Implicit Causality and Pronoun Resolution in Intersubjective Discourse Relations

Lyu, Siqi; Wang, Luming

doi:10.3389/fpsyg.2022.866103

ORIGINAL RESEARCH article

Front. Psychol., 10 May 2022

Sec. Psychology of Language

Volume 13 - 2022 | https://doi.org/10.3389/fpsyg.2022.866103

Implicit Causality and Pronoun Resolution in Intersubjective Discourse Relations

$\r\nSiqi Lyu$ Siqi Lyu¹ $Luming Wang*\r\n$ Luming Wang^2*

¹Institute of Psychology, University of Tartu, Tartu, Estonia
²College of Foreign Languages, Zhejiang University of Technology, Hangzhou, China

Interpersonal verbs like disappoint and praise in Lucy disappointed/praised Mary because she… bias the potential cause of the event to one of the antecedent noun phrases (henceforth NPs) (e.g., Lucy for disappoint whereas Mary for praise). Using Chinese as its materials, this study investigated how verb-based implicit causality affects online pronoun resolution in backward concession (e.g., Lucy disappointed/praised Mary although she…), an intersubjective discourse relation where the subordinate although-clause forms an indirect relationship with the preceding main clause. Experiment 1 was a baseline experiment with the typical structure where implicit causality is found to be effective, i.e., backward causality. Results showed a clear modulation effect of implicit causality on pronoun resolution such that as verb bias strength decreases, participants were faster in processing sentences that disambiguate the pronoun to the verb-inconsistent NP. However, this modulation effect was not observed in Experiment 2 where we used the same verbs but replaced because with although. There was no preference for the pronoun to be disambiguated toward the verb-consistent NP or the verb-inconsistent NP in backward concession. The results of Experiments 1 and 2 were replicated in Experiment 3 where we directly compared causal and concessive relations. We suggest that the absent effect of verb-based implicit causality in backward concession could be attributed to the intersubjective nature of the concessive relation. Discourse devices such as although indicate speakers’ subjective perspective and comprehenders are able to quickly accommodate the speaker’s point of view during online discourse processing.

Introduction

Successful communication relies on the correct interpretation of pronouns like it, he/she, and they in context. Comprehenders make use of various sources of information in discourse to retrieve the correct referent of pronouns (Kehler, 2002). Certain interpersonal verbs have statistically reliable preferences about their causal antecedents (“Implicit Causality”; Garvey and Caramazza, 1974). For example, interpersonal verbs like disappoint in (1a) and praise in (1b) have an implicit feature of attributing the potential cause of the event to the first noun phrase (NP1) [e.g., Lucy in (1a)] or the second noun phrase (NP2) [e.g., Mary in (1b)], respectively. Whereas it is likely that Mary was disappointed because of something Lucy did, there is a preference that the reason why Lucy praised Mary is because of something Mary has done.

(1) (a) Lucy_i disappointed Mary because she_i… (NP1-biased verb).

(b) Lucy praised Mary_i because she_i… (NP2-biased verb).

This implicit feature of interpersonal verbs has been observed in other discourse relations than causality. For instance, an opposite pattern is observed in sentences that describe the potential consequence of an event with a subordinate so-clause (‘Implicit Consequentiality’; see Au, 1986; Stewart et al., 1998; Cheng and Almor, 2017; Garnham et al., 2021; among others). Whereas the pronoun in Lucy disappointed Mary so she… is most likely to refer to Mary, that in Lucy praised Mary so she… tends to refer to Lucy. In another study, Xu et al. (2019) examined pronoun resolution in different discourse relations and found a distinctive feature of concessive relation such that compared with other relations like so-clause, and-clause, or two independent clauses (without any connectives in between), although-clause was processed with greater uncertainty.

Concession has been considered as a negated causal relation (König, 1991; König and Siemund, 2000; Louwerse, 2001) but differs from causality in that it involves the speaker’s subjective attitude (Verhagen, 2005; Lyu et al., 2020). The aim of the present study is to investigate how verb-based implicit causality affects pronoun resolution in backward concession, an intersubjective discourse relation where the subordinate concessive clause (henceforth C2) forms an indirect relationship with the preceding main clause (henceforth C1). Below we first introduce subjectivity and intersubjectivity on theoretical accounts, and then move to previous findings on the influence of verb-based implicit causality on online pronoun resolution, followed by a review of previous experiments on the processing of concession in discourse and an overview of experiments in the current study.

Subjectivity, Intersubjectivity, and Concession: Theoretical Accounts

A coherent discourse consists of various types of relations between events. For instance, (2a) describes a causal relation where the causal event she worked hard (p) and its consequence she passed the oral English test (q) form a typical cause-consequence relationship p → q in the real world. On the other hand, what is described in (2b) concerns the speaker’s reasoning such that the observation that she passed the oral English test (p) is an argument for the claim that she worked hard (q). Relations like (2a) are considered as an objective relation and those like (2b) subjective, respectively (Sanders and Spooren, 2015; Kleijn et al., 2021; see also “content domain” and “epistemic domain”, respectively, by Sweetser, 1990).

(2) (a) She passed the oral English test today, because she worked hard. (content domain)

(b) She worked hard, because she passed the oral English test today. (epistemic domain)

Concession has always been considered as a negated causal relation (König, 1991; König and Siemund, 2000; Louwerse, 2001). Whereas the two events work hard (p) and pass the oral English test (q) in (2a) form a causal relation p → q as indicated by the connective because, the causal link between events is rejected in a concessive relation like (3a), where the concessive connective although leads to a negated consequential event fail the oral English test (q). Similarly, whereas (2b) describes an epistemic causal relation between passed the oral English test (p) and work hard (q), a concessive relation like (3b) claims a negated conclusion, i.e., she didn’t work hard (q).

(3) (a) She failed the oral English test today, although she worked hard. (content domain)

(b) She didn’t work hard, although she passed the English test today. (epistemic domain)

Concession additionally features an intersubjective coordination between the speaker and the addressee. A wide range of linguistic phenomena involve “connecting, differentiating, and ‘tailoring’ the contents of points of view with respect to each other (rather than organizing a connection to the world)” (Verhagen, 2005, p. 4), of which concession is one. On this account, the concessive connective in q although p evokes two mental spaces (Fauconnier, 1994) with different epistemic stances, one in which the speaker acknowledges a potentially valid inference p → q, and the other in which the speaker displays a negative epistemic stance toward the proposition p herself. In saying a concessive sentence like (3a), for instance, while the speaker acknowledges that given p (i.e., she worked hard) there may be good reasons to adopt q (i.e., she passed the oral English test), she nevertheless invites the addressee to adopt a contradictory result, i.e., she failed, which is incompatible with q. Even though a concessive relation like (3a) describes events in the content domain that is usually considered objective, it comes with an inherent intersubjective feature that the speaker invites her addressee to accept a contradictory conclusion.

In some cases, although doesn’t express a real-world relationship between two states of affairs in the way because does (Iten, 2005, p. 163). Take (4) as an example, where the two clauses are relatively independent from each other and do not necessarily form a (negated) cause-consequence relationship. The two clauses connected by although in (4) are not based on the real-world knowledge that if one lost a flight ticket (p) then she will not fail the oral English test (q), which is not a plausible logic in the real world. Rather, (4) is likely to be interpreted as a plausible sentence such that the first clause is a claim made by the speaker and the second is a retreating statement to it. In such cases, the two events are not directly related. The speaker adds some unrelated information that she wants the addressee to accept despite the fact that she acknowledges in the first clause, and the addressee is very likely to tolerate the sentence and accept whatever the speaker provides.

(4) She failed the oral English test today, although she lost a flight ticket last weekend.

A backward causal relation, however, expresses a “tighter” cause-consequence relationship between two clauses. In (5), for instance, the causal connective because cues readers of a real-world cause for the event described in the first clause. In this case, comprehenders will try to build up a cause-consequence relationship if one lost a flight ticket (p) then she will likely pass the oral English test (q) when reading (5). However, given that such a cause-consequence relationship is illogical in the real world, the sentence is likely to be considered implausible.

(5) *She passed the oral English test today, because she lost a flight ticket last weekend.

As Verhagen (2005, p. 170) puts it, backward concession can “easily produce an effect of retreating and weakening the original claim.” Different from a backward causal relation, backward concession is by default an intersubjective relation. The two clauses of backward concession form an indirect relationship where the concessive clause provides the speaker’s retreating statement to the proposition claimed in the preceding main clause.

Implicit Causality and Pronoun Resolution

When describing the relationship between two entities in structures like NP1 V NP2 because…, the interpersonal verb biases the potential cause of the event toward one of the antecedent NPs (Garvey and Caramazza, 1974). Take (1) as an example, repeated here as (6). There tends to be an overall preference for comprehenders to consider Lucy as the underlying cause of the event of disappointing in (6a). By contrast, in (6b), readers will tend to consider Mary as the implicit cause of the event of praising. When verb’s bias is defined as the proportion of completions in which the first noun is the antecedent, bias values range continuously from 1.0 (NP1 is the unanimous antecedent) to 0.0 (NP2 is the unanimous antecedent) (Caramazza et al., 1977). The verb implicit causality can thus be presented as a continuum, to the one end standing the NP1-biased verbs like disappoint in (6a), and to the other end the NP2-biased verbs like praise in (6b).

(6) (a) Lucy_i disappointed Mary because she_i… (NP1-biased verb).

(b) Lucy praised Mary_i because she_i… (NP2-biased verb).

Verb bias is not an absolute constraint on the interpretation of the pronoun (Koornneef and van Berkum, 2006; Hartshorne and Snedeker, 2013). For example, both sentences in (7) are plausible and coherent despite being inconsistent with the verbs’ biases. That is, the referential preference based on the verb bias can be overridden by the disambiguating information in C2 without rendering the sentence ungrammatical or incoherent.

(7) (a) Lucy disappointed Mary_i because she_i had a high expectation.

(b) Lucy_i praised Mary because she_i was very satisfied.

Studies have consistently shown that comprehenders make use of the information of verb-based implicit causality to resolve the pronoun during real-time processing (e.g., Caramazza et al., 1977; Vonk, 1985; Guerry et al., 2006). In an experiment where participants were asked to read sentences like (8) and indicate their choice for pronoun assignment by saying the appropriate person’s name out loud, Caramazza et al. (1977) found that compared with (8a) in which C2 disambiguated the pronoun to a direction consistent with the verb in C1, participants needed more time to read the entire sentence like (8b) where the verb in C1 and the information in C2 disambiguated the pronoun toward different directions. The results suggest that verbs’ implicit feature of specifying the causal relationship between the participants of an action has an effect on comprehenders’ real-time interpretation of the sentence.

(8) (a) Tom scolded Bill_i because he_i was annoying. (verb-consistent condition)

(b) Tom_i scolded Bill because he_i was annoyed. (verb-inconsistent condition)

Most of recent psycholinguistic research have focused on when verb-based implicit causality is applied during online processing of pronouns, among which inconclusive results have been drawn with regard to two competing accounts: integration account vs. focusing account. On the one hand, a number of studies found that the information of verb-based implicit causality is only used during sentence-final clausal integration (Garnham et al., 1996; Stewart et al., 2000; Garnham, 2001). For instance, in a series of probe recognition tasks, Garnham et al. (1996) showed that the effect of implicit causality emerges only after the presentation of disambiguating information, i.e., during the integration of the two clauses. On the other hand, more recent studies have adopted the visual-world paradigm and event-related potentials (ERP) and showed that implicit causality was used before the disambiguating information was available (Koornneef and van Berkum, 2006; van Berkum et al., 2007; Cozijn et al., 2011; Järvikivi et al., 2017) or even before participants had encountered the causal connective (Pyykkönen and Järvikivi, 2010). On this account, a sentence fragment like Lucy praised Mary because… immediately brings Mary into focus at the expense of Lucy, and comprehenders prefer to relate a personal pronoun to the most focused antecedent.

Despite the large amount of research on the time course of the implicit causality effect, few studies have zoomed into the subtle difference in the strength of verb bias and examine how it interplays with the disambiguating information in C2 in affecting real-time pronoun resolution. As we explained above, implicit causality is a continuum, the two ends of which are the NP1-biased verbs and NP2-biased verbs, respectively (Caramazza et al., 1977). In an offline experiment, Ferstl et al. (2011) systematically measured the implicit causality of 305 English verbs using a sentence completion task and provided the bias score of each verb that ranges from 100 (all completions NP1) to –100 (all completions NP2). As illustrated in (9a) and (9b), attract and abandon are both NP1-biased verbs but differ in terms of their bias strength. Thus it can be seen that not every NP1-biased verb affects the pronoun resolution to the same extent, and so does the NP2-biased verb.

(9) (a) Lucy_i attracted Mary because she_i… (bias score 87).

(b) Lucy_i abandoned Mary because she_i… (bias score 33).

While both the sentence fragments in (9) are likely to be completed with Lucy, as shown in (10), (9a) still has a higher probability than (9b) due to the stronger bias strength of attract than abandon. From a processing perspective, sentences with strong-bias verb like attract in (10a) will likely be easier to process than those with weak-bias verbs like abandon in (10b).

(10) (a) Lucy_i attracted Mary because she_i was beautiful.

(b) Lucy_i abandoned Mary because she_i was annoyed.

It is thus worth taking into account the referential probability of each verb when examining the implicit causality effect. In light of Ferstl et al. (2011), in the present study we will index the strength of the verb bias with the referential probability obtained in a sentence completion task (see section “PRE-TEST”). We will treat the verb-based implicit causality as a continuum and include verbs’ bias scores as a continuous factor in our statistical model, so that a more precise picture of how the pronoun is resolved as the strength of the verb bias changes can be provided.

Processing Concession in Discourse

Concession has always been studied together with causality and found to be harder to comprehend than causal relations. For example, studies have shown that concessive sentences are processed more slowly than causal sentences, and the recall is worse for concessive sentences than for causal sentences (Townsend, 1983; Caron et al., 1988; Köhne and Demberg, 2012, 2013; Xu et al., 2018; Lyu et al., 2020).

Recently, in a self-paced reading experiment with Chinese materials, Lyu et al. (2020, Experiment 2) found a distinct feature of concession such that although certain “seemingly” implausible words like kaiche ‘driving’ in (11) induced processing difficulty at the initial stage [i.e., at the critical region, underlined in (11)], it later became as acceptable as the plausible words like yingyu ‘English’ [i.e., at the post-critical region, e.g., douzhidao ‘everybody knows it’ in (11)]. The authors suggest that concession is an (inter)subjective relation that is restricted to the speaker’s cognitive domain and that during online processing, participants can quickly shift their perspective to accommodate the speaker’s point of view and accept the “implausible” content at a later stage.

(11) suiran Ahui yuyantianfu henqiang

although Ahui language talent very good

danshi buxihuan xue yingyu/kaiche

but not like learn English/driving

dajia douzhidao

everybody all know

‘Although Ahui has a talent for language, he doesn’t like learning English/driving. Everybody knows it.’

There are other studies that have demonstrated an indirect relationship between the two clauses in concession. For instance, Townsend and Bever (1982) asked participants to listen to sentence fragments that contain singular morphological bias [e.g., the boxer wants in (12)] and then read aloud the interrupting target word (IS or ARE) on the screen as soon as possible. They found no difference in terms of participants’ responses to the singular and the plural target following a concessive fragment led by though. The results suggest that the interpretation of incomplete phrases in the later clause of concession is unaffected by the content of an initial concessive clause.

(12) Though the boxer wants to avoid unnecessary injuries, dodging punches… IS/ARE

In another experiment, Morera et al. (2017) presented concessive sentence like (13) in pages (page breaks indicated by “||”). They asked participants to first listen to a fragment concessive clause, after which participants were presented an emotional icon (emoticon) that either matched or did not match the emotional valence of the concessive clause, and then they had to choose the correct continuation of the sentence [indicated in brackets in (13)]. The authors found that the emoticon has no effect on the choice of the correct continuation of a concessive relation, which corroborates Townsend and Bever’s (1978) findings in suggesting that initial concessive clauses open up the possibility for “anything” to follow and are not effective in providing a context with which the subsequent main clause can be integrated.

(13) Although the pupil studied a lot, —— ☺/☹ —— [he passed/he failed] the exam.

In a recent ERP study with backward concession, Xu et al. (2019) showed that resolving the pronoun in a concessive sentence with although, as in (14), elicits larger sustained negativity (time-locked to she) than sentences containing an explicit consequential connective so, a coordinator and, or sentences with a full stop between the two clauses. The larger sustained negativity in concession was interpreted as greater uncertainty as to which antecedent would be retrieved as the referent of the pronoun. That is, by the time participants reached she in a concessive relation like (14), they were not able to make use of the implicit causality information in the previous context to successfully refer the pronoun to the correct antecedent.

(14) Lily_i disappointed Nina, although she_i quit the business. (verb-consistent)

While Xu et al.’s (2019) results showed that the real-time comprehension of the pronoun itself (i.e., by the time participants reached the pronoun) in backward concession was less dependent on the implicit causality information in C1, it remains to be answered whether implicit causality will start to play a role after the disambiguating information in C2 has unfolded. If implicit causality in concession plays a role at a later stage (e.g., the integration account; Garnham et al., 1996), then the time taken to integrate C2 after it has been completely unfolded should be longer in verb-inconsistent condition than in verb-consistent condition. However, if C1 and C2 in backward concession are relatively independent from each other (i.e., the implicit causality information carried by the verb in C1 has limited effects on pronoun resolution in C2), then there should be no difference in comprehending the verb-consistent and verb-inconsistent conditions even after the completion of the whole sentence. This is the main hypothesis to be tested in the present study.

Overview of Experiments

The present study aimed at examining how verb-based implicit causality is applied in the online pronoun resolution in an intersubjective discourse relation called backward concession. Experiment 1 tested the validity of the implicit causality effect in Chinese using a structure of [_C1 NP1 V NP2] [_C2 because Pro…], a typical discourse relation where implicit causality is found to be effective. As a baseline experiment, Experiment 1 only adopted causal relations in the content domain. In a 2 × 2 design, we used both NP1-biased verbs and NP2-biased verbs and manipulated the disambiguating information in C2 as either consistent or inconsistent with the verb bias. Experiment 2 investigated how verb-based implicit causality affects online pronoun resolution in backward concession. We adopted the same design as Experiment 1 but replaced because with although. In Experiment 3, we crossed discourse relation (causality vs. concession) and consistency of the disambiguating information in C2 (verb-consistent vs. verb-inconsistent) in a within-subject design. By doing so, we are able to directly compare causality and concession and provide further evidence for how implicit causality affects online pronoun resolution in the two closely related discourse relations.

Pre-Test

In order to obtain the prototypical NP1-biased and NP2-biased verbs in Chinese, a pre-test was conducted in which participants were asked which person Ta “he, she” refers to in sentence fragments like (15). The fragments were created with 150 Chinese implicit causality verbs translated from the English implicit causality verb corpus (Ferstl et al., 2011). All verbs were two-character lexical causatives. The verb was placed between two Chinese nicknames a-X ‘‘dear X’’ or xiao-X ‘‘little X.’’ Each sentence fragment was followed by three potential answers, including nickname 1, nickname 2, and ‘‘others,’’ presented visually using the Tencent online questionnaire application¹. Thirty-nine participants (13 men, 26 women, mean age 22.56 years, range 18--35 years) participated in the pre-test and were instructed to choose the best answer that first came to their mind. The referential probability of each verb was calculated, based on which 32 NP1-biased verbs and 32 NP2-biased verbs were selected for the processing experiments². The two types of verbs were matched in terms of their referential probabilities (p > 0.2)³. The mean referential probability as well as the probability range of the two types of verbs are given in Table 1⁴. See Supplementary Materials for a full list of verbs and their referential probabilities.

(15) A-huang zhenjingle A-hui, yinwei Ta…

A-huang_[nickname1] surprised A-hui_[nickname2], because Ta…

TABLE 1

Table 1. Mean referential probability of NP1-biased verbs and NP2-biased verbs.

Thirty-two pairs of two-character Chinese names were created and to be used in the experiment as NP1 and NP2, respectively. In order to exclude the potential gender effect brought by the names, in the real experiments we presented critical sentences like (15) in the auditory form, where the Chinese third-person singular pronoun Ta cannot be differentiated in pronunciation (both pronounced as /tā/). Given that in the real experiments, the comprehension questions following each trial with NP1 and NP2 as its potential answers were to be presented visually, we matched NP1 and NP2 in terms of their homophone frequency and number of stokes. The summed homophone frequency and total number of strokes of the two characters in each name were calculated, and separate t-tests showed no significant differences between NP1 and NP2 (ps > 0.60).

Experiment 1

Experiment 1 examined the modulation effect of implicit causality information on online pronoun resolution with the typical structure where implicit causality has been found to be effective, i.e., backward causality. We included both NP1-biased verbs and NP2-biased verbs and manipulated the disambiguating information in C2 as either verb-consistent or verb-inconsistent, resulting in a 2 × 2 design.

Methods

Participants

Seventy-five (29 men, 46 women, mean age 19.69 years, range 18–24 years) native speakers of Mandarin Chinese who did not participate in the pre-tests were paid to participate in Experiment 1. Informed consent was obtained from all participants. All participants had normal or correct-to-normal vision and normal hearing.

Materials

A total of 32 sets of sentences were constructed with the verbs and names selected from the pre-test. All sentences consisted of two clauses, i.e., C1 and C2. C1 always began with a temporal adverb, followed by the structure NP1 V NP2, where the verb was either biased to NP1 or NP2. C2 was the causal clause led by yinwei “because,” followed by an ambiguous Ta and a six-character predicate that helped resolve the co-reference of Ta. The information in C2 disambiguated the pronoun to either the verb-consistent NP or the verb-inconsistent NP. See Table 2 for an example set of stimuli. All materials are available in the Supplementary Materials.

TABLE 2

Table 2. Example stimuli in Experiments 1 and 3.

A number of factors were controlled for the six-character predicates in C2. First, we extracted the log-frequencies of each word/segment in C2 from the SUBTLEX-CH corpus (Cai and Brysbaert, 2010). The mean log-frequency of all segments in a predicate were calculated and treated as the log-frequency of that predicate. Second, the syntactic complexity of C2 and the semantic role of Ta in C2 were manually coded by a trained Chinese linguist. In terms of syntactic complexity, predicates with more than two components [e.g., noun phrase (NP), adjective phrase (AP), verb phrase (VP)] were treated as complex structures and coded as “1,” and those with two or less components were treated as simple structures and coded as “0.” As for semantic roles, predicates where Ta was agent were coded as “1” and those where Ta was non-agent were coded as “0.” The descriptive data of these factors are provided in Table 3, with linear/logistic regression results showing no significant differences across conditions in terms of each variable (ps > 0.07).

TABLE 3

Table 3. Parameters of predicates in C2.

In order to examine whether the disambiguating predicates we constructed can successfully resolve the pronoun, we conducted a norming test by asking 12 native speakers (3 men, 9 women, mean age 20.42 years, range 18–23 years) to read the sentences that were to be used in the real experiment and judge which person Ta refers to. None of this group of participants participated in any other experiments reported in this study. We divided the materials into four Latin square lists and each participant was assigned to one of the four lists. The accuracy was above 92% for all conditions with logistic regression results showing no significant differences among them (ps > 0.2), suggesting that the pronouns in the constructed materials were able to be resolved as expected⁵.

The 32 sets of items, together with the materials of our other study that had a similar design (not reported here), were divided into eight counter-balanced lists for the online processing experiment. Each list consisted of four sets (altogether 16 trials) of critical items of the current study. In addition to the 16 critical trials of the current study and the 16 trials from our other study, we added 36 two-clause fillers with various syntactic structures to each list, resulting in a full list of 68 sentences for each participant. Each critical sentence was followed by a comprehension question that was identical across conditions, i.e., “Who do you think Ta in the last sentence refers to?”, with NP1 and NP2 as its potential answers. All fillers also came with a comprehension question with two potential answers, one of which was the correct answer. Part of the filler questions were the same as critical items and the others were targeted at other contents in the sentence.

All sentences were recorded by a woman native speaker at the frequency of 44,100 Hz in a quiet room using Praat (Boersma, 2001). The two clauses in a sentence were recorded as one audio file so as to maintain the natural intonation of the sentence. There was a natural stop between two clauses for all experimental items and fillers, which varied between 190 and 620 ms. 95% of the experimental items had an interval between 300 and 500 ms.

Procedure

Participants were instructed to sit in a quiet room and randomly assigned to one of the eight lists. Before each trial, there was a fixation “+” on the screen. Participants pressed the SPACE bar on the keyboard to proceed to the next page, on which they looked at a blank screen while listening to the audio of the whole sentence over headphones. After the audio stopped, a comprehension question appeared visually on the screen with its two potential answers. The position of the correct answer was balanced across all trials. Participants were instructed to use F or J on the keyboard to make the choice as fast and as accurately as possible. Participants’ decision and their response time (i.e., the time from the onset of the comprehension question till participants’ press of the key) were automatically recorded by the program. All trials were presented pseudo-randomly using E-Prime 3 (Psychology Software Tools, Pittsburgh, PA, United States). Participants went through four practice trials before the real experiment so as to get familiar with the experimental procedure. The whole experiment took about 15 min to complete.

Data Analysis

Only the trials where participants successfully reached the expected referent of the pronoun were analyzed. Linear mixed effects models were fit to the response time data using lme4 package version 1.1–27.1 (Bates et al., 2015) in R version 4.1 (R Core Team, 2021). The lmerTest package version 3.1–3 (Kuznetsova et al., 2017) in R was used for demonstrating the p-value. The fixed effects were consistency and verb, both of which were manually sum coded (verb-consistent as –0.5 and verb-inconsistent as +0.5; NP1-biased verb as –0.5 and NP2-biased verb as +0.5). To more precisely observe how verb bias strength modulates online pronoun resolution, we included the referential probability of the bias-congruent referent for each verb as a covariate called strength, which was scaled and centered for NP1-biased verbs and NP2-biased verbs respectively using scale() function in R. The response times were log-transformed to stabilize variance and achieve approximately normal residuals (Box and Cox, 1964). As for random effects, since each set of item contains a particular verb that comes with a specific bias strength, which has already been included as a covariate in our model, only subject was included as a random effect. Each model was initially built with maximum random intercepts and random slopes for subjects, and the random slope was eliminated stepwise if the model failed to converge. Post hoc tests were conducted following significant interactions with Bonferroni correction.

Results

The overall comprehension accuracy was 93.69% (SD = 0.24) for all experimental trials and fillers, suggesting that participants generally performed well in the task. Experimental trials to which participants answered incorrectly were excluded for analysis (accounting for 14.3% of the total experimental trials). For the remaining trials, response times were replaced by NAs for values more than two standard deviations above or below the mean (accounting for 3% of the remaining observations). The mean response times of each condition are shown in Figure 1, and statistical results are given in Table 4.

FIGURE 1

Figure 1. Mean log-transformed response times in each condition of Experiment 1. Error bars represent one standard error.

TABLE 4

Table 4. Model output of Experiment 1.

We found a main effect of verb showing that NP2-biased sentences were processed with significantly less time than NP1-biased sentences, which replicated the results of previous studies on other language than Chinese (e.g., Dery and Bittner, 2016). In addition, we found a main effect of consistency showing that verb-inconsistent sentence took longer time to process than verb-consistent sentence and a consistency*strength interaction. Post hoc tests with Bonferroni correction showed that whereas verb-inconsistent sentences became significantly easier to process as the strength of verb bias decreases (β = 0.06, t = 2.72, adjusted p < 0.05), verb bias strength had no effect on the processing time of verb-consistent sentences (adjusted p > 0.1), which we suggest was due to a ceiling effect. That is, participants were always very fast to respond to verb-consistent sentences such that their response time to verb-consistent sentences could not be significantly faster as the verb bias becomes stronger. The effect of verb bias strength has reached a top limit in affecting/facilitating participants’ processing of the verb-consistent sentences.

Discussion

In this experiment we adopted the structure [_C1 NP1 V NP2] [_C2 because Pro…] and included both NP1-biased verbs and NP2-biased verbs and manipulated the disambiguating information in C2 as either verb-consistent or verb-inconsistent. Results showed that verb-inconsistent sentences need more time to process than verb-consistent sentences, which replicated the results of previous studies that verb bias direction plays an important role in online pronoun resolution (e.g., Caramazza et al., 1977; Vonk, 1985; Guerry et al., 2006). Importantly, with the inclusion of the verb bias strength in our model, we found an interaction between the consistency of the disambiguating information in C2 and the verb bias strength in C1 such that whereas verb-consistent sentences remained easy to process regardless of the strength of verb bias, which we suggest was due to a ceiling effect, verb-inconsistent sentences became significantly easier to process as verb bias strength decreases. This experiment provided a more precise picture that implicit causality is a continuum and that the subtle difference in the strength of verb bias has an effect on online pronoun resolution. Using Chinese as its materials, Experiment 1 provided evidence that implicit causality is valid across languages and cultures (Hartshorne et al., 2013).

Experiment 2

Experiment 2 aimed at investigating whether and how the verb-based implicit causality in C1 of backward concession affects pronoun resolution in the subsequent C2. Different from backward causality that contains a real-world cause-consequence relationship shared in both the speaker’s and the hearer’s common knowledge, backward concession is an intersubjective relation in which the two clauses are in an indirect relationship. We therefore expect the implicit causality to be less effective on pronoun resolution in backward concession.