Introduction

Chinese relative clauses processing

Chinese has a Subject-Verb-Object (SVO) structure, but the word order in Chinese relative clauses is different from most other languages with SVO structure. Unlike English or French, whose relative clause structure is head-initial, Chinese has head-final relative clauses, with the relative clause preceding the head noun (see (1)). Thus, in the Chinese subject-relative clause (SRC), the word order is “V-N(patient)-de-N(agent)” while in object-relative clause (ORC), the word order is “N(agent)-V-de-N(patient).” On one hand, the distance between the relative clause verb and the head noun is longer in SRCs, which may cause longer integration and processing time for SRCs. In addition, the canonical agent-first word order is reversed in SRCs, potentially also increasing processing difficulty. On the other hand, the de in the ORC is temporarily ambiguous. The relativizer de also appears as a post-nominal genitive (possessive) marker. This ambiguity in ORCs is not resolved until the main clause verb. Thus, it may take longer for readers to process and reanalyze the de in ORCs to reach the correct interpretation. In both cases, the main clause verb is a critical point of integration and/or disambiguation (cf. Staub, Dillon, & Clifton, 2017).

(1) a. 传唤(summon)证人(witness)的(that)法官(judge)喜欢(like)京剧(Beijing opera).
  The judge that summoned the witness liked Beijing opera. (SRC, plausible)
  b. 传唤(summon)法官(judge)的(that)证人(witness)喜欢(like)京剧(Beijing opera).
  The witness that summoned the judge liked Beijing opera. (SRC, implausible)
  c. 法官(judge)传唤(summon)的(that)证人(witness)喜欢(like)京剧(Beijing opera).
  The witness that the judge summoned liked Beijing opera. (ORC, plausible)
  d. 证人(witness)传唤(summon)的(that)法官(judge)喜欢(like)京剧(Beijing opera).
  The judge that the witness summoned liked Beijing opera. (ORC, implausible)
  e. The judge/witness summoned the judge/witness. The witness/judge liked Beijing opera (True/False)

Given these characteristics, there is controversy about whether SRCs are easier to process in Chinese (as they are in English), or whether ORCs are easier. Researchers have found that Chinese SRCs are easier to process and comprehend than ORCs in a series of self-paced reading and eye-tracking experiments (e.g., Jäger, Chen, Li, Lin, & Vasishth, 2015; Lin & Bever, 2006; Vasishth, Chen, Li, & Guo, 2013). They argue that the findings support experience-based or frequentist theories: because SRCs are more frequent than ORCs in Chinese, readers have been exposed to SRCs more and thus process and comprehend Chinese SRCs more quickly and easily (Gennari & MacDonald, 2008; Hsiao & MacDonald, 2013; Levy, 2008).

Conversely, other researchers have observed that SRCs are processed more slowly than ORCs, as evidenced by inflated reading times on the first noun, relative clause verb, relativizer de, and second noun in several self-paced reading studies (Chen, Ning, & Dunlap, 2008; Gibson & Wu, 2011; Hsiao & Gibson, 2003). Based on these findings, these authors propose working-memory-based theories (e.g., Gibson, 1998) account for the difference: when reaching the head noun in a Chinese relative clause, readers have more difficulty retrieving the relative clause verb from their working memory in SRCs than ORCs, because the distance between the relative clause verb and the head noun is greater in SRCs than ORCs. Specifically, the working-memory-based theory predicts this difficulty in retrieving the relative clause verb should occur at the relativizer de and head noun.

Of course, a third possibility exists: Chinese SRCs and ORCs might be relatively equivalent in terms of parsing difficulty, for different reasons. And in fact, a third research camp has reported that no preference exists between relative clauses among Chinese readers. Chen and colleagues (2006) observed no preference for either SRCs or ORCs in a Chinese sentence complexity rating task. As described above, there are grounds to suspect that both structures present parsing difficulty for different reasons. To be more specific, compare Chinese RCs to English RCs. In English, ORCs are the dispreferred structure for multiple reasons: they are less frequent than SRCs (Wells, Christiansen, Race, Acheson, & MacDonald, 2009), they impose a greater working-memory burden (Gibson, 1998), and thematic role assignment is non-canonical (e.g., Christianson, Luke, & Ferreira, 2010; Ferreira, 2003; Lim & Christianson, 2013; Townsend & Bever, 2001; Zhou & Christianson, 2016a, b; Zhou, 2017). Given that these various factors leading to processing difficulty are distributed between SRCs and ORCs in Chinese, there is reason to question whether an a priori assumption that Chinese must have one highly preferred RC structure is warranted. Nevertheless, precisely this assumption as has largely been either explicitly or implicitly made in the previous, conflicted literature.

One potential factor contributing to these mixed results is the plausibility of the thematic role assignments (i.e., agent and patient roles) in the sentences that have been used in previous studies. Since Crain and Steedman (1985), it has been recognized that relative plausibility can modulate processing speed and accuracy of difficult syntactic structures. Although some previous studies of Chinese RC processing have controlled for plausibility (e.g., Gibson & Wu, 2011; Jäger et al., 2015), there has been no attempt to pit plausibility against syntactic structure explicitly. The closest example to this sort of a design was Wu, Kaiser, & Anderson (2012), who manipulated the animacy of the RC heads. They found that processing preferences for the two RC structures could be reversed, depending on the animacy configurations of the two nouns (animate-inanimate, inanimate-animate). This finding strongly suggests that neither memory load nor the relative frequency of the structures can account for the previously observed processing asymmetries. If in fact neither relative clause structure is strongly preferred, i.e., if both structures contain characteristics that could complicate parsing and/or comprehension, then the semantic content might weigh heavily in guiding the online behaviors taken as indicators of parsing and/or the offline comprehension of the structures. In other words, it is possible that previous conflicting results are attributable to semantic confounds in the materials. This is the hypothesis we explore here. In order to do so, we employ an experimental paradigm that has been used over the past 15 years within the Good Enough Processing literature (cf. Christianson, 2016; Ferreira, Christianson, & Hollingworth, 2001; Ferreira & Patson, 2007), which pits asymmetrically plausible thematic role assignments against (purportedly) asymmetrically complex syntactic structures. We describe below how this manipulation results in strong additive interactions in structures such as English passives vs. actives and ORCs vs. SRCs. Specifically, when implausible thematic assignments occur in syntactic structures that are more difficult to process, online processing speed is slowed and offline comprehension accuracy is reduced. The working hypothesis that we test here, then, is that if one Chinese RC structure is in fact preferred over the other, we should find interactively additive effects of plausibility and structure on the dispreferred RC structure. If, however, the two RC structures are approximately equally difficult to process (for different reasons), then we should find only an effect of plausibility such that implausible sentences cause slower processing and worse comprehension accuracy.

Good-Enough (GE) Processing theory

Traditional psycholinguistic theories hold that readers comprehend sentences by computing syntactic representations relatively effortlessly and automatically, especially for arguments of main clause verbs (e.g., Frazier & Clifton, 1996; see Ferreira et al., 2001). From these syntactic structures, semantic representations are derived, based on syntactically licensed interpretations of the verbatim content of the sentences.

Misinterpretations do often happen though, and moreover, they can occur in systematic and predictable ways, not just haphazardly (e.g., due to inattention; see Christianson, 2016). Evidence suggests that misinterpretations are derived from competition between the syntactic structure of the verbatim input and other information sources, i.e., various processing heuristics that can, under certain circumstances, overwhelm the syntactic structure (e.g., Christianson et al., 2001, 2010; Christianson & Luke, 2011; Christianson, Williams, Zacks, & Ferreira, 2006; Ferreira, 2003; den Ouden, Dickey, Anderson, & Christianson, 2016; Ferreira, Bailey, & Ferraro, 2002; Hussey, Ward, Christianson, & Kramer, 2015).

Good-Enough (GE) Processing theory focuses on accounting for how misinterpretations arise (Christianson, 2016; Ferreira et al., 2002; Ferreira & Patson, 2007). It suggests that readers often rely on a heuristic processing route to build representations that are judged to be good enough to handle the current language comprehension task, but often not good enough to reach an accurate interpretation of the actual language input. GE processing-related phenomena have been observed when readers (both native and non-native speakers) process garden-path sentences (Christianson et al., 2001, 2006; Lassotta, Omaki, & Franck, 2016), active/passive sentences (Christianson et al., 2010; Christianson & Lim, 2013a; Ferreira, 2003), and plausible/implausible relative clauses (RCs) (Lim & Christianson, 2013b; Zhou & Christianson, 2016a, b).

When native English speakers read garden-path sentences, such as (2a), they often misinterpret deer as the object of the subordinate verb hunted and answer “yes” to (2b).

(2) a. While the man hunted the deer ran into the woods.
  b. Did the man hunt the deer?
  c. While the man hunted the deer paced in the zoo.

Although this interpretation is certainly available to inferential processes, it is not strictly licensed by the syntax. Furthermore, most readers also assign deer to the subject position of ran, yet still maintain the syntactically unlicensed interpretation that the man hunted the deer and the deer ran into the woods (Patson, Darowski, Moon, & Ferreira, 2009). Note also that readers are far less likely to answer “yes” to questions such as (2b) after reading sentences like (2c), despite the fact that (2c) presents the same structural difficulties as (2a). Thus, these readers derive an interpretation that is good enough to satisfy a general constraint on plausibility, but not good enough to derive an interpretation that is strictly faithful to the content of the input (i.e., that the man was hunting in general, but not necessarily the particular deer that was running away).

Lassotta and colleagues (2016) have observed the same good-enough processing among native French speaking adults and children when these readers (mis)interpreted garden-path sentences in French. In garden-path sentences like Où est-ce qu’Aline a expliqué dans le salon qu’elle allai attraper des papillons? (“Where did Aline explain in the living room that she was going to catch butterflies?”), the filled-gap prepositional phrase (in the living room) should help readers to disambiguate the interpretation and associate the wh-phrase to the embedded-clause verb. However, French adults and children still preferentially attached the wh-phrase to the first verb in the sentence (main clause verb) rather than the embedded-clause verb. The authors attributed the main clause misinterpretation to good-enough syntactic representations. When both the wh-phrase and the filler-gap prepositional phrase are competing for attachment to the main clause verb, readers maintain the incomplete, and ultimately incorrect, syntactic representation that they first build and fail to revise it correctly to reach the true interpretation.

More evidence of good enough processing has been found in processing sentences containing thematic role assignment difficulties, such as active/passive sentences (Ferreira, 2003). When readers are asked to determine the thematic roles of plausible active sentences (e.g., “The dog bit the man”), implausible active sentences (e.g., “The man bit the dog”), plausible passive sentences (e.g., “The man was bitten by the dog”), and implausible passive sentences (e.g., “The dog was bitten by the man”), they make the most errors in response to the implausible passive sentences because the canonical noun–verb–noun syntactic order in English tends to map onto AGENT-verb-PATIENT thematic structure (Townsend & Bever, 2001), and this order generates a plausible interpretation. As a result, the interpretation that should be derived from the noncanonical (passive) syntactic structure is over-ridden by a plausible semantic interpretation, and thematic roles are reversed (cf. Christianson et al., 2010, in which implausible active sentences primed the production of passives, but implausible passives did not). The results suggest that readers rely on a word-order heuristic as well as a plausibility heuristic, which conspire to cause miscomprehension. The same processing pattern has been observed in second language (L2) English speakers’ comprehension as well (Jacob & Felser, 2016; Lim & Christianson, 2013a).

Recently, Zhou and Christianson (2016a) found that readers engage in good-enough processing when reading unambiguous RCs with similarly competing syntactic and semantic information. Computation of the more challenging ORC structure in (3a) is interfered by the implausible semantic information, causing longer fixation durations, longer whole sentence reading times, more regressions in sentence processing, and lower comprehension accuracy in a paraphrase verification task (3b).

(3) a. The cat that the mouse chased was fast.
  b. The mouse chased the cat. The cat was fast. (Paraphrase verification. T/F)

Zhou and Christianson found that readers were more heavily influenced by the semantic plausibility information in computing the meanings of these sentences compared to when the same implausible thematic roles appeared in SRCs. This pattern suggests that when the algorithmic syntactic route runs into increased computational complexity, and its output conflicts with the output of the heuristic semantic route, readers tend to rely on the heuristic route to guide attentional allocation during parsing and in the offline interpretation. This “fast and frugal” processing strategy often yields an incorrect interpretation when sentences are purposely constructed in such a way so as to pit structure and semantics against one another. It could be argued that the application of heuristics is not particularly “fast,” given that reading times are often reported to be elevated in more difficult semantic sentences. However, the “fast and frugal” characterization (Ferreira, 2003) should be evaluated in the full context of the processing routines hypothesized within the GE framework. Slattery, Sturt, Christianson, Yoshida, and Ferreira (2013) present evidence that, contra an earlier proposal by Christianson et al. (2001, 2006), syntactic structures might be fully revised. These revised structures and the semantic representations they generate enter into competition with previously built, or partially built, structures and their representations. Adjudication between the resulting semantic representations, including full inhibition of one of the representations if required by the task at hand, requires time and effort (cf. Hussey et al., 2015).

“Good-enough” processing is usually quite effective and accurate in normal linguistic environments; however, when materials are constructed to contain temporary ambiguities or to pit structure and semantics against one another (as is the case here), comprehension errors are frequently observed. Importantly for the present study, previous results within the good-enough framework predict the following interaction of structural difficulty and plausibility. When one structure is strongly dispreferred (i.e., when it is non-canonical due to less frequent thematic role assignment mappings), implausible thematic role assignments create even more difficulty in both online parsing and offline comprehension, resulting in an additive interaction between structure and plausibility such that implausible non-canonical structures are most difficult to process and comprehend.

In the case of Chinese relative clauses, however, the prediction is potentially different. If one relative clause structure is strongly preferred, the same interaction as in English should hold. But if there is not a strongly preferred structure, plausibility alone should affect online and offline measures. In other words, if both SRCs and ORCs present parsing difficulties (for different reasons), one might expect a relatively equal application of semantic plausibility information across both structures as a strategy for extracting the meaning in a hopefully “good-enough” fashion. Recall that the previous literature disagrees on which structure is preferred, and that the one previous manipulation of plausibility information pushed around the structural preference (Wu et al., 2012). As such, it might be that what has contributed to the conflicting results is plausibility, which sometimes has been controlled, but not manipulated. In this case, we would expect main effects of plausibility only, as structural cues are difficult to compute in both relative clause structures.

Current Study

This study examined structural preferences in Chinese relative clauses by examining how Chinese readers read and comprehend these structures when syntactic structure and semantic plausibility information point toward conflicting interpretations. Because previous studies have shown that animacy significantly affects relative clause processing in Chinese (consistent with our suspicions about the importance of plausibility in Chinese relative clause processing) (Wu, 2009; Wu et al., 2010, 2012), we used only animate nouns as the objects and subjects in our study. Specifically, we used sentences that were similar to those in Zhou and Christianson (2016a), manipulating both syntactic (relative clause types) and semantic complexity (plausibility). According to Good-Enough Processing theory, the relative clause structure that is more difficult to process should show more evidence of plausibility heuristics influencing both the online processing and ultimate interpretations. If the two relative clauses are relatively equal in processing difficulty, plausibility should significantly affected interpretations, and plausibility should not interact with structure. Finally, if initial interpretations are derived solely from the licit, verbatim, syntactic structure, plausibility should have no effect on early first-pass eye movement measures, which are traditionally interpreted as indices of syntactic structure-building, independent of semantic content or sentence-level plausibility or pragmatics (e.g., Rayner, Carlson, & Frazier, 1983); later online measures and offline comprehension might still be influenced by plausibility, however.

In addition, we anticipate online processing effects, realized as inflated reading times on regions that require more syntactic processing effort. Specifically, we expect a significantly shorter processing time for ORC at the second noun, relativizer de, and main clause verb if working memory capacity affects the retrieval of the subject/object in relative clauses (Gibson, 1998; Hsiao & Gibson, 2003). On the other hand, we expect faster processing of SRCs if frequency and experience with relative clauses dominate readers’ processing (Hsiao & MacDonald, 2013; Jäger et al., 2015; Lin & Bever, 2006; Vasishth et al., 2013). We will examine these effects at the relativizer de, second noun, and main clause verb regions. The rationale for analyzing these regions will be elaborated upon in the Results section. In the Discussion, we will discuss the results in connection to previous Chinese relative clause processing studies, as well as general theories of language processing, including Good Enough Processing theory.

Methods

Participants

Fifty-six native Chinese speakers (33 females) who were undergraduate and graduate students at a Midwestern American university participated in an eye-tracking experiment. Their ages ranged from 18 to 30 years old (M = 24; SD = 4.38). They received either $10 or 1 research credit. Six participants’ data were excluded due to either fatigue or difficulty in calibrating their eyes, leaving data from 50 participants in the final analyses. Before the experiment, an additional 22 native Chinese speakers from the same university completed a plausibility norming task for the target sentences. These participants did not participate in the eye-tracking experiment.

Materials

The Chinese sentence reading task included 55 target1 sentences (see Appendix A), most of which were adapted from Ferreira (2003) and Lim and Christianson (2013b), and 180 filler sentences. Structure and plausibility were manipulated in the target sentences, resulting in a 2 (SRC, ORC) × 2 (plausible, implausible) fully-crossed, within-participant and within-item design. Examples of the four sentence conditions are presented in (1). The plausibility of all target sentences (whether adapted or original) was normed for Chinese readers in plausible (e.g., 法官judge传唤summon证人witness) and implausible (e.g., 证人witness传唤summon法官judge) forms on a 1-to-7 Likert scale in simple sentence structures (1= very implausible; 7 = very plausible; plausible: M = 6.86, SD = 0.25; implausible: M = 2.27, SD = 0.94; p < 0.001; see Appendix B). Materials were assigned pseudo-randomly for each participant across four lists in a Latin-square design, so that each participant saw each item only once.

Apparatus

Participants’ eye movements and offline comprehension accuracy were collected using an SR Research EyeLink 1000 remote desktop eye tracker running SR Research Experiment Builder software and sampling at a rate of 1000Hz. We used a chin and forehead rest to stabilize participants’ heads. Sentences were displayed in black SimSun font (15pt) on a white background on a computer monitor, which was approximately 70 cm away from the participants. All viewing was binocular, but data were recorded from only the right eye.

Procedure

After providing informed consent, as specified and approved by the university’s Office for the Protection of Subjects Internal Review Board (IRB), participants sat in front of the eye tracker’s computer monitor. After calibration, they were instructed to read the sentences at a normal reading speed while their eye movements were recorded. After each sentence, a paraphrase verification probe was used to measure comprehension. Readers were asked to decide whether the verification probe was true or false based on the sentence they had just read. The verification probe always contained statements about both the relative clause and the main clause, as illustrated in (3e). The answer was always “True” for the target items, but the number of True and False correct responses was equal across all items in the experiment.

Results

All reading time data points that were three standard deviations from the condition mean and shorter than 80ms were excluded (0.07%). Additional data trimming was performed for each numeric measure in the interest area analyses. Table 1 provides the descriptive statistics for the whole sentence reading time and response accuracy. Additionally, we analyzed the first fixation durations, gaze durations, go-past times, total dwell times, regressions-in, regressions-out, and skipping rates in three interest areas: the relativizer de (的), the second noun in the RC, and the main clause verb.2 These interest areas in ORC (4a) and SRC (4b) are exemplified below.

(4) a. 法官judge传唤 summon   that  证人 witness   喜欢 like 京剧 Beijing opera.
                       Relativizer de            Second Noun Main Clause Verb
  b. 传唤summon 法官judge  that  证人 witness   喜欢 like 京剧 Beijing opera.
                       Relativizer de            Second Noun Main Clause Verb

Table 1

Means and Standard Deviations (in parentheses) for Whole Sentence Reading Time and Comprehension Probe Accuracy.

Whole Sentence Reading Time Accuracy

Implausible Plausible Implausible Plausible

Object-RC 4184.97 (2686.73) 3350.27 (1972) 0.73 (0.44) 0.87 (0.42)
Subject-RC 3994.65 (2499.85) 3169.42 (1764.78) 0.78 (0.33) 0.89 (0.32)

Note: Values in whole sentence reading time are in milliseconds (msec); values for accuracy are proportion correct.

The relativizer de (roughly similar to ‘that’) was selected because it is the relative clause marker in Chinese. The second noun was examined because it is the end of the relative clause. Previous self-paced reading studies (e.g., Chen et al., 2008; Gibson & Wu, 2013) have found that the differences between SRC and ORC were salient at the relativizer de and the second noun region. With respect to the main clause verb, we should see wrap-up effects as well as how readers integrate the relative clauses into the whole sentence (Staub et al., 2017). In addition to these regions, previous studies also examined reading times on the first noun and the relative clause verb. We did not investigate these two regions because they showed up in the first positions of the sentences, prior to the relevant structure.

First fixation duration refers to the length of the initial fixation on an interest area (word). It is measured from the onset of fixation until the eyes move again in either direction, and is generally taken as a measure of lexical recognition (Rayner, 1998). Gaze duration is the sum of total fixations on the interest area before the eyes move out of it in either direction. It is generally taken as a measure of lexical access (Rayner, 1998). First fixation duration and gaze duration are considered to be “early” measures of eye movements, indexing readers’ cognitive processes when they recognize the orthographic properties of a word and access its meaning, respectively. Go-past time, also called regression path duration, is measured from the time a word is first fixated until the eyes move past the word in a forward direction, including all fixations resulting from regressions back into the prior text. Go-past time is taken as a measure of the difficulty of integrating the currently fixated word into the unfolding global sentence representation (structural and semantic) (Rayner, 1998). Total time is the sum of all fixations in the interest area until the offset of the stimulus, which is also a later measure, and which is generally taken to index total processing time on a given word (Rayner, 1998). Regressions-in is the proportion of trials in which readers regress into the interest area, while regressions-out is the proportion of trials they regress out of the region. Skipping shows when readers skip the interest area. Regressions-in and regressions-out reflect processing and integration difficulties. Skipping of a region suggests that the region was identified to some level of recognition prior to fixation. This is generally assumed to be accomplished through parafoveal processing, and is typical for short, frequent, and/or highly predictable words (cf. Drieghe, Rayner, & Pollatsek, 2005). Regressions-in, regressions-out, and skipping are binomial measures, whereas the rest are continuous measures.

Linear mixed-effects modeling (Bates, Mächler, Bolker, & Walker, 2014) with a maximal random effect structure was applied to analyze the continuous reading data, including whole sentence reading times, as well as first fixation durations, gaze durations, go-past times, and total reading times in each interest area. Plausibility, relative clause structure, and their interaction (reference coded) were added as fixed main effects. Logistic mixed-effects modeling with the same main effects and random effects structure was used to analyze the binomial accuracy data as well as regressions-in, regressions-out, and skipping in the interest areas.

Whole sentence reading time

Table 1 presents the descriptive data for whole sentence reading times and response accuracy. Model results for the whole sentence reading time are provided in Table 2. Analyses revealed that plausibility was the only significant predictor of whole sentence reading times. Readers spent more time processing the implausible sentences than the plausible sentences. Structure did not significantly influence whole sentence reading times. The difference in both syntactic conditions between implausible and plausible conditions was approximately 800 msec.

Table 2

Estimates, SEs, t-Value, and p-Value of Fixed Effects for Whole Sentence Reading Time.

Predictor b SE t p

Intercept 0.10 0.09 1.10 0.28
Plausibility –0.30 0.05 –5.95 <0.001
Structure –0.04 0.06 –0.80 0.43
Plausibility × Structure –0.01 0.06 –0.11 0.92

Note: Significant effects are bolded.

Interest Area Analyses

Relativizer de

Trials in which readers skipped the relativizer de were excluded (36%) from all analyses except the skipping data analysis. Table 3 contains the descriptive statistics for the seven interest area measures for the relativizer de. Table 4 provides linear mixed-effects model outputs for the first fixation durations, gaze durations, go-past times, and total dwell times on the relativizer de. Plausibility significantly affected go-past times. Readers had longer go-past times on the relativizer de in implausible sentences than in plausible sentences. No variable significantly predicted first fixation durations, gaze durations, or total dwell times on the relativizer de.

Table 3

Means and Standard Deviations (in Parentheses) for First Fixation Durations, Gaze Durations, Go-past times, Total Dwell Times, Regressions-in, Regressions-out, and Skipping Rates on the Relativizer de Interest Area.

ORC SRC

Implausible Plausible Implausible Plausible

First Fixation Duration 216.41 (95.16) 229.88 (96.35) 222.94 (95.90) 217.4 (89.69)
Gaze duration 219.96 (98.39) 232.1 (109.01) 238.79 (139.93) 220.14 (91.78)
Go past 484.98 (430.53) 396.75 (109.01) 450.79 (385.02) 417.31 (306.19)
Total Time 300.13 (331.47) 238.6 (271.01) 277.07 (343.54) 205.87 (250.95)
Regressions in 0.44 (0.50) 0.46 (0.50) 0.41 (0.49) 0.33 (0.47)
Regressions out 0.25 (0.43) 0.20 (0.40) 0.23 (0.42) 0.29 (0.46)
Skip 0.64 (0.48) 0.64 (0.48) 0.63 (0.48) 0.62 (0.49)

Table 4

Estimates, SEs, t-Values, and p-Values of Fixed Effects for Continuous Reading Measures on the Relativizer de Interest Area.

Predictor First fixation duration Gaze duration Go-past times Total dwell time

b SE t p b SE t p b SE t p b SE t p

Intercept –0.06 0.08 –0.8 0.43 –0.06 0.08 –0.84 0.41 0.09 0.08 1.19 0.24 0.04 0.09 0.47 0.64
Plausibility 0.15 0.1 1.53 0.13 0.11 0.1 1.19 0.24 –0.22 0.09 –2.36 <0.05 –0.11 0.11 –0.96 0.34
Structure 0.07 0.1 0.7 0.49 0.11 0.1 1.12 0.27 –0.07 0.1 –0.73 0.47 –0.02 0.1 –0.22 0.83
Plausibility × Structure –0.22 0.16 –1.35 0.19 –0.23 0.14 –1.67 0.10 0.19 0.13 1.44 0.15 –0.1 0.15 –0.67 0.50

Table 5 provides logit mixed-effects model outputs for the regressions-in, regressions-out, and skipping on the relativizer de. Results showed a marginal interaction of plausibility and structure for regressions-out of the relativizer de such that in the ORC condition, readers regressed out the de in implausible sentences more than plausible ones; in the SRC condition, readers regressed out of the de in the plausible sentences more than the implausible ones. No variable significantly predicted regressions-in or skipping on the relativizer de.

Table 5

Estimates, SEs, z-Values, and p-Values of Fixed Effects for Binomial Measures on the Relativizer de Interest Area.

Predictor Regressions in Regressions out Skip

b SE z p b SE z p b SE z p

Intercept –0.32 0.18 –1.75 0.08 –1.25 0.23 –5.49 <–0.01 0.6 0.13 4.51 <–0.01
Plausibility 0.09 0.21 0.46 0.65 –0.35 0.26 –1.33 0.18 –0.01 0.18 –0.04 0.97
Structure –0.16 0.21 –0.78 0.44 –0.12 0.26 –0.47 0.64 –0.04 0.18 –0.25 0.81
Plausibility × Structure –0.41 0.3 –1.37 0.17 0.71 0.37 1.90 0.06 –0.04 0.25 –0.17 0.86

The high skipping rate of de, a short, frequent, morphosyntactic marker, suggests that it was identified in the parafovea on over one-third of the trials and skipped, irrespective of structure; however, the marginal interaction suggests that readers were aware of its importance to the ongoing parse, and its relationship to the plausibility of the thematic roles of the surrounding noun phrases. Because de is so short, it is almost certain that both the upcoming noun and the previous noun were in the parafovea while de was fixated (Rayner, 1998). This rather wide and complex collection of information sources would explain why all of the fixation measures on de, especially the measures associated with integration into the unfolding parse (total time, go-past time), were relatively large. A precise explanation of how the various complexities of both SRCs and ORCs drive integration of de (as indexed here by the marginally significant interaction) requires further investigation.

Second Noun

Trials in which readers skipped the second noun were removed (7.5%) from analyses except for the skipping data analyses. Table 6 contains the descriptive statistics for the seven interest area measures for the second noun. Table 7 presents the model results for continuous measures on the second noun. Results showed that plausibility was a significant predictor for go-past times and total reading times on the second noun, with implausible sentences displaying inflated times in all measures. It appears that the level of disruption observed at the relativizer de had been resolved by the time the second noun was fully fixated, and plausibility alone dictated the relatively long later measures associated with integration into the unfolding parse (total time, go-past time).

Table 6

Means in Msec and Standard Deviations (in Parentheses) for First Fixation Durations, Gaze Durations, Go-past times, Total Dwell Times, Regressions-in, Regressions-out, and Skipping Rates on the Second Noun.

ORC SRC

Implausible Plausible Implausible Plausible

First Fixation Duration 234.89 (93.33) 231.76 (90.67) 237.38 (86.90) 240.12 (102.68)
Gaze duration 273.37 (140.20) 278.16 (166.87) 290.89 (165.82) 278.53 (154.78)
Go past 694.19 (715.61) 502.67 (427.23) 694.69 (691.98) 530.75 (427.56)
Total Time 709.99 (605.14) 538.16 (495.96) 656.11 (473.61) 499.44 (408.50)
Regressions in 0.32 (0.47) 0.33 (0.47) 0.29 (0.45) 0.25 (0.43)
Regressions out 0.37 (0.48) 0.28 (0.45) 0.37 (0.48) 0.33 (0.47)
Skip 0.15 (0.36) 0.23 (0.42) 0.17 (0.37) 0.20 (0.40)

Table 7

Estimates, SEs, t-Values, and p-Values of Fixed Effects for Continuous Reading Measures on the Second Noun Interest Area.

Predictor First fixation duration Gaze duration Go-past times Total dwell time

b SE t p b SE t p b SE t p b SE t p

Intercept –0.01 0.06 –0.14 0.89 –0.03 0.06 –0.43 0.67 0.09 0.08 1.13 0.26 0.15 0.11 1.29 0.20
Plausibility –0.04 0.08 –0.45 0.65 –0.02 0.09 –0.24 0.82 –0.27 0.10 –2.85 <0.05 –0.29 0.12 –2.44 <0.05
Structure 0.03 0.08 0.43 0.67 0.09 0.08 1.09 0.28 0.05 0.09 0.55 0.59 –0.01 0.1 –0.06 0.95
Plausibility × Structure 0.04 0.11 0.31 0.75 –0.07 0.12 –0.58 0.57 0.05 0.12 0.38 0.70 –0.11 0.13 –0.82 0.41

Plausibility also significantly modulated readers’ regressions-out and skipping on the second noun; they made fewer regressions out of the second noun and skipped it more in plausible sentences compared to implausible sentences. Table 8 presents the model results for the binomial measures on the second noun.

Table 8

Estimates, SEs, z-Values, and p-Values of Fixed Effects for Binomial Measures on the Second Noun Interest Area.

Predictor Regressions in Regressions out Skip

b SE z p b SE z p b SE z p

Intercept 0.89 0.2 4.38 <0.01 –0.59 0.16 –3.63 <0.01 –2.06 0.25 –8.12 <0.01
Plausibility 0.03 0.18 0.17 0.87 –0.44 0.19 –2.3 <0.05 0.58 0.24 2.4 <0.05
Structure 0.17 0.19 0.91 0.37 0.01 0.18 0.07 0.94 0.1 0.25 0.39 0.69
Plausibility × Structure 0.24 0.27 0.91 0.36 0.26 0.27 0.95 0.34 –0.34 0.34 –1 0.32

Main Clause Verb

Trials in which readers skipped the Main Clause Verb were removed (4.1%) from analyses except for the skipping data analyses. Table 9 contains the descriptive statistics for the seven interest area measures for the main clause verb. Table 10 presents the model results for the continuous measures on the main clause verb. Analysis of go-past times showed a significant plausibility effect: plausible sentences led to shorter go-past times. There was also a marginal effect of structure on total time, with SRC main verbs exhibiting approximately 50 msec faster reading total times than ORC main verbs.

Table 9

Means and standard deviations (in parentheses) for first fixation durations, gaze durations, go-past times, total dwell times, regressions-in, regressions-out, and skipping rates on the main clause verb.

ORC SRC

Implausible Plausible Implausible Plausible

First Fixation Duration 249.18 (113.99) 234.48 (91.20) 248.07 (109.63) 243.79 (101.77)
Gaze duration 435.63 (335.60) 451.37 (337.85) 445.88 (345.42) 493.73 (424.12)
Go past 2368.4 (2110.7) 1881.19 (1604.63) 2264.37 (2055.8) 1648.78 (1372.46)
Total Time 952.54 (753.96) 875.54 (749.26) 911.89 (852.59) 814.18 (674.92)
Regressions in 0.01 (0.09) 0.02 (0.14) 0.01 (0.08) 0.01 (0.11)
Regressions out 0.83 (0.38) 0.82 (0.39) 0.85 (0.36) 0.80 (0.40)
Skip 0.02 (0.14) 0.05 (0.22) 0.05 (0.21) 0.03 (0.17)

Table 10

Estimates, SEs, t-Values, and p-Value of Fixed Effects for Continuous Reading Measures on the Main Clause Verb Interest Area.

Predictor First fixation duration Gaze duration Go-past times Total dwell time

b SE t p b SE t p b SE t p b SE t p

Intercept 0.04 0.07 0.51 0.62 –0.05 0.08 –0.66 0.52 0.13 0.13 0.97 0.34 0.05 0.12 0.45 0.66
Plausibility –0.1 0.09 –1.12 0.27 0.03 0.09 0.33 0.74 –0.25 0.12 –2.08 <0.05 –0.11 0.11 –0.98 0.33
Structure –0.01 0.08 –0.06 0.95 0.02 0.09 0.27 0.79 –0.03 0.11 –0.32 0.75 –0.01 0.1 –0.06 0.95
Plausibility × Structure 0.08 0.12 0.67 0.51 0.05 0.13 0.43 0.67 –0.05 0.15 –0.32 0.75 –0.09 0.15 –0.58 0.56

Table 11 shows model results for the continuous measures on Main Clause Verb. No variable significantly affected the early measures, regressions-in, or regressions-out of the main clause verb. With respect to skipping, plausible sentences displayed a marginally higher skipping rate over implausible sentences, and the interaction between plausibility and structure was significant. What seems to be driving the interaction is that readers tended to skip the main clause verb in implausible SRCs more than in implausible ORCs; however, all skipping rates were =/<5%, so it is not clear how robust this effect is.

Table 11

Estimates, SEs, z-Values, and p-Values of Fixed Effects for Binomial Measures on the Main Clause Verb Interest Area.

Predictor Regressions in Regressions out Skip

b SE z p b SE z p b SE z p

Intercept –11.66 2.66 4.38 <0.01 2.06 0.29 7.20 <0.001 –5.16 0.63 –8.13 <0.01
Plausibility 0.58 2.64 0.22 0.83 0.16 0.28 0.57 0.57 1.07 0.57 1.89 0.06
Structure –0.51 3.41 0.15 0.88 0.19 0.29 0.66 0.51 0.90 0.58 1.56 0.12
Plausibility × Structure 0.05 4.27 0.01 0.99 0.25 0.40 0.62 0.54 –1.58 0.79 –2.01 <0.05

Comprehension accuracy

Table 12 presents the model outputs for comprehension accuracy analyses. The results revealed that both plausibility and structure significantly affected the response accuracy to the paraphrase verification probes. Readers comprehended plausible sentences more accurately than implausible sentences. They also showed higher comprehension accuracy for SRCs than ORCs. There was no interaction between plausibility and structure.

Table 12

Estimates, SEs, z-Values, and p-Values of Fixed Effects for Response Accuracy.

Predictor b SE z p

Intercept 1.28 0.36 3.52 <0.001
Plausibility 2.46 0.42 5.79 <0.001
Structure 0.95 0.41 2.34 <0.05
Plausibility × Structure 0.10 0.63 0.17 0.87

Note: The significant effects are in bold.

Discussion

This study examined the online processing and offline comprehension of relative clauses in Chinese (Mandarin). Previous research on Chinese relative clause processing has returned three basic results. As detailed in the introduction, some studies claim to show that subject-relative clauses (SRCs) are easier to process and comprehend, as predicted by experience-based processing theories (e.g., Jäger et al., 2015; Lin & Bever, 2006; Vasishth et al., 2013). Other studies claim to show that object-relative clauses (ORCs) are easier, as predicted by working-memory-based theories (e.g., Chen et al., 2008; Gibson & Wu, 2011; Hsiao & Gibson, 2003). A smaller third set of findings suggests that there is no clearly preferred relative clause structure in Chinese (Chen et al., 2006). This lack of preference is explained by appealing to the competing theories of Chinese relative clause processing. Specifically, there is reason to believe that both structures present processing challenges of different sorts, and there is no a priori reason to assume that one or some combination of these challenges should produce a clearly “preferred” structure. Furthermore, a survey of the existent literature reveals that although several previous studies have controlled for plausibility, only two manipulated plausibility (operationalized as animacy) as an independent variable. In those studies (Chen et al., 2006; Wu et al., 2012), relative clause preferences were made to flip depending on the relative plausibility of the nouns. Thus, we applied a processing principle from Good-Enough Processing (cf. Christianson, 2016; Ferreira et al., 2003) to explore Chinese relative clause processing patterns. Specifically, Good-Enough Processing predicts that heuristic (i.e., non-syntactic) information, such as plausibility, should exert more influence on the online processing and offline interpretation of syntactic structures that are more difficult to parse, compared to those that are easier to parse.

In order to test the importance of plausibility in Chinese relative clause processing and comprehension, we crossed SRC and ORC structures with plausible and implausible thematic role assignments in Chinese in a fully-crossed 2 × 2 design. We hypothesized that if Chinese does in fact have a strongly preferred relative clause structure, then the addition of implausible thematic role assignments to the more difficult (dispreferred) structure should produce an additive interaction such that the implausible, dispreferred structure becomes extremely difficult to process and comprehend. This prediction derives from several studies in English relative clause processing and passive sentence processing resulting in just such additive interactions (e.g., Christianson et al., 2010; Ferreira, 2003; Lim & Christianson, 2013b; Zhou & Christianson, 2016a). Alternatively, if Chinese relative clauses are relatively equivalent in their difficulty, perhaps for different reasons, we expected to observe a significant main effect for plausibility alone (Chen et al., 2006; cf. Stoops, Luke, & Christianson, 2014, on the importance of animacy information in Russian processing). This approach represents a novel contribution to the current debate in the literature as to whether SRCs or ORCs are easier to process in Chinese (Chen et al., 2006; Gibson & Wu, 2011; Jäger et al., 2015; Vasishth et al., 2013).

The results showed that plausibility alone significantly affected whole sentence reading times. Readers spent less time reading overall when processing plausible sentences compared to implausible sentences, regardless of relative clause structure. The results are consistent with those of Chen et al. (2006), who manipulated the animacy of the nouns in Chinese relative clauses (an indirect manipulation of plausibility). Chen also found no significant effects of structure, only a significant effect of animacy, and concluded that, “…semantic cues play a more significant role than syntactic cues in Chinese subjects’ sentence processing…” (Chen et al., 2006, p. 8). Importantly, the results are inconsistent with both predictions generated by experience-based processing theories and working-memory-based theories, because Chinese readers did not show a reading time advantage for either of the relative clause structures.

Interest area analyses generally yielded data consistent with the whole sentence reading results. There were no effects of plausibility or structure in early reading measures, first fixation durations and gaze durations, on the relativizer (relative clause marker) de, the second noun, or the main verb. However, there were clear signs of processing disruption on de, as syntactic structure and semantic plausibility information appeared to influence eye movement measures in a not altogether clear pattern. Overall, though, readers seemed relatively insensitive to structural considerations in all IAs. There was no significant main effect of structure in early or late measures, in any IA analysis, contrary to syntax-first, two-stage models (e.g., Rayner, Carlson, & Frazier, 1983). Rather, readers seemed more sensitive to plausibility information largely with respect to integration of the contents of the various IAs, once they had been lexically identified and accessed, as suggested by significant plausibility effects on go-past times on de, the second noun, and the main clause verb. Readers also appeared to rely primarily on plausibility information when reanalyzing and integrating the words into the structure and deriving an interpretation. Thus, the IAs results also failed to support experience-based processing theories or working-memory-based theories of Chinese RC processing. Rather, the IA data suggest that readers employed a good-enough processing strategy to follow a semantic route to comprehension.

Both plausibility and structure modulated readers’ offline comprehension; however, these effects were notably non-interactive. Plausible sentences were easier to comprehend than implausible ones. SRCs were easier to comprehend than ORCs. This result provides some evidence in support of experience-based processing accounts of Chinese relative clause processing (e.g., Jäger et al., 2015; Lin & Bever, 2006; Vasishth et al., 2013), as SRCs are more frequent. However, in the previous experience-based processing studies, none of the comprehension differences reached significance in the three self-paced reading experiments in Vasishth et al. (2013). Similarly, Jäger et al. (2015) failed to find any comprehension accuracy differences between ORCs and SRCs in either their self-paced reading experiment or their eye-tracking experiment. This offline preference for SRCs in the current study was notably slight, and not consistently observed in the online processing time measures. Taken together, these results suggest that any a priori assumptions about the strong canonicity of either of the relative clause structures in Chinese should be viewed with caution.

Finally, these results are consistent with observations in other languages regarding the way in which competing syntactic and semantic information is integrated. Within the Good Enough framework (Christianson, 2016; Ferreira et al., 2002), it is proposed that input is subjected to “algorithmic” syntactic parsing as well as “heuristic” processing. The syntactic parse is fragile, however (Ferreira, 2003; Sachs, 1967), and in cases in which the output from the two processing routes conflict, it may be overwhelmed by the output of the heuristic route. Such a system predicts that structures that are non-canonical – less frequent or otherwise more costly to compute – should be more vulnerable to intrusions from heuristic processing. As a result, implausible sentences are more likely to end up misinterpreted with their thematic roles reversed, thereby generating a plausible but syntactically unfaithful interpretation. In English, ORCs are also more susceptible to these thematic role reversals (e.g., Zhou & Christianson, 2016a). Unlike the case of English relative clauses, however, plausibility and structure did not interact in the current Chinese relative clauses in online or offline measures. The significantly lower accuracy rates for offline comprehension were numerically in the expected direction, though: lower accuracy in implausible sentences, and lowest accuracy in implausible ORCs, in which both meaning and, according to some of the results here (mostly marginal effects in the online measures), structure conspire to flip the thematic roles. Of particular interest to future research is the broader pattern observed here, however, suggesting that semantic plausibility appears to have a much stronger effect on online processing measures than do structural considerations, and that these factors do not interact in Chinese relative clause structure. If this pattern were to be replicated, and expanded to other Chinese structures, it would suggest that Chinese language processing might rely on “good-enough” routines more so than English or, say, French, in which both structure and plausibility have been shown to have early significant and interactive online effects (Lassotta et al., 2016; Zhou & Christianson, 2016a, b). If this observation is borne out in future research, it is possible Chinese susceptibility to good-enough processing could be related to the extreme paucity of inflectional morphology in the language and occasional indeterminacy with respect to word boundaries (Lin, Anderson, Ku, Christianson, & Packard, 2011) in the unspaced orthography, which may induce further processing load in complex structures.

In sum, our reading time data demonstrated that readers showed no preference for either Chinese relative clause structure, which is inconsistent with the predictions from experience-based processing accounts and working memory-based accounts. Although there was a slightly advantage in comprehending SRCs, supporting experience-based accounts, this SRC comprehension advantage has not been consistently observed in previous studies and may not be reliable. Both online reading and offline comprehension data showed strong plausibility effects, suggesting that Chinese readers employ a heuristic, good-enough processing strategy when processing relative clauses manipulated to pit plausibly against structure. This study thus provides a novel explanation for the debate in the Chinese relative clause processing literature, and it is the first to test Good-Enough Processing theory in Chinese.

Data Accessibility Statement

All the stimuli can be found in the Appendix A. The plausibility norming results for all the stimuli can be found in Appendix B. Data available upon request.