Fifty Mandarin native speakers participated over the internet. The participants were all above 18 and had normal or corrected-to-normal vision. The studies reported in this paper were approved by the Institutional Review Board (IRB) at the University of Southern California.

Offline experiments were hosted on Qualtrics. On each trial, the topic sentence was presented first, followed by the target sentence on the next screen. This was done to match the format of the self-paced reading task. After reading each sentence, participants answered a comprehension question (e.g. ‘who wrote the academic paper?’) with answer choices consisting of the matrix subject (e.g. ‘Xiaoming’) and the local subject (‘Prof. Wang’ in Non-blocker conditions and ‘I’ in Blocker conditions). See (11) for an example. Thus, participants’ answers indicate how they interpret ziji (e.g. as referring to ‘Xiaoming’, or to ‘Prof. Wang’ or ‘I’). Our set-up with the answer choices follows He & Kaiser (2016). The order of the antecedent choices was counterbalanced.

All three linguistic approaches (see Section 2) predict ziji to show a stronger preference for local binding with subject blockers than with non-blockers, although whether the blocking effect is absolute is an empirical question. Thus, a main effect of Blocker type is expected. A main effect of Verb bias is also predicted such that self-directed verbs will elicit more local choices than other-directed verbs. Crucially, if verb bias semantics modulates the blocking effect, then between the two blocker conditions, we predict higher percentages of local binding with self-directed than with other-directed verbs.

Given the binary nature of the data, we fit mixed-effect logistic models (with glmer) in R (R core team 2018) implemented by the R package lme4. Two contrasts were fit for the factors Blocker type (Blocker coded as +0.5, Non-blocker coded as –0.5) and Verb bias (Other-directed verb coded as +0.5, Self-directed verb coded as –0.5), respectively. T-values were transformed into p-values using Satterthwaite approximation implemented by the lmerTest package (Kuznetsova et al. 2015). For all experiments reported in this paper, models were first fit with random intercepts and random slopes. If the model failed to converge, we simplified the model by following Bates et al. (2015). A parsimonious model is to be preferred over a more complex one if model comparison suggests that the two models do not differ significantly.

The antecedent choice data is displayed in Figure 2. The y-axis shows the mean percentage of local coreference for ziji. For statistics, see Table 1. We found main effects of Blocker type and Verb bias (ps < 0.001): participants preferred local interpretations of ziji more often in the blocker than in the non-blocker conditions, and in the self-directed than in the other-directed verb conditions. The Blocker type main effect suggests that the external POV center ‘I’ competes with the internal POV center (the discourse topic), supporting claims in the literature. The Verb bias main effect indicates that verb semantics plays an important role in the interpretation of ziji: participants chose local antecedents more often in the self-directed than in the other-directed verb condition, even in the presence of a blocker (68% vs. 46%, β = –1.26, SE = 0.22, t = –5.61, p < 0.001). Finally, we found an unpredicted Blocker type by Verb bias interaction (p < 0.001). Pairwise comparisons suggest that the blocking effect reached significance in the other-directed verb conditions (β = 1.65, SE = 0.24, t = 6.96, p < 0.001), but not in the self-directed verb conditions (β = –0.04, SE = 0.22, t = –0.21, p = 0.84). We return to this asymmetry in the Discussion section.

Forty-eight Mandarin native speakers above 18 participated remotely on Ibex Farm (Drummond 2013). None had participated in Experiment 1a.

The topic sentence and the target sentence were presented separately. Participants first read the topic sentence presented as a whole. They then pressed the space bar to read the target sentence word by word. Each key press revealed a new word, while the previous word was masked with a dash. Each sentence was followed by a comprehension question. On target trials, 12 questions out of 20 asked about who ziji refers to.2 The other 8 items asked unambiguous questions about context sentences to ensure that participants paid attention to the context. All filler items were followed by unambiguous questions as well. Participants’ responses to all the unambiguous questions were used to calculate their comprehension accuracy. The order of answer choices was randomized.

The standard and the structure-based processing models make the same predictions: there should be a local binding bias in the blocker conditions and a LD binding bias in the non-blocker conditions. Recall that in online studies we use verb bias to diagnose local vs. LD interpretations. Therefore, we predict an interaction of Blocker type and Verb bias at or after the critical region ziji: In the blocker conditions, if non-local binding is blocked (i.e. local binding is preferred), we predict longer RTs with other-directed verbs which conflict with local readings of ziji. Conversely, in the non-blocker conditions, we predict longer RTs with self-directed verbs as they are incompatible with non-local readings.

Before data analysis, we decided that participants with less than 70% accuracy on comprehension questions would be excluded. Since everyone satisfied this criterion (mean accuracy: 93.75%), none were excluded. RTs shorter than 80ms or longer than 4000ms were removed. So were RTs longer than 2.5 SDs above the mean by region and condition. This resulted in the removal of 3.37% of the original data. We used mixed-effects models in R to analyze the data. Here, we report statistical results based on raw RTs. Analyses based on log-transformed RTs showed similar results (except for one region in Experiment 2c; we note the minor difference in the text).

The mean RTs by condition and by region are plotted in Figure 3. See Table 2 for statistics. The critical region is Region 7, ziji. At the pre-critical regions, there was a main effect of Blocker type at Region 2, the matrix subject pronoun, probably a spurious effect as this word is constant across all conditions and Experiments 2b/2c did not show this effect. We also observed main effects of Blocker type at Region 4, the embedded subject, and Region 5, the adverbial (ps < 0.001): non-blocker conditions were read more slowly than the blocker conditions. This is expected as the third-person NP (e.g. Prof. Wang) has more syllables than the monosyllabic wo (‘I’). Intriguingly, at the embedded verb (Region 6) and the critical region ziji (Region 7), the main effect of Blocker type reversed (ps < 0.001). Now, the blocker conditions were read more slowly than the non-blocker conditions. We attribute this to the cost of perspective shift from the matrix subject to the first-person pronoun. This main effect of Blocker type persists till the end of the sentence (ps < 0.005).

At the critical region ziji, only the Blocker type main effect was significant (p < 0.001). Crucially, at Region 8, the first spillover region, a Blocker type by Verb bias interaction emerged (p = 0.002). Pairwise comparisons by Blocker type indicate that in the blocker conditions, other-directed verbs caused longer RTs (β = 36.41, SE =13.99, t = 2.60, p < 0.01), but in the non-blocker conditions, self-directed verbs lead to longer RTs (β = –24.86, SE = 11.72, t = –2.12, p = 0.035). This suggests that Mandarin natives preferred local binding when a subject blocker ‘I’ intervenes (hence struggle when other-directed verbs push for non-local binding) but preferred non-local binding in the non-blocker conditions (hence struggle when self-directed verbs push for local binding). This interaction persists in the next two regions (ps < 0.001). At the penultimate region, the verb bias effect in the non-blocker conditions did not reach significance (β = –11.72, SE = 9.82, t = –1.19, p = 0.23) although other-directed verbs lead to slowdowns in the blocker conditions (β = 98.86, SE = 24.88, t = 3.97, p < 0.001). At the final region, the contrasting verb bias effects were significant in both blocker conditions (β = 394.47, SE = 70.29, t = 5.61, p < 0.001) and non-blocker conditions (β = –103.34, SE = 43.34, t = –2.38, p = 0.018), matching the pattern at Region 8.

Experiment 1 examined the blocking effect in Mandarin by using both online and offline methods. Experiment 1a aims to verify the blocking effect with the subject blocker ‘I’. Its secondary goal is to see whether verb bias semantics modulates the binding preference even in the presence of a blocker. Experiment 1b probes native speakers’ reading times to shed light on their real-time processing of ziji. In the offline forced-choice Experiment 1a, we operationalized the blocking effect as the local coreference advantage in the blocker conditions relative to the non-blocker conditions. The offline results support the view that the subject blocker ‘I’ makes LD binding more difficult. However, local binding was preferred only 46% of the time when the verb was other-directed. This indicates that verb semantics plays a crucial role in modulating the blocking effect.

Interestingly, Experiment 1a also showed a Blocker type by Verb bias interaction, such that the blocking effect only reached significance in the other-directed verb conditions but not in the self-directed conditions. We view this asymmetric blocking effect as providing additional evidence for the impact of verb semantics: the subject blocker ‘I’ will only be effective when participants have a strong tendency to establish LD dependencies. In the self-directed verb conditions, however, as participants mostly preferred local binding anyway, the blocker was not as effective. Only in the other-directed verb conditions was the subject blocker more effective in inducing blocking. One might notice that even in the self-directed/blocker condition, the mean proportion of local coreference (68%) is not close to ceiling. We attribute this to the presence of the topic sentence: According to the Topic Empathy Hierarchy, we expect the topic sentence to boost the likelihood of interpreting the non-local matrix subject as the antecedent of ziji.3 Note that the lower-than-expected proportion of local coreference choices does not impact our conclusions, which are based on relative differences between conditions.

In the self-paced reading Experiment 1b, we relied on verb bias effects as diagnostics for detecting local vs. LD binding preferences. The results revealed that antecedent retrieval is sensitive to discourse-level cues: the reading time differences between the self-directed and other-directed verb conditions in the non-blocker conditions signal that participants attempted to bind ziji to the discourse topic, the internal POV center. This is predicted by the standard cue-based and structure-based retrieval models alike, but inconsistent with a strong “syntax-first” approach which assumes that non-structural cues are invisible to the language parser at the early stage. We realize that this verb effect in Experiment 1b appeared at spillover regions and could be regarded as a late-stage effect (Sturt 2003). However, a structural effect, reflected as longer RTs with other-directed verbs in non-blocker conditions, is not observed at all (and we definitely see no signs of such an effect preceding the discourse-level effect). In addition, as we will see, in Experiment 2c, the discourse-level effect appears slightly earlier at the critical region in non-blocker conditions (see Badecker & Straub (2002) for evidence that slight variation can exist regarding the onset of predicted effects during self-paced reading). Therefore, we tentatively conclude that prominent discourse-level cues are immediately available for antecedent retrieval during the real-time processing of ziji.

Furthermore, the reading times provide strong evidence that when a local subject blocker ‘I’ is present, participants prefer local binding. This is shown by reading slowdowns at the spillover regions preceded by other-directed verbs (where verb semantics biases non-local binding) compared to self-directed verbs (where verb semantics biases local binding). At this stage, we speculate that local binding is preferred presumably due to two reasons. First, the blocker ‘I’ matches the [+POV] cue encoded on ziji (as we shall see later, the syntactic [+SUB] cue matters as well) and thus interferes with LD binding by a non-local matrix subject (internal POV center) which matches the same set of cues. Second, according to linguistic accounts (e.g. Kuno 1987; Pan 2001; Charnavel et al. 2017), people are more prone to take the perspective of first-person pronouns than third-person NPs. In other words, the blocker is relatively more activated than the matrix subject,4 hence the preference for local binding.

Overall, Experiment 1a indicates that the blocking effect is influenced by both blocker type and verb semantics. Thus, the “blocking effect” may be better viewed as a defeasible constraint. Furthermore, results from Experiment 1b suggest that online antecedent retrieval is sensitive to discourse-level cues, and that subject blockers bias participants towards local binding of ziji. However, Experiment 1b does not inform us whether it is the discourse-level cue [+POV] alone that induces blocking, because the subject blocker ‘I’ also encodes [+SUB]. It could be that the [+POV] cue is visible only because it is in a syntactically prominent subject position. What will happen if the blocker is in an object position? We turn to this question in Experiment 2.

Forty-eight Mandarin native speakers above 18 participated in the experiment. None had participated in any previous experiment.

The procedure was the same as Experiment 1a.

The syntax-based account predicts that an object blocker will not induce the blocking effect. In contrast, the perspective-based and the unified accounts predict a main effect of Blocker type. However, the latter two differ regarding the strength of blocking. The perspective-based account does not predict any difference between subject and object blocking, while the unified account predicts stronger blocking with the subject blocker ‘I’ than with the object blocker ‘me’. We also predict a main effect of verb bias.

Data analysis for this experiment was the same as in Experiment 1a.

Figure 4 displays participants’ mean preference for local binding of ziji across conditions. See Table 3 for statistical analysis. We found main effects of Blocker type and Verb bias (ps < 0.001). The blocker conditions lead to higher percentages of local binding regardless of verb bias. The overall low tendency for local binding even in the self-directed/blocker condition echoes Experiment 1a, where we argue that the context sentence boosts non-local interpretations (thus suppressing local interpretations) to some extent. Note that the symmetrical pattern in Figure 4 differs from the asymmetric pattern in Experiment 1a, a point we will return to in the discussion. The main effect of Verb bias indicates that verb semantics plays a crucial role in determining the binding of ziji (p < 0.001). As in Experiment 1a, a verb bias effect emerged between the two blocker conditions as well: other-directed verbs lead to a decreased tendency for local binding compared to self-directed verbs (31% vs. 62%, β = –1.53, SE = 0.21, t = –7.35, p < 0.001). This strengthens our view that verb semantics modulates the blocking effect. The Blocker type by Verb bias interaction was not significant. Overall, Experiment 2a shows that the object blocker ‘me’ induces a blocking effect, which is unexpected under a syntactic account.

Next, we compare the strengths of subject blocking (Experiment 1a) and object blocking (Experiment 2a). To this end, we included Blocker position (subject (Exp1a) vs. object (Exp 2a)) as a between-experiment factor (contrast coding: Object = +0.5, Subject = –0.5). See Table 4 for statistics. We found a three-way interaction of Blocker type, Verb bias, and Blocker position (p < 0.001). To better understand the interaction, we examined the blocking effect by blocker positions within each verb bias type. A significant interaction of Blocker type and Blocker position would suggest that the strength of blocking effect varies as a function of the blocker position. Statistical analysis indicated that with self-directed verbs, the blocking effect of the syntactically less prominent ‘me’ seems to be stronger than ‘I’ due to a marginally significant interaction (β = 0.56, SE = 0.29, t = 1.94, p = 0.05). This seems to contradict the unified account which predicts the opposite. But as we mention in the discussion below, this is presumably related to the matrix verb difference (e.g. ‘heard’ vs. ‘told’) in Experiments 1 and 2. If we follow Huang and Liu (2001) by assuming that LD ziji shows logophoric properties, this result falls out naturally. Crucially, statistical analysis also showed the Blocker type by Blocker position interaction to be significant in the other-directed verb conditions (β = –0.99, SE = 0.32, t = –3.11, p = 0.002). This suggests that the syntactically prominent subject blocker induced a stronger blocking effect than the object blocker when the verb semantics biases non-local antecedents, matching the prediction of the unified account but not that of the perspective-based account.

Seventy-six Mandarin native speakers above 18 participated in this experiment. None had not participated in any previous experiment.

The procedure was the same as in Experiment 1b.

As we previously discovered that the blocking effect has a syntactic component (i.e. subject blocking stronger effect than object blocking), the standard cue-based and the structure-based retrieval models make different predictions regarding native speakers’ online processing patterns (see Figure 1 in Section 3). The standard model predicts that LD binding should be preferred in the non-blocker conditions but less preferred or even unavailable in the blocker conditions. Therefore, a significant Blocker type by Verb bias interaction at the critical region ziji is expected. In contrast, the structure-based model predicts that object blockers in syntactically illegitimate positions will not interfere with LD binding of ziji. Therefore, the structure-based model only predicts a main effect of Verb bias (i.e. longer RTs with self-directed verbs) at ziji.

As in Experiment 1b, we removed participants whose comprehension accuracy fell below the 70% threshold. Eight participants failed to meet this criterion. The remaining 68 participants had a mean accuracy rate of 90.91%. We then filtered out RTs 2.5 SDs above the mean by region and condition, which resulted in the removal of 2.45% of the original data. We used mixed-effect regression to analyze the RTs.

The mean RTs by condition and region are plotted in Figure 5. See Table 5 for statistics. The critical region is Region 8, ziji. Among all the regions, only the pre-critical Region 7, the biased verb, showed a Blocker type by Verb bias interaction (p < 0.001). The early timing of this effect was not predicted although it is not surprising given similar findings from self-paced reading (Chen et al. 2012) and eye-tracking (Jäger et al. 2015; Chang et al. 2020), presumably because participants anticipated the onset of ziji at the biased verb region due to the predictability of the stimuli after repeated trials.

However, a closer look at this interaction shows that it is not the one predicted by the standard cue-based approach. In the blocker conditions, self-directed verbs caused RT slowdowns (β = 47.51, SE = 15.04, t = 3.16, p = 0.002), while in the non-blocker conditions other-directed verbs that lead to slowdowns (β = –22.39, SE = 10.77, t = –2.08, p = 0.038). This pattern suggests that participants preferred non-local binding when the (potential) object blocker ‘me’ was present (hence struggled when self-directed verbs pushed for local binding) and preferred local binding when there was no first-person blocker (hence struggled when other-directed verbs pushed for non-local binding).

This shows that the (potential) blocker ‘me’ actually did not induce blocking during real-time processing (as we have evidence of participants considering non-local readings in the blocker conditions). This is consistent with the predictions of the structure-based account, but not the standard cue-based retrieval account.

Second, the mysterious “blocking effect” in the non-blocker conditions is unexpected, but it’s worth noting that this effect is not significant in the log-transformed-RT analysis (β = –0.03, SE = 0.02, t = –1.28, p = 0.20), so we should not read too much into it. We suggest an explanation in the Discussion below, which will motivate a modified Experiment 2c.

Experiment 2a and 2b test the blocking effect with the object blocker ‘me’. The syntax-based account predicts that a blocker in object position cannot block LD binding of ziji by a discourse topic. In contrast, the perspective-based and unified accounts acknowledging the logophoric properties of ziji (i.e. sensitive to POV centers) allow for blocking effects with object blockers. The forced-choice judgment results from Experiment 2a support the latter accounts. However, by comparing the offline results in Experiments 1a and 2a, we found that the blocking effect is stronger with subject blockers. We take this as evidence that both perspective-taking and syntactic prominence are involved in blocking, in line with the unified account.5

There are two other discoveries worthy of notice. First, the strength of the blocking effect depends on verb semantics. Experiment 2a showed – echoing Experiment 1a – that if the verb is other-directed, even with an object blocker, the mean preference for local antecedents was only 31%.

Second, the blocking effect in Experiment 2a is symmetrical, unlike the asymmetric pattern in Experiment 1a. In Experiment 1a, we suggested that if the verb is self-directed, the blocker is largely ineffective in blocking LD binding as participants mostly only considered local binding. In Experiment 2a, the matrix verb was changed to a source verb ‘told’ instead of a mixture of perceiver and source verbs (e.g. ‘heard’, ‘said’, ‘noticed’, etc.). Since ziji was interpreted logophorically given a discourse topic, a Source conceivably attracted some of the antecedent choices in Experiment 2a even when the verb was self-directed. This increased tendency for non-local binding was neutralized in the blocker conditions but not in the non-blocker conditions, hence the symmetrical blocking pattern in Experiment 2a. In fact, previous experimental work has corroborated this verb-related logophoricity effect in the case of English reflexive himself/herself (Kaiser et al. 2009; Sloggett 2017). Although we are not aware of any related work in Mandarin, it is reasonable to expect this effect to also occur with perspective-sensitive ziji given its logophoric properties (e.g. Huang & Liu, 2001; Huang et al., 2009).

The self-paced reading Experiment 2b tests whether the discourse-level cue [+POV] alone suffices to influence real-time antecedent retrieval in a context where the [+POV] antecedent lacks the [+SUB] feature, as it is an object. The standard retrieval model predicts different verb bias effects for the object blocker and non-blocker conditions, while the structure-based model predicts similar verb bias effects. Intriguingly, the reading time patterns do differ between the blocker and non-blocker conditions, but not in a way predicted by the standard model. We found that native speakers preferred non-local binding in the blocker conditions and local binding in the non-blocker conditions. Nevertheless, the absence of blocking by a (potential) object blocker during real-time processing is consistent with the prediction of the structure-based retrieval model, but not the standard cue-based retrieval model. Note that the absence of the blocking effect in online parsing contrasts with the weak blocking effect in the offline data from Experiment 2a. This difference suggests that a pure discourse-level blocking effect may emerge later on during processing. (The structure-based model does not speak directly to this, see Section 4).

However, neither model predicts the tendency for local binding with non-blockers. This may be related to working memory burden due to similarity-based interference and dependency length. The accessibility of NPs is known to depend on whether there is interference from other similar NPs (Gordon et al. 2001; 2004; 2006) and referential complexity (Warren & Gibson 2002). In the non-blocker conditions, there are three third-person NPs (e.g. ‘he’, ‘others’, ‘Prof. Wang’) whereas in the blocker conditions, there are only two. Thus, the presence of three third-person NPs in working memory potentially makes sentence processing challenging. This challenge might be compounded by the long distance between the discourse topic and the reflexive (7 words compared to 6 in Experiment 1b). Therefore, in Experiment 2c, we would like to replicate Experiment 2b with a slightly simpler design by removing the adverbial region.

Sixty-two adult Mandarin native speakers participated. None had participated in any previous experiment.

The procedure for Experiment 2c was the same as in Experiment 2b.

If the removal of the adverbial alleviates participants’ working memory burden, we predict a main effect of Verb bias across (non-)blocker conditions such that self-directed verbs lead to reading slowdowns at either the pre-critical verb region or the critical region ziji.

No participant was removed as all participants passed the 70% threshold with a mean accuracy of 92.51%. After the removal of RTs 2.5 SD above the mean by region and condition, 97.27% of the original data were entered into statistical analysis.

The mean RTs by condition and region are in Figure 6. See Table 6 for statistics. Prior to the critical region ziji, there were no significant effects. At the critical region, a main effect of Verb bias emerged (p < 0.001): self-directed verbs lead to longer RTs than other-directed verbs, consistent with the prediction of the structure-based retrieval model. For the spillover regions, Region 10 showed a Verb bias main effect (p < 0.001) and a Blocker type by Verb bias interaction (ps < 0.005). This late interaction should not be taken as supporting evidence for the standard cue-based retrieval model, so we will not discuss it in detail. It appeared because self-directed verbs lead to significant reading slowdowns in the non-blocker conditions (β = –202.95, SE = 37.97, t = –5.35, p < 0.001), but no verb bias effect was found within the blocker conditions (β = –35.18, SE = 46.71, t = –0.75, p = 0.45). Overall, Experiment 2c replicates Experiment 2b on the absence of object blocking.

Experiment 2c improved upon Experiment 2b by eliminating the adverbial region to reduce participants’ potential working memory burden. This move proved to be effective: we now find that in the non-blocker conditions, participants preferred non-local binding very early at the critical region ziji, as expected. This suggests that the unexpected (and statistically unstable) result for verb bias effects in the non-blocker conditions in Experiment 2b is indeed likely to be due to memory load or some other factor unrelated to blocking. Crucially, for the blocker conditions, Experiment 2c replicates the reading time patterns from in Experiment 2b, in stark contrast with the reading time results in Experiment 1b with subject blockers. In summary, together with Experiment 1b and 2b, Experiment 2c provides evidence supporting the unified approach and the structure-based retrieval model.