2. Experiment 1

Experiment 1 was designed to investigate the interpretational biases of pro and overt pronouns in contextually neutral sentences. As illustrated in (9), the experimental sentences were created to be globally ambiguous such that the subject (e.g. Li Gang _m) and the object (e.g. Wang Qiang _m) could both potentially be an antecedent for the pronoun (e.g. pro or an overt pronoun). (The subscript _m indicates ‘male’ name bias; in later examples, the subscript _f is used to indicate ‘female’ name bias.)

That is, in (9) the event ‘Li Gang _m being in an office’ is just as plausible as the event ‘Wang Qiang _m being in the office’.

Given the results of previous studies of languages with pro, it was predicted that pro was likely to refer to the NP in the subject position. This is consistent with Accessibility theory (Ariel Reference Ariel1990), in that as the highest accessibility marker, pro refers to a discourse antecedent with high mental accessibility. On the other hand, the prediction for the overt pronoun was less clear. Studies of Italian and Spanish, for example, showed division of labor between pro and overt pronouns with a clear object bias for the overt pronoun and a subject bias for pro (Alonso-Ovalle et al. Reference Alonso-Ovalle, Fernández-Solera, Frazier and Clifton2002, Carminati Reference Carminati2002, Filiaci et al. Reference Filiaci, Sorace and Carreiras2014). This is also consistent with Accessibility theory because as a lower accessibility marker, an overt pronoun was predicted to refer to an antecedent with lower mental accessibility, such as an NP in the object position in our experiment. However, previous studies of Chinese (as well as of those of Korean and Japanese) suggest a subject bias both for pro and overt pronouns, even though Yang et al. (Reference Yang, Gordon, Hendrick and Wu1999, Reference Yang, Gordon, Hendrick, Wu and Chou2001), in fact, claimed that there was no need to distinguish pro and overt pronouns in Chinese in the way that they contributed to discourse coherence. Thus, given these previous studies, we predicted that overt pronouns would also show subject bias similarly to pro. However, given the results of Experiment 4 of Yang et al. (Reference Yang, Gordon, Hendrick and Wu1999) and the predictions of Accessibility theory, which had been supported by studies of Italian and Spanish, we also predicted that the degree of the subject bias would be stronger for pro than for overt pronouns.

2.1 Method

2.1.1 Participants

Twenty native speakers of Mandarin participated in Experiment 1 (six males; average age: 19 years). At the time of the study, they were enrolled at a university in China. The participants were born and raised in China and had never lived abroad. No one reported a native-like proficiency in a language other than Chinese.

2.1.2 Materials

Thirty-two sets of complex sentences consisting of a main clause, preceded by a subordinate clause similar to those in (9) were constructed. Pronominal elements (pro or overt pronouns) occurred in the main clause. The subordinate clause contained two same-gender referents appearing respectively in the subject and object positions. These same-gender referents were equally distributed between males and females using typical Chinese proper names such as Li Gang _m and Ding Lan _f. To make sure that all the names we used were clear in their associated gender, we first ran a norming study. Ten college students (five males; average age: 19 years) who did not take part in other experiments rated 90 Chinese names for their implied gender on a scale from 1 (clearly male) to 5 (clearly female). Based on the results, we identified 84 names (42 male names) that received average ratings of lower than 2 (likely to be male) or higher than 4 (likely to be female) and used only those names in the following experiments (see Appendix). Two lists were created based on the Latin Square design such that a participant would see only one condition from each pair. The 32 experimental items were combined with an additional 60 fillers. Fillers consisted of two clauses containing one or two referents involving pro, overt pronouns, or reflexives, as shown in (10) (the gloss LE indicates currently relevant state of the example).

Thus, each list contained 92 sentences that were pseudo-randomized such that experimental items from the same condition would not appear in the same row. Every sentence in a list was followed by two choices, where the pronominal element in a sentence was interpreted as referring to the NP in the subject or object position, as shown in (11) below.

The order of the choices was counterbalanced such that the subject antecedent appeared as option A for half of the items and as option B for the remaining half.

In addition, participants were required to rate how confident they were of their choices on a Likert scale from 1 (very unsure) to 5 (absolutely certain), as shown in (12).

Since anaphoric resolution in our experiment is an issue of preference but not of grammaticality, confidence rating is an important indicator of the participants’ gradient difference when they make a forced choice between two readings (see Christianson et al. Reference Christianson, Hollingworth, Halliwell and Ferreira2001, Carminati Reference Carminati2002).

2.1.3 Procedure

The experiment was carried out in a group in a classroom. Participants were given questionnaires along with written instructions with examples. There was no time limit.

2.1.4 Data analysis

Two experimental items were removed from analyses, as they contained transfer-of-possession verbs whose tense could bias the interpretation of ambiguous pronouns (see Rohde Reference Rohde2008). For the remaining 30 sets of experimental stimuli, we first analyzed choice responses using a generalized Linear Mixed Effect (LME) model with a binomial distribution (Baayen, Davidson & Bates Reference Baayen, Davidson and Bates2008). The models included the experimental condition (Pronoun type: pro vs. overt) as a fixed effect and crossed random effects for participants and items. They were built to have maximum random effects and were only simplified in cases of non-convergence (Barr et al. Reference Barr, Levy, Scheepers and Tily2013). Random slope parameters included in the model are reported in the ‘slope’ column of Table 2 below, along with coefficients, standard errors and z-values for the fixed effects. We calculated p-values with Monte Carlo Markov Chaining (pvals.fnc function, Bates & Maechler Reference Bates and Maechler2010). In addition, confidence ratings were analyzed based on Linear Mixed Effect Regression (LMER) (Baayen Reference Baayen2008, Baayen et al. Reference Baayen, Davidson and Bates2008, Jaeger Reference Jaeger2008), using the lme4 R package (Bates et al. Reference Bates, Maechler, Bolker and Walker2015; version 1.1-8). The experimental conditions (Pronoun type: pro vs. overt pronouns) and Choice response (subject vs. object), in addition to their interactions, were included as fixed effects. For random effects, intercepts for subject and item as well as by-subject and by-item random slope for both effects were included. The analysis yielded coefficients, standard errors, and t-values for each fixed effect and interaction. For the linear models, the coefficient was considered significant at α = 0.05 if the absolute value of t exceeded 2 (Baayen Reference Baayen2008).

2.2 Results and discussion

Experimental results are summarized in Table 1 and Figure 1, and statistical analysis results are presented in Tables 2 and 3. The results showed that there were more subject choices (448 out of 600 total responses; 74.7%) than object choices (152 responses; 25.3%) regardless of the pronominal type. However, while both pronominal types showed subject bias, the Overt pronoun condition elicited a higher number of object responses (104 out of 300 total responses) than the Pro condition (48 out of 300 total responses). These observations were confirmed with a significant main effect of the Pronoun type (p < 0.0001), suggesting that the subject antecedent preference was stronger for pro (84% of the total responses) than for overt pronouns (65.3%).

Table 1 Antecedent choice responses and mean confidence ratings (standard deviation in parentheses) in Experiment 1.

Table 2 Generalized linear mixed effects results for antecedent choices in Experiment 1. The asterisk indicates that the effect is significant at p < .05 (based on the |t| > 2 criterion).

Table 3 Linear mixed effect model results for confidence ratings in Experiment 1. The asterisk indicates that the effect is significant at p < .05 (based on the |t| > 2 criterion).

Figure 1 Percentage of antecedent choices.

Similarly, analyses of confidence ratings showed a significant main effect of the Pronoun type, suggesting that participants were more confident about their interpretations of pro (4.26 out of 5) than about those of overt pronouns (3.95) (t = 3.424). There was also a main effect of Choice, indicating that participants were more confident when they responded with the subject antecedent interpretation than the object antecedent interpretation (t = 3.256). Finally, even though there was no significant interaction between the Pronoun type and the Choice, we ran planned paired comparisons using the Tukey test (glht function of multcomp package: Hothorn, Bretz & Westfall Reference Hothorn, Bretz and Westfall2008; version 1.4-1) in R (R Core Team 2018). The results showed that with the subject antecedent interpretation, the Pro condition received significantly higher confidence ratings than the Overt pronoun condition (p = .01). In contrast, with the object antecedent interpretation, there was no difference between Pro and Overt pronoun conditions (n.s.).

Overall, both pro and overt pronouns displayed a subject preference. This is consistent with previous studies of discourse-oriented languages such as Korean (Kwon & Polinsky Reference Kwon and Polinsky2011, Kim et al. Reference Kim, Theres and Schafer2013) and Japanese (Ueno & Kehler Reference Ueno and Kehler2016) but not with the studies of Spanish (Alonso-Ovalle et al. Reference Alonso-Ovalle, Fernández-Solera, Frazier and Clifton2002, Filiaci et al. Reference Filiaci, Sorace and Carreiras2014), Italian (Caminati Reference Carminati2002), Catalan (Mayol & Clark Reference Mayol and Clark2010) or Greek (Papadopoulou et al. Reference Papadopoulou, Peristeri, Plemenou, Marinis and Tsimpli2015). In addition, these results are consistent with our observation of Yang et al.’s data (their Experiment 4) discussed in Section 1 above (see also Li Reference Li2014, Li et al. Reference Li, Mak and Sanders2016). Figure 3 in Yang et al. suggests reading times were faster for the Overt pronoun condition than for the Pro condition when the pronoun referred to the object NP. Similarly, our results showed that the Overt pronoun condition was more likely to refer to the object NP than the Pro condition was even though both pronominal types were subject-biased. These results suggest that while both pro and overt pronouns were subject-biased in Chinese, pro had a stronger subject bias and/or a stronger dispreference to refer to a non-subject NP than overt pronouns did.

To summarize, the results of Experiment 1 showed that both pro and overt pronouns were subject-biased in semantically neutral contexts. However, this subject bias was stronger for pro than for overt pronouns. In Experiment 2, we investigated the interpretational preferences of pro and overt pronouns in semantically biasing contexts.

3. Experiment 2

Experiment 1 showed clear subject bias both for pro and overt pronouns in semantically neutral contexts. In Experiment 2, we examined the interpretational biases of pro and overt pronouns in semantically biasing contexts using a naturalness rating task. Thus, by comparing the results of Experiment 2 with those of Experiment 1, we aimed to examine the role of sentential contexts in referential resolution of pro and overt pronouns in Chinese.

As illustrated in (13), the sentences in Experiment 2 consisted of a main clause and a preceding subordinate clause (the gloss pfv indicates perfective aspect marker).

The main clause contained either pro or an overt pronoun, and the subordinate clause contained two potential antecedents of the same gender. Crucially, the sentences were devised such that the antecedent of a pronominal element would be semantically signaled in a given sentence. For example, in (13a) and (13b), 吴军 ‘Wu Jun _m’ is the person who makes a call and is also likely to be the person making a mistake in the process. Thus, pro in (13a) and 他 ‘he’ in (13b) are likely to refer to the subject NP (i.e. 吴军 ‘Wu Jun _m’). On the other hand, in (13c) and (13d), as 周斌 ‘Zhou Bin _m’ is the person who receives a phone call, he is likely to be the person who answers the phone with some delay. Thus, pro in (13c) and 他 ‘he’ in (13d) are likely to refer to the object NP (i.e. 周斌 ‘Zhou Bin _m’).

Given the results of Experiment 1, we predicted that the subject bias of pro and overt pronouns would interact with sentential contexts. Thus, sentences would sound more natural when the sentential contexts coincided with the interpretational biases of a pronominal than when the contexts conflicted with them. Given that both pro and overt pronouns were subject-biased, it was predicted that sentences with subject-biased contexts ((13a) and (13b)) would sound more natural than sentences with object-biased contexts ((13c) and (13d)). In addition, given that pro had a stronger subject bias than overt pronouns, for the subject-biased context conditions, sentences with pro would sound more natural than sentences with overt pronouns. On the other hand, for the object-biased contexts, we predicted the reverse to be true; sentences with overt pronouns would sound more natural than sentences with pro.

3.1 Method

3.1.1 Participants

Twenty undergraduates (eight males; average age: 19 years) enrolled in a university in China were paid to participate in Experiment 2. They were all native speakers of Mandarin Chinese and did not participate in any other experiment in this study.

3.1.2 Materials

Thirty-two sets of target sentences similar to those in (13) above were created. As in Experiment 1, every sentence was followed by two questions. First, participants were asked to select their preferred interpretation for each sentence. For example, sentences (13a) and (13b) were followed by statement (14), while sentences (13c) and (13d) were followed by statement (15).

In order to prevent an order effect, we balanced the number of the subject (e.g. Wu Jun _m) and object NP (e.g. Zhou Bin _m) occurring as option A or B. Thus, for half of the sentences, the subject NP and object NP were presented as option A and B respectively; for the remaining half of the sentences, the order was reversed.

Participants were also asked to rate the naturalness of the sentences given their interpretation of choice using the Likert scale from 1 (very awkward) to 5 (absolutely natural), as shown in (16).

Four lists were created based on the Latin square design such that a participant would see only one condition from the same item in a given list. In addition, sixty filler sentences were added to the lists. Fillers consisted of two clauses containing one or two different-gender referents and involved pro, an overt pronoun or a reflexive, as shown in (17).

These 92 sentences in each list were displayed in a pseudo-random order, preventing two experimental sentences from the same condition from appearing successively.

3.1.3 Procedure

The procedure of this experiment was the same as that in Experiment 1.

3.1.4 Data analysis

Interpretational preference was analyzed using a generalized Linear Mixed Effect (LME) model with a binomial distribution, and naturalness ratings were analyzed using a Linear Mixed Effect Regression (LMER) analysis (Baayen Reference Baayen2008, Baayen et al. Reference Baayen, Davidson and Bates2008, Jaeger Reference Jaeger2008). The models incorporated the pronoun type and contextual bias, as well as their interaction, as fixed factors and crossed random effects for participants and items. The remaining analysis procedures were the same as the ones reported in Experiment 1.

3.2 Results and discussion

The results of Experiment 2 are summarized in Table 4 and Figure 2 below, and the results of statistical analyses are presented in Tables 5 and 6 below. Overall, our predictions were partially confirmed.

Table 4 Selection rate of the intended interpretations (raw counts in parentheses) and naturalness rating of Experiment 2 (standard deviations in parentheses).

Table 5 Results of statistical analyses of preferred interpretation of Experiment 2. The asterisk indicates that the effect is significant at p < .05 (based on the |t| > 2 criterion).

Table 6 Linear mixed effect model results for naturalness rating. The asterisk indicates that the effect is significant at p < .05 (based on the |t| > 2 criterion).

Figure 2 Naturalness ratings across all conditions.

For the analysis of preferred interpretations, we coded whether the intended interpretation was selected for a given context. The results showed the significant main effect of the Contextual bias (p = 0.001). Thus, although both pro and overt pronouns were interpreted to refer to the subject or object NP depending on semantic manipulation, participants accepted the intended referents more readily in the Subject-biased context condition than in the Object-biased context condition (subject: 315 vs. object: 299 out of 320 responses each) (Table 4 and Table 5). This means that there was a stronger bias for subject interpretation than for object interpretation. On the other hand, there was no significant main effect of the Pronoun type, suggesting that sentences were created equally acceptable for Pro and Overt pronoun conditions.

For the naturalness ratings, we initially predicted that the Subject-biased context condition would receive higher naturalness ratings than the Object-biased context condition, as pro and overt pronouns both showed subject bias in Experiment 1. This prediction was confirmed with a significant main effect of the Contextual bias (Table 6). Sentences were rated to be more natural when a pronominal element referred to the subject NP (4.49) than to the object NP (3.89) regardless of the pronominal type.

There was also a main effect of the Pronoun type, with higher naturalness ratings for the Overt pronoun condition (4.3) than for the Pro condition (4.08) (Figure 2). However, this was due to the naturalness rating results in the Object-biased context condition, as indicated by a significant interaction of the Contextual bias and the Pronoun type. We predicted that the Pro condition would receive higher naturalness ratings than the Overt pronoun condition for the Subject-biased context condition but that ratings would be reversed for the Object-biased context condition. However, as shown in Figure 2, our prediction was confirmed only for the Object-biased context condition, not for the Subject-biased context condition. Indeed, post-hoc pairwise comparisons showed the Pro condition received significantly lower naturalness ratings than the Overt pronoun condition only in the Object-biased context condition (p = 0.044). In the Subject-biased context condition, these two conditions did not differ (n.s.).

The results of Experiment 2 confirmed the subject bias of both pro and overt pronouns found in Experiment 1, which is also consistent with Yang et al. (Reference Yang, Gordon, Hendrick and Wu1999, Reference Yang, Gordon, Hendrick, Wu and Chou2001), Li (Reference Li2014), Li et al. (Reference Li, Mak and Sanders2016) and Simpson et al. (Reference Simpson, Wu and Li2016). However, contrary to the argument in Yang et al. that pro and overt pronouns did not differ in their processing mechanisms, the results of Experiment 2 showed different interpretational mechanisms for pro and overt pronouns in Chinese. These results are compatible with the results of Experiment 1, suggesting that pro had a stronger preference for the subject NP antecedent and/or a stronger dispreference for an object NP antecedent than overt pronouns did.

Given these results, Experiment 3 investigated whether interpretational biases of pro and overt pronouns constrained the processing of these pronominal elements during online sentence processing.

4. Experiment 3

Experiments 1 and 2 yielded consistent results that both pro and overt pronouns were subject-biased. However, the results also suggest that overt pronouns had weaker subject bias and were more likely to refer to object NPs than pro was. In Experiment 3, we examined online processing of Chinese pro and overt pronouns using a self-paced reading time method.

The experimental sentences involved either pro or an overt pronoun at the subject position of the main clause (R5) (R is short for Region, thus R followed by a number was used to indicate different regions in Experiment 3), and gender-biased Chinese names in the subject (R1) and the object (R3) position of the preceding subordinate clause, as seen in (18) (e.g. 孙兰 ‘Sun Lan _f’ and 刘军 ‘Liu Jun _m’).

Accordingly, the referent of the overt pronoun could be inferred at the main clause subject position (R5) based on the disambiguating gender information of the overt pronouns. The referent of pro, however, stayed ambiguous until R9, where such information became available based on gender stereotypes. For example, 芭蕾舞女 ‘ballerina’ at R9 in (18) indicates that the pronoun in the subject position is likely to be a female, referring to the subject in (18a) and (18b) and the object in (18c) and (18d).

Thus, the critical regions of our interest were R5 for the Overt pronoun condition and R9, 芭蕾舞女 ‘ballerina’, for all the conditions. We predicted that the reading times of these regions would be modulated as a function of the interpretational bias of pro and overt pronouns. Specifically, given that both pro and overt pronouns were subject-biased in Chinese, it was predicted that reading times of the overt pronoun at R5 and 芭蕾舞女 ‘ballerina’ at R9 in (18a) and (18b) would be shorter than those in (18c) and (18d). In addition, given that the subject bias was stronger or the dispreference of an object antecedent was stronger for pro than for overt pronouns, the reading times of the Pro condition (18a) would be shorter than those of the Overt pronoun condition (18b) for the subject NP-biased conditions at R9. On the other hand, for the object NP-biased condition, the Pro condition (18c) would be read more slowly than the Overt pronoun condition (18d). Thus, it is predicted that there would be a main effect of the antecedent type and an interaction of the antecedent type and the pronoun type at R9.

4.1 Method

4.1.1 Participants

Thirty-six native speakers of Mandarin (eight males; average age = 19 years) attending a university in China were paid to participate in the experiment. They were born and raised in China and none of them participated in the other experiments in this study. No participants reported a native-like fluency in a language other than Chinese.

4.1.2 Materials and design

Thirty-six items similar to the sentences in (19) were constructed. As in Experiments 1 and 2, four lists were created based on the Latin Square design. In addition, 108 fillers of similar length and complexity were constructed and added to the lists. Fillers consisted of two clauses containing those Chinese names and nouns appearing at R9 that were not strongly gender-biased.

4.1.3 Procedure

Experiment 3 was run individually in a quiet room using Linger (Doug Rohde MIT). Participants silently read sentences for comprehension, which were presented phrase by phrase in a non-cumulative moving-window fashion. A yes/no comprehension question was presented after each sentence. Comprehension questions did not probe the resolution of the pronouns, as we did not want participants to develop strategic processing. For example, a sentence like those in (18) above was followed by question equivalent to (19).

There were six practice trials. The experiment lasted approximately 30 minutes.

4.1.4 Data analysis

The mean comprehension accuracy for the experimental items was 91.2%. The reading time analyses reported below included both correctly and incorrectly answered trials. Reading times that were three standard deviations beyond the mean were excluded from further analyses. Overall, 4.73% of data were removed before analysis. Reading time data were then analyzed with Linear Mixed Effect Regression models (Baayen Reference Baayen2008, Baayen et al. Reference Baayen, Davidson and Bates2008, Jaeger Reference Jaeger2008). The experimental conditions (Pronoun type: pro vs. overt pronoun) and Antecedent (subject vs. object), in addition to their interaction, were included as fixed effects. For random effects, intercepts for subject and item as well as by-subject and by-item random slope for both effects were included. The remaining analysis procedures were analogous with those reported in Experiment 1.

4.2 Results and discussion

There was no reliable main effect or interaction in any region in the first subordinate clause (R1 through R4). The reading results of the main clause (R5 through R12) are summarized in Table 7 and Figure 3, and the results of statistical analyses are presented in Table 8 and Table 9.

Table 7 Average reading times (standard deviations) for each region in the main clause, by condition.

Table 8 Linear Mixed Effects results for reading times from R5 to R8.

Table 9 Linear Mixed Effects results for reading times from R9 to R12. The asterisk indicates that the effect is significant at p < .05 (based on the |t| > 2 criterion).

Note: Coefficients, standard errors, and t-values are reported for the main effect of Pronoun type and Antecedent, as well as for the interaction of these two factors.

Figure 3 Mean reading times of the main clause across all conditions. The error bars represent standard error above/below the mean of the participant mean.

Our prediction at R5 was not confirmed. As discussed already, from R5 to R8, we only focused on the Overt pronoun condition, as disambiguating gender information was not yet available for the Pro condition. There were no reliable effects in any of these regions, however, as can be seen from Figure 3 and Table 8. Thus, reading times at R5 did not confirm the subject bias of overt pronouns found in Experiments 1 and 2.

On the other hand, the analyses of reading times at R9 confirmed our predictions. We predicted that the subject NP-biased condition would be read faster than the object NP-biased condition, as both pro and overt pronouns displayed a subject preference. This was confirmed with a main effect of Antecedent at R10 (t = −3.09), at R11 (t = −2.49) and R12 (t = −2.03). In all these regions, the reading times of the subject NP-biased condition were shorter than those of the object NP-biased condition regardless of the pronominal type. Also, as we predicted, there was a significant interaction of the Pronoun type and Antecedent at R9 (t = −2.06) and R10 (t = −2.41). Post-hoc pairwise comparisons at these positions showed that the reading times of the Pro condition were significantly longer than those of the Overt pronoun condition for the object NP-biased conditions (p < .05). For the subject-biased condition, however, the reading times of these two conditions did not differ (n.s.).

Although the effects were slightly delayed, which is common in self-paced reading time experiments (see Chow, Lewis & Phillips Reference Chow, Lewis and Phillips2014), the overall results are compatible with our offline experiments. Both pro and overt pronouns were subject-biased, with pro displaying a stronger dispreference for the object antecedent than overt pronouns.

5. General discussion and conclusion

In this study, we examined the interpretational biases of pro and overt pronouns in Chinese during offline and online sentence processing. Overall, the experimental results suggest that both pro and overt pronouns were subject-biased. In semantically neutral contexts, both pro and overt pronouns were judged as more likely to refer to subject NPs than object NPs, although this tendency was stronger for pro (Experiment 1). The subject bias of pro and overt pronouns was also confirmed in semantically biased contexts (Experiment 2). Despite context manipulation, there was a stronger preference for the subject antecedent interpretation than for the object antecedent interpretation regardless of the pronominal type. In addition, naturalness ratings were higher with the subject interpretation than with the object interpretation regardless of the pronominal type. The results of Experiment 2 also showed that naturalness ratings were lower for sentences with pro than for those with overt pronouns when the referential expressions referred to the object NP, suggesting that dispreference of object antecedents was stronger for pro than for overt pronouns. Finally, the results of Experiment 3 confirmed that these interpretational biases constrained the processing of pro and overt pronouns during online sentence processing. Thus, sentences were easier to process with the subject antecedent interpretation than with the object antecedent interpretation regardless of the pronominal type. In addition, with the object antecedent interpretation, sentences were more difficult to process with pro than with overt pronouns. While these results showed certain similarity between pro and overt pronouns in Chinese, both online and offline results suggest that the mechanisms that underlie the interpretation of null and overt pronouns were not the same. The results thus provided strong evidence against the argument by Yang et al. (Reference Yang, Gordon, Hendrick and Wu1999) that pro does not differ from overt pronouns in Chinese in the way that it contributes to discourse coherence.

Overall, our results converge with previous results of discourse-oriented pro-drop languages (Korean: Kim et al. Reference Kim, Theres and Schafer2013, Kwon & Polinsky, Reference Kwon and Polinsky2011, Japanese: Ueno & Kehler Reference Ueno and Kehler2016, Chinese: Yang et al. Reference Yang, Gordon, Hendrick and Wu1999, Reference Yang, Gordon, Hendrick, Wu and Chou2001; Li Reference Li2014; Li et al. Reference Li, Mak and Sanders2016; Simpson et al. Reference Simpson, Wu and Li2016) in that both pro and overt pronouns were subject-biased. The results are also consistent with Accessibility theory (Ariel Reference Ariel1990, Reference Ariel, Ted, Schilperoord and Spooren2001), which argues for the form-function correlations on the accessibility marking scale such that referring expressions signal a specific degree of mental accessibility. According to the proposal, the highest accessible marker (e.g. pro) is more likely to be used to refer to the most salient entity (e.g. the subject) than a lower accessibility marker (e.g. overt pronoun) is. This was confirmed by a stronger subject preference of pro in our findings. Although overt pronouns in Chinese also showed a subject bias similarly to pro, they were still more likely to refer to an object antecedent than pro was.

Our results, however, diverge from those of other pro-drop languages with rich verbal inflection such as Spanish (Alonso-Ovalle et al. Reference Alonso-Ovalle, Fernández-Solera, Frazier and Clifton2002, Filiaci et al. Reference Filiaci, Sorace and Carreiras2014), Italian (Carminati Reference Carminati2002), Catalan (Mayol & Clark Reference Mayol and Clark2010) and Greek (Papadopoulou et al. Reference Papadopoulou, Peristeri, Plemenou, Marinis and Tsimpli2015), which showed a clear division of labor between pro and overt pronouns. A remaining question is how to account for these cross-linguistic differences in anaphoric resolution. One possibility is that different levels of accessibility are associated with pronominal elements in different languages (Ariel Reference Ariel1990). That is, the relative differences in degree of accessibility between referring forms along the accessibility marking scale may not be exactly the same across different languages. Therefore, even formally-equivalent referring expressions from different languages might refer to antecedents with different degrees of accessibility in each language (Ariel Reference Ariel1990, Filiaci et al. Reference Filiaci, Sorace and Carreiras2014). We argue that the relative difference between pro and overt pronouns in discourse-oriented languages is smaller than in Indo-European pro-drop languages. Thus, pro and overt pronouns share more similarity, both favorably referring to the subject antecedent.

Related to this is the argument that cross-linguistic differences in anaphoric resolution could be due to different levels of sensitivity to different sources of information in different languages (Kwon & Sturt Reference Kwon and Sturt2013). That is, in a discourse-oriented language such as Chinese, Korean, or Japanese (Huang Reference Huang1984), discourse cues could take priority over other cues such as morpho-syntactic information. Indeed, examining the processing of pro in Korean, Kwon & Sturt (Reference Kwon and Sturt2013) showed that pro interpretation was more sensitive to discourse manipulation such that when the referential ambiguity of pro was resolved through a discourse topic NP, the parser did not evaluate additional morpho-syntactic cues when they became available. This was different from English, where referential ambiguity resolution of overt pronouns was equally sensitive to discourse as to morpho-syntactic cues (Liversedge & Van Gompel Reference Liversedge and van Gompeln.d.). These results suggest that referential resolution is guided by different levels of sensitivity to different cues in a given language such that an interpretation of an anaphoric expression based on a cue of greater importance is preferred over one that is based on a cue of less importance. If so, the strong subject bias of pro and overt pronouns of Chinese found in this study (also in studies of Korean and Japanese) could be due to the prominence of topic as a cue in Chinese (Li & Thompson Reference Li and Thompson1981), given that the subject position is likely to be associated with a discourse topic. In fact, this account of the results is also consistent with Accessibility theory, which argues that discourse salience as well as grammatical functions of antecedents affects their mental accessibility. That is, according to Accessibility theory, a (global or local) discourse topic is argued to maintain a higher degree of accessibility and thus is more likely to be referred to with a high accessibility marker (e.g. pro) than a nontopical NP is (Ariel Reference Ariel, Ted, Schilperoord and Spooren2001). Thus, the effect of discourse salience on mental accessibility should be further investigated. For example, it is an empirical question whether pro will show as strong a dispreference to refer to a topical object NP as it does to refer to a non-topical object NP. It is unlikely, given that Chinese is a discourse-oriented language. However, we leave this to a future study.

Various other factors may also contribute to different levels of anaphoric biases. Kaiser & Truswell (Reference Kaiser and Trueswell2008) argued that different forms of anaphoric expression could be subject to different kinds of cues after finding that Finnish pronouns and demonstratives showed different sensitivity to features of antecedents. Likewise, Ueno & Kehler (Reference Ueno and Kehler2016) reported that in Japanese, pro was primarily regulated by grammatical role whereas overt pronouns were more sensitive to verb aspect. In addition, different historical origins of pronouns in different languages might contribute to cross-linguistic variations of anaphoric resolution (Carminati Reference Carminati2002, Filiaci et al. Reference Filiaci, Sorace and Carreiras2014). For example, different historical origins of Spanish pronouns and Italian pronouns may lead to different levels of anaphoric biases. The frequency of referring expressions used in a specific language may also affect the anaphoric biases. In Chinese, overt pronouns are widely used, almost as frequently as pro. (e.g. Tai Reference Tai and Hinds1978, Chen Reference Chen1986, Tomlin & Pu Reference Tomlin and Pu1991, Tao Reference Tao and Fox1996, Christensen Reference Christensen2000). Therefore, the overt pronoun might be the standard, unrestricted form and therefore used to refer to both subject and object, while pro is restricted to refer to the subject only. Likewise, our experimental results align with frequency results reported in previous studies. For example, using the Chinese version of the pear stories, Chen (Reference Chen1986: 129) showed that a preverbal subject/topic NP was more likely to be referred to by pro (50%) than by overt pronouns (30%) or full NPs (20%). Conversely, an object NP was more likely to be referred to by full NPs (45%) than by overt pronouns (35%) or pro (25%). Wang & Liang (Reference Wang and Liang2020) also reported similar results based on BCC Corpus (Beijing Language and Culture University Corpus Center, a corpus of 15 billion Chinese words). They showed that a subject NP was overwhelmingly referred to by pro (n = 231) rather than by overt pronouns (n = 29) or full NPs (n = 4), while a non-topic object was more likely to be referred to by full names (n = 34) than by overt pronouns (11) or pro (n = 0).

Several questions remain unanswered. First, while we mainly discussed the relation between referring expressions and antecedents in terms of mental accessibility, previous studies have shown relevance of other factors such as animacy, thematic roles, etc. (see Christianson & Cho Reference Christianson and Cho2009 for a related discussion of pro in Odawa in the context of Optimality Theory). This calls for systematic investigations of potential interactions among these different factors. Second, our study was restricted to intra-sentential reference and thus it is an empirical question whether the findings would extend to inter-sentential reference. According to the Accessibility theory (Ariel Reference Ariel1990, Reference Ariel, Ted, Schilperoord and Spooren2001), the relation between an antecedent and a referring expression is tighter within a unit (e.g. within a clause, sentence, paragraph) than across boundaries (e.g. across clauses, sentences, paragraphs), so the degree of accessibility of relevant mental representations is correspondingly higher in the former than the latter. Thus, we predict greater overall use of overt pronouns for inter-sentential reference, although the choice of pro vs. overt pronouns will still be affected by discourse salience as well as grammatical functions of antecedents. We leave this question to a future study.

In conclusion, both pro and overt pronouns in Chinese showed a strong subject bias during offline and online sentence processing. They were both more likely to refer to a subject NP than to an object NP. However, the bias was stronger for pro, which showed a stronger dispreference for object NP antecedents than overt pronouns did. Our data are compatible with Accessibility theory in that less informative anaphoric expressions were less likely to refer to less accessible antecedents than more informative anaphoric expressions were. We also argue that anaphoric resolution across languages may be sensitive to cues of different importance in a given language, such as discourse topic.