentences and filler sentences used in this study. A total of 117 test sentences were equally divided into three lists. 78 filler sentences were randomly distributed across the three lists. In each list, half of the filler sentences were grammatical and another half were ungrammatical. Therefore, each participant encountered a questionnaire made of 39 test stimuli along with 26

(45)

fillers among which 26 sentences were grammatical and the left were ungrammatical.

2.3.3 Inclusion criteria for test sentences

For Semantic Rating Test, the most strict cutoff was adopted for the inclusion of verb-noun pairs. The mean rating of phrases under A-not-A and A-not-B1 conditions should be no less than 4.0 but for the last one condition, averaged value of its semantic rating should be lower than 3.5. The inclusion criterion for Sentence Acceptability Test was identical as used in Yue (2012). The acceptability rate should be larger or equal to 75% for grammatical condition; accordingly, rate of acceptance for ungrammatical sentences should be lower or equal to 25%. After these procedures, 30 verb-noun combinations and 90 sentences were preserved as test sentences in the subsequent experiment session. A further statistical analysis on materials after items removal revealed that semantic relatedness between critical verbs and its object nouns did not differ significantly under baseline and semantically acceptable manipulated condition (A-not-A condition: mean = 4.61, SD = 0.798; A-not-B1 condition: mean = 4.54, SD = 0.894; t = 1.166, p = .24). Similarly, ungrammatical sentences did not manifest significant difference between the mean acceptability rate in the Sentence Acceptability Test (A-not-B1 condition: mean = 0.126, SD = 0.041; A-not-B2 condition: mean = 0.105, SD = 0.049; t = 1.852, p = .07).

2.4 Procedure

(46)

Farm, an online experiment platform. Texts (contextual sentences and A-not-A questions) were presented visually in the center of a computer screen. Each sentence was divided into frames. Each frame consisted of one or two characters defined by word segmentation in Chinese. Participants were instructed to read one frame and press the space key to continue when they finished reading it. Then another frame would appear on the screen for them to read. Such procedure would continue until the end of the sentence. Each trial begins with the presentation of a fixation cross (500ms) and end with a 500ms break. Subsequently, participants were required to complete an offline task by answering a comprehension question related to the preceding experimental sentence they have seen. This question would be presented on the screen as a whole and without spaces between words. The task was counterbalanced across participants. Each participant only saw one of the three lists of stimuli. The experimental session began with a short training session followed by three experimental blocks comprising 30 sentences each, between which the participants took short breaks. The entire experiment lasted approximately 20 minutes.

2.5 Data analyses

The current study used R software package (R Core Team, 2019) for statistical analysis on data collected from self-paced reading and sentence comprehension tasks. For the sentence comprehension task, accuracy was calculated for each participant based on their performance in answering the comprehension questions after reading both test and filler sentences. In this study, only data from participants with high

(47)

accuracy were analysed as their high accuracy indicated that they were concentrated when reading experimental sentences. All the response from participants who failed to achieve an accuracy above 80% were removed from further analysis.

For outlier removal, reaction time data which is greater than 3000 ms or less than 100 ms was excluded for each target word region: the first verb (V1), the second verb (V2), the word immediately after the second verb (PV2). Moreover, values below the mean value minus 3 times the standard deviation or above the mean value plus 3 times the standard deviation were dropped from the analyses as outliers (Howell, 1998). Altogether, these procedures resulted in a loss of 38 data points at the first verb position (approximately 2.2% of the data), 40 data points at the second verb position (approximately 2.3% of the data) and 46 data points for the first word at the post second-verb position (approximately 2.7% of the data).

The remaining data for the three target regions were analysed separately. Raw reading times met the precondition of normal distribution, therefore no log-transformation was needed.

Reaction time data obtained at the second verb position where we did manipulations were analysed first. A one-way ANOVA was conducted to compare reaction times of the second verb when reading sentences under different conditions.

Considering individual difference and potential priming effects of a preceding word on the reading times of the following word (Mitchell, 1984; Vasishth, 2006), reaction times at the first verb position were also included to see relative changes in

(48)

time on reading a one-character verb in Chinese. A 3*2 repeated measures ANOVA with Condition (A-not-A, A-not-B1 & A-not-B2) and Position (first & second) as dependent variables was carried out to see whether reaction times of verbs within A-not-A constructions interact with manipulated conditions.

In addition, previous research suggested that facilitation effect may be behaviorally delayed which can only be observed in later stages (Mitchell, 1984; Bicknell et al., 2010). One-way ANOVA was thus conducted to make comparisons between reaction times of words at the post second-verb position.

For normally distributed data sets with unequal variances, p-values were calculated using Welch's correction (1951) available in R to compensate for lack of homogeneity of variance in one-way ANOVA.

(49)

Chapter 3 Results

3.1 Comprehension Accuracy

Most participants had relatively high accuracy and only one participant’s score was extremely low (32.2%). The mean proportion of correctly answered comprehension questions for all the participants was 87.9% (SD = 10.4). Data from 6 participants (list1: 1 male; list2: 1 male, 2 female; list3: 2 male) were removed due to low accuracy in the sentence comprehension task.

3.2 Reaction time data analysis

Table 3.1 shows the results of descriptive statistics for reading times of the first verbs, the second verbs and the word immediately after second verbs under A-not-A, A-not-B1 and A-not-B2 conditions. As shown in Table 3.1, there were descriptively differences in reading times between sentences under different conditions, especially at the post second-verb position between A-not-A and A-not-B2 conditions.

(50)

In the following, separate analyses for the reaction times at three regions of interest are presented. Reaction time data obtained at the second verb position were analysed first. After that, reaction times at the first and second verb position were analysed together to see relative changes in time on reading a one-character verb. Finally, reaction times of words at the post second-verb position were compared under different conditions.

Table 3.1

Descriptive statistics for reaction times at target word positions across conditions

Condition Word Positions

V1 V2 PV2

A-not-A 319.7 (128.0) 307.4 (124.5) 338.1 (143.7) A-not-B1 326.0 (138.4) 307.8 (130.0) 369.0 (175.6) A-not-B2 322.2 (127.8) 310.6 (132.5) 375.2 (174.2)

Note. V1 = the first verb, V2 = the second verb, PV2 = the first word at post second-verb position. Standard deviations were shown in brackets.

3.2.1 Second verbs within A-not-A constructions

The first analysis only included reaction time data obtained at the second verb position within A-not-A constructions where we made manipulations for test sentences. From Table 3.1, we can clearly see that participants’ mean reaction times at the second-verb position within test sentences belonging to three various conditions

(51)

were similar to each other, especially for A-not-A and A-not-B1 conditions. A one-way ANOVA analysis revealed that reaction times that participants required to read the second verbs of A-not-A constructions did not differ significantly across original and manipulated conditions (F(2, 1667) = .10, p = .904).

3.2.2 First and second verbs within A-not-A constructions

Considering individual difference in reaction time, this analysis included reaction time data collected from not only the second verb position but also the first verb position. Figure 3.1 visually illustrates the mean reaction times obtained from the first and second verb positions within A-not-A constructions under three conditions.

Figure 3.1

Mean reaction times at verb positions across conditions

From Table 3.1, it is clear that participants tended to read the first verbs longer compared to the second verb within A-not-A constructions regardless of condition. A 3*2 repeated measures ANOVA revealed a significant main effect of Position (F(1,52)

(52)

= 11.67, p = .001). Further post-hoc tests indicated that reaction time at the second verb position were significantly shorter compared to reaction time at the first verb position. No significant main effect of Condition or interaction effect between Position and Condition was found (main effect of Condition: F(2,52) = .74, p = .48; interaction effect of Condition and Position: F(2,52) = 1.29, p = .28).

3.2.3 Words after the second verbs of A-not-A constructions

In this analysis, reaction times of words after the second verbs were extracted under original and manipulated condition. Figure 3.2 visually illustrates the mean reaction times obtained after the second verb position under three conditions.

Figure 3.2

Mean reaction times at the post second-verb position across conditions

As we can see in Table 3.1 and Figure 3.2, reaction time showed a graded pattern for three conditions. Participants required the least time to read post-verb words under A-not-A conditions (grammatical), followed by A-not-B1 (ungrammatical but

(53)

semantically congruent) and A-not-B2 conditions (ungrammatical and semantically unacceptable). One-way ANOVA revealed that the graded effect observed in descriptive statistics was significant (A-not-A: mean = 338.1, SD = 143.7; A-not-B1: mean = 369.0, SD = 175.6; A-not-B2: mean = 375.2, SD = 174.2; F(2, 1661) = 9.15, p = .0001). Further post-hoc tests suggested participants read the words after A-not-A constructions significantly faster within grammatical sentences (A vs. B1: p < .005; A vs. B2: p < .001). Furthermore, there was no significant difference between reaction times of target words under ungrammatical conditions (B1 vs. B2: p = .53).

Noticing that in some of our experimental sentences, the words after the second verbs were the last word of the sentence, another post-hoc analysis was conducted to examine the possibility that the significance that we got may be a result of wrap-up effects (Hagoort, 2003) rather than facilitation effect of context on semantic processing, another analysis was carried out with identical data but excluding items which the post second-verb word was also at the sentence-final position. Statistical results still showed a significant main effect of Condition (A-not-A: mean = 327.8, SD = 132.9; A-not-B1: mean = 349.7, SD = 150.3; A-not-B2: mean = 361.5, SD = 161.8; F(2, 1446) = 6.77, p = .001). Likewise, post-hoc tests showed significant difference between the mean reaction time of words under A-not-A (baseline) condition and two other conditions (A vs. B1: p < .05; A vs. B2: p = .001) but did not reveal significance between reaction times of words within ungrammatical sentences (B1 vs. B2: p = .21).

(54)

Chapter 4 Discussion

The current study was conducted to investigate whether semantics of a constituent word integrate with the global context when syntactic information is not successfully processed during reading the A-not-A construction, a special syntactic structure in Mandarin Chinese. More specifically, a word-by-word self-paced reading experiment was carried out to examine whether participants’ reading times directly at or immediately after the critical region were affected by semantic congruence between the second verbs and the object nouns in sentences violating A-not-A construction rules. Three conditions were constructed as follows: syntactic correct and semantically acceptable (A-not-A), with a syntactic violation but semantically acceptable (A-not-B1), and syntactically correct but semantically unacceptable (A-not-B1). Separate analyses of reaction times at different target word positions revealed that (1) facilitation effects represented by shorter reading times in the grammatically congruent A-not-A condition compared to the two other ungrammatical conditions were found at the words immediately following the critical

(55)

second verb, and (2) no reliable difference was observed between reaction times of any region of interest in the sentences containing only semantic violations and those with combined syntactic and semantic violations. This chapter will first introduce main findings in this study focusing on semantic and syntactic processing in A-not-A constructions, and then reveal its implications, limitations and future directions.

4.1 No semantic effect at critical verb region

In the current study, mean reaction times of critical verbs for semantically acceptable A-not-B1 condition were shorter than those of semantically unacceptable A-not-B2 condition, but the difference failed to reach the significance level, consistent with the weak and not significant N400 effect observed in Yue’s experiment (2012), suggesting that semantic context did not contribute significantly in the processing of Chinese A-not-A questions. The absence of semantic processing in this study is in line with previous research which provided supporting evidence for the syntax-first hypothesis in which comprehenders appear to initially rely on information about structural constraints; semantic congruence between the target word and the global context should be processed later (Frazier, 1987; Friederici, 2002; Hahne & Friederici, 2002; Friederici et al., 2004; Isel et al., 2007).

The lack of semantic integration with the global context also contrast with the results of previous studies exploring the relationship between syntactic and semantic processing during language comprehension of Chinese syntactic structures in which claim that semantic integration is not dependent on successful syntactic analysis and

Semantic processing when syntactic requirement is not satisfied during reading Chinese A-not-A questions

during reading Chinese A-not-A questions

during reading Chinese A-not-A questions

ACKNOWLEDGEMENTS

ABSTRACT

TABLE OF CONTENTS

LIST OF TABLES

LIST OF FIGURES

Chapter 1

Introduction

1.1 Theoretical models of sentence comprehension

1.2 Previous research on the syntax-semantics interaction

1.3 The present study

Chapter 2

Method

2.1 Participants

2.2 Materials

2.3 Pretests

2.4 Procedure

2.5 Data analyses

Chapter 3

Results

3.1 Comprehension Accuracy

3.2 Reaction time data analysis

Chapter 4

Discussion

4.1 No semantic effect at critical verb region