Sung et al._Tu_2015

Investigating the Processing of Relative Clauses in

Mandarin Chinese:Evidence from Eye-Movement Data

Yao-Ting Sung1,2·Jih-Ho Cha2·Jung-Yueh Tu3·

Ming-Da Wu2·Wei-Chun Lin2

?Springer Science+Business Media New York2015

Abstract A number of previous studies on Chinese relative clauses(RC)have reported con-?icting results on processing asymmetry.This study aims to revisit the prevalent debate on whether subject-extracted RCs(SRC)or object-extracted RCs(ORC)are easier to process by using the eye-movement technique.In the current study,the data are analyzed in terms of the gaze duration and regression of eye-movement in three critical areas:head noun,embed-ded verb,and RC-modifying noun phrase as subject.The results show an ORC preference for the processing of RC structures,which supports the word-order account and the Depen-dency Locality Theory,and a better cross-clausal integration for SRC,which supports the perspective-shift account.The processing asymmetry in Chinese RCs are discussed under relevant theoretical accounts,such as structure-based,memory-based,and perspective shift accounts.We argue that the?ndings are associated with the syntactic nature of Chinese(a head-initial language with pre-nominal RCs).

Keywords Mandarin Chinese·Relative clauses·Eye-movement·Processing asymmetry·Sentence complexity

Introduction

This study revisited a long-debated issue regarding processing asymmetry of relative clauses (RCs)in Chinese.An RC is a subordinate clause that modi?es a noun and is embedded within a noun phrase.There are two types of RCs.One is the subject-extracted RC(SRC),

B Jung-Yueh Tu

jungyueh.tu@https://www.wendangku.net/doc/b218399925.html,

1Department of Educational Psychology and Counseling,National Taiwan Normal University, 162,Sec.1,Ho-Ping E.Rd.,Taipei,Taiwan

2Center of Learning Technology for Chinese,National Taiwan Normal University,

162,Sec.1,Ho-Ping E.Rd.,Taipei,Taiwan

3International Chinese Education Center,School of Humanities,Shanghai Jiao Tong University, No.1954,Hua Shan Rd.,Shanghai200030,China

in which the noun that the RC modi?es is extracted from the subject position.The other is the object-extracted RC(ORC),in which the noun that the RC modi?es is extracted from the object position.A considerable amount of previous research has demonstrated that an SRC is easier to process than an ORC in head-initial languages such as English(Gordon et al. 2006),Dutch(Frazier1987;Mak et al.2002),and French(Holmes and O’Regan1981),as well as in head-?nal languages such as Japanese(Ueno and Garnsey2008),Korean(Kwon et al.2010),and German(Schriefers et al.1995).However,the apparent universality of SRC preference has been challenged by Chinese RCs.

An SRC preference in Chinese has been found in some studies(Lin and Bever2006a,b, 2007,2011;Wu2009;Vasishth et al.2013)whereas others have found an ORC preference (Hsiao and Gibson2003;Hsu and Chen2007;Gibson and Wu2013;Lin and Garnsey2011). This discrepancy,which means there is no consensus about the processing preference of Chinese RCs,may be attributable to the nature of Chinese syntax.Chinese is a head-initial language with a dominant subject–verb–object(SVO)word order(Greenberg1963),while the RC-modifying noun phrase is in a head-?nal structure where the RC precedes its head. In other words,Chinese is a head-initial language that exhibits a head-?nal pattern in RCs.

In order to approach the dif?culty of processing Chinese RCs,we adopted an eye-movement-tracking technique to provide detailed online processing information,which can enrich what is already known from the of?ine data obtained in previous research on Chinese RCs,using methods such as self-paced reading tasks,corpus analysis,and computational modeling.Our aim was to obtain more data that would help to resolve this controversial issue.

Two Types of RCs in Chinese

There are two types of RCs in Chinese,SRC and ORC:

In(1),the noun xiaozhang,extracted from the subject position of the embedded verb (EV)jieshao,serves as the head noun(HN)of the RC introduced by the relativizer DE.The noun xiaozhang leaves an empty position,called a“gap.”The extracted noun xiaozhang is co-indexed with the gap and is called the“?ller”because it should?ll the gap.In(2),the noun laoshi is the extracted HN from the object position,forming an ORC.In both sentences, the RC-modifying noun phrase(including the RC and the HN)functions as the subject of the main https://www.wendangku.net/doc/b218399925.html,prehending and integrating RCs require the dependency between the?ller and the gap to be developed in harmony.

Processing Preference in Chinese RCs

Several studies have investigated the processing of Chinese RCs.Some studies found an SRC preference while others showed an ORC preference.The?nding of both preference types in Chinese indicates that there is not a universal SRC preference.The presence of processing asymmetry in Chinese RCs raises the issues of whether the processing patterns of Chinese RCs are language-speci?c,and whether the discordant?ndings are related to the syntactically mixed pattern in Chinese RCs.

Among those studies supporting the presence of a universal SRC preference,Lin and Bever(2006a,b,2007,2011)reported on a series of studies of RC processing preference that used self-paced reading tasks of singly embedded and doubly embedded RCs.They also compared the two conditions of RC modi?cation:subject-modifying RC versus object-modifying RC.They found that participants spent signi?cantly shorter reading times(RTs) on both the relativizer and the HN in SRCs than in ORCs irrespective of whether the RCs modi?ed the subject or the object of the main clauses.Their results suggested an effect of an SRC preference,which is in line with?ndings across other languages.

That SRCs are universally easier to process,however,is challenged by the results of Hsiao and Gibson(2003).They conducted a self-paced task on singly embedded and doubly embedded RCs with an RC-modifying subject of main clauses.They found that in doubly embedded RCs,the participants showed slower RTs on the HN and the EV in SRCs than in ORCs.They demonstrated a preference for ORCs in Chinese,which is a head-initial language with prenominal RCs.

Theoretical Accounts for RC Processing

Previous research has shown that readers sometimes experience dif?culty when processing either of the two types of RCs and that this dif?culty can be attributed to various factors, resulting in many theories and processing models being proposed.

Structure-Based Accounts

The structure-based account generally emphasizes the role of syntactic information in sen-tence comprehension,and invokes the syntactic position or syntactic knowledge as the prominent factor in sentence comprehension(see Lin et al.2005;Lin and Bever2006a). Keenan and Comrie(1977)proposed a universal tendency,called the Noun Phrase Acces-sibility Hierarchy(NPAH),which ranks the syntactic positions in a sentence as follows: subject,direct object,indirect object,oblique object,possessor,and object of comparison. Accordingly,a language that can relativize a given position in the hierarchy can also rela-tivize all antecedent positions.Such a universal tendency implies that the use of SRCs makes sentence processing easier.

In addition,the word-order account(Bever1970;MacDonald and Christiansen2002) focuses on how readers analyze the structure as they read the sentence conveyed by the sequence of word order,from left to right.The underlying assumption is how canonical any given sequence of words is.The basic word order in Chinese is SVO,while the word order of SRCs is VO(DE)S and that of ORCs is SV(DE)O.The word-order account would therefore favor any given canonical structure.

Another structure-based account,incremental minimalist parser(Lin and Bever2006a), emphasizes the role of syntactic position in sentence processing.This account argues that a

gap located at a higher structural position is to be reached earlier than one located at a lower position.In the case of RCs,the gap in an SRC(i.e.,the subject)is located higher than the gap in an ORC(i.e.,the object).According to this account,Chinese SRCs are easier to process than ORCs because the gap in the SRC,which is the extracted subject,is located higher than the gap in the ORC,which is the extracted object.The gap positions for the SRC and the ORC in hierarchical representations are shown below.

(3)The hierarchical representation of SRC

(4)The hierarchical representation of ORC

Memory-Based Accounts

The memory-based account emphasizes that functional factors such as cognitive resources or working memory load will constrain sentence comprehension,and that the processing dif?culty increases with the complexity of a given structure.The Dependency Locality Theory (DLT)(Gibson1998,2000)belongs to this account type.Its underlying assumption is based on sentence comprehension involving a series of words being input one at a time and readers using their available cognitive resources to integrate the current words into built structures, and storing these structures.The available computational resources that readers rely on for processing are the storage cost and the integration cost.

In particular,the integration-cost metric of the DLT is useful for explaining the ORC advantage in Chinese.In the case of RCs,this theory claims that the cognitive cost of inte-grating the HN extracted out of the RC and the gap it leaves is re?ected by the number of discourse referents,introduced by NPs and VPs,appearing between them.The integration cost for the HN in an SRC is thus higher,as the number of referents between the extracted

HN and the gap in an SRC is greater than that in an ORC,indicating greater comprehension effort being required.The location of the intervening elements helps predict the degree of comprehension dif?culty,with dependency over a long distance leading to both higher dif?-culty and higher cost for the integration.The gap-?ller distances for SRC and ORC in linear representations are given below.

Perspective-Shift Account

The perspective-shift account is associated with the conceptual or semantic information and the effect of discourse cues in processing(MacWhinney1977,1982;MacWhinney and Pleh 1988).MacWhinney(2005)proposed that examining the role of perspective shifting(usually the syntactic subject)in sentence comprehension may help in understanding how cognitive factors and syntax may work together to facilitate sentence comprehension.Considering the structure of RCs,the HN of either an SRC or an ORC could be the subject of the main clause. When processing a sentence with an RC,the HN,as the subject or object of RCs,needs to be integrated with the main clause.

This account argued that processing is easier when readers maintain a consistent perspec-tive,as in SRC,than when they shift perspective,as in ORC(MacWhinney1977,1982; MacWhinney and Pleh1988).For instance,in processing English SRC(e.g.,‘The dog that chased the cat kicked the horse’),the subject of the main clause(the dog)is also the subject of the relative clause(the dog).Thus,a consistent perspective is maintained.However in processing English ORC(e.g.,‘The dog that the cat chased kicked the horse’),readers need to move from the subject of the main clause(the dog)to the subject of the relative clause(the cat)when reading the EV chased,and move again back to the subject of the main clause(the dog)when reading the matrix verb kicked.Hence,the readers’perspective shifts.Following this,in the integration of RCs and the main clause,there may be perspective shift involved. The complexity of sentence processing may increase with the number and types of shifts in the perspective.

The Current Study

This study focused on three major questions related to the con?icting research?ndings about processing the two kinds of RCs:(a)which type of RC is more dif?cult to process;(b)which of the theoretical accounts better explains the results;and(c)how does an RC-modifying subject integrate with the main clause?We address the third of these questions for two reasons. First,the head dependency of RCs in Chinese is not consistent with head-initial patterns,and this question asks if the inconsistency makes the integration dif?cult.Second,most studies

of RCs focus on their structure.However,considering that a head-?nal RC structure exists in a head-initial language,we would like to know how an RC-modifying subject is integrated with the main clause.

Advantages of an Eye-Movement-Tracking Technique

The experiments performed in previous studies of the processing of Chinese RCs mostly involved self-paced reading tasks,in which reading times for each region were analyzed as the comprehension measures.People generally read more slowly than normal when perform-ing such a task.Moreover,in such tasks the readers are not allowed to regress to ambiguous regions.In contrast,the reading is performed in a normal condition when using the eye-movement-tracking technique,and the researchers can record multidimensional data and better measures of initial interpretation processes(e.g.,Garnsey et al.1997;Traxler et al. 2002;Staub2010).Thus,the eye-tracking paradigm re?ects a combination of initial process-ing and later stage of processing,which can provide a direct measure of comprehension and reveal more subtle cognitive processes(Henderson and Ferreira1990).It has been previously used to examine the processing dif?culty of RCs in many languages(for Dutch,see Mak et al.(2002);for English,see Traxler et al.(2002);for French,see Holmes and O’Regan (1981);for Korean,see Kwon et al.2010).In the present study we performed real-time eye-movement tracking and determined if eye-movement data would provide useful evidence about processing asymmetry.

Hypothesis and Interest Areas

The interest areas,based on previous research,included HN,EV,and RC-modifying subject noun phrase(S-NP,hereafter).These areas were examined in order to identify which type of RC is easier to process and where processing dif?culties arise.We also examined the regression from the main clause to S-NP in order to determine if different types of S-NP in?uence how readers integrate the main clause.

The predictions of different theories about the RC,HN,EV and main clause vary according to the factors emphasized,as follows(summarized in“Appendix1”):

(a)The NPAH would predict an SRC preference because the subject is easier to relativize

than the object,and participants would spend less time on the HN of SRCs than on that of ORCs.That is,the processing time for HN in SRCs would be shorter than that in ORCs.

(b)The word-order account would predict ORCs to be easier to process because this account

follows the canonical word order in Chinese.Since the word order in ORCs is the canonical word order(SV…O)in Chinese,the processing time for NP in SRCs would be expected to be longer than that in ORCs.

(c)The incremental minimalist parser would predict an SRC preference because the gap in

SRCs is hierarchically located higher than the gap in ORCs.Thus,the gap in SRCs(i.e., the subject)is reached more easily.So,it is hypothesized that the processing time for HN and S-NP in SRCs would be shorter than that in ORCs.

(d)The DLT would also predict an ORC preference because the HN(e.g.,xiaozhang in(1)

above)is further from the gap in SRCs than in ORCs.In addition,two discourse referents

(e.g.,jieshao and laoshi in(1))situated between the gap and the?ller is present in SRCs,

but does not exist on the same path in ORCs.This means that readers must retrieve the HN over the intervening discourse referent in the SRC.Readers therefore do not need to spend more resources integrating the gap and the?ller,and they store a discourse

referent when processing ORCs.Hence,it is expected that the processing time for the HN and the S-NP in SRCs would be longer than that in ORCs.

(e)The perspective-shift account would predict that SRCs are easier to process because

readers maintain the same perspective from the subject of the RC(e.g.,xiaozhang in

(1))to the subject of the main clause.For instance,in processing(1),the subject of RC

xiaozhang is the subject of both RC and the main clause,so the readers’perspective remains constant.However,in processing ORCs,readers need to rapidly move their perspective from the subject of RC xiaozhang to the subject of the main clause laoshi, thereby shifting their perspective.Thus,the processing time for SRCs is hypothesized to be shorter than that for ORCs.

Methods

Participants

Forty-one participants,who were all college students(18–22years old)in Taiwan,were recruited in this experiment.All were native speakers of Mandarin Chinese and had normal or corrected-to-normal vision.The data from40participants were used(one was eliminated since the participant did not?nish the experiment).The participants include30females and 10males.

Apparatus

Stimuli were displayed on a19-inch LCD monitor(CHIMEI CMV A902).Eye movements were recorded using an eye tracker(EyeLink1000Desktop,SR Research)at a sampling rate of1000Hz.The experiment was performed using two desktop computers with Intel Core i53.2-GHz CPUs:one was used to control experimental procedure and to display the stimuli,and the other monitored and recorded eye movements.Possible errors due to head movements were avoided by requiring the participants to lean on a chin rest.Viewing was binocular,but only the left eye was recorded.

The programming was performed using Experiment Builder1.10.1,and data were ana-lyzed using EyeLink Data Viewer1.11.1and SPSS.

Materials and Design

The experiment had a2×2within-subject design,with the independent variables being clause type and distance.The clause types were SRCs and ORCs.The distance independent variable referred to the distance between the gap and the?ller.Distance was dichotomized into long distance(6–10characters)and short distance(1–5characters).The dependent variables were the accuracy in the reading comprehension test,gaze duration,regression-path duration,total viewing time,and regression rate.

Eye-Movement-Tracking Task

The eye-movement-tracking task involved120sentences,comprising60experimental sen-tences(sentences with RCs)and60?llers(sentences without RCs).Each of the sentences was presented in single line horizontally from left to right in traditional Chinese on the screen.The experimental sentences were made up of30SRCs and30ORCs,with15short-distance and

long-distance sentences,respectively.The distance was controlled by the variable lengths of the modi?ers preceding the critical HN.The critical HNs were composed of two characters. The average number of strokes for HNs was9.09.The average lexical frequency of HNs in SRCs was549.73and that in ORCs was764.6,1difference between which did not reach statistically signi?cance[t(28)=.48,p=.638].Examples of the experimental sentences are as follows(the complete stimuli are upon request):

Sixty?llers were implemented in the study.All sentences were displayed as single lines in the middle of the LCD monitor.The lengths of the sentences ranged from12to21characters. The characters had a size of36pixels×36pixels and were separated by10pixels×36 pixels.Participants were seated70cm from the computer monitor,which resulted in each character spanning a visual angle of1.06degrees.

Measurement Tools

For the eye-movement data,we employed four measures related to the structural processing of RCs:gaze duration,regression-path duration,total viewing time,and regression rate(which here refers to what is also called regressions in).The gaze duration is the total amount of time spent on all?rst-pass?xations on a region before the eyes move out of the region either right-or left-ward(Rayner1998).This measure is generally regarded as a measure of initial sentence processing.The regression-path duration is the total time spent?xating on all of the target and pretarget regions,from the?rst?xation on a target region to?xation to the right of 1The lexical frequency was computed by word list with accumulated word frequency in Sinica Corpus http:// https://www.wendangku.net/doc/b218399925.html,.tw/eng_teaching.html.

the target region(Rayner and Duffy1986;Liversedge et al.1998).This measure is considered to be sensitive for detecting the dif?culty at later stages of processing.The total viewing time is the total time spent?xating on the target region.The regression rate in our study(which refers to regressions in)is corresponding to the probability of rereading the target(Yen et al. 2008),i.e.,the probability of regressions back into the target region after it has already been read.Please note that for regression rate in the current study,‘regressions in’is analyzed instead of regressions out.We used two measures related to the integration of RCs and the main clause:total viewing time and regression rate.Regarding the comprehension test,the accuracy rate was measured.

Procedure

The eye-movement-tracking task was conducted for about60min.The participants were?rst asked to sign the consent form and provide background information.The participants were given the instruction by demonstration before the practice session and the experiment ses-sion.Before each session the experimenter calibrated the eye-tracker.During the experiment session,there was a short break after the?rst half of the session.The short break was designed to minimize fatigue possibly caused by the infrared rays illuminated on participants’eyes over a long period of time.The eye-movement-tracking task was then conducted for about 30min.

The participants sat70cm from the monitor with their head leaning on the chin rest. The task started with a13-point calibration,followed by5practice trials that were in the same format as the normal trials.Participants were instructed to?xate on a dot(to enable drift correction)located at the position where the?rst character of the sentence would be subsequently displayed.The participants were then instructed to read the sentence in their most natural way when the stimuli appeared on the monitor.When the participants had ?nished reading,they pressed the SPACE key to begin a true/false comprehension statement. True and false statements were equally distributed across conditions.Participants responded by pressing‘F’(with a circle sticker)for true and‘J’(with a cross sticker)for false.Each testing sentence was followed by a reading comprehension true/false question to ensure that the For example,one of the comprehension statements

for(5)was‘The teacher talked in a very polite manner.’There were 120trials in total,with a break at every40trials.A13-point calibration was administered after each break.

Results

Accuracy of Comprehension Test

The overall accuracy rate for RCs was95.7%.The test results for participants(F1)and items(F2)were analyzed by two-way repeated-measures ANOV As.As given in Fig.1,the analysis showed a main effect of clause type only in the participants’analysis[F1(1,39)= 15.86,mean square error(M SE)=.002,p<.001,η2=.29;F2(1,14)=1.06,M SE= .011,p=.320,η2=.07].The main effect of distance was also signi?cant only in the participants’analysis[F1(1,39)=7.33,M SE=.003,p<.05,η2=.16;F2(1,14)= 1.90,M SE=.004,p=.192,η2=.12].However,there was no signi?cant interaction

ORC

long distance short distance

Clause types

Fig.1Accuracy of the comprehension test

of clause type and distance(both ps>.70).These results indicate that the sentences with SRCs were easier to understand than those with ORCs.2

Eye-Movement Data

This study focused on the processing of RC itself and the integration between RC and the main clause.Eye-movement-tracking analyses were performed for HN,EV,and S-NP.

Two-way repeated-measures ANOV As were conducted to analyze the eye-movement data for participants(F1)and items(F2).Fixations shorter than80ms or longer than1200ms (5.08%of total?xations)were not included in the analyses(Drieghe et al.2008;White 2008).Table1lists the descriptive statistics for the indices described below.

Head Noun

HNs were measured using gaze duration,regression-path duration,total viewing time,and regression rate;the analysis results are presented below.

Gaze Duration The main effect of clause type on gaze duration was signi?cant[F1(1,39)= 19.34,MSE=1129,p<.001,η2=.33;F2(1,14)=17.95,MSE=416,p< .001,η2=.56].The gaze duration was signi?cantly longer for HNs in SRCs(M= 278ms,S D=52ms)than for those in ORCs(M=254ms,S D=46ms).However, there was no signi?cant main effect of distance or signi?cant interaction of clause type and distance(all ps>.10)(see Fig.2).

2One of the reviewers mentioned that lexical repetition might reduce effect size,and suggested the linear mixed effect model(LMM)with the trial sequence as a?xed factor for alternative analysis.The results of ANOV As and LMM indicated that the main effects of clause type,distance,and clause type x distance remained the same overall pattern.As for sequence,its interactions with clause type showed that the difference between SRCs and ORCs became bigger over time only for the accuracy rate of comprehension test,gaze duration in S-NP,and regression rate in S-NP,which were noted here.Mainly,the LMM results conformed to the major ?ndings in the current discussions.

Table1Mean(SD)values of eye-movement indices for each language component

Head noun Embedded verb Subject noun phrase

Gaze duration

ORC-L249(44)262(45)1518(446)

ORC-S260(47)269(53)852(232)

SRC-L279(49)264(70)1813(549)

SRC-S276(54)261(58)1023(284) Regression path duration

ORC-L340(87)N/A N/A

ORC-S377(103)N/A N/A

SRC-L551(154)N/A N/A

SRC-S446(102)N/A N/A

Total viewing time

ORC-L407(110)N/A2299(545)

ORC-S453(98)N/A1368(296)

SRC-L515(118)N/A2694(603)

SRC-S480(102)N/A1504(280) Regression rate

ORC-L.34(.18).51(.20).46(.15)

ORC-S.33(.18).62(.20).41(.15)

SRC-L.32(.18).58(.26).41(.16)

SRC-S.31(.16).54(.27).36(.14)

long distance short distance

Clause types

ORC

Fig.2Gaze duration for head nouns in RCs

Regression-Path Duration The main effect of clause type on regression-path duration was signi?cant[F1(1,39)=78.00,M SE=10056,p<.001,η2=.67;F2(1,14)= 71.05,M SE=4189,p<.001,η2=.84].The regression-path duration was signi?-cantly longer for HNs in SRCs(M=498ms,S D=141ms)than for those in ORCs

long distance short distance

Clause types

Fig.3Regression-path duration for head nouns in RCs

(M=358ms,S D=97ms).The main effect of distance was signi?cant only in the par-ticipants’analysis[F1(1,39)=4.33,M SE=10645,p<.05,η2=.10;F2(1,14)= 4.05,M SE=3397,p=.064,η2=.22].The interaction between clause type and distance was also signi?cant[F1(1,39)=24.15,M SE=8287,p<.001,η2= .38;F2(1,14)=33.75,M SE=2081,p<.001,η2=.71].A simple main effect test showed that in long-distance RCs,the regression-path duration was signi?cantly longer for HNs in long-distance SRCs(M=551ms,S D=154ms)than for those in long-distance ORCs(M=340ms,S D=87ms)[F1(1,78)=96.87,M SE= 9172,p<.001,η2=.56;F2(1,28)=104.79,M SE=3135,p<.001,η2=.79]. In short-distance RCs,the regression-path duration was signi?cantly longer for HNs in short-distance SRCs(M=446ms,S D=102ms)than for those in short-distance ORCs (M=377ms,S D=103ms)[F1(1,78)=10.47,M SE=9172,p<.01,η2= .12;F2(1,28)=12.55,MSE=3135,p<.01,η2=.31](see Fig.3).

Total Viewing Time The main effect of clause type on total viewing time was signi?cant [F1(1,39)=25.95,M SE=7013,p<.001,η2=.40;F2(1,14)=7.77,M SE= 8651,p<.05,η2=.36].The total viewing time was signi?cantly longer for HNs in SRCs (M=498ms,S D=112ms)than for those in ORCs(M=430ms,S D=107ms). The main effect of distance was not signi?cant(both ps>.50).The interaction between clause type and distance was also signi?cant[F1(1,39)=15.57,M SE=4178,p< .001,η2=.29;F2(1,14)=7.84,MSE=2881,p<.05,η2=.36].A simple main effect test showed that in long-distance RCs,the total viewing time was signi?cantly longer for HNs in long-distance SRCs(M=515ms,SD=118ms)than for those in long-distance ORCs(M=407ms,S D=110ms)[F1(1,78)=41.52,M SE=5595,p<.001,η2= .35;F2(1,28)=14.54,M SE=5766,p<.001,η2=.34].However,in short-distance RCs,the total viewing time was not signi?cantly longer for HNs in short-distance SRCs than for those in short-distance ORCs(both ps>.10)(see Fig.4).

Regression Rate The main effect of clause type on regression rate was not signi?cant [F1(1,39)=2.10,M SE=.008,p=.161,η2=.05;F2<1],nor were the main effect of distance(both ps>.60)and the interaction of clause type and distance(both ps>.90).

long distance short distance

Clause types

Fig.4Total viewing time for head nouns in RCs

The results for HNs showed that gaze duration,regression-path duration,and total viewing time on SRCs were all longer than ORCs.These results support the prediction of DLT that the processing time for SRCs would be longer than that for ORCs.

Embedded Verb

EVs were measured using gaze duration and regression rate;the analysis results are presented below.Note that the measures of regression-path duration and total viewing time were not used since the positions of the EVs in the two types of RCs are different.In particular,the EV of SRC is in sentence-initial position so that its regression time may be underestimated. Thus,the two measures involving regression time were excluded for the analysis of EV. Gaze Duration The main effect of clause type was not signi?cant(both ps>.60),as was the main effect of distance(both ps>.70)and the interaction of clause type and distance [F1<1;F2(1,14)=1.32,M SE=419,p=.270,η2=.09].

Regression Rate The main effect of clause type was not signi?cant(both ps>.40),as was the main effect of distance(both ps>.06).However,the interaction between clause type and distance was signi?cant[F1(1,39)=9.43,M SE=.023,p<.01,η2=.20;F2(1,14)= 23.27,M SE=.003,p<.001,η2=.62].A simple main effect test showed that in short-

distance RCs,the regression rate was signi?cantly higher for EVs in short-distance ORCs (M=.62,S D=.20)than for those in short-distance SRCs(M=.54,S D=.27)only in the items’analysis[F1(1,78)=2.86,M SE=.041,p=.120,η2=.04;F2(1,28)= 9.06,M SE=.007,p<.01,η2=.24].However,in long-distance RCs,the regression rate was not signi?cantly higher for EVs in long-distance SRCs than for those in long-distance ORCs(both ps>.10)(see Fig.5).

The results from these two indicators cannot differentiate the processing dif?culty between SRCs and ORCs.The interpretation for such?ndings will be discussed in the next section.

long distance short distance

Clause types

Fig.5Regression rate for embedded verbs in RCs

Subject Noun Phrase

S-NPs were measured using gaze duration,total viewing time and regression rate.Note that the measure of regression-path duration was not applicable to the analysis of S-NPs in that S-NPs in RCs are sentence-initial,which leaves nothing preceding to be calculated.The analysis results are presented below.

Gaze Duration The main effect of clause type on gaze duration was signi?cant[F1(1,39)= 84.10,M SE=25844,p<.001,η2=.68;F2(1,14)=173.29,M SE=5088,p< .001,η2=.93].The gaze duration was signi?cantly longer for S-NPs in SRCs(M= 1418ms,S D=585ms)than for those in ORCs(M=1185ms,S D=484ms).The main effect of distance was signi?cant[F1(1,39)=270.96,M SE=78251,p<.001,η2= .87;F2(1,14)=520.74,M SE=14857,p<.001,η2=.97].The gaze duration of S-NPs was signi?cantly longer for long-distance sentences(M=1665ms,S D=515ms) than for short-distance sentences(M=937ms,S D=270ms).The interaction between clause type and distance was also signi?cant[F1(1,39)=5.23,MSE=29195,p< .05,η2=.12;F2(1,14)=9.74,M SE=8867,p<.01,η2=.41].A simple main effect test showed that in long-distance RCs,the gaze duration was signi?cantly longer for S-NPs in long-distance SRCs(M=1813ms,S D=542ms)than for those in long-distance ORCs(M=1518ms,S D=440ms)[F1(1,78)=63.20,M SE= 27520,p<.001,η2=.45;F2(1,28)=108.92,M SE=6978,p<.001,η2=.80]. In short-distance RCs,the gaze duration was signi?cantly longer for S-NPs in short-distance SRCs(M=1023ms,S D=281ms)than for those in short-distance ORCs (M=852ms,S D=229ms)[F1(1,78)=21.33,M SE=27520,p<.001,η2= .22;F2(1,28)=29.82,M SE=6978,p<.001,η2=.52](see Fig.6).

Total Viewing Time The main effect of clause type on total viewing time was signi?cant [F1(1,39)=42.05,M SE=66841,p<.001,η2=.52;F2(1,14)=17.51,M SE= 54317,p<.001,η2=.56].The total viewing time was signi?cantly longer for S-NPs in SRCs(M=2099ms,S D=758ms)than for those in ORCs(M=1834ms,S D= 639ms).The main effect of distance was signi?cant[F1(1,39)=376.43,M SE=

ORC

long distance short distance

Clause types

Fig.6Gaze duration for S-NPs

ORC

long distance short distance

Clause types

Fig.7Total viewing time for S-NPs

119521,p<.001,η2=.91;F2(1,14)=343.75,M SE=50322,p<.001,η2=.96]. The total viewing time of S-NPs was signi?cantly longer for long-distance sentences (M=2497ms,S D=608ms)than for short-distance sentences(M=1436ms,S D= 296ms).The interaction between clause type and distance was also signi?cant[F1(1,39)= 10.86,M SE=61679,p<.01,η2=.22;F2(1,14)=7.17,M SE=36326,p< .05,η2=.34].A simple main effect test showed in long-distance RCs,the total viewing time was signi?cantly longer for S-NPs in long-distance SRCs(M=2694ms,S D=603ms) than for those in long-distance ORCs(M=2299ms,S D=545ms)[F1(1,78)= 48.44,M SE=64260,p<.001,η2=.38;F2(1,28)=24.34,M SE=45322,p< .001,η2=.47].In short-distance RCs,the total viewing time was signi?cantly longer for S-NPs in short-distance SRCs(M=1504ms,S D=280ms)than for those in short-distance ORCs(M=1368ms,S D=296ms)only in the participants’analysis [F1(1,78)=5.73,M SE=64260,p<.02,η2=.07;F2(1,28)=2.38,M SE= 45322,p=.143,η2=.09](see Fig.7).

ORC

long distance short distance

Clause types

Fig.8Regression rate for S-NPs

Regression Rate There was a main effect of clause type on regression rate only in the par-ticipants’analysis[F1(1,39)=4.55,M SE=.019,p<.05,η2=.11;F2(1,14)= 3.33,M SE=.007,p=.089,η2=.19].The main effect of distance was also signi?-cant only in the participants’analysis[F1(1,39)=4.98,M SE=.019,p<.05,η2= .11;F2(1,14)=2.15,M SE=.012,p=.168,η2=.13].The interaction between clause type and distance was not signi?cant(both ps>.80)(see Fig.8).

The results for S-NPs showed that gaze duration and total viewing time on SRCs were both longer than on ORCs,which support the predictions of the word-order account and DLT. On the other hand,the regression rate for S-NPs in SRCs was lower than in ORCs.This result implied that the S-NPs in SRCs were easier than ORCs to integrate with the main clauses, which supports the prediction of the perspective-shift account.3

Discussion

This study investigated the dif?culty of processing RCs in Chinese,and the integration of the RC and the main clause.The results showed that ORCs were easier to process within RCs. However,the integration of the S-NP and the main clause was faster for SRCs.Our?ndings were consistent with the predictions of the DLT,word-order account,and perspective-shift account.Below we discuss our?ndings for theoretical accounts relative to the research questions that we were aiming to address.

ORC Processing Preference Within an RC-Modifying Noun Phrase

The preference for and easier processing of ORCs was evident from?ve results:(a)the gaze duration was shorter for HNs in ORCs than for those in SRCs;(b)the regression-path duration was shorter for HNs in ORCs than for those in SRCs;(c)the total viewing time was shorter for HNs in ORCs than for those in SRCs;(d)the gaze duration for S-NPs in ORCs is shorter than that in SRCs.;and(e)the total viewing time was shorter for S-NPs in ORCs than for those in SRCs.These results matched the predictions of the DLT and word-order account. 3See footnote2.

According to the DLT(Gibson1998),dependency requires greater cognitive resources when two dependent elements are further apart,and readers tend to prefer shorter local dependency and syntactic relationships.In SRCs,readers need greater cognitive resources to integrate the extracted HN and the gap because the HN is far from its gap and readers need to keep in memory the discourse referents(e.g.,jieshao and laoshi in(1))in between while integrating the processed information and upcoming information(which corresponds to a storage cost according to the DLT).In other words,the gap?ller in SRCs formed a longer dependency and there was an intervening discourse referent,so that SRCs had greater integration and storage costs.In ORCs,the gap and its HN form a shorter distance,re?ecting local dependency. Thus,our?ndings suggest that ORCs are preferred in Chinese.

This result can also be explained by the word-order account(MacDonald and Chris-tiansen2002),which argues that sentences with noncanonical word orders may cause a greater processing load for readers.Recall that the dominant word order in Chinese is SVO. Comparing SRCs and ORCs,the clausal word order in SRCs(i.e.,VO…S)is noncanonical, while that in ORCs(i.e.,SV…O)is consistent with the dominant word order.Thus,SRCs should be structurally more dif?cult than ORCs.

The gaze duration of S-NPs in ORCs is shorter than that in SRCs,which also supports the aforementioned result that Chinese ORCs are easier to process than SRCs.

Integration Between RC and Main Clause

The integration of RC with the main clause can be observed by two indicators.The?rst is the accuracy of the comprehension test.The result from the comprehension test showed that the accuracy rate for SRCs was higher than that for ORCs in the participant’s analysis,which may imply that readers understand the sentence with SRC as the subject better than its ORC counterpart because the comprehension questions focused on the interpretation of HN in RC as the subject of the main clause.The second is the regression rate in S-NP.The result of the regression rate in S-NP indicates that readers need less time to integrate the main clause and the S-NP in SRC sentence when they are interpreting the whole sentence.Those results above can be accounted for by the perspective shift,which hypothesizes that perspective changes consume processing resources and therefore posit the processing dif?culty.In the SRC sentence,the HN(e.g.,xiaozhang in(5a))serves as the subject for the RC and also as the subject for the main clause.Thus,in SRCs there are no perspective shifts.However, in the ORC sentence,the HN serves both as the object for the RC and as the subject for the main clause.This means that the perspective switches from the object of the RC(e.g., laoshi in(5b))to the subject of the main clause(e.g.,laoshi in(5b)).When encountering the relativizer DE,readers realized that what they had read so far possibly formed an RC;when they encountered the matrix verb(e.g.,shouhua in(5))they started to consider the object of the RC as the subject of the main clause,and therefore the perspective shift occurred in ORCs but not in SRCs.The perspective in SRCs did not shift and therefore a sentence with SRC is easier to process than its ORC counterpart.

The results from the current study suggest that in Chinese the S-NP in SRC sentences is easier to integrate with the main clause,which can be explained by the perspective-shift account.Analyzing the regression data obtained from an eye-movement-tracking system made it possible to explore the integration between the S-NP and the main clause,which has not received suf?cient attention in previous studies,except for Staub’s(2010)study,which has already investigated to what extent different eye movement measures can dissociate processing patterns for RCs in English.

The Role of Modi?ers in Chinese RCs

In addition to the main?ndings,it is worth mentioning that the modi?ers of the HN in RCs may in?uence how readers process sentences.In the long-distance RCs,the modi?ers of the HN in the ORCs convey information that allows readers to predict the upcoming head, whereas those in SRCs seemingly do not provide such information about the HN.

We found that the total viewing time of the HN was longer in SRCs than in ORCs for the long-distance RCs but not for the short-distance RCs.This difference may be due to the modi?ers in the long-distance RCs.The information conveyed by the modi?ers in ORCs directly modi?es the HN and helps readers to predict the upcoming head.In contrast,the same modi?ers in SRCs modify the object of the RCs instead of the HN.Therefore,the information from the modi?ers does not help readers to predict the upcoming head.Furthermore,the long-distance SRCs in our stimuli are likely to form a nested dependency(where an RC appears within another RC)with a high structural complexity,which could result in incorrect initial parsing.On top of that,the marker DE has multiple functions,including those of a possessive marker,an adjective marker,a relativizer,and a nominalizer,which may cause readers to spend more time processing,since they may?rst parse DE as a possessive marker rather than as a relativizer.This is perhaps also the reason why readers spent more time on the long-distance SRCs than on the long-distance ORCs.

The result indicates that modi?ers function as a prominent cue for the processing of long-distance ORCs.

Processing Chinese RCs as Language Speci?c

Many psycholinguistic studies have found an SRC preference in both head-initial and head-?nal languages.However,previous studies of the processing of RCs in Chinese produced discrepant results;for example,Lin and Bever(2006a)reported an SRC preference while Gibson and Hsiao(2003)reported an ORC preference.This might be due to Chinese present-ing a mixed typological pattern,being a head-initial language with a head-?nal RC structure (Kwon et al.2010).

The current study used an eye-movement-tracking technique to revisit this issue.Our results revealed an ORC preference within an RC and a better cross-clausal integration for the sentence with SRC.Building upon previous studies,from the present results we propose that the ORC preference in Chinese is language-speci?c.An ORC preference is found within RCs,which is probably due to the syntactic nature of Chinese.For example, the word order in ORCs is also the canonical word order in Chinese.Also,the modi?ers in RCs help to predict the HN of ORCs but not that of SRCs.On the other hand,a better cross-clausal integration for the sentence with SRC was found when integrating the RC and the main clause due to no perspective shifts.The sentence with ORC,however,induces perspective shift from the object in the RC to the subject of the main clause.A perspective shift features general cognitive processing,rather than language-speci?c one.Dealing with perspective shifts requires cognitive loads beyond linguistic structures.The more perspective shifts occur in a sentence,the greater dif?culties emerge in the sentence processing.Hence, one perspective shift occurring in the ORC sentence causes more cognitive loads and greater processing dif?culty,whereas no perspective shift emerges in the SRC sentence,leading to a better cross-clausal integration.

The results of the present investigation of processing asymmetry between SRCs and ORCs suggest that ORCs are structurally preferred in Chinese and that SRCs are easier to integrate with the main clause.Our examination of the gaze duration and regression of eye movement

involving parsing and integration made it possible to obtain these results.We argue that those ?ndings suggest an ORC preference within RC and a better SRC cross-clausal integration. Conclusion

In this study we employed an eye-movement-tracking technique to investigate Chinese RC processing.Although this issue has been examined in previous studies,the methods they used were unable to reveal the re-reading patterns that occur when processing becomes dif?cult.Our approach made it possible to detect the dif?culty encountered at later stages of processing,which has uncovered the regression patterns in sentence processing and provided a better understanding of how Chinese RCs are processed and integrated.While most of the previous studies focused on the processing of RC structures,we investigated the effects of RC structures on the integration of the RC and the main clause.The results of this study suggest that there is an ORC preference for the processing of RC structures,which supports the word-order account and the DLT,and a better cross-clausal integration for SRC,which supports the perspective-shift account.The reported?ndings may help to clarify the controversial issue of RC preference in Chinese.

Acknowledgments This research is partially supported by the“Aim for the Top University Project”and “Center of Learning Technology for Chinese”of National Taiwan Normal University(NTNU),sponsored by the Ministry of Education,Taiwan,R.O.C.and the“International Research-Intensive Center of Excellence Program”of NTNU and Ministry of Science and Technology,Taiwan,R.O.C.under Grant Nos.NSC102-2511-S-003-001-MY3;101-2511-S-003-047-MY3;103-2911-I-003-301.

Appendix1

See Table2.

Appendix2

There are15sets of stimuli in the experiment.Every set contains two pairs of RCs,with the ?rst pair short-distance RCs and the second long-distance RCs.In each pair,the?rst sentence is SRC and the other is ORC.

Table2The predictions of processing preference in each interest area based on the accounts Accounts Predictions

HN S-NP Clausal integration

NPAH SRC N/A N/A

Word-order account N/A ORC N/A

IMP SRC SRC N/A

DLT ORC ORC N/A Perspective-shift N/A N/A SRC