Masaya Hosoda- Department of English, Seijo University, Setagaya, Japan
The existing psycholinguistic research suggests that verbs' implicit causality (IC) elicits two types of bias: a coreference bias, which favors re-mentioning the causally implicated entity of the event (she = Mary in Mary annoyed Lisa because she. ..), and a coherence bias, which leads speakers to expect an explanation in the upcoming discourse (Mary annoyed Lisa is continued with Mary sang loudly). Of these two biases, previous second-language (L2) studies have predominantly focused on coreference bias in contexts where an upcoming explanation is explicitly signaled (Mary annoyed Lisa because...). The present study advances the L2 literature by examining both coherence and coreference biases in L2 comprehension. Eye-tracking and story-continuation experiments revealed that L2 learners are fundamentally weaker than native speakers in terms of coherence bias. As a result, an upcoming explanation must be explicitly signaled for IC to trigger coreference bias during online L2 processing. The findings suggest that while the underlying mechanism of IC bias functions similarly in both L1 and L2 comprehension, there is a pronounced L1–L2 difference in the ease with which an implicit explanation relation can be activated through expectation-based processing. The findings are discussed in terms of the source and time course of IC, as well as theoretical accounts of L2 prediction.
1 Introduction
Discourse comprehension involves complex cognitive processes that use explicit and implicit cues. Verbs' implicit causality (IC) bias is a well-studied implicit cue in psycholinguistic research. For example, in the sentence fragment (1a), the causal connective because explicitly signals a coherence relation between clauses, with the dependent clause explaining the event in the matrix clause, which is referred to as the explanation relation. As the cause of the event denoted by the verb annoyed is imputed to the subject position (NP1 [first noun phrase]) entity (e.g., Mary was an annoying person, which was why she annoyed Lisa), she would typically be interpreted as Mary. Conversely, in fragment (1b), with punished as the verb, she is usually interpreted as Lisa because, in this case, the event cause is imputed to the object position (NP2 [second noun phrase]) entity (e.g., Lisa did something wrong, which was why Mary punished her).
(1) a. Mary annoyed Lisa because she…
b. Mary punished Lisa because she…
Previous studies on first-language (L1) comprehension have demonstrated that this coreference bias of IC impacts moment-to-moment (i.e., online) comprehension processes, as shown in self-paced reading and eye-tracking tasks (e.g., Koornneef and Vanberkum, 2006; Pyykkönen and Järvikivi, 2010). Notably, IC coreference bias has been observed in various languages (English, German, Spanish, Russian, Japanese, Mandarin, and Korean; Bott and Solstad, 2014; Hartshorne et al., 2013), suggesting that IC plays a language-universal role in guiding coreference processing.
It has been demonstrated that IC induces another discourse-level bias in addition to the coreference bias. Kehler et al. (2008) discovered that speakers provided an explanation continuation (She sang loudly) when prompted to continue sentences with an IC verb (Mary annoyed Bob) in 60% of the instances. By contrast, 25% of the continuations from sentences with a non-IC verb (Mary saw Bob) were explanations. Thus, IC verbs create a stronger-than-normal expectation for an explanation in the subsequent discourse, a phenomenon known as (explanatory) coherence bias. The observation of coherence bias indicates that IC is used not only for integrative processing but also for expectation-based processing regarding the upcoming discourse.
Furthermore, IC has increasingly become the focus of second-language (L2) research (Cheng and Almor, 2017; Contemori and Dussias, 2019; Hijikata, 2021; Hosoda, 2023; Kim and Grüter, 2021). Studies have demonstrated that L2 learners use IC bias for online coreference processing during comprehension. Meanwhile, coreference bias may be reduced or absent in L2 learners with limited exposure to IC verbs in L2 or when a disparity occurs in verb usage between L1 and L2 (Cheng and Almor, 2017; Hosoda, 2020). This study focused on IC processing by Japanese learners of English. Although Japanese learners are highly sensitive to the IC coreference bias in Japanese, they often show reduced sensitivity to NP1 coreference bias in English (Hijikata, 2021; Hosoda, 2020). This may be due to differences in how English and Japanese NP1-IC verbs encode causation at the morphemic level. Specifically, NP1 verbs explicitly mark causation with overt morphemes in Japanese (e.g., Mary ha Bob wo iradata-seru), whereas causation is integrated into the lexical semantics of the verb without any explicit marker in English (e.g., Mary annoyed Bob). Examining Japanese learners' sensitivity to IC bias in English may hence elucidate the cross-linguistic influence of L1 on discourse processes.
Despite these findings, previous L2 studies have predominantly examined coreference bias in contexts with an explicit explanation relation signaled by a causal connective (e.g., Mary annoyed Bob because). Considering that IC verbs invoke coherence bias, which leads speakers to expect an explanation in the upcoming discourse, investigating the coreference and coherence biases of IC without an explicit explanation provides a theoretically sound approach to exploring discourse expectations in L2 comprehension.
Various L2 theoretical models suggest differences in expectation-based processing between native speakers and L2 learners (two-stage model, Corps et al., 2023; prediction-byproduction model, Amos and Pickering, 2020; and interface hypothesis, Sorace, 2011). In particular, the RAGE hypothesis posits that the differences in L1 and L2 comprehension emerge directly from L2 learners' reduced ability to generate expectations regarding the upcoming discourse (Grüter et al., 2017; Grüter and Rohde, 2021). Cognitive factors that affect language processing in general (e.g., working memory capacity, attentional resources, task-induced processing, and the quality of L2 lexical representations) may be the cause of the reduced expectation in L2 learners (see Kaan, 2014; Schlenter, 2023 for reviews). Accordingly, L2 learners might struggle more than native speakers in expecting an explanation from IC verbs, potentially showing weaker IC effects when the explanation is not signaled. Given this context, this study aimed to advance our understanding of IC processing in L2 learners by addressing the following unresolved questions:
1. Does IC invoke coreference bias during online L2 processing without the explicit explanation relation?
2. Is the coherence bias of IC, previously observed in native speakers, present in L2 learners?
1.1 Source of the implicit causality bias
Previous IC studies have provided numerous explanations for the mechanisms underlying the IC coreference bias. The major question is whether IC bias is inherent in the linguistic properties of verbs or stems from speakers' general world knowledge. The verb-based account argues that IC bias is determined by the verb's semantics and the thematic roles of its arguments; the stimulus (Mary in Mary annoyed Bob) is more likely to be associated with the event cause than the experiencer (Bob; Crinean and Garnham, 2006). The world-knowledge-based account argues that IC bias emerges from speakers' inferences regarding the typical causes and effects of the event (Hartshorne et al., 2015). For example, one could infer that Mary annoyed Bob is caused by Mary based on the typical interpersonal relationship associated with such situations (Mary sang loudly, which annoyed Bob).
Although these accounts describe some aspects of the effects of IC, they explain only the coreference bias that occurs in a single sentence. To address this, Bott and Solstad (2014) integrated the verb-based and world-knowledge accounts to develop the empty-slot theory (see also Bott and Solstad, 2021; Solstad and Bott, 2022). What distinguishes this theory from the others is that it provides a unified account of the coreference and coherence biases that occur beyond the sentence boundary. Specifically, it posits that the IC bias is fundamentally attributable to the lexical semantics of IC verbs (as in verb-based accounts). Crucially, IC verb semantics triggers the expectation that an explanation will follow in the subsequent discourse because of its explanatory underspecified content, called a slot. In Mary annoyed Bob, the semantics of the verb annoyed has a slot (e.g., Mary did something annoying, which explains Bob's annoyance with her). Critically, this slot is unfilled because the proper name Mary provides no specification. Because speakers follow a general processing strategy to avoid underspecification, this slot causes them to expect the upcoming discourse to provide an explanation (coherence bias). As Mary is the main target of this explanation, coreference bias arises, favoring re-mentioning this referent in the produced explanation.
This relationship between the IC verb semantics, explanatory expectation, and coreference bias leads to the following prediction: If speakers expect an explanation from the IC verb and maintain it during comprehension, the expected explanation allows IC to cause coreference bias, even when the explanation is not explicitly signaled. Recall that in the empty-slot theory, coreference bias results from coherence bias toward explanatory expectations. This means that the manifestation of coreference bias in the absence of an explicit explanation requires speakers to expect an explanation. From this perspective, the observation of coreference bias in contexts that lack an explicit explanation serves as a theoretically supported indicator of discourse-level expectations.
1.2 Time course of implicit causality
Another persistent debate in IC literature concerns the time course of IC—when and how IC affects online comprehension processes (Koornneef et al., 2016; Koornneef and Sanders, 2013; Koornneef and Vanberkum, 2006; Pyykkönen and Järvikivi, 2010). A well-known manifestation of online IC effects is the pronoun-inconsistency effect, whereby comprehension, operationalized by reading time or eye fixation, is delayed by a pronoun that contradicts the coreference bias of IC. For example, in (2), reading times slow down upon encountering the pronoun he because it refers to the NP2 entity (Bob), which is inconsistent with the NP1 bias of the verb's IC.
(2) Mary annoyed Bob because he…
Two opposing accounts for this effect have been proposed. The immediate-focusing account posits that IC information immediately brings one of the verb's arguments into focus at the expense of the other (McKoon et al., 1993). Hence, this account predicts that an inconsistency effect emerges immediately after encountering a critical pronoun. Oppositely, the clausal-integration account posits that IC is used retroactively at the end of the sentence, where interpretations of the two clauses are integrated (Stewart et al., 2000). Accordingly, inconsistency effects are assumed to manifest in the final region of the sentence. Early studies on the time course of IC employed methods such as the probe task or word-by-word self-paced reading. Some studies reported inconsistency effects in the middle of the sentence, a finding consistent with the immediate-focusing account. By contrast, other studies observed these effects at the end of the sentence, thereby supporting the clausal-integration account.
More recently, IC studies have primarily used eye-tracking, a method that captures a wide range of comprehension processes by monitoring various eye movements. Notably, these studies consistently observed the IC effects immediately after or at the introduction of the bias-inconsistent pronoun (Koornneef and Sanders, 2013; Koornneef and Vanberkum, 2006, Experiment 2). The early effects of IC are further supported by research using the visual-world paradigm, which showed that native speakers rapidly fixate on IC-biased referents (Pyykkönen and Järvikivi, 2010).1
Most relevant to the current purpose, Koornneef et al. (2016) reported early IC effects in contexts that lack an explicit explanation. In their eye-tracking study, stories were presented without a causal connective (e.g., David apologized to Linda. She, according to the witness, was the one to blame. [translated from Dutch]). L1-Dutch speakers' reading times were delayed five words after a pronoun that was inconsistent with the IC coreference bias (the underlined was in the above example). This finding indicates that native speakers expect an explanation from IC verb semantics during comprehension, which allows IC to bias online coreferential processing (see also Koornneef and Sanders, 2013 for the offline evidence of IC effects without an explicit explanation).
To explain the early IC effects, Koornneef and Sanders (2013) and Koornneef et al. (2016) developed the incremental integration account. This model synthesizes components from the immediate-focusing and clausal-integration accounts and recognizes their applicability to different phases of IC processing. Specifically, it suggests that the IC first brings the causally implicated entity into focus (consistent with the immediate-focusing account). Crucially, speakers are assumed to incrementally integrate pronouns with the focused entity on a word-by-word basis. Because the pronoun encounter triggers the use of IC, the IC effects are predicted to emerge immediately after the pronoun, which is consistent with the mid-sentence IC effects frequently observed in the literature.
1.3 Implicit causality in L2 comprehension
Recently, IC coreference bias has become a focal point of investigation in L2 studies (Hijikata, 2021; Hosoda, 2020, 2023; Kim and Grüter, 2021; Wang and Gabriele, 2022). These studies show that IC rapidly influences L2 processes; however, IC effects are often smaller or delayed in L2 comprehension than in L1 comprehension. According to Wang and Gabriele's (2022) self-paced reading experiment, L1-Chinese learners of English showed native-like pronoun-inconsistency effects at or immediately after the critical pronoun, indicating that L2 learners use IC for online coreference processing. However, the story-continuation experiment revealed that learners produced significantly more references inconsistent with the IC bias than native speakers, indicating that learners' sensitivity to the IC coreference bias was weaker than that of native speakers. Kim and Grüter (2021) found online inconsistency effects in intermediate to advanced L1-Korean learners of English using the visual-world paradigm. Specifically, IC affected learners' eye movements 1,000 to 1,500 ms after pronoun offset (approximately two to three words after the pronoun; Nathan disturbed Owen all the time because he needed help with his homework). However, these effects were more limited in timing and size than those observed for native English speakers, which persisted from because to 1,500 ms after the pronoun offset (Nathan disturbed Owen all the time because he needed help with his homework). These findings suggest that, although IC biases online coreferential processing in L1 and L2 comprehension, its effects are slower and weaker in L2 comprehension.
To the best of my knowledge, the only study that has reported comparable IC effects between L1 and L2 speakers is that of Contemori and Dussias (2019). Their study showed that highly proficient L1-Spanish and L2-English bilinguals, immersed in L2 from an early age, exhibited native-like IC effects in online self-paced reading and offline story-continuation tasks. Thus, very high L2 proficiency, combined with early L2 exposure, might be necessary for IC processing comparable to that of native speakers.
Notably, as stated in the Introduction, these L2 studies used materials that included an explicit explanation signaled by a causal connective (e.g., because). This methodological feature may have contributed to the observed IC effects by making the explanation relation readily available for learners' online processing. The present study directly addressed this issue by comparing contexts with and without an explicit explanation.
The reduced coreference bias among L2 learners can be observed in specific bias directions (NP1 or NP2). For example, Cheng and Almor's (2017, 2018) story-continuation experiments revealed that L1-Chinese learners of English showed a weaker NP2 bias than native English speakers, thereby producing more NP1 references after NP2 verbs. Conversely, the NP1 bias was attenuated in L1-Japanese learners of English. Hosoda (2020) reported that, although Japanese learners generally produced bias-consistent story continuations, the effects of NP1 bias were consistently weaker when the task was conducted in L2-English than L1-Japanese. In a self-paced reading study, Hosoda (2023) found that NP2 bias immediately influenced online coreferential processing, whereas the effects of NP1 bias did not emerge until the sentence-final region (see Hijikata, 2021 for another evidence of the difficulty experienced by Japanese learners with NP1 bias).
A possible explanation for these observations is related to cross-linguistic differences. The weaker NP2 bias among L1-Chinese learners can be attributed to the fact that Chinese has fewer NP2 verbs than English (Cheng and Almor, 2017). The greater prevalence of NP1 verbs in Chinese may cross-linguistically lead L1-Chinese learners to favor NP1 references.
For the weaker NP1 bias among L1-Japanese learners, a possible explanation is the morphemic differences between English and Japanese verbs. As mentioned in the Introduction, Japanese NP1 verbs have overt causative morphemes (e.g., -seru or -saseru) that explicitly mark causation (e.g., konran-saseru [confuse] and shitsubou-saseru [disappoint]). Conversely, English NP1 verbs do not have overt morphemes with causation being implicitly conveyed through the verb semantics. This difference negatively influences the application of the lexical knowledge of NP1 verbs to discourse processes (Hosoda, 2020, 2023). Specifically, even when L1-Japanese learners knew the meaning of English NP1 verbs (as confirmed by the translation task showing that English verbs were correctly translated into their L1 counterparts), they exhibited reduced NP1 bias in discourse processes, such as referential resolution and the prediction of upcoming referents (Hijikata, 2021; Hosoda, 2022). This observation differs from that of NP2 verbs, for which causation is not explicitly marked morphemically in either Japanese or English (e.g., konomu [like] and bassuru [punish]). Consequently, Japanese learners can use NP2 bias in English to a similar extent as native English speakers and as they do in their own L1 (Hosoda, 2020, 2022). Based on these existing findings, this study predicted that L1-Japanese learners of English would show reduced effects of coherence and coreference biases for NP1 verbs compared with native speakers.
1.4 Focus of the present study
Existing L2 studies have focused solely on coreference bias in contexts with an explicitly signaled explanation (e.g., Mary annoyed Bob because he...). According to the empty-slot theory, the manifestation of coreference bias without an explicit explanation requires speakers to expect an explanation from IC verbs (Bott and Solstad, 2014, 2021; Solstad and Bott, 2022). Considering L2 accounts positing that L2 learners are more limited in their ability to generate expectations than native speakers (Grüter and Rohde, 2021; Schlenter, 2023), they may struggle to make an explanatory expectation from IC verbs (coherence bias). In such a scenario, IC will not invoke coreference bias during L2 processing, specifically when the explanation is not explicitly signaled. This study investigated this possibility.
2 Experiment 1
Experiment 1 addressed the following research question (RQ): Is the occurrence of early IC effects during L2 processes, reported in existing L2 literature, extended to a context in which the explanation relation is not explicitly signaled?
To address this RQ, a full-stop condition was set, whereby the stimuli were presented without a causal connective (Mary annoyed Bob. He...). The experiment used eye-tracking to analyze the time course of IC.
Notably, this study used eye-tracking in reading rather than listening (as in the visual-world paradigm). This is because the listening mode is used much less frequently than reading with the L2 learner population tested in this study, primarily because these learners rarely use L2 in everyday communication. This lower familiarity with listening would have confounded the results. It should also be noted that Experiment 1 did not include native speakers in the control group. It is extremely difficult to recruit a sufficient sample of native English speakers for an in-person eye-tracking experiment who match the L2 participants in age, socioeconomic status, and educational backgrounds in environments where English is rarely used daily. This is a limitation of this study that will be revisited in the General Discussion.
2.1 Method
2.1.1 Participants
Forty-two L1-Japanese university students participated in the study (24 females; Mage = 20.02; age range: 18–24). They had been learning English in Japan for six or more years. None of them had studied abroad in an English-speaking country. All lived in environments where English is used as a foreign language, meaning that it is not used for communication in daily life. All participants had normal or corrected-to-normal visual acuity. Their L2 proficiency was estimated to be at the CEFR A2–B1 levels or 28–80 on the TOEFL iBT test (Papageorgiou et al., 2015), based on their scores on the TOEIC IP test (M = 474.62, SD = 127.37). This was conducted as a placement test for the university English course 1–2 months before the experiment. Data from two participants were excluded owing to major losses in their eye-tracking data.
2.1.2 Materials
2.1.2.1 Implicit causality verbs
This study used 24 IC verbs (12 NP1 and 12 NP2 verbs), originally from the 300 English verbs identified by Ferstl et al. (2011) in a norming study on IC bias in English verbs.2 The coreference bias of these verbs was piloted with separate Japanese university students (N = 20), who completed the story-continuation task in Japanese and English. The bias was determined using the rate of NP1 references (number of NP1 references/[number of NP1 + NP2 references]). Verbs eliciting 70% or more NP1 references in English and Japanese were categorized as NP1 verbs, whereas those eliciting 30% or fewer NP1 references in both languages were categorized as NP2 verbs. Accordingly, the experimental IC verbs were confirmed to cause the coreference bias of similar strength in the same referential directions in Japanese and English. This suggests that the bias was similar in these languages; hence, cross-linguistic differences in the bias should not interfere with learners' use of IC in L2 English. Furthermore, verbs eliciting fewer than 65% references to NP1 or NP2 referents in Japanese and English were categorized as non-IC verbs, which were used in Experiment 2.3
All verbs were within Rank 3,000 or below on the New Japan Association of College English Teachers (JACET) List of 8,000 Basic Words (JACET Basic Word Revision Committee, 2016), which lists the 8,000 English words that Japanese students learn between elementary school and university based on frequency. English words of this rank have been introduced in junior high or high schools in Japan. Therefore, the IC verbs used in this study were assumed to be familiar to the participants.
Notably, this study did not differentiate between the types of IC verbs (action, psychological, and state verbs) for three main reasons. First, the inclusion of a diverse set of IC verbs allowed the use of verbs whose strength and direction of bias were shared in English and Japanese. Second, this study aimed to maximize the number of items to ensure adequate statistical power. Third, this study's primary interest was to examine IC effects in a heterogeneous set of verbs, rather than effects attributable to specific verb types.
2.1.2.2 Story stimuli
The IC verbs were used in three-sentence story stimuli constructed following L1 studies on the time course of IC (Koornneef et al., 2016). All stimuli comprised English vocabulary and grammatical structures introduced in junior high schools in Japan, thereby ensuring that the participants did not struggle excessively with lexical or grammatical processing.
Table 1 shows an example of the stimulus. The first sentence sketched a story and introduced two characters (one female and one male) with their names.4 The second sentence continued the story while mentioning the characters with they to keep them similarly salient.
Table 1. Example NP1 stimuli.
The third sentence was the target sentence in the format NP1 IC verb-ed NP2 because he... The presence of the causal connective was manipulated by replacing because with a full stop.
The target sentence established consistent and inconsistent conditions based on the consistency of the pronoun he with the coreference bias of IC. Pronoun consistency was manipulated by switching characters in the NP1 and NP2 positions instead of changing he to she to avoid potential confounding due to word differences. At least five words after the pronoun were held in common across the conditions to accommodate spillover IC effects. After this common region, the consistent and inconsistent stories diverged, resulting in information that made the overall story coherent.
Each stimulus was accompanied by a comprehension question to encourage the participants to read carefully. The question targeted explicit information in the story that was irrelevant to the interpretation of the pronoun. Half of the questions were correctly answered with “Yes” and the rest with “No.” Four sets were constructed, crossing 2 (pronoun consistency: consistent, inconsistent) × 2 (connective: because, full stop) conditions. The assignment of stimuli to the conditions was counterbalanced across the sets.
2.1.2.3 Translation task
A translation task was conducted to confirm whether the participants had semantic knowledge of IC verbs. This task was necessary because IC effects would not emerge unless the participants knew the meaning of the IC verbs. The participants were presented with a matrix clause of the 24 target sentences (e.g., “Mary annoyed Bob.”) and asked to translate it into Japanese.
2.1.3 Apparatus and procedure
The eye movements were recorded using the Tobii Pro Spectrum (300 Hz). The experiment was performed using the Tobii Pro Lab. A chin rest was used to minimize head movement. The participants were tested individually in a silent room. The experimenter explained the procedure to the participants and obtained their written informed consent. Participants sat approximately 60 cm from the 23.8-inch monitor and received written and oral instructions. Subsequently, they read two practice stories (Courier New 20-point font with double spacing) presented in their entirety, each accompanied by a comprehension question.
The eye-tracker was calibrated using a standard nine-point grid. When the error in any gaze position exceeded 1.0°, the eye-tracker was recalibrated until the average error was smaller than 0.5°. The participants read the experimental stimuli, and their eye movements were recorded. The order in which the stimuli were presented was randomized for each participant. After each stimulus, they answered a comprehension question by fixating on the “Yes” or “No” mark displayed on the monitor. Each session consisted of three blocks. The participants had a 5-min break between blocks, and the eye-tracker was recalibrated. After all stimuli were presented, the translation task was administered.
2.1.4 Coding and data treatment
2.1.4.1 Translation task
Two raters coded the responses as “correct” or “incorrect” by matching them with the intended meaning of the IC verb. In particular, the raters carefully checked whether the NP1 verbs were translated with a transitive causative meaning because the differences between Japanese and English are evident in this usage. Other meanings or usages (e.g., passive interpretations) were coded as “incorrect.” Conversely, the raters were flexible regarding wording variations as long as the translations captured the gist of the intended meaning (e.g., for disappointed, shitsubou-saseta and gakkari-saseta were coded as “correct”). If any ambiguity was noted, the response was coded as “incorrect.” The inter-rater agreement rate was 96%. All disagreements were resolved through discussion.
2.1.4.2 Eye-tracking measures
This study conducted a region-by-region, rather than word-by-word, analysis. This is because function words are often skipped and fixated on 35% of the time, whereas content words receive fixations 85% of the time (Carpenter and Just, 1983).
Three regions were set, each comprising two words (“Steve hurt Hanako because/he had/always by/nature been/an aggressive person.”). Region 1 included the critical pronoun and one word after it (“.../he had/always by/nature been/...”). Region 2 consisted of two and three words after the pronoun (“.../he had/always by/nature been/...”). Region 3 included four and five words after the pronoun (“.../he had/always by/nature been/...”). Four eye-tracking measures were computed to capture the IC effects in the initial and late processes. First-fixation and first-gaze durations were computed as measures of the initial processes (e.g., initial access to the word's semantic information). First-fixation duration was the duration of the first fixation on a region. First-gaze duration was the sum of the fixation durations on one region before the participant moved forward or returned to another region.
Regression-path and right-bound durations were computed to measure late processes (e.g., processing difficulty associated with the integration of a target word into the previous context). The regression-path duration summed all fixation durations when the participant fixated on a region until they moved on to the following region. This means that the measure included all fixation durations of looking back at the prior region of the text after the target region had been fixated on.5 The right-bound duration was the sum of all fixation durations on a region before the participant left the region in the forward direction. Right-bound duration differed from regression-path duration as it did not include fixation durations while looking back at prior regions.
The data from these measures were converted into per-syllable measures to account for differences in word length. Subsequently, the data were log-transformed to correct for right skewness. Among the trials, 1% were excluded because of tracker losses or eye blinks. The data from incorrectly translated items (16%) were also excluded. Fixation durations that were more than 2 SD from the mean in any experimental condition were treated as missing data (< 10% for all measures).
2.1.5 Statistical analysis
Linear mixed-effects models were used to analyze eye-tracking measures. The fixed effects were bias direction (NP1 and NP2), pronoun consistency (consistent and inconsistent), connective (because and full stop), all sum-coded, and Direction × Consistency × Connective interaction. For random effects, the best-fitting and most parsimonious structures were selected through a backward model comparison. Specifically, a random slope was successively removed from the maximal model (with random intercepts of participants and items as well as the by-participant random slopes of consistency, bias direction, connective, and Direction × Consistency × Connective interaction), unless its removal significantly decreased the model fit. This was done by referring to the AIC (Akaike Information Criterion) and p values of the model comparisons using the anova function. The selected models had random intercepts of participants and items and a by-item random slope of consistency (formula: eye-tracking measure ~ direction * consistency * connective + [1 | participant] + [1 + consistency | item]).
2.2 Results
2.2.1 Translation task
The participants produced more erroneous translations for NP1 verbs (error rates = 22%) than for NP2 verbs (error rates = 10%; β = 1.92, SE = 0.92, z = 2.07, p = 0.039), indicating that they had less semantic knowledge of NP1 than NP2 verbs. In particular, in 65% of the erroneous trials, the participants interpreted NP1 verbs in a passive sense (e.g., Mary ha Bob ni shitsubou-saserareta [“Mary was disappointed by Bob”] for Mary disappointed Bob), rather than in a transitive causative sense. The implications of this finding are presented in the General Discussion.
2.2.2 Eye-tracking measures
The scores for the comprehension questions were generally high (M = 90.44, SD = 12.98), confirming that the participants carefully read the stimuli (no participant showed a correct percentage of 70% or less). Table 2 presents the descriptive statistics of the eye-tracking measures in the because and full-stop conditions. Figure 1 illustrates the mean duration of the eye-tracking measures.
Table 2. Eye-tracking measures (in ms per syllable) as a function of IC-Bias direction, pronoun consistency, and region.
Figure 1. Mean durations of eye-tracking measures (± SEM bars).
This section focuses on the results concerning the pronoun-inconsistency effects that are relevant to the RQs.6 None of the eye-tracking measures showed significant effects associated with pronoun consistency in Region 1 (p > 0.050). However, significant Direction × Consistency × Connective interactions emerged for the regression-path duration in Regions 2 (β = 0.04, SE = 0.02, t = 2.07, p = 0.039) and 3 (β = 0.04, SE = 0.02, t = 2.07, p =0.039). Follow-up tests revealed that the regression-path duration was delayed by the bias-inconsistent pronoun when the NP2 verb was followed by because (Region 2: β = −0.25, SE = 0.11, t = −2.33, p = 0.026; Region 3: β = −0.26, SE = 0.11, t = −2.36, p = 0.025). None of the eye-tracking measures were affected by pronoun consistency when because was absent or the verb had an NP1 bias (p > 0.050).7
2.3 Discussion
The analyses revealed that the connective condition and IC bias directions interactively determined the occurrence of early IC effects during L2 processes. Specifically, NP2 verbs elicited significant inconsistency effects for the regression-path duration in Regions 2 and 3. Notably, these effects were observed only when the explanation relation was signaled by because. The measures showed no inconsistency effect in the full-stop condition or after NP1 verbs.
The null effect of NP1 bias matches the evidence from prior research, showing that L1-Japanese learners less effectively use NP1 bias than NP2 bias (Hijikata, 2021; Hosoda, 2023). However, prior L2 research on the time course of IC has never contrasted the connective and full-stop contexts. The present findings provide initial evidence that an upcoming explanation must be explicitly signaled for IC to invoke a mid-sentence influence on online L2 processing.
The empty-slot theory posits that the occurrence of coreference bias in the absence of an explicit explanation requires speakers to expect an explanation from IC verb semantics. This account explains the present finding that the L2 participants struggled to expect explanations from IC verbs in the full-stop condition. This was addressed in Experiment 2.
3 Experiment 2
The main RQ of Experiment 2 was as follows: Are L2 learners' expectations of an explanation from IC verbs lower than those of native speakers? To address this, a story-continuation task was used, in which the participants created continuations of the story sentence, including an IC (Mary annoyed Bob) or non-IC (Mary saw Bob) verb. As mentioned in the Introduction, native speakers favor explanation continuations for IC-verb stories relative to non-IC verb stories. Experiment 2 compared this coherence bias between L2 learners and native speakers. Based on the results of Experiment 1, this study predicted that L2 learners would show a lower coherence bias than native speakers. Because this study was specifically interested in whether an explanation was expected from IC verbs, the because condition in which an upcoming explanation is signaled was omitted (only the full-stop condition was set).
In addition, Experiment 2 asked the following RQ: When the explanation relation is not signaled, is L2 learners' reference to the IC-biased referent reduced compared with that of native speakers? Weakened coherence bias should lead to a reduction in coreference bias considering the assumption that coreference bias results from coherence bias (Bott and Solstad, 2014, 2021; Solstad and Bott, 2022). To investigate this, the L1 and L2 groups were compared in terms of consistency between the entities mentioned in the continuation and IC coreference bias (henceforth, reference-IC bias consistency).
It must be noted that the story-continuation task is an offline measure; continuations are produced after a sentence has been comprehended. Therefore, the data do not provide direct information on whether the participants expect an explanation during comprehension. Despite this limitation, this study used the story-continuation task because it is the most widely adopted method to test the expectation of coherence relations in L1 (Bott and Solstad, 2021; Kehler et al., 2008) and L2 (Grüter et al., 2017) studies. Employing this task was necessary to ensure the comparability of the results with those in the literature.
3.1 Method
3.1.1 Participants
In the L1 group, 56 monolingual English speakers were recruited through Amazon Mechanical Turk (27 females; Mage = 21.80; age range: 19–42). All participants had completed a university-level education. Ten participants who failed to appropriately complete an instructional manipulation check that aimed to ensure that they had read the instructions carefully were excluded. The analysis excluded another six participants because they did not provide an appropriate answer to “catch” items, which aimed to ensure that the participants carefully comprehended the stimuli. These procedures were necessary because participants in the online survey often “satisfice,” meaning that they try to complete the task without concentrating. The remaining 40 participants were included in the final analysis. They were paid US $7.00 for their participation.
In the L2 group, 40 Japanese university students participated (24 females; Mage = 18.71; age range: 18–20). None of them had been tested in Experiment 1 or had study-abroad experience in an English-speaking country. Their L2 proficiency was estimated to be at the CEFR A2–B1 levels or 28–80 on the TOEFL iBT test, based on their scores on the TOEIC IP test (M = 459.38, SD = 101.69), which was administered 3 weeks before the experiment. These scores suggested that the L2 participants' proficiency was almost similar to that in Experiment 1.
3.1.2 Materials
3.1.2.1 Verbs and the story-continuation task
In addition to the 24 IC verbs used in Experiment 1, 24 non-IC verbs from the norming study reported in Experiment 1 were used (see Method in Experiment 1). These verbs generated fewer than 65% of the references to the NP1 or NP2 referents in Japanese and English. This confirmed that non-IC verbs in English and Japanese do not cause a coreference bias toward the NP1 or NP2 direction.
The context stimuli were constructed using the 48 experimental verbs in the format of NP1 verb-ed NP2. Examples of the stimuli are presented in Table 3. The stimuli had two familiar English names for different genders. Half of the stimuli had a female NP1, whereas the other half had a male NP1.
Table 3. Example story stimuli in experiment 2.
In this task, the participants were instructed to write a continuation of the context stimulus that naturally came to mind in English. They were asked not to worry about grammar or spelling mistakes while avoiding humor. The exact instruction (presented in English and Japanese to the L1 and L2 participants, respectively) was as follows: “You will create continuations of stories in English. Please read the English sentences carefully to understand the story. Then, write a continuation of the story that comes to mind.”
Two material sets were constructed with the 48 experimental items and 48 fillers used in another experiment for the interpretation of the relative clause. The gender of the NPs was counterbalanced across the sets. The order of presentation of the items was randomized for each participant.
3.1.2.2 Translation task
The translation task was administered to the L2 participants in the same format as in Experiment 1. The items included 24 IC verbs and 24 non-IC verbs.
3.1.3 Procedure
The L1 participants were recruited through Amazon Mechanical Turk and directed to the survey website Qualtrics. They completed a demographic questionnaire and an instructional manipulation check. Subsequently, a story-continuation task was performed. One to four participants were tested simultaneously in the L2 group. The participants completed the story-continuation task in English, and the translation task was administered.
3.1.4 Coding and data treatment
3.1.4.1 Translation task
Two raters coded the responses (including the author) as either “correct” or “incorrect,” using the same procedure as in Experiment 1. The inter-rater agreement rate was 92%. All disagreements were resolved through discussion.
3.1.4.2 Story-continuation task
Prior to coding, continuations were excluded if they did not make sense (< 1% of the data) or if L2 participant provided an incorrect translation of the corresponding item (14% of the L2 data).
3.1.4.3 Coherence bias
Two raters coded the continuations (including the author) for whether they explained the cause of the event in the stimulus (“explanation”) or not (“non-explanation”). The raters used two tests based on previous story-continuation studies on coherence relations to identify explanation continuation (Grüter et al., 2017; Kehler et al., 2008). First, the raters considered whether the continuation answered the why question of the stimulus event. Second, they checked whether the causal connective because could felicitously relate stimulus and continuation without changing the gist of the story. The continuation was coded as “non-explanation” if it did not meet either of these criteria.
For example, the continuation Bob broke the rule for the stimulus Mary punished Bob is coded as “explanation” because it answers the why question (Why did Mary punish Bob?) and can be felicitously related to the stimulus by because (Mary punished Bob because Bob broke the rule). Conversely, Mary worked for the same company as Bob would be coded as “non-explanation” because it does not answer the why question but describes details of the event (i.e., an elaboration relation). Continuations were coded as “non-explanation” when they were interpreted as the result (e.g., Bob felt sorry), temporal succession (e.g., Mary taught him what to do), or an unexpected outcome of the event (e.g., But Bob continued to make trouble).
To identify the coherence relation, the raters drew on transition words (e.g., connectives and adjectives) if they were present in the continuation (e.g., Mary punished Bob. Because he broke the rule). However, the raters carefully checked whether the story continued reasonably using the relations denoted by such words. For example, even though because was used, some continuations can more felicitously be interpreted as elaborations rather than explanations (e.g., Mary punished Bob. Because Bob was her coworker.). In this case, the continuation was coded as “non-explanation.” When two or more types of coherence relations could be inferred, the continuation was coded as “unclear” and excluded from the analysis (5% of the data). Data were also excluded when the continuations did not make sense (< 1% of the data). The inter-coder agreement was 90%. All disagreements were resolved through discussion.
3.1.4.4 Coreference bias
Two raters (including the author) annotated the continuations produced in the IC (NP1 and NP2) contexts for which NP1 or NP2 entity was mentioned. When the pronoun was used, the raters used its gender as a cue. However, the raters carefully confirmed whether the referent referred to by the pronoun made sense considering the rest of the story. Continuations were coded as “unclear” when they had any ambiguity or neither the NP1 nor NP2 entity was mentioned (e.g., mentioning an entity that is not present in the stimuli or mentioning both NPs with a conjoined noun phrase [e.g., Mary and Bob] or a plural pronoun [they]). “Unclear” continuations (7% of the data) were excluded from the analysis. The inter-rater agreement rate was 92%. All disagreements were resolved through discussion.
3.1.5 Statistical analysis
3.1.5.1 Coherence bias
The continuation data (explanation vs. non-explanation) were submitted to a mixed-effects logistic regression model to test coherence bias, which caused more explanation continuations in the IC (NP1 and NP2) contexts than in the non-IC context. The fixed effects were bias direction (NP1, NP2, and non-IC [reference level]), group (L1 and L2), and the Direction × Group interaction. The bias direction was dummy-coded with the other sum-coded. This is because the difference between non-IC verbs, serving as the reference level, and IC (NP1 and NP2) verbs (i.e., simple effects) was relevant to the RQs, rather than the main effect of bias direction. The random-effects structure was determined through a backward model comparison using the same procedure as in Experiment 1. The selected model had random intercepts of participants and items as well as the by-participant random slope of the bias direction and by-item random slope of the group (formula: continuation ~ direction * group + [1 + direction | participant] + [1 + group | item]).
3.1.5.2 Coreference bias
The reference-IC bias consistency was computed by matching the referent mentioned in the continuation with the IC coreference bias (consistent vs. inconsistent). This consistency was analyzed using a mixed-effects logistic regression model with the fixed effects of bias direction (NP1 and NP2) and group (L1 and L2), both sum-coded, and the Direction × Group interaction. The random-effects structure was determined using a backward model comparison, as in other analyses. The selected model had random intercepts of participants and items as well as the by-participant random slope of the bias direction and by-item random slope of the group (formula: consistency ~ direction * group + [1 + direction | participant] + [1 + group | item]).
3.2 Results
3.2.1 Translation task
L2 participants produced more erroneous translations for NP1 verbs (error rates = 27%) than NP2 (error rates = 9%; β = −2.14, SE = 0.95, z = −2.56, p = 0.024) and non-IC (error rates = 10%; β = −1.77, SE = 0.80, z = −2.21, p = 0.027) verbs. Similar to Experiment 1, L2 participants often interpreted NP1 verbs with a passive meaning (61% of the erroneous trials). The error rates were not significantly different between the NP2 and non-IC verbs (β = −0.40, SE = 0.88, z = −0.45, p = 0.651).
3.2.2 Coherence bias
The descriptive statistics of the explanation rates are presented in Table 4.8 As illustrated in Figure 2, significant Direction × Group interactions emerged in NP1 and NP2 contexts (NP1: β = −0.93, SE = 0.18, z = −5.15, p < 0.001; NP2: β = −0.70, SE = 0.17, z = −3.99, p < 0.001).
Table 4. Means and standard deviations of rates of explanation continuations.
Figure 2. Explanation continuation rates for the L1 and L2 groups (± SEM bars).
Follow-up tests revealed that L1 participants showed coherence bias (more explanations in the IC context than the non-IC context) in NP1 and NP2 contexts (NP1: β = 1.97, SE = 0.43, z = 4.55, p < 0.001; NP2: β = 2.88, SE = 0.44, z = 6.53, p < 0.001). L2 participants showed coherence bias only in the NP2 context (β = 1.49, SE = 0.33, z = 4.56, p < 0.001) with no significant difference between the NP1 and non-IC contexts (β = 0.13, SE = 0.35, z = 0.39, p = 0.700).
The model showed no significant L1–L2 group difference in the non-IC context (β = 0.11, SE = 0.13, z = 0.90, p = 0.371). Conversely, NP1 and NP2 contexts elicited more explanations from L1 than L2 participants (NP1: β = −0.82, SE = 0.12, z = −6.73, p < 0.001; NP2: β = −0.53, SE = 0.11, z = −4.96, p < 0.001).
3.2.3 Coreference bias
Table 5 presents the descriptive statistics of reference-IC bias consistency rates. Figure 3 illustrates the results. Overall, the consistency was lower among L2 participants than L1 participants and in the NP1 context than the NP2 context, as indicated by the significant main effects of group (β = −0.29, SE = 0.11, z = −2.52, p = 0.012) and bias direction (β = −0.82, SE = 0.17, z = −4.93, p < 0.001), respectively.
Table 5. Means and standard deviations of reference-IC bias consistency rates.
Figure 3. Reference-IC bias consistency rates for the L1 and L2 groups (± SEM bars).
Additionally, a significant Direction × Group interaction was observed (β = −0.25, SE = 0.12, z = −2.02, p = 0.044). The NP1 context elicited higher consistency from the L1 group than the L2 group (β = −0.51, SE = 0.14, z = −3.53, p < 0.001). Conversely, the NP2 context showed no significant L1–L2 group difference (β = −0.07, SE = 0.16, z = −0.45, p = 0.651).9
3.3 Discussion
The results confirmed that explanatory expectations from IC verbs were more reduced in L2 learners than in native speakers. Unlike L1 participants, who showed higher explanation rates in both NP1 and NP2 contexts than in the non-IC context,10 L2 participants showed this coherence bias only in the NP2 context. Moreover, the explanation rates were consistently lower in L2 than L1 participants. These results indicated that IC verbs invoked coherence bias in L2 participants only in the limited condition (after the NP2 verb) and to a limited extent compared with L1 participants.
Experiment 2 also found that L2 participants were limited in terms of coreference bias, as reflected by their lower reference-IC bias consistency than that of L1 participants in the NP1 context. Because L2 participants failed to expect an explanation from the NP1 verbs, the explanation relation was not sufficiently operative to allow them to refer to the causally implicated referent, thereby creating the intergroup differences specifically in the NP1 context.
Finally, the NP2 context showed no significant L1–L2 differences in terms of coreference bias. This differs from the results for coherence bias, where L2 participants were limited relative to L1 participants in both NP1 and NP2 contexts. It might be that its recency to the end of the stimulus sentence made the NP2 entity more salient than the NP1 entity in the participants' mental models when the continuation was created.11 This might have increased the NP2 references to mask possible intergroup differences in the NP2 context. This idea was supported by the fact that the main effect of bias direction was significant (as reported in the first paragraph of the coreference bias results), indicating that the participants generally favored NP2 references over NP1 references.11
4 General discussion and conclusion
This study investigated how verbs' IC affects L2 learners' coreference and coherence processing. Experiment 1 revealed that NP2 bias affected learners' online coreference processing. Furthermore, this effect was observed only when an upcoming explanation was explicitly signaled. Subsequently, Experiment 2 showed that L2 learners were limited compared with native speakers in terms of coherence bias, whereby speakers expect explanations from IC verbs. Additionally, NP1 verbs failed to cause either coreference or coherence biases in learners, which matched the absence of the NP1 bias effect on online L2 processing, as shown in Experiment 1.
Prior to discussing the findings, we must be careful regarding the comparison of L1 and L2 comprehension in terms of the time course of IC because Experiment 1 did not contrast L2 learners with native speakers. However, Experiment 2 directly compared L1 and L2 participants and found that the IC effects in the L2 group were significantly more limited than those in the L1 group. Additionally, the difficulty experienced by L2 learners in the offline task was unlikely to be alleviated in the online task, considering that the online task poses higher cognitive demands: When performing an online task, the incoming linguistic information must be continuously processed within a limited time constraint, whereas linguistic knowledge can be retroactively or strategically used in an offline task. Considering these observations, it is reasonable to presume that using IC for online processing is more difficult for L2 learners than native speakers. From this perspective, the following sections discuss the findings in terms of the source and time course of IC.
4.1 Source of implicit causality bias in L2 comprehension
The empty-slot theory posits that the coreference bias of IC emerges as an epiphenomenon of coherence bias toward explanatory expectations (Bott and Solstad, 2014, 2021; Solstad and Bott, 2022). Thus, the finding of Experiment 1 that IC failed to cause coreference bias in the absence of an explicitly signaled explanation suggests that L2 learners were essentially limited in expecting an explanation from IC verbs. This idea was supported by Experiment 2, which showed that the L2 group exhibited a significantly weaker preference for explanation continuations after IC verbs than the L1 group. These findings suggest that L2 learners' fundamental weakness in IC processing is in coherence bias; they were significantly weaker than native speakers in generating discourse explanatory expectations from IC verb semantics. Owing to this limited explanatory expectation, the upcoming explanation must be explicitly signaled for learners to use the IC coreference bias online.
These findings indicate that the source of IC bias do not vary qualitatively between native speakers and L2 learners. In either case, coreference bias results from the explanatory coherence bias, which stems from IC verb semantics. Therefore, in the L2 context, this study supports the accounts (including the empty-slot theory) that attribute the source of IC bias to verb semantics (Bott and Solstad, 2014, 2021; Crinean and Garnham, 2006).
Furthermore, this study discovered that a significant L1–L2 distinction was located in the ease with which an implicit explanation relation can be activated through expectation-based processing. Hence, this study supports growing evidence indicating that L2 comprehension—compared to L1 comprehension—is characterized by reduced expectations (Cheng and Almor, 2018; Grüter et al., 2017; Kim and Grüter, 2021; Lew-Williams and Fernald, 2010).
The reduced expectation effects observed among the L2 participants may be related to the quality of their lexical representations. Specifically, the underspecified quality of L2 lexical representations constrained the retrieval of L2 semantic information, negatively affecting the generation of semantically driven expectations (Kaan, 2014; Kim and Grüter, 2021). In support of this reasoning, recent reviews of L2 prediction have indicated that the accuracy or consistency of lexical representations is a major factor in the reduced prediction effects in L2 (Kaan, 2014; Schlenter, 2023). In relation to the empirical evidence of IC, Kim and Grüter (2021) cited L2 learners' underspecified lexical representations as a potential cause of their weaker online IC effects compared with the stronger and more persistent IC effects in native speakers. Recall that, in the present study, L2 participants were confirmed by the translation task to know the meanings of IC verbs. Thus, their reduced expectations likely resulted from the utilization of the existing knowledge of IC verbs rather than a lack of that knowledge. Specifically, although L2 learners knew the meaning of IC verbs (as confirmed by the translation task), their representations of the knowledge of IC verbs were less detailed than those of native speakers. Consequently, the learners utilized their lexical knowledge of IC verbs less effectively than native speakers to expect an explanation.
From a theoretical standpoint, this perspective aligns with the lexical bottleneck hypothesis. Originally, this account posits that low-quality lexical representations constrain the integration of lexical and syntactic information (Hopp, 2013, 2018). However, the present study argues that it can be extended to incorporate expectation-based processing, in line with Kim and Grüter's (2021) argument that underdeveloped lexical representations can impede the effective use of lexical semantics, which is crucial for generating expectations. Accordingly, less developed lexical representations are assumed to limit learners' IC-based expectations of how the global discourse unfolds.
An alternative (not mutually exclusive) explanation for the reduced expectation effects involves neurocognitive components such as working memory and attentional resources. Specifically, in L2 comprehension, a substantial percentage of attentional resources in working memory is devoted to local-level processing (e.g., the recognition of individual words and parsing of the current clause). This leaves few resources for retrieving semantic information from the prior discourse and integrating multiple pieces of information, both of which are necessary for discourse-level expectations (e.g., in the case of IC coherence bias, combining verb semantics with the knowledge of event cause). Particularly, forming explanatory expectations from NP1 verbs is assumed to demand extensive attentional resources for the current L2 participants because these verbs are morphemically different from their counterparts in their L1. Supporting this idea, Experiment 2 found that L2 participants showed no coherence bias originating from NP1 verbs.
Weaker explanatory expectations in the L2 group also align with L2 accounts pointing to L1–L2 differences in predictive processing (Amos and Pickering, 2020; Corps et al., 2023; Grüter et al., 2017; Hopp, 2018). Specifically, the RAGE hypothesis states that the different processing patterns observed between native speakers and L2 learners result from the learners' reduced ability to generate expectations. Previous empirical studies have supported this view in terms of word- or phrase-level expectations (Grüter and Rohde, 2021; Lew-Williams and Fernald, 2010). The current findings corroborate the idea of reduced L2 expectations and extend it to clause- or sentence-level expectations by demonstrating that L2 learners are less sensitive to the IC coherence bias than native speakers.
In this regard, Grüter et al. (2017) reported seemingly contrasting findings to those of this study. In their research, native speakers and L2 learners expected different types of coherence relations from sentences with perfective and imperfective verbs. Notably, verb aspects are overtly marked by auxiliary verbs and morphemes (e.g., John handed/was handing a book to Bob), rendering the distinction between the different aspects explicit. By contrast, both IC coherence and coreference biases lack explicit linguistic markers and are (in agree with the “both IC coherence and coreference biases”) implicated in the semantics of the verb. Owing to its less deterministic nature, IC bias may be more challenging for L2 learners to use than verb aspects.
4.2 Time course of implicit causality in L2 comprehension
Experiment 1 found that only contexts in which NP2 verbs were followed by an explicitly signaled explanation (i.e., the NP2-because condition) elicited early IC effects during online L2 comprehension. The absence of online effects from the NP1 bias aligns with previous IC research on L1-Japanese learners (Hijikata, 2021; Hosoda, 2023) and may be attributed to cross-linguistic differences. In Japanese, NP1 verbs include overt causative morphemes (e.g., konran-saseru and shitsubou-saseru) that explicitly signal causation, whereas English NP1 verbs lack such morphemes (e.g., confuse and disappoint), thereby leaving causation implicitly encoded. This disparity likely hindered learners' ability to use IC information from NP1 verbs to guide online coreference processing. Another notable observation from the translation task was that L2 participants often incorrectly translated NP1 verbs by assigning them passive instead of transitive causative meanings. Although these mistranslated items were excluded from the analysis, this finding suggests that L2 participants' representations of causation expressed by English NP1 verbs were less developed.
Together, the findings of this study suggest that an explicit explanation relation and similar linguistic features in L1 are necessary for IC to influence online L2 coreference processing. When the explanation relation was not signaled (in the full-stop condition), learners use the IC only in the offline task. When verbs are differently encoded in L1 (NP1 verbs), IC is available in neither the online nor offline task.
This conclusion indicates that the difficulty L2 learners experience in applying IC bias to discourse processes primarily stems from cross-linguistic interference at the word level. As discussed above, the key difference between Japanese and English with respect to IC bias lies in the presence or absence of an overt causative morpheme in NP1 verbs. NP2 verbs do not explicitly mark causation in either language, and both languages share similar discourse-level coherence relations, which can be explicitly signaled by connectives (e.g., nazenara in Japanese, because in English) or inferred from context. According to the empty-slot theory (Bott and Solstad, 2014, 2021; Solstad and Bott, 2022), Japanese learners of English are less sensitive to the explanatory empty slot due to their underspecified representations of causation in English NP1 verbs, which can be attributed to cross-linguistic interference. As a result, they are less likely to expect that events denoted by NP1 verbs will be followed by an explanation in the upcoming discourse, manifesting as the absence of coherence bias. Moreover, the underspecified representations likely increase the cognitive demands required to compute the causally implicated referent during real-time comprehension. Consequently, online IC effects were observed only for NP2 bias, while NP1 bias effects did not emerge—even in the because condition, where the explanation relation was explicitly signaled.
These findings indicate that IC's entry into online L2 comprehension is determined by both discourse coherence relations and learners' L1 backgrounds. Apparently, the time course of IC in L2 cannot be explained simply by traditional focusing or integration account. The focusing account proposes that IC enters comprehension processes immediately after the bias-consistent pronoun (McKoon et al., 1993), whereas, according to the integration account, IC exerts effects only at the end of the sentence (Stewart et al., 2000). Neither account adequately explains the modulation of discourse or learner factors in the time course of IC. Thus, this study agrees with the incremental integration account, which integrates the focusing and integration phases to describe the full spectrum of IC effects (Koornneef et al., 2016; Koornneef and Sanders, 2013). This model maintains that the focusing phase determines the timing at which IC begins to affect comprehension, whereas the integration phase determines how linguistic factors and speakers' individual differences modulate the IC effects.
This account explains the present findings that the integration phase of IC is conditioned by coherence relations and L1 linguistic properties. Specifically, L2 learners can use IC bias online under the condition that the explanation relation is explicitly signaled and the IC verbs are similarly represented in L1 and L2. Regarding the focusing phase, the current study's findings, however, do not provide definitive conclusions. This is because under the current methodology, online IC effects are observable only at or after the pronoun is encountered as IC effects are operationalized by the consistency between the pronoun and IC bias. Therefore, whether the starting point of the IC effect is the participants' proactive prediction of the referent prior to the pronoun or their incremental integration after the pronoun cannot be discerned.
Consequently, I avoid making definitive claims regarding when IC begins to affect L2 learners' comprehension. It is, still, certain that L2 learners do not wait until the end of the sentence to start using IC information; Experiment 1 has found the pronoun inconsistency effects in Regions 2 and 3 that are located in the middle of the sentence. Learners used IC immediately after encountering the pronoun at the latest, in line with the accumulating evidence of early IC effects during L2 comprehension (Contemori and Dussias, 2019; Hosoda, 2023; Kim and Grüter, 2021; Wang and Gabriele, 2022). Accordingly, this study contributes to the expanding L2 literature showing that L2 learners can use IC coreference bias not only retroactively but also incrementally during comprehension. The novelty of this study is that it specifies the conditions under which such online IC effects occur in L2 processing as well as when and why L2 learners fail to use IC for coreference and coherence processing. These discoveries provide a finer-grained picture of the cognitive mechanism of IC processing in L2.
To conclude, several limitations of this study should be discussed to guide future research. First, the sample size was relatively small, resulting in a narrow range of L2 proficiency among participants. A critical limitation is that both experiments tested basic-level L2 learners. If reduced expectations are a characteristic of L2 learners' comprehension, as posited by the RAGE hypothesis, we would expect weakened expectations to persist even among highly proficient learners who typically possess richer L2 lexical representations and are less constrained by attentional resources. Further empirical investigations are required to clarify this aspect.
Second, this study did not compare online IC processing of L2 learners with that of native speakers. Although online IC processing can be assumed to pose greater difficulty for L2 learners than for native speakers (see the second paragraph of the General Discussion), further research comparing L1 and L2 groups is necessary to corroborate this view.
Third, although eye tracking can measure the time course of comprehension processes, it does not directly reveal the specific cognitive operations that participants engaged in during the task. For example, L2 learners might strategically code-switch or code-mix while processing L2 materials, which could confound their processing time. To capture L2 learners' IC processing more thoroughly, future research should complement the present findings with neurophysiological measures, such as event-related potentials.
Finally, the extent to which L2 participants generated explanatory expectations during real-time comprehension is unclear because Experiment 2 used the offline story-continuation task. Future research should supplement this study's findings with online measures.
Nevertheless, the fact that L2 learners showed reduced expectation effects even in the offline task does not necessarily undermine the conclusions of this study. Considering that the online task involves higher cognitive demands, it seems unlikely that L2 learners would generate explanatory expectations online if they already fail to do so in the offline task.
Data availability statement
Publicly available datasets were analysed in this study. This data can be found here: https://osf.io/3qnz6/?view_only=a4e46c7da585467e994632f4f57aee4b.
Ethics statement
The studies involving humans were approved by Hokkaido University of Education. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.
Author contributions
MH: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.
Funding
The author(s) declare that financial support was received for the research and/or publication of this article. This study was supported by JSPS Grant-in-Aid for Early-Career Scientists (Nos. 20K13127 & 24K16137).
Conflict of interest
The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/flang.2025.1494500/full#supplementary-material
Footnotes
1. ^A persistent debate exists regarding whether native speakers proactively use IC to predict the referent online because evidence regarding the anticipatory looks to the biased referent before the critical pronoun is inconsistent.
2. ^All the materials, datasets, analysis codes, and statistical results in this study are available at https://osf.io/3qnz6/?view_only=a4e46c7da585467e994632f4f57aee4b.
3. ^Non-IC verbs included stimulus–experiencer verbs (e.g., surprised and encouraged) and an experiencer–stimulus verb (forgot), which could be classified as NP1 and NP2 verbs, respectively. However, norming data from Ferstl et al. (2011) and this study showed that the percentage of NP1 references for these verbs ranged from 42% to 55%, indicating that these verbs do not induce a bias toward NP1 or NP2 references. Based on this evidence, this study categorized these verbs as non-IC verbs. I appreciate the reviewer's comment on this issue.
4. ^Japanese university students from the same population (N = 32) confirmed that the genders of the names were easy to identify.
5. ^Because regression-path duration is the most inclusive eye-tracking measure, some studies deemed it as an early measure (Clifton et al., 2007; Cunnings and Sturt, 2018). This study treated it as a late measure because it indexes processing difficulty, which occurs in later stages of comprehension.
6. ^All statistical results are provided in the Supplementary material.
7. ^L2 proficiency test scores (centered) were added to the maximal converging models to examine the potential effects of the participants' L2 proficiency on the results. The model comparisons revealed that the addition of L2 proficiency did not significantly contribute to the model fit, meaning that the difference in the participants' L2 proficiency did not influence the results.
8. ^All statistical results are provided in the Supplementary material.
9. ^L2 proficiency test scores (centered) were added to the maximal converging models of L2 data. The addition of L2 proficiency did not significantly improve the model fit in either coherence or coreference bias. This means that the difference in the participants' L2 proficiency did not influence the results.
10. ^One may argue that the explanation rates for the L1 group were not high (70% at best). However, these explanation rates were generally similar to those in L1 research in which native speakers provided explanations about 60% of the time after IC verbs (Kehler et al., 2008).
11. ^In the non-IC context, the analysis showed no L1–L2 group difference in NP2 reference rates (β = −0.18, SE = 0.23, z = −0.78, p =.437), indicating that the L1 and L2 participants were affected by the recency of the NP2 entity, if any, to a similar extent. Considering this, the reduced coreference bias in the L2 participants in the NP1 context cannot be explained by the idea that the L2 participants were more strongly affected by the recency than the L1 participants.
References
Amos, R. M., and Pickering, M. J. (2020). A theory of prediction in simultaneous interpreting. Biling.: Lang. Cogn. 23, 706–715. doi: 10.1017/S1366728919000671
Bott, O., and Solstad, T. (2014). “From verbs to discourse: a novel account of implicit causality,” in Psycholinguistic Approaches to Meaning and Understanding Across Languages, eds. C. Fabricius-Hanse, B. Hemforth, and B. Mertins (Cham: Springer), 213–251.
Bott, O., and Solstad, T. (2021). Discourse expectations: Explaining the implicit causality biases of verbs. Linguistics 2, 361–416. doi: 10.1515/ling-2021-0007
Carpenter, P. A., and Just, M. A. (1983). “What your eyes do while your mind is reading,” in Eye Movements in Reading: Perceptual and Language Processes, eds. K. Rayner (San Diego, CA: Academic Press), 275–307.
Cheng, W., and Almor, A. (2017). The effect of implicit causality and consequentiality on nonnative pronoun resolution. Appl. Psycholinguist. 38, 1–26. doi: 10.1017/S0142716416000035
Cheng, W., and Almor, A. (2018). A Bayesian approach to establishing coreference in second language discourse: Evidence from implicit causality and consequentiality verbs. Bilingualism: Lang. Cognit. 22, 456–475. doi: 10.1017/S136672891800055X
Clifton, C., Staub, A., and Rayner, K. (2007). “Eye movements in reading words and sentences,” in Eye Movements: A Window on Mind and Brain, eds. R. V. Gompel, M. Fisher, W. Murray, and R. L. Hill (Amsterdam: Elsevier), 341–371.
Contemori, C., and Dussias, P. E. (2019). Prediction at the discourse level in Spanish–English bilinguals: an eye-tracking study. Front. Psychol. 10:956. doi: 10.3389/fpsyg.2019.00956
Corps, R. E., Liao, M., and Pickering, M. J. (2023). Evidence for two stages of prediction in non-native speakers: a visual-world eye-tracking study. Bilingualism: Lang. Cognit. 26, 231–243. doi: 10.1017/S1366728922000499
Crinean, M., and Garnham, A. (2006). Implicit causality, implicit consequentiality and semantic roles. Lang. Cogn. Process. 21, 636–648. doi: 10.1080/01690960500199763
Cunnings, I., and Sturt, P. (2018). Retrieval interference and semantic interpretation. J. Memory Lang. 102, 16–27. doi: 10.1016/j.jml.2018.05.001
Ferstl, E. C., Garnham, A., and Manouilidou, C. (2011). Implicit causality bias in English: a corpus of 300 verbs. Behav. Res. Methods 43, 124–135. doi: 10.3758/s13428-010-0023-2
Grüter, T., and Rohde, H. (2021). Limits on expectation-based processing: use of grammatical aspect for co-reference in L2. Appl. Psycholinguist. 42, 51–75. doi: 10.1017/S0142716420000582
Grüter, T., Rohde, H., and Schafer, A. J. (2017). Coreference and discourse coherence in L2: The roles of grammatical aspect and referential form. Linguist. Approach. Bilingual. 7, 199–229. doi: 10.1075/lab.15011.gru
Hartshorne, J. K., O'Donnell, T. J., and Tenenbaum, J. B. (2015). The causes and consequences explicit in verbs. Lang. Cognit. Neurosci. 30, 716–734. doi: 10.1080/23273798.2015.1008524
Hartshorne, J. K., Sudo, Y., and Uruwashi, M. (2013). Are implicit causality pronoun resolution biases consistent across languages and cultures? Exp. Psychol. 60, 179–196. doi: 10.1027/1618-3169/a000187
Hijikata, Y. (2021). The time course of the effects of implicit causality bias on anaphora resolution by Japanese learners of English. Ann. Rev. English Lang. Educ. Japan (ARELE) 32, 1–16. doi: 10.20581/arele.32.0_1
Hopp, H. (2013). Grammatical gender in adult L2 acquisition: Relations between lexical and syntactic variability. Second Lang. Res. 29, 33–56. doi: 10.1177/0267658312461803
Hopp, H. (2018). The bilingual mental lexicon in L2 sentence processing. Second Lang. 17, 5–27. doi: 10.11431/secondlanguage.17.0_5
Hosoda, M. (2020). Establishing coreference in Japanese EFL learners' using verbs' implicit causality: A sentence completion study. ARELE 31, 193–208.
Hosoda, M. (2022). “Expecting coherence relations from verbs' implicit causality: a comparison between L2 learners and native speakers,” in Proceedings of the 47th Japan Society of English Language Education Conference in Hokkaido, 130–131.
Hosoda, M. (2023). Time course of verbs' implicit causality during L2 comprehension: An extended replication of Hijikata 2021 Japanese EFL learners. ARELE 34, 97–112.
JACET Basic Word Revision Committee (2016). The New JACET list of 8000 Basic Words. Tokyo: Kirihara Shoten.
Kaan, E. (2014). Predictive sentence processing in L2 and L1: what is different? Linguist. Approach. Bilingual. 4, 257–282. doi: 10.1075/lab.4.2.05kaa
Kehler, A., Kertz, L., Rohde, H., and Elman, J. L. (2008). Coherence and coreference revisited. J. Semant. 25, 1–44. doi: 10.1093/jos/ffm018
Kim, H., and Grüter, T. (2021). Predictive processing of implicit causality in a second language: a visual-world eye-tracking study. Stud. Second Lang. Acquisit. 43, 133–154. doi: 10.1017/S0272263120000443
Koornneef, A., Dotlačil, J., van den Broek, P., and Sanders, T. (2016). The influence of linguistic and cognitive factors on the time course of verb-based implicit causality. Quart. J. Exp. Psychol. 69, 455–481. doi: 10.1080/17470218.2015.1055282
Koornneef, A. W., and Sanders, T. J. (2013). Establishing coherence relations in discourse: The influence of implicit causality and connectives on pronoun resolution. Lang. Cogn. Process. 28, 1169–1206. doi: 10.1080/01690965.2012.699076
Koornneef, A. W., and Vanberkum, J. (2006). On the use of verb-based implicit causality in sentence comprehension: evidence from self-paced reading and eye tracking. J. Mem. Lang. 54, 445–465. doi: 10.1016/j.jml.2005.12.003
Lew-Williams, C., and Fernald, A. (2010). Real-time processing of gender-marked articles by native and non-native Spanish speakers. J. Mem. Lang. 63, 447–464. doi: 10.1016/j.jml.2010.07.003
McKoon, G., Greene, S. B., and Ratcliff, R. (1993). Discourse models, pronoun resolution, and the implicit causality of verbs. J. Exp. Psychol. 19, 1040–1052. doi: 10.1037//0278-7393.19.5.1040
Papageorgiou, S., Tannenbaum, R. J., Bridgeman, B., and Cho, Y. (2015). The Association Between TOEFL iBT® Test Scores and the Common European Framework of Reference (CEFR) Levels. Princeton, NJ: Educational Testing Service.
Pyykkönen, P., and Järvikivi, J. (2010). Activation and persistence of implicit causality information in spoken language comprehension. Exp. Psychol. 57, 5–16. doi: 10.1027/1618-3169/a000002
Schlenter, J. (2023). Prediction in bilingual sentence processing: How prediction differs in a later learned language from a first language. Biling.: Lang. Cogn. 26, 253–267. doi: 10.1017/S1366728922000736
Solstad, T., and Bott, O. (2022). On the nature of implicit causality and consequentiality: the case of psychological verbs. Lang. Cognit. Neurosci. 37, 1311–1340. doi: 10.1080/23273798.2022.2069277
Sorace, A. (2011). Pinning down the concept of ‘interface' in bilingualism. Linguist. Approach. Bilingual. 1, 1–33. doi: 10.1075/lab.1.1.01sor
Stewart, A. J., Pickering, M. J., and Sanford, A. J. (2000). The time course of the influence of implicit causality information: focusing versus integration accounts. J. Mem. Lang. 42, 423–443. doi: 10.1006/jmla.1999.2691
Keywords: implicit causality, L2 processing, eye-tracking, coherence relations, coreference processing, expectation-based processing
Citation: Hosoda M (2025) Verbs' implicit causality in coreference and coherence processing during L2 comprehension. Front. Lang. Sci. 4:1494500. doi: 10.3389/flang.2025.1494500
Received: 11 September 2024; Accepted: 26 February 2025;
Published: 09 April 2025.
Edited by:
Pia Knoeferle, Humboldt University of Berlin, GermanyReviewed by:
Sofiana Lindemann, Transilvania University of Braşov, RomaniaDalia Elleuch, University of Sfax, Tunisia
Mohamed Taiebine, UEMF, Morocco
Chao Sun, Peking University, China
Copyright © 2025 Hosoda. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Masaya Hosoda, aHptMTI3QGdtYWlsLmNvbQ==