How referential uncertainty is modulated by conjunctions: ERP evidence from advanced Chinese–English L2 learners and English L1 speakers

Abstract

Conjunctions play a crucial role in the construction of a coherent mental representation by signaling coherence relations between clauses, especially for second language users. By using event-related potentials (ERPs), this study aimed to investigate how different conjunctions (so, and, although, or a full stop) affect the interpretation of a following ambiguous pronoun for both native and non-native speakers, in sentences such as Lily disappointed Nina, so she …. ERP results showed that relative to so, and, and full stop sentences, the pronoun in although clauses elicited a larger Nref (sustained negativity) response in both native (L1) readers and second language (L2) readers, irrespective of whether the verb in the first clause biased towards a particular noun phrase (NP) referent. Moreover, larger Nrefs to pronouns were seen in L2 than L1 readers when clauses were connected by so, although or a full stop. Additionally, larger Nref responses were evoked by pronouns in NP2- than NP1-biased conditions when the clauses were connected by the conjunction so or when sentences contained no overt conjunctions (full stop). These findings indicate that different conjunctions exert different modulating effects on resolving referential uncertainty/ambiguity. Relative to native speakers, non-native speakers are more likely to encounter referential uncertainty when the sentences are conjoined by conjunctions with more complex semantics.

Keywords

conjunction ERPs implicit causality pronoun resolution referential uncertainty

I Introduction

Conjunctions play an important role in constructing a mental representation of a text, as they can direct readers’ expectations regarding how upcoming statements relate to previous ones (Murray, 1995). In particular, when there is referential uncertainty in discourse due to more than one referent being simultaneously available for a referring expression (e.g. pronoun), the presence of an overt conjunction can help establish a coherent discourse representation, and as a consequence, may reduce referential uncertainty (Hobbs, 1979; Kehler, 2002; Kehler et al., 2008; Rohde et al., 2007; Wolf and Gibson, 2004). Conjunctions are particularly important for second language discourse processing because second language learners depend more than native speakers on such overt pragmatic cues to establish discourse coherence (Pan and Felser, 2011; Roberts et al., 2008). Although a number of psycholinguistic studies have examined how different conjunctions influence pronoun resolution in native speakers (Ehrlich, 1980; Koornneef and Sanders, 2013; Mak and Sanders, 2010; Stevenson et al., 1994, 2000; Wolf et al., 2004; Xu et al., 2015), few studies have thus far explored this question from the perspective of second-language processing (e.g. Demel, 1990). The purpose of the present study was to investigate, first, how referential uncertainty during pronoun resolution is affected by coherence relations signaled through different conjunctions (e.g. so, and, although, or a full stop); second, how this effect is modulated by verb-based semantic bias, and third, whether advanced second language (L2) learners and native speakers (L1) depend on the same information and cognitive processes to resolve referential uncertainty.

1 Ambiguity/uncertainty during referential processing

Referential ambiguity/uncertainty may occur when there is more than one antecedent available for a referential expression. For example, in context (1) two girls are introduced, which will lead to an uncertainty as to whom the referential expression (the girl) refers to. This is in contrast to the context in (2) in which only one girl is introduced. Previous event-related potential (ERP) studies have consistently shown a larger sustained negative shift (Nref) right at the ambiguous word, in response to referential ambiguity, emerging at about 300–400 ms over the front of the head and without an obvious peak (Nieuwland, 2014; Nieuwland and Van Berkum, 2006, 2008a, 2008b; Van Berkum et al., 1999).

(1) David had told the two girls to clean up their room before lunch time. But one of the girls had stayed in bed all morning, and the other had been on the phone all the time. David told the girl …

(2) David had told the boy and the girl to clean up their room before lunch time. But the boy had stayed in bed all morning, and the girl had been on the phone all the time. David told the girl …

Unlike the N400 or P600, which are typically evoked by semantic or syntactic anomalies, the Nref is an ERP component which is mainly correlated with referential processing. The size of the Nref effect is associated with the degree of referential ambiguity/uncertainty and has been found to be positively correlated with readers’ working memory capacity (e.g. Nieuwland and Van Berkum, 2006).

A number of factors, including syntactic, semantic, and even discourse-level constraints, can decrease the referential uncertainty by increasing the prominence of an entity (Kehler et al., 2008). First, there is a preference to assign a pronoun to antecedents occupying the sentence topic/subject position (e.g. Cowles et al., 2007; Crawley et al., 1990; Kaiser, 2011; Xu, 2015; Xu and Zhou, 2016), or to antecedents which play the same grammatical role as the pronoun (Ellen et al., 1978; Kehler et al., 2008). Second, implicit causality, as a property of a group of interpersonal verbs, in which one of those verbs’ arguments is usually implicated as the underlying cause of a certain action or attitude, also helps resolve the referential ambiguity (e.g. Garvey and Caramazza, 1974; Garvey et al., 1975; Greene and McKoon, 1995; Koornneef and Van Berkum, 2006; Stewart et al., 2000). Van Berkum et al. (2007) found that a bias-inconsistent pronoun elicited a larger P600 than a bias-consistent pronoun, probably because implicit causality can be used as a cue to predict which entity would be more likely to be mentioned subsequently. Moreover, coherence conjunctions like because, so, and, and although can signal how one proposition relates to the other in discourse. The type of conjunction could therefore strengthen or weaken the implicitly expected co-referential relations between clauses, which could change the degree of referential uncertainty accordingly.

2 The influence of conjunctions on sentence/referential processing

Previous studies suggest that being able to use discourse markers is a beneficial reading strategy for both L1 and L2 readers (Degand and Sanders, 2002; Jung, 2003; Murray, 1995). In the following sections, we will review literature of experimental studies of conjunctions in L1 and L2, respectively.

a Conjunctions for L1 speakers

To get to the meaning of a discourse, readers are required not only to grasp the meaning of each individual proposition, but also to build a coherence relation between propositions (Kintsch, 1988; Kintsch and van Dijk, 1978; Xu, et al., 2015). Different coherence relations can be either implicitly conveyed without any conjunctions, or explicitly conveyed through various types of conjunctions, such as the causal conjunction because, the consequential conjunction so, the additive conjunction and, or the concessive conjunction although. As cohesive devices, conjunctions can be indicative of the coherence relations between propositions, which in turn can provide information as to which antecedent a subsequent referential expression is likely to refer to.

Previous studies have shown that texts can be comprehended better and faster when there is a conjunction such as because, compared to when there are no overt conjunctions (Cozijn et al., 2011; Degand and Sanders, 2002; Millis and Just, 1994; Xu et al., 2018). This is especially the case for less skilled readers (e.g. Van Silfhout et al., 2015), as overt conjunctions can signal how one part of a text is related to another.

Among the various coherence conjunctions, causal conjunctions such as because have received the most attention. The presence of a causal conjunction following a statement typically conveys a consequence-cause relation, focusing on the cause of the previously described event (Garvey and Caramazza, 1974; Garvey et al., 1975; Greene and McKoon, 1995; Koornneef and Van Berkum, 2006; Stevenson et al., 1994, 2000; Stewart et al., 2000). The consequential conjunction so differs from because, as so represents a reversed causal relation, namely ‘cause-consequence’ rather than ‘consequence-cause’ relation, and hence puts the consequence of an event in focus. Research on native speakers has shown that these causal and consequential conjunctions differently affect how referential ambiguity of a following referential expression is resolved (Fukumura and van Gompel, 2010; Stevenson et al., 1994, 2000). For example, Stevenson et al. (1994) found that while readers preferred to interpret an ambiguous pronoun he as referring to the first noun phrase (NP), John, in because-linked sentences (e.g. John disappointed Bill because he …), they preferred to interpret the pronoun as referring to the second NP, Bill, in so-linked sentences (e.g. John disappointed Bill so he …). A possible explanation is that the conjunction so provokes a referential preference for the character associated with the consequential part of the event in the first clause.

The conjunction and is mainly characterized as expressing an additive relation between two clauses, where the second clause elaborates in a nonspecific manner of the content of the first (Murray, 1997), leaving room for all possible inferential enrichments compatible with its semantics (Mauri and van der Auwera, 2012). However, and can also be causally interpreted, as it can convey a cause-consequence relation, like the conjunction so (Ehrlich, 1980; Halliday and Hasan, 1976; Knott and Sanders, 1988; Martin, 1992; Sanders et al., 1992). Some (Evers-Vermeul and Sanders, 2009; Goldman and Murray, 1992; Sanders, 2005) have argued that an additive relation is the simplest relation because it exerts the least constraints on the semantic relationship between clauses, and is always acquired earlier than the other relations (e.g. concessive relation although, causal relation because/so). In a sentence completion study, Stevenson et al. (2000) found that both conjunctions so and and direct attention to the endpoint or consequence of the described actions (e.g. Joseph hit Patrick , so /and he …) or states (e.g. Joseph impressed Patrick , so /and he …), leading to a clear preference for pronoun resolution.

Finally, the conjunction although denotes a consequence-concession relation. Taking sentence (4) as an example, this conjunction introduces a clause describing a denial of the expected causes of an event, which is different from the causal relation as expressed in sentence (3).

(3) Jack got a promotion because he worked hard.

(4) Jack got a promotion although he didn’t work hard.

According to the taxonomic analyses of coherence relations (Sanders et al., 1992, 1993), concessive conjunction although differs from other causal conjunctions such as because, so, and even and in several important primitives including polarity (negative relation vs. positive relation) and continuity (continuous relation vs. discontinuous relation).

This distinction in taxonomic properties among conjunctions could possibly lead to dissociations in real-time processing. Indeed, behavioral studies showed that processing concessive relations is much slower (Xu et al., 2018) or less accurate (Kӧhne and Demberg, 2012) than processing causal relations. Using ERPs, Xu et al. (2015) observed a larger P600 at the critical word (warm) for concessive constructions (e.g. Grandma moved from Southern to Northern China although she likes the South, where the winter is warm) relative to causal constructions (e.g. Grandma moved from Northern to Southern China because she likes the South, where the winter is warm), presumably because the additional process of negating a presupposed casual relation makes a concessive relation more cognitively demanding (Verhagen, 2000).

Coherence relations can also be constructed even when there are no explicit devices signaling the relation between clauses/sentences. The implicit coherence relations, in such cases, may be causally interpreted, as a causal relation is typically the default (Hoek and Zufferey, 2015; Mak and Sanders, 2013; Oudega, 2011; Sanders, 2005). By using cross-modal probe recognition, McDonald and MacWhinney (1995) found that the observed subject preference of pronoun resolution (also with first-mentioned advantage) in NP1-biased sentence (with Stimulus-Experiencer verbs such as amaze) disappeared in the NP2-biased sentence (with Experiencer-Stimulus verbs such as admire), irrespective of whether the causal relation was expressed explicitly or not, suggesting that the implicit causality relation is not necessarily dependent on an explicit conjunction. However, although a number of studies showed that the absence of a conjunction in texts compared to its presence leads to longer processing time or worse text comprehension (e.g. Haberlandt, 1982; van Silfhout et al., 2015; Xu et al., 2018), other studies showed a more complex picture. For example, Millis and Just (1994) showed that the absence of a conjunction (e.g. because) in comparison to its presence increases the probe recognition time at the end of the sentence but decreases the time at an earlier probe position. In an eye-tracking study, Cozijn et al. (2011), however, showed that the absence of a conjunction slows down the processing of the immediately following words, but speeds up the processing of the sentence final words. These inconsistent findings indicate that whether the absence of conjunctions can lead to difficulty in sentence processing is determined by a number of factors including the strength of implicit causality, the research techniques, and even the position of the test point.

b Conjunctions for L2 learners

Non-native speakers (L2 learners) may benefit more from coherence markers, as they may depend more on the superficial cues on how to integrate separate clauses/sentences (e.g. Crosson and Lesaux, 2013; Chung, 2000; Eshagh, 2001; Horiba, 1996; McClure and Steffensen, 1985; Ozono, 2002; Zorana, 2013; but, for a different perspective, see Droop and Verhoeven, 2003). In particular, previous behavioral research has demonstrated that non-native speakers may be guided more strongly by pragmatic cues in on-line referential processing compared with native speakers (Pan and Felser, 2011; Roberts et al., 2008). However, there are very few experimental studies probing into the relation between conjunction type and referential processing in L2 speakers. Existing empirical L2 studies on conjunctions in text mainly focus on whether L2 learners can process discourse conjunctions in a native-like way (e.g. Crosson, and Lesaux, 2013) or whether the presence of a coherence conjunction can facilitate text understanding (e.g. Eshagh, 2001; Ozono, 2002). For example, Ozono (2002) reported that Japanese learners of English as a foreign language (EFL), including both high- and low-proficiency learners, showed better text comprehension when the logical relations between sentences were indicated explicitly than when they were indicated implicitly. Similarly, Eshagh (2001) found that Iranian EFL learners performed worse in a reading comprehension task when the conjunctions were omitted than when they were present. However, although conjunctions are particularly beneficial for L2 readers, understanding and using conjunctions is not always easy: some conjunctions (e.g. although) introduce relations that are relatively complex, are acquired late, and are resource-demanding (Degand and Sanders, 2002; MacLean and d’Anglejan, 1986). Using off-line questionnaire tests, several studies (e.g. Ozono, 2002; Ozono and Ito, 2003) examined how L2 text comprehension could be affected by adversative (e.g. however) and other conjunctions (e.g. causal: because/therefore). They found that adversative relations (e.g. however) were more difficult to process than other relations (e.g. because) for low-proficiency learners. In particular, only advanced L2 learners could benefit from coherence conjunctions (Degand and Sanders, 2002; Geva, 1986, 1992), presumably because L2 beginners allocate most of their resources to processing lexical and syntactic information, and may not have sufficient resources for higher level text processing such as the logic relationship of discourse. Moreover, L2 learners may even misuse or misinterpret certain conjunctions (Alexander, 1980; Zufferey et al., 2015), for example, they may not be able to detect the incorrect use of a logical conjunction (e.g. since, if) between two sentences (Horiba, 1996; Zufferey et al., 2015); only with increased proficiency can L2 learners improve their ability to utilize and infer logical relationships between clauses/sentences (Geva, 1986, 1992). Consistent with this argument, recent psycholinguistic evidence has shown that, with increases in proficiency, advanced L2 learners can be about as sensitive as native speakers in detecting the misuse of conjunctions (e.g. if, when; Zufferey et al., 2015).

3 The influence of verb-based semantic bias on referential processing

The influence of conjunctions on resolving ambiguous pronouns can be modulated by implicit causality. For example, in the sentence ‘John disappointed Bill because he …’ the verb ‘disappoint’ implicates that ‘John’ might be responsible for ‘Bill’s disappointment’ (verbs like ‘disappoint’ are labeled as NP1-biased verbs as they regularly induce a preference to interpret a following pronoun as referring to the subject of the verb), while in the sentence ‘John praised Bill because he …’, the verb ‘praise’ biases ‘Bill’ towards the causer of ‘John’s praise’ (verbs like ‘praise’ are labeled as NP2-biased verbs as they induce an object referential preference for the pronoun). An alternative explanation (Kehler, 2002; Kehler et al., 2008; Rohde et al., 2006, 2007), however, suggests that implicit causality is not simply a lexical feature, but rather is a side effect of constructing a coherent discourse representation, as an implicit causality verb generally triggers the expectation for an explanation.

Although implicit causality is considered to be universal across languages, which verbs show which biases may vary. Previous studies have shown that while English and Chinese use similar NP1-bias verbs to express causation, they use different ways to realize NP2-bias causation. For example, both Chinese and English use similar NP1-bias psychological verbs to express causation, but Chinese only has a limited number of NP2-bias psychological verbs in comparison to English. Therefore, Chinese mainly uses periphrastic (Lily shi Nina shiwang / ‘Lily makes Nina disappointed’) rather than lexical (e.g. disappoint) structures to express NP2-bias causation (Cheng and Almor, 2017). In periphrastic causatives, causation is explicitly expressed in another verb shi (‘make’) rather than implicitly incorporated into the psychological verb itself (e.g. ‘disappoint’; Cheng and Almor, 2017). These cross-language variations could exert different influences on second language referential processing.

In an early behavioral study, Ehrlich (1980) investigated how conjunctions and verb types (NP1-bias vs. NP2-bias) interact to modulate referential uncertainty. They found that the pronoun was preferentially resolved towards the referent which is consistent with the direction of bias of the verb in because sentence (i.e. the first noun in NP1-biased sentences and second noun in NP2-biased sentences), but an opposite pattern was obtained in but sentence (i.e. the second noun in NP1-biased sentences and the first noun in NP2-biased sentences), since the adversative conjunction but can reverse the bias of the verb. Interestingly, the author failed to observe the biasing effect of verb-based implicit causality in clauses linked by additive conjunction and. Xu et al. (Xu and Zhou, 2016; Xu et al., 2017) recently investigated how pronoun resolution was influenced by topic structure and verb-based semantic bias in because-linked sentences, using the ERP technique. They found an early positivity (P200) in response to topic-shift relative to topic-continuation for the NP2-biased verbs. Importantly, they also found a larger P600 in response to topic-shift relative to topic-continuation regardless of whether NP1- or NP2-biased verbs were used. This seems to suggest that topic shift requires more attentional resources and therefore evoked an early positive modulation; it also suggests that verb-based semantic bias has limited influence on pronoun resolution when the two sentences were overtly connected by additive or causal conjunctions.

Apart from the influence of additive/causal conjunctions and implicit causality on pronoun resolution, little is known about how pronoun resolution in clauses/sentences connected by other conjunctions (e.g. although, so) is modulated by verb type (NP1 bias vs. NP2 bias). In particular, the existing relevant studies mainly focused on native speakers, it is thus not clear how non-native speakers deal with the same issue, that is, how discourse conjunctions interact with verb type to modulate referential uncertainty.

4 The present study

The current study examined, first, how referential uncertainty during pronoun resolution is influenced by different conjunctions (i.e. so, and, although and full stop/no conjunction); second, whether and how the influence of conjunctions on referential uncertainty is modulated by the verb-based semantic bias (NP1- vs. NP2-biased verbs); and third, and more importantly, whether native and non-native speakers will differ in the use of the conjunction and verb information to resolve referential uncertainty. Although both native and non-native speakers can benefit from coherence conjunctions during discourse processing, native speakers are more likely to automatically infer the missing link; non-native speakers, on the other hand, may depend more on cohesive devices (e.g. conjunctions) to facilitate comprehension by regarding them as an additional orienting tool to interpret the message and to reduce the probability of misinterpretation (Zorana, 2013).

Table 1 illustrates the experimental conditions. Items contained an NP1-biased (5a–5d) or NP2-biased verb (6a–6d) (with the bias based on the conjunction because, as determined by a pretest described below). Each sentence consisted of two clauses which were linked by the conjunction so, and, or although, or had no conjunction (full stop condition). The initial clause described an event or behavior involving two protagonists of the same gender introduced by proper names (e.g. Lily, Nina), which acted as subject or object, respectively. The second clause started with a conjunction (except in (d)), which was immediately followed by a pronoun (he or she) that could refer to either of the aforementioned protagonists (or a third entity). This allowed us to test to what extent the referential preference could be modulated by verb bias and the type of conjunction.

Table 1.

Exemplar sentences with NP1-biased verbs from (5a) to (5d) and NP2-biased verbs from (6a) to (6d).

Condition	Examples
so	(5a) Lily disappointed Nina, so she quit the business. (6a) Lily suspected Nina, so she finished the task alone.
and	(5b) Lily disappointed Nina, and she quit the business. (6b) Lily suspected Nina, and she finished the task alone.
although	(5c) Lily disappointed Nina, although she quit the business. (6c) Lily suspected Nina, although she finished the task alone.
full stop	(5d) Lily disappointed Nina. she quit the business. (6d) Lily suspected Nina. she finished the task alone.

Given that quite a number of studies have found an Nref in response to referential ambiguity (e.g. Nieuwland, 2014; Nieuwland and Van Berkum, 2006, 2008a, 2008b), we expect to observe Nref responses for situations where referential uncertainty occurred. In particular, the size of Nref effect could be related to the degree of referential uncertainty: the more ambiguous a referential pronoun is, the more effort will be required to interpret it, and the larger the Nref responses will be (Boudewyn et al., 2015). Besides Nref, based on existing relevant studies (e.g. Xu et al., 2017), we also predict some early positive modulations of the ERP elicited by the pronoun, as the attentional requirements for the reader could be different across conditions.

Similar to so, the conjunction and can also indicate a cause-consequence relation in addition to additive relations (Halliday and Hasan, 1976; Knott and Sanders, 1988; Martin, 1992; Sanders et al., 1992). However, as addressed previously, although both so and and encode cause-consequence relations, and differs from so in a number of aspects, it is thus not clear whether L1 and L2 readers will be sensitive to these differences. Different from the additive and causal relations, however, processing a concessive relation (e.g. although) requires additional computing processes (e.g. computing negation; Verhagen, 2000), which would probably increase the difficulty of constructing a coherent discourse representation. Readers may have fewer resources left to keep two referential representations in the although condition and may therefore encounter more referential uncertainty, leading to a larger Nref. Alternatively, it is also possible that, when fewer resources are left to maintain two referents in their discourse representation, readers might then quickly commit to one referential interpretation to reduce difficulty (Karimi and Ferreira, 2016). This interpretation will predict less referential uncertainty, and a smaller Nref for although versus the other conditions.

With regard to the full stop condition, given that conjunctions can provide explicit signals about how two propositions could be related, its absence may disturb discourse coherence, which in turn could lead to more referential uncertainty. Thus, relative to these conditions with explicit conjunctions, the full stop condition may give rise to the largest Nref response.

As for the verb semantics, given that previous studies (Ehrlich, 1980; Xu et al., 2016, 2017) have found that the preferential interpretation of a pronoun remains unchanged in overtly conjoined causal sentences (e.g. Tom praised Lisa, because he …) or additive sentences (e.g. Steve blamed Frank and he spilt the coffee; Ehrlich, 1980), irrespective of the direction of verb bias (NP1 vs. NP2), we expect no significant influence of verb type on the ERP responses to the ambiguous pronoun in so- and and-linked sentences. Also for concessive sentences, we do not expect an effect of verb-bias: as addressed above, processing concessive relation requires additional resources. The impact of verb-based semantics, under such circumstances, could be greatly weakened or even disappear. Finally, for full stop sentences, there are two possibilities. First, the Nref responses elicited by the ambiguous pronoun may fail to show difference between the NP1- and NP2-biased conditions, similar to previous ERP findings in the overtly connected causal relations (Xu et al., 2016, 2017). The other possibility is that the Nref response elicited by the ambiguous pronoun in the NP2-biased condition would be larger than that in NP1-biased condition, given that the superficial cues such as first mention/subject preference effect would be more prominent when overt devices are not available, and in particular, the structural preference effect is consistent with the verb bias in NP1-biased condition but not in NP2-biased condition.

Finally, for the between-group difference, L2 learners may show larger Nrefs than L1 readers when sentences are connected by more complex conjunctions like although or when no overt coherence devices are presented (full stop condition), because second language learners may have more difficulties maintaining two referential representations in sentences with complex coherence relations or without any overt coherence markers. Alternatively, it is also possible that second language learners may quickly commit to one referent. If this is the case, a smaller rather than larger Nref should be observed.

II Method

1 Participants

Twenty-six advanced Chinese–English L2 learners and 23 English native speakers took part in the ERP experiment. Data from two L1 speakers were removed because of too many EEG artifacts. All participants had normal or corrected-to-normal vision and none of them had history of neurological or psychiatric impairments. Each participant gave informed consent before the formal experiment and was compensated for participation.

a Chinese–English L2 speakers

All L2 participants were proficient Chinese–English learners. To rule out the potential influence of participants’ English language proficiency on the experiment (e.g. not being able to understand the sentences), two individual factors (English learning background and vocabulary size) were considered when selecting eligible participants. First, 35 English major postgraduates, who had passed the TEM8 (Test for English Majors-Level 8) and had on average 10 years of experience in learning English, were recruited from Nanjing Normal University. All participants were instructed to complete the LexTALE test, which is a standardized lexical test (e.g. Kristin and Mirjam, 2012; Lemhöfer and Broersma, 2012). The test includes 60 trials, each of which consists of a string of letters. For each trial, the participants were asked to decide whether the letter string was an existing English word, regardless of whether or not they knew its exact meaning (in total, 40 trials were existing English words, while 20 trials were non-words). Then each participant’s LexTALE score was calculated according to a fixed formula: ((number of correct words/40*100) + (number of correct non-words/20*100))/2. Based on this, 26 participants with scores higher than 75 (78.7 on average; 7 males, age ranging from 22 to 25 years with mean age of 23 years) were selected for the ERP experiment.

b English native speakers

The native speakers of English came to China as college exchange students or high school English teachers. Eleven were from the USA, five from the UK, three from Canada, one from Zimbabwe, and one from Kenya.

2 Design and materials

As indicated in Table 1, this study was a 4 × 2 design, which manipulated the coherence relations between the two clauses (so, and, although, full stop) and the implicit causality of the verb in the first clause (NP1- or NP2-biased verbs). Materials were constructed and evaluated on the basis of three separate pretests: a gender rating test, and two forced-choice tests to rate the implicit bias induced by the verbs and the conjunctions. Participants in these pretests were Chinese–English L2 speakers, drawn from the same population and with the English experience and the level of English proficiency as those L2 speakers who participated in the ERP experiment. A different group of participants was recruited for each of the three pretests.

The gender rating test aimed to examine whether the proper names used in the materials were typically male or female. Twenty participants rated 240 proper names (120 male names and 120 female names) on gender typicality, using a 7-point Likert scale with 1 corresponding to ‘most typically male’ and 7 to ‘most typically female’. As shown in Table 2, the average score of male names was close to 1 while the average score of female names was close to 7, suggesting that the names selected for the study were indeed considered typically male or typically female.

Table 2.

Mean scores and standard deviations in the gender rating test.

	Mean	SD
Male names	1.4	0.5
Female names	6.6	0.5

The aim of the second pretest was to select typical NP1- and NP2-biased verbs. Three-hundred English verbs (including 150 NP1-biased verbs and 150 NP2-biased verbs) were selected and adapted from previous research on verb-based implicit causality (Au, 1986; Brown and Fish, 1983; Garvey and Caramazza, 1974; Garvey et al., 1975; Greene and McKoon, 1995; Koornneef and Van Berkum, 2006; Stewart et al., 2000). Each item in the pretest contained the causal conjunction because, and was truncated after the ambiguous pronoun (e.g. Lily disappointed Nina because she …). Twenty-four participants were asked to identify which one of the two potential antecedents (e.g. Lily or Nina) the pronoun referred to, or to choose a third referent not mentioned in the context. On the basis of this pretest, 240 English verbs were selected, including 120 NP1-biased verbs (e.g. disappoint) and 120 NP2-biased verbs (e.g. punish). As shown in Table 3, the referential bias for these NP1-biased and NP2-biased verbs did not substantially differ from each other (73.9% NP1 referents for NP1-biased verbs vs. 76.8% NP2 referents for NP2-biased verbs, p > 0.5). The frequency (based on the Corpus of Contemporary American English) and the length of these two types of verbs were kept similar (frequency: t(119) = 0.44, p = 0.62; word length: t(119) = 1.54; p = 0.13).

Table 3.

Mean percentages and standard deviations of NP1 reference for NP1-biased verbs and of NP2 reference for NP2-biased verbs.

	Mean (%)	SD (%)
NP1-biased verbs	73.9	11.3
NP2-biased verbs	76.8	12.4

The aim of the third pretest was to see to what extent the pronoun resolution could be modulated by the verb bias and the conjunction. In this test, all experimental items (240 sentences, Latin squared over eight lists, 30 per condition per list) were truncated after the pronoun (e.g. Lily disappointed Nina, so she …). Twenty-three participants were each randomly assigned to one of the counterbalanced lists. As in the second pretest, participants were asked to indicate whether the pronoun referred to the first NP, the second NP or an unmentioned referent. Results are given in Table 4. The arcsine-transformed proportions of pronoun responses were analysed in an ANOVA with conjunction, verb type, and choice (NP1 vs. NP2 vs. a third referent) as within-participant variables.

Table 4.

Means (standard deviations) of the pronoun referential probability test in NP1- and NP2-biased sentences.

	NP1 (%)		NP2 (%)		A third person (%)
	NP1-bias	NP2-bias	NP1-bias	NP2-bias	NP1-bias	NP2-bias
so sentences	31.5 (6.5)	67.1 (5.5)	67.5 (6.7)	32.2 (5.2)	0.9 (1.9)	0.7 (1.7)
and sentences	34.5 (9.5)	66.1 (6.7)	64.4 (8.3)	33.2 (7.7)	1.2 (2.6)	0.7 (2.5)
although sentences	51.4 (4.5)	49.1 (3.6)	48.3 (4.4)	50.6 (3.9)	0.3 (1.4)	0.3 (1.0)
full stop sentences	44.3 (8.2)	29.3 (5.8)	34.3 (13.2)	49.1 (3.6)	21.4 (9.5)	21.6 (4.8)

Note. The top row is the response category, and the second row is the verb type.

There were a main effect of conjunction F(3,66) = 153.36, p < 0.001, and a significant two-way interaction between conjunction and referent type (NP1 vs. NP2 vs. a third referent), F(6,132) = 151.55. There was also a three-way interaction between conjunction, verb type, and referent type, F(6,132) = 112.3, p < 0.001. To resolve the three-way interaction, follow-up analyses were carried out in NP1- and NP2-biased sentences, separately.

The results from NP1-biased verbs indicate that readers preferentially interpret the ambiguous pronoun as referring to NP2 rather than to NP1 in both so and and conditions (Table 4), ps < 0.001. However, in contrast to so and and conditions, the full stop condition showed a reversed referential preference, that is, a referential preference for NP1 rather than NP2, t(22) = 2.98, p = 0.007. Interestingly, the percentage of out-context reference was much higher in the full stop condition than the other three conditions, ps < 0.001, suggesting that the lack of conjunctions significantly increases the possibilities of resolving the ambiguous pronoun towards an uncertain third person. More importantly, different from the other three conditions, pronouns in the although condition had no clear preference towards NP1 or NP2, t(22) = 1.68, p > 0.1 indicating that the presence of the concessive conjunction although leads to more referential uncertainty.

The results from NP2-biased verbs showed that, different from NP1-biased verbs, readers preferentially linked the ambiguous pronoun to NP1 rather than NP2 in both the so and and conditions, ps < 0.001. In contrast to the so and and conditions (see Table 4), however, the full stop condition showed a referential preference for NP2 rather than for NP1, t(22) = 10.48, p < 0.001. Similar to the results from NP1-biased verbs, the percentage of out-context reference was much higher in the full stop condition than in the other three conditions, ps < 0.001, and pronouns had no clear preference towards NP1 or NP2 in the although condition, t < 1, p > 0.3, indicating again that the presence of a concessive conjunction although increases referential uncertainty.

In the ERP study, the 240 critical stimulus items selected on the basis of the pretests (120 for each verb type) were distributed across 8 different testing lists according to a Latin square procedure, with each list containing an equal number of items (30 items) per condition. For each list, one half of the stimuli included only male protagonists while the other half included only female protagonists. In addition to the critical items, 60 two-clause filler items were formulated in order to encourage participants to read the sentences naturally, without adopting certain processing strategies. The 60 filler items were similar in construction to the experimental items and consisted of: 20 sentences in which the two clauses were connected by other conjunctions (e.g. but, because); 20 sentences with the first clause containing verbs without semantic bias, and 20 sentences with two protagonists of different biological genders in the main clause. The filler sentences were the same in each list. The order of the sentences was pseudo-randomized, with the restriction that no more than three consecutive sentences were of the same condition and no more than three consecutive sentences had comprehension questions with the same intended answer (e.g. YES response). Participants were randomly assigned to the testing lists.

3 Procedures

Participants were seated comfortably in a chair in a dimly-lit and sound-attenuated room, and were instructed to read every sentence attentively. Each trial began with a fixation point (‘+’) at the center of the screen for 500 ms, which was followed by a blank screen for 500 ms. Then the first clause containing two candidate antecedents was presented as a whole on the screen. After reading it, the participant pressed the space bar to initiate the second clause that was presented word by word at the center of the screen. Each word was presented for 400 ms followed by a 400 ms blank screen. After that, participants were required to answer a yes/no comprehension question by pressing one of two keys. None of the questions probed the resolution of the referential ambiguity (e.g. Did someone quit the business?). The questions were designed to probe participants’ understanding of the sentences. The assignment of left/right hand to yes/no response was counterbalanced across participants. Each participant saw a practice block of 20 sentences, which were similar in structure to the test stimuli. The stimuli were divided into five blocks; the participant had a break of about 5 minutes on average between each block. The study took about 2 hours per participant, including electrode preparation.

4 EEG recording and data analysis

EEG activity was recorded from 63 electrodes mounted in an elastic cap (Electro-cap International). The EEG was referenced online to the tip of the nose and rereferenced offline to the algebraic average activity measured in the left and right mastoids (TP9 and TP10). The vertical electrooculogram (VEOG) was monitored from an electrode located above the right eye and the horizontal electrooculogram (HEOG) from an electrode located at the outer canthus of the left eye. Electrode impedance was kept below 5 kΩ. EEG signals were filtered using a band-pass of 0.016–70 Hz, and digitized at a sampling rate of 500 Hz.

For each trial, the 800 ms ERP epoch was extracted, time-locked to the onset of pronoun in the second clause, with a prestimulus baseline of 200 ms. Trials with incorrect responses to the comprehension questions were eliminated from data analysis; the remaining trials were screened for drift artifacts. In the offline data analysis, trials with ocular artifacts were corrected using the independent component analysis approach (Jung et al., 1998; Makeig et al., 1996), and trials with absolute values greater than 60 µV were excluded from further analysis. For NP1-biased verbs, the mean number of trials included for EEG analysis was 27.6 for so sentences, 26.3 for and sentences, 25.4 for although sentences, and 25.8 for full stop sentences. For NP2-biased verbs, the mean number of trials included for EEG analysis was 27.0 for so sentences, 25.9 for and sentences, 24.6 for although sentences, and 24.3 for full stop sentences.

Based on visual inspection and relevant literature (e.g. Nieuwland, 2014; Nieuwland and Van Berkum, 2006, 2008a, 2008b; Van Berkum et al., 1999), the 150–250 ms and the 300–800 ms time windows were selected for statistical analysis. ANOVAs were conducted on mean ERP amplitudes in these time windows with conjunction (so vs. and vs. although vs. full stop), verb type (NP1-biased verbs vs. NP2-biased verbs) and topographical factors as within-participant variables and language group as a between-participant variable. For the midline analysis, the topographical factor was electrode with three levels: anterior (FM: Fz and FCz), central (CM: Cz and CPz), and posterior (PM: Pz and POz). For the lateral analysis, the topographic factors were region (three levels: anterior vs. central vs. posterior) and hemisphere (two levels: left vs. right). The region and hemisphere factors were crossed, resulting in six regions of interest (ROIs): left frontal (FL: F1, F3, F5, FC1, FC3, FC5), left central (CL: C1, C3, C5, CP1, CP3, CP5), left posterior (PL: P1, P3, P5, PO3, PO7), right frontal (FR: F2, F4, F6, FC2, FC4, FC6), right central (CR: C2, C4, C6, CP2, CP4, CP6), and right posterior (PR: P2, P4, P6, PO4, PO8). Mean amplitudes over electrodes in each region of interest were entered into ANOVAs. Bonferroni correction was used for multiple comparisons. The Greenhouse-Geisser correction was applied when appropriate (Geisser and Greenhouse, 1959).

III Results

1 Behavioral results

As shown in Table 5, the mean proportion of correct responses to probe questions for each experimental condition was more than 90%, indicating that participants could understand each sentence well. Repeated-measures ANOVA, with verb type and conjunction as within-participant variables and group as a between-participant variable, showed a significant main effect of conjunction, F(3, 141) = 8.20, p < 0.001, η² = 0.149. Follow-up pairwise comparisons showed lower accuracy for the although condition relative to the so condition (p < 0.01, d = 0.794) and the and condition (p < 0.01, d = 0.623). Moreover, there was also a significant two-way interaction between verb type and group, F(1,47) = 5.98, p = 0.018, η² = 0.113. Further analyses showed that NP2-biased sentences had higher accuracy than NP1-biased sentences in the L2 group, t(25) = 2.97, p = 0.007, d = 0.587, whereas there was no difference of verb type in the L1 group.

Table 5.

Mean percentage accuracy for each condition in sentence comprehension in NP1- and NP2- biased sentences for the L2 and the L1 groups (standard deviations in parentheses).

	L2		L1
	NP1	NP2	NP1	NP2
so sentences	96.7 (3.7)	97.8 (2.7)	93.2 (6.0)	92.8 (5.7)
and sentences	95.6 (2.5)	96.8 (4.2)	91.7 (4.7)	92.6 (6.0)
although sentences	91.7 (4.6)	93.2 (6.1)	90.3 (6.1)	92.0 (5.8)
full stop sentences	94.1 (5.2)	94.7 (4.1)	94.4 (5.0)	90.0 (8.3)

2 Electrophysiological results

The grand-average ERPs, time-locked to the ambiguous pronoun he or she in non-native (L2) and native (L1) speakers are shown in Figure 1 and Figure 2, respectively.¹ The although sentences appeared to elicit a larger sustained negativity (Nref) than so, and, and full stop sentences in both L2 and L1 groups (see Figure 1 and 2). Visual inspection also suggested that larger Nrefs were evoked in L2 than L1 groups for sentences connected by so, although or sentences with no overt conjunctions (full stop), whereas there was almost no group difference in sentences conjoined by and. Moreover, an increased Nref was found for pronouns in NP2-biased sentences than in NP1-biased sentences when sentences were linked by the conjunction so or when there were no conjunctions linking the sentences (see Figure 3). In addition to the Nrefs (300–800 ms), there seem to be some early ERP activities (150–250 ms). These findings were confirmed by the following statistical analyses.

Figure 1.

The grand-average ERPs, averaged over all electrodes in each ROI, time-locked to the critical pronouns for the so, and, although and the full stop conditions, collapsed over verb type, in L2 group (FL = left frontal ROI, FM = middle frontal ROI, FR = right frontal ROI, CL = left central ROI, CM = middle central ROI, CR = right central ROI, PL = left posterior ROI, PM = middle posterior ROI, PR = right posterior ROI).

Figure 2.

The grand-average ERPs, averaged over all electrodes in each ROI, time-locked to the critical pronouns for the so, and, although and the full stop conditions, collapsed over verb type, in L1 group (FL = left frontal ROI, FM = middle frontal ROI, FR = right frontal ROI, CL = left central ROI, CM = middle central ROI, CR = right central ROI, PL = left posterior ROI, PM = middle posterior ROI, PR = right posterior ROI).

Figure 3.

Topographic maps for difference (at the pronoun position) between the NP2 verbs and the NP1 verbs in the so condition, the and condition, the although condition, and the full stop condition collapsed over groups.

a 150–250 ms time window

Mean amplitudes recorded at the lateral and midline electrodes were analysed separately. In the early time window, there was a significant two-way interaction between conjunction and group (in the lateral F(3,135) = 5.49, p = 0.004, η² = 0.109, and in the midline F(3,135) = 6.02, p = 0.002, η² = 0.118). Further analyses showed that for the L2 group, larger positivities were evoked by so and and conditions relative to the although condition ps < 0.001, or the full stop condition ps < 0.001 for both lateral and midline analyses. For the L1 group, however, neither the main effect of conjunction nor the interactions between conjunction and other factors were significant, ps > 0.1.

b 300–800 ms time window

At lateral sites, the repeated measures ANOVA showed a significant four-way interaction between verb type, conjunction, region, and hemisphere, F(6, 270) = 3.13, p = 0.013, η² = 0.065, though the three-way interaction between verb type, conjunction, and region was also significant, F(6,270) = 2.43, p = 0.05, η² = 0.051. Importantly, there was also a three-way interaction between conjunction, region and group F(6,270) = 2.90, p = 0.05, η² = 0.06. At midline sites, conjunction interacted with region, F(6, 270) = 13.92, p < 0.001, η² = 0.236, and group, F(1,45) = 2.97, p = 0.048, η² = 0.062.

To measure the effect of conjunction on referential uncertainty, the four-way interaction was resolved according to the verb type. For NP1-biased verbs, referential uncertainty evoked a larger sustained negativity for the although condition relative to the so condition (left-frontal/central and the right-frontal/central, ps < 0.001, ds > 0.8), the and condition (left-frontal, p < 0.001, d = 0.65, left-central, p < 0.001, d = 0.71, and the right-central, p = 0.016, d = 0.56), and the full stop condition (left-frontal/central and the right-frontal/central, ps < 0.001, ds > 0.8). Moreover, the full stop condition evoked a less negative response (i.e. larger positivity) than the and condition (right-frontal, p = 0.043, d = 0.47). For NP2-biased verbs, similarly, post-hoc comparisons showed a larger sustained negativity for the although condition relative to the so condition (left-frontal p < 0.001, d = 0.71, left central p = 0.014, d = 0.43, right-frontal, p < 0.001, d = 0.49), the and condition (left-frontal p < 0.001, d = 0.70, left central p = 0.05, d = 0.38, right-frontal p = 0.008, d = 0.60, and right central p = 0.033, d = 0.41), and the full stop condition (left-frontal/central, ps < 0.001, ds > 0.7, and right central p = 0.018, d = 0.48). Moreover, the full stop condition evoked a reduced negativity compared to the and condition (left-central p = 0.017, d = 0.44) and the so condition (marginally in the left-central region, p = 0.09, d = 0.42). Over the midline sites, follow-up analyses to resolve the interaction between conjunction and region showed a larger sustained negativity for the although condition relative to the so condition (frontal, p < 0.001, d = 0.91, central, p < 0.001, d = 0.66), the and condition (frontal p < 0.001, d = 0.67, and central, p = 0.006, d = 0.48), and the full stop condition (frontal, p < 0.001, d = 0.92, central, p < 0.001, d = 0.68). In addition, the full stop condition evoked a less negative response than the and condition in frontal region p = 0.05, d = 0.45.

As for verb type, the resolution of the four-way interaction showed that NP2-biased verbs resulted in larger Nrefs than that in NP1-biased verbs for the so condition (left-central, F(1,45) = 4.11, p = 0.49, η² = 0.084, right frontal, F(1,45) = 6.22, p = 0.016, η² = 0.121, right-central, F(1,45) = 9.28, p = 0.004, η² = 0.171, and right-posterior, F(1,45) = 8.49, p = 0.006, η² = 0.159), and the full stop condition (right-frontal, F(1,45) = 13.62, p < 0.001, η² = 0.232, and right-central, F(1,45) = 4.21, p = 0.046, η² = 0.085). However, there was no effect of verb type for the although condition ps > 0.4, or the and condition ps > 0.1 (see Figure 3).

Finally, to measure the group effect on different conjunctions, follow-up analyses were carried out to resolve the three-way interaction between conjunction, region and group. Over lateral sites, the L2 group showed larger Nrefs than the L1 group in the so condition (central, F(1,45) = 8.18, p = 0.006, η² = 0.154), the although condition (central, F(1,45) = 4.84, p = 0.033, η² = 0.097), and the full stop condition (posterior, F(1,45) = 13.66, p < 0.001, η² = 0.233, and marginally significant in central F(1,45) = 2.97, p = 0.092, η² = 0.062). However, there was no significant group difference across regions with respect to the and condition, ps > 0.3. Similarly, at midline sites, the L2 group showed larger Nrefs than the L1 group in the so condition (central, F(1,45) = 3.70, p = 0.061,η² = 0.076, posterior F(1,45) = 7.03, p = 0.011, η² = 0.135), the although condition (central F(1,45) = 2.98, p = 0.091, η² = 0.062, posterior F(1,45) = 4.78, p = 0.034, η² = 0.096), and the full-stop condition (central, F(1,45) = 5.60, p = 0.022, η² = 0.111, posterior F(1,45) = 13.51, p < 0.001, η² = 0.231). Again, there was no significant group difference in the and condition, ps > 0.4.

IV Discussion

This study investigated how the resolution of referential uncertainty is influenced by conjunctions and verb-based semantic bias. The behavioral results showed that the accuracy of probe questions was much lower for the although condition than for the so condition or the and condition. Furthermore, questions were more correctly answered in NP2-biased sentences than in NP1-biased sentences for the L2 group, but not the L1 group. This indicates that understanding although sentences is more difficult than understanding so and and sentences irrespective of language proficiency, and that L2 learners have more difficulties understanding NP1-biased sentences than NP2-biased sentences. The modulation of verb-bias on comprehension accuracy in L2 but not L1 is probably related to the potential influence of the learners’ L1 on L2. Since Chinese has only a limited set of NP1-biased psychological verbs (e.g. disappoint) as compared with English, Chinese favors periphrastic causatives (e.g. Lily shi Nina shiwang / ‘Lily makes Nina disappointed’) rather than lexical causatives (e.g. disappoint) to express causation (Cheng and Almor, 2017). These cross-language variations could make Chinese L2 learners of English less effective than English native speakers at applying implicit causality information to language comprehension.

With regard to the ERP findings, the although condition evoked larger sustained negativities than the other three conditions in both L1 and L2 readers irrespective of the semantic bias of the verb. Moreover, larger Nrefs were found for pronouns in the L2 group than in the L1 group in the so, the although, and the full stop conditions, whereas no group difference was observed in the and condition. As for verb type, pronouns in NP2-biased sentences elicited larger Nrefs than those in NP1-biased sentences in the so condition and the full stop condition, whereas there was no clear effect of verb type in the and condition or the although condition. Overall, although we found a cross-group difference concerning verb bias in behavioral performance, we did not find such modulation on ERP data. The divergences seem to suggest that the cross-language differences in expressing implicit causality (especially the NP2-biased causation) only affect L2 learners’ untimed offline judgments but not the real-time pronoun interpretation.

With regard to early effects, the although and the full stop conditions evoked a smaller early positive response than the so or the and condition in L2 group. Previous studies concerning pronoun resolution (Xu, 2015; Xu et al., 2016, 2017) have shown that the early positivity (P2) evoked by pronouns could be related to an increased attentional engagement in activating the pronoun’s potential referents. The reduced early positivity in the although and the full stop conditions relative to the other two conditions might suggest that, for L2 learners, less attentional resources are needed to activate the referents when attention has been already attracted by the unexpected coherence marker or the absence of any coherence markers. However, it is also plausible that the reduced positivity evoked in the although condition is the early onset of a sustained Nref component, though this explanation does hold for the full stop condition.

1 Conjunctions and referential uncertainty

Previous studies have shown that ambiguous pronouns resulted in greater Nref responses compared with unambiguous ones, because ambiguous pronouns cause the comprehender to entertain two referential interpretations whereas the unambiguous ones do not (Nieuwland, 2014; Nieuwland, and Van Berkum, 2006; Van Berkum et al., 2004). In this study, an enlarged Nref response was observed in the although condition as compared with the other three conditions (so, and, and full stop conditions), irrespective of verb type (NP1 bias vs. NP2 bias) and language background (L1 vs. L2), probably because the presence of the concessive marker although blocks a presumed causal assumption induced by the biasing verbs and thus leads to more uncertainty as to which referential candidate will be retrieved as the referent of the pronoun. This interpretation is consistent with the forced-choice test, which showed that the two referential candidates (NP1 vs. NP2) were more similar in probability in the although condition (for NP1 verbs: 51.4% vs. 48.3%, and for NP2 verbs: 49.1% vs. 50.6%) compared with the so or the and condition. Alternatively, the increased uncertainty in the although condition could also be associated with differences in terms of cognitive complexity, as indicated by the behavioral results (although condition has lower accuracy relative to so condition or and condition) and by taxonomic analyses of coherence relations (Sanders, 2005; Sanders et al., 1992, 1993). Relative to the other relations, processing concessive relations involves additional computations, that is, the negation of a presumed causal assumption and the reconstruction of an expected discourse relation (e.g. Iten, 1998; König, 1985; Kӧhne and Demberg, 2012; Kӧnig and Siemund, 2000; Verhagen, 2000). This reconstruction is resource consuming (Xu et al., 2015, 2018), leading to fewer resources left for maintaining two referential interpretations. Participants, under such circumstances, might quickly commit to one referential interpretation to reduce difficulty (Karimi and Ferreira, 2016). This hypothesis, however, is not supported by the ERP findings, nor is it supported by the off-line pretest, since the forced-choice test showed that NP1 and NP2 were equally accessible for the pronoun in although sentences. Therefore, although concessive relation is cognitively more complex, this does not mean that processing concessive relation will necessarily reduce the degree of ambiguity.

Compared with and or so (NP2-biased) sentences, the full stop sentences resulted in less negative activity in the central regions. One possible explanation for this unexpected deflection could be related to the probability of entertaining one versus two referential interpretations. Specifically, there are two sources of information when so or and is present (i.e. conjunctions and verb semantics), and these two sources go in opposite directions: NP1-biased verbs that are paired with these two conjunctions result in more NP2 interpretations, and NP2-biased verbs result in more NP1 interpretations. However, only one source of information is available in the full stop condition (i.e. verb semantics). It is more likely that both referential interpretations are entertained during online processing of the ambiguous pronoun when there are two opposite-going sources of information than when there is only one cue for referential resolution. This interpretation is also in line with the increased context-external referential interpretations in the full stop condition, namely, the probability that both referential interpretations are entertained is reduced when a third, novel referential candidate is considered as the referent. Yet another possible explanation could be related with the different statistical environment for the full stop sentences and the other types of sentences, since the sentences which had conjunctions are three times more frequent in the experimental context than those that did not (i.e. full stop sentences). This could lead to an unexpectedness effect immediately upon encountering the pronoun and therefore gives rise to a reduced negativity (i.e. larger positivity). Thus, the pattern of results in the full stop condition might have been caused by probability of entertaining one versus more than one referential interpretation, though at the same time we cannot rule out the possible influence of the experimental environment.

2 Verb semantic and referential uncertainty

There was an effect of verb type in the full stop condition and the so condition, with larger Nref responses for pronouns in the NP2-biased sentences than in NP1-biased sentences. It is plausible that without overt conjunctions, readers are more prone to resort to the superficial cues such as syntactic/structural constraints, to resolve the ambiguous pronoun. For example, they may assign the pronoun to the first-mentioned entity/the superficial subject (the ‘first mentioning preference’ effect; Gernsbacher and Hargreaves, 1988; Järvikivi, et al., 2005). This first-mention/subject preference effect, however, is incompatible with the direction of the bias induced by the NP2-biased verbs, though it is compatible with that induced by the NP1-biased verbs. As a consequence, the force of referential bias induced by NP2 verbs was attenuated and referential uncertainty, in turn, was increased, resulting in an enlarged Nref response. This interpretation, however, is difficult to explain the similar results found in so-linked sentences. The significant effect of verb type on ERP responses in the so sentences is inconsistent with our hypothesis; here we only offer some tentative explanations. As indicated by previous studies (e.g. Fukumura and van Gompel, 2010; Stevenson et al., 2000), the presence of a conjunction so typically biases the pronoun to refer to the thematic role associated with the consequences, as so directs readers’ attention to the endpoint of an action or a state. Indeed, the forced-choice test showed a clear object (NP2) preference for NP1-biased verbs but a subject (NP1) preference for NP2-biased verbs. It is noteworthy that although the off-line results showed that the percentage of object reference is comparable to that of subject reference, these two cases may not result in equal size of referential bias in the real-time processing. The strength of referential bias for NP1 referent (induced by NP2-biased verbs) may not be as strong as that for NP2 referent (induced by NP1-biased verbs), because relative to the first entity (NP1/agent), the second one (NP2/patient) is more likely to be interpreted as the endpoint (or consequence) of an event when the sentence is incrementally presented. Therefore, more referential uncertainty was experienced in NP2-biased sentences than in NP1-biased sentences. Further research is required to elucidate how different conjunctions interact with semantic factors to affect referential ambiguity, especially those involving unambiguous pronouns (with different gender antecedents).

3 L2/L1 differences in processing conjunctions and referential relations

The L2 group showed larger Nrefs than the L1 group in so, although, and full stop sentences except the and sentences. The real-time processing differences could be related with different cognitive complexity underlying these different coherence relations. According to the cognitive complexity account (Goldman and Murray, 1992; Sanders, 2005; Sanders et al., 1992; Spooren and Sanders, 2008), the additive relation is considered to be less complex than causal or adversative relations, because additive relation only involves additivity whereas causal/concessive relations involve causality (i.e. causal relation) or negative causality (i.e. the negation of a causal relation) in addition to additivity. The differences in cognitive complexity are consistent with the development studies showing that additive relations are acquired before causal and adversative relations (Bloom et al., 1980; Clark, 2003; Spooren and Sanders, 2008). Second language learners may encounter greater difficulties than native speakers to entertain two referential representations when they encounter sentences connected by cognitively more complex causal or adversative relations. However, they can be as efficient as native speakers at maintaining both referential representations in the more basic additive relation. Furthermore, the absence of an overt coherence marker in full stop sentences would exert different influence on native and non-native speakers: L2 learners might be more dependent on coherence markers than L1 speakers and therefore they may encounter more referential ambiguities than L1 speakers (Degand and Sanders, 2002; Geva, 1986, 1992; Goldman and Murray, 1992).

The cross-group ERP differences concerning processing different coherence conjunctions are consistent with the existing L2 ERP literature on processing simple and complex dependencies. For example, it has demonstrated that L2 learners can achieve a native-like level in the processing of local dependencies such as subject–verb agreement (Ojima et al., 2005), whereas L2 learners have greater difficulties to process nonlocal dependencies such as wh-question structures (Felser and Roberts, 2007; Marinis et al., 2005). One interpretation for this dissociation is the resource limitation hypothesis (Clahsen and Felser, 2006). L2 learners are more susceptible to resource limitation than L1 speakers, especially for complex sentences, as processing complex sentences makes greater demands on working memory (Clahsen and Felser, 2006).

V Conclusions

To conclude, by comparing the Nref responses elicited by ambiguous pronouns embedded in sentences connected by different conjunctions, we found that the although condition evoked larger sustained negativities than the other three conditions in both L1 and L2 readers, irrespective of whether NP1- or NP2-biasd verbs were adopted. However, despite these similarities, non-native speakers did not show the same level of sensitivity as native speakers to those conjunctions which are cognitively more complex (e.g. so, although). Moreover, increased referential uncertainty was generated by pronouns after NP2-biased verbs than after NP1-biased verbs when sentences were linked by the conjunction so or when there were no conjunctions linking the sentences. These findings indicate that different conjunctions exert different modulating effects on resolving referential uncertainty, and relative to native speakers, non-native speakers are more likely to encounter difficulties and therefore produce more referential uncertainty when the sentences are conjoined by conjunctions with more complex semantics.

Footnotes

Appendix 1

Appendix 2

Appendix 3 Acknowledgements

We thank Edith Kaan, Stephen Politzer-Ahles and five anonymous reviewers for their suggestions and comments concerning earlier versions of the manuscript.

Declaration of Conflicting Interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received the following financial support for the research, authorship, and/or publication of this article: This study was supported by grants from the Natural Science Foundation of China (31300929), the Natural Science Foundation of the Higher Education Institutions of Jiangsu Province (16KJD180003), and the project supporting the top-100 young leading researchers in Nanjing Normal University, China. It was also supported by the ‘Blue Project’ for outstanding young teachers of Nanjing Normal University as well as by the Academic Development Priority Program of Jiangsu Higher Education Institutions awarded to School of Foreign Languages and Cultures, Nanjing Normal University.

Notes

References

Alexander

(1980) A learning to learn perspective on reading in a foreign language. System 8: 113–19.

(1986) A verb is worth a thousand words: The causes and consequences of interpersonal events implicit in language. Journal of Memory and Language 25: 104–22.

Bloom

Lahey

Hood

Lifter

Fiess

(1980) Complex sentences: Acquisition of syntactic connectives and the semantic relations they encode. Journal of Child Language 7: 235–61.

Boudewyn

Long

Traxler

et al . (2015) Sensitivity to referential ambiguity in discourse: The role of attention, working memory, and verbal ability. Journal of Cognitive Neuroscience 27: 2309–23.

Brown

Fish

(1983) The psychological causality implicit in language. Cognition: International Journal of Cognition Science 14: 237–73.

Cheng

Almor

(2017) The effect of implicit causality and consequentiality on nonnative pronoun resolution. Applied Psycholinguistics 38: 1–26.

Chung

JSL

(2000) Signals and reading comprehension-theory and practice. System 28: 247–59.

Clahsen

Felser

(2006) How native-like is non-native language processing? Trends in Cognitive Sciences 10: 564–70.

Clark

(2003) First language acquisition. Cambridge: Cambridge University Press.

10.

Cowles

Walenski

Kluender

(2007) Linguistic and cognitive prominence in anaphor resolution: Topic, contrastive focus and pronouns. Topoi 26: 3–18.

11.

Cozijn

Noordman

LGM

Vonk

(2011) Propositional integration and world-knowledge inference: Processes in understanding ‘because’ sentences. Discourse Processes 48: 475–500.

12.

Crawley

Stevenson

Kleinman

(1990) The use of heuristic strategies in the interpretation of pronouns. Journal of Psycholinguistic Research 19: 245–64.

13.

Crosson

Lesaux

(2013) Does knowledge of connectives play a unique role in the reading comprehension of English learners and English-only students? Journal of Research in Reading 36: 241–60.

14.

Degand

Sanders

(2002) The impact of relational markers on expository text comprehension in L1 and L2. Reading and Writing: An Interdisciplinary Journal 15: 739–57.

15.

Demel

(1990) The relationship between overall reading comprehension and comprehension of coreferential ties for second language readers of English. TESOL Quarterly 24: 267–92.

16.

Droop

Verhoeven

(2003) Language proficiency and reading ability in first- and second-language learners. Reading Research Quarterly 38: 78–103.

17.

Ehrlich

(1980) Comprehension of pronouns. The Quarterly Journal of Experimental Psychology 63: 3–15.

18.

Ellen

William

Alfonso

(1978) Parallel function strategy in pronoun assignment. Cognition 6: 117–33.

19.

Eshagh

(2001) The role of conjunctions in reading comprehension and in textual sequence for Iranian Students. The International Journal of American Linguistics 5: 1–16.

20.

Evers-Vermeul

Sanders

(2009) The emergence of Dutch connectives: How cumulative cognitive complexity explains the order of acquisition. Journal of Child Language 36: 829–54.

21.

Felser

Roberts

(2007) Processing wh-dependencies in a second language: A cross-modal priming study. Second Language Research 23: 9–36.

22.

Fukumura

van Gompel

RPG

(2010) Choosing anaphoric expressions: Do people take into account likelihood of reference? Journal of Memory and Language 62: 52–66.

23.

Garvey

Caramazza

(1974) Implicit causality in verbs. Linguistic Inquiry 5: 459–64.

24.

Garvey

Caramazza

Yates

(1975) Factors influencing assignment of pronoun antecedents. Cognition 3: 227–43.

25.

Geisser

Greenhouse

(1959) On methods in the analysis of profile data. Psychometrika 24: 95–112.

26.

Gernsbacher

Hargreaves

(1988) Accessing sentence participants: The advantage of first mention. Journal of Memory and Language 27: 699–717.

27.

Geva

(1986) Reading comprehension in a second language: The role of conjunctions. TESL Canada Journal 3: 85–96.

28.

Geva

(1992) The role of conjunctions in L2 text comprehension. TESOL Quarterly 26: 731–47.

29.

Goldman

Murray

(1992) Knowledge of connectors as cohesion devices in text: A comparative study of native-English and English-as-a-second-language speakers. Journal of Educational Psychology 84: 504–19.

30.

Greene

McKoon

(1995) Telling something we can’t know: Experimental approaches to verbs exhibiting implicit causality. Psychological Science 6: 262–70.

31.

Haberlandt

(1982) Readers expectations in text comprehension. In: Le Ny

Kintsch

(eds) Language and language comprehension (pp. 239–49). Amsterdam: North-Holland.

32.

Halliday

MAK

Hasan

(1976) Cohesion in English. London: Longman.

33.

Hobbs

(1979) Coherence and coreference. Cognitive Science 3: 67–90.

34.

Hoek

Zufferey

(2015) Factors influencing the implication of discourse relations across languages. Proceedings 11th Joint ACL-ISO workshop on Interoperable Semantic Annotation (ISA-11) (pp. 39–45). Tilburg: TiCC, Tilburg Center for Cognition and Communication.

35.

Horiba

(1996) Comprehension processes in L2 reading: Language competence, textual coherence, and inferences. Studies in Second Language Acquisition 18: 433–73.

36.

Iten

(1998) The meaning of although: A relevance the theoretic account. UCL Working Papers in Linguistic 10: 81–108.

37.

Järvikivi

Van Gompel

RPG

Hyӧnä

Bertram

(2005) Ambiguous pronoun resolution: Contrasting the first mention and subject-reference accounts. Psychological Science 16: 260–64.

38.

Jung

(2003) The role of discourse signaling cues in second language listening comprehension. The Modern Language Journal 87: 562–77.

39.

Jung

Humphries

Lee

et al . (1998) Extended ICA removes artifacts from electroencephalographic recordings. In: Jordan

Kearns

Solla

(eds) Advances in Neural Information Processing Systems 10 (pp. 894–900). Cambridge, MA: MIT Press.

40.

Kaiser

(2011) Focusing on pronouns: Consequences of subjecthood, pronominalisation, and contrastive focus. Language Cognitive Processes 26: 1625–66.

41.

Karimi

Ferreira

(2016) Informativity renders a referent more accessible: Evidence from eyetracking. Psychonomic Bulletin and Review 23: 507–25.

42.

Kehler

(2002) Coherence, reference, and the theory of grammar. Stanford, CA: CSLI.

43.

Kehler

Kertz

Rohde

Elman

(2008) Coherence and coreference revisited. Journal of Semantics 25: 1–44.

44.

Kintsch

(1988) The use of knowledge in discourse processing: A construction-integration model. Psychological Review 95: 163–82.

45.

Kintsch

van Dijk

(1978) Towards a model of text comprehension and production. Psychological Review 85: 363–94.

46.

Knott

Sanders

(1988) The classification of coherence relations and their linguistic markers: An exploration of two languages. Journal of Pragmatics 30: 135–75.

47.

Kӧhne

Demberg

(2012) Incremental and predictive discourse processing based on causal and concessive discourse markers: A visual world study. Abstract presented at the 25th Annual CUNY Conference on Human Sentence Processing, New York.

48.

König

(1985) On the history of concessive connectives in English: Diachronic and synchronic evidence. Lingua 66: 1–19.

49.

Kӧnig

Siemund

(2000) Causal and concessive clauses: Formal and semantic relations. In: Couper-Kuhlen

Kortmann

(eds) Cause – condition – concession – contrast (pp. 341–60). Berlin: Mouton de Gruyter.

50.

Koornneef

Sanders

TJM

(2013) Establishing coherence relations in discourse: The influence of implicit causality and connectives on pronoun resolution. Language and Cognitive Processes 28: 1169–1206.

51.

Koornneef

Van Berkum

JJA

(2006) On the use of verb-based implicit causality in sentence comprehension: Evidence from self-paced reading and eye tracking. Journal of Memory and Language 54: 445–65.

52.

Kristin

Mirjam

(2012) Introducing LexTALE: A quick and valid lexical test for advanced learners of English. Behavior Research 44: 325–43.

53.

Lemhöfer

Broersma

(2012) Introducing LexTALE: A quick and valid lexical test for advanced learners of English. Behavior Research Methods 44: 325–43.

54.

MacLean

d’Anglejan

(1986) Rational cloze and retrospection: Insights into first and second language reading comprehension. Canadian Modern Language Review 42: 814–26.

55.

Mak

Sanders

TJM

(2010) Incremental discourse processing: How coherence relations influence the resolution of pronouns. In: Everaert

Lentz

De Mulder

Nilsen

Zondervan

(eds) The linguistic enterprise: From knowledge of language to knowledge in linguistic (pp. 167–82). Amsterdam: John Benjamins.

56.

Mak

Sanders

TJM

(2013) The role of causality in discourse processing: Effects of expectation and coherence relations. Language and Cognitive Processes 28: 1414–37.

57.

Makeig

Bell

Jung

Sejnowski

(1996) Independent component analysis of electroencephalographic data. Advances in Neural Information Processing Systems 8: 145–51.

58.

Marinis

Roberts

Felser

Clahsen

(2005) Gaps in second language processing. Studies in Second Language Acquisition 27: 53–78.

59.

Martin

(1992) English text: System and structure. Amsterdam: Benjamins.

60.

Mauri

van der Auwera

(2012) Connectives. In: Jaszczolt

Allan

(eds) Cambridge handbook of pragmatics (pp. 377–401). Cambridge: Cambridge University Press.

61.

McClure

Steffensen

(1985) A study in the use of conjunctions across grades and ethnic groups. Research in the Teaching of English 19: 217–36.

62.

McDonald

MacWhinney

(1995) The time course of anaphor resolution: Effects of implicit verb causality and gender. Journal of Memory and Language 34: 543–66.

63.

Millis

Just

(1994) The influence of connectives on sentence comprehension. Journal of Memory and Language 33: 128–47.

64.

Murray

(1995) Logical connectives and local coherence. In: Lorch

O’Brien

(eds) Sources of coherence in reading (pp. 107–25). Hillsdale, NJ: Lawrence Erlbaum.

65.

Murray

(1997) Connectives and narrative text: The role of continuity. Memory and Cognition 25: 227–36.

66.

Nieuwland

(2014) ‘Who’s he?’ Event-related brain potentials and unbound pronouns. Journal of Memory and Language 76: 1–28.

67.

Nieuwland

Van Berkum

JJA

(2006) Individual difference and contextual bias in pronoun resolution: Evidence from ERPs. Brain Research 1118: 155–67.

68.

Nieuwland

Van Berkum

JJA

(2008a) The interplay between semantic and referential aspects of anaphoric noun phrase resolution: evidence from ERPs. Brain and Language 106: 119–31.

69.

Nieuwland

Van Berkum

JJA

(2008b) The neurocognition of referential ambiguity in language comprehension. Language Linguistic Compass 2: 603–30.

70.

Ojima

Nakata

Kakiqi

(2005) An ERP study of second language learning after childhood: Effects of proficiency. Journal of Cognitive Neuroscience 17: 1212–28.

71.

Oudega

(2011) How default is causality-by-default? Unpublished MA thesis, Utrecht University, Utrecht, Netherlands.

72.

Ozono

(2002) An empirical study on the effects of logical connectives in EFL reading comprehension. Annual Review of English Language Education in Japan 13: 61–70.

73.

Ozono

Ito

(2003) Logical connectives as catalysts for interactive L2 reading. System 31: 283–97.

74.

Pan

Felser

(2011) Referential context effects in L2 ambiguity resolution: Evidence from self-paced reading. Lingua 121: 221–36.

75.

Roberts

Gullberg

Indefrey

(2008) On-line pronoun resolution in L2 discourse: L1 influence and general learner effects. Studies in Second Language Acquisition 30: 333–57.

76.

Rohde

Kehler

Elman

(2006) Event structure and discourse coherence biases in pronoun interpretation. In: Proceedings of the 28th Annual Conference of the Cognitive Science Society (pp. 697–702), Vancouver, July 26–29.

77.

Rohde

Kehler

Elman

(2007) Pronoun interpretation as a side effect of discourse coherence. In: Proceedings of the 29th Annual Conference of the Cognitive Science Society (pp. 617–22). Memphis, TN.

78.

Sanders

TJM

(2005) Coherence, causality and cognitive complexity in discourse. In: Aurnague

Bras

Le Draoulec

Vieu

(eds) Proceedings of the First International Symposium on the Exploration and Modelling of Meaning (SEM-05) (pp. 105–144). Toulouse: University of Toulouse-le-Mirail.

79.

Sanders

TJM

Spooren

Noordman

LGM

(1992) Toward a taxonomy of coherence relations. Discourse Processes 15: 1–35.

80.

Sanders

TJM

Spooren

Noordman

LGM

(1993) Coherence relations in a cognitive theory of discourse representation. Cognitive Linguistics 4: 93–133.

81.

Spooren

WPW

Sanders

TJM

(2008) The acquisition of coherence relations: On cognitive complexity in discourse. Journal of Pragmatics 40: 2003–26.

82.

Stevenson

Crawley

Kleinman

(1994) Thematic roles, focus and the representation of events. Language and Cognitive Processes 9: 519–48.

83.

Stevenson

Knott

Oberlander

McDonald

(2000) Interpreting pronouns and connectives: Interactions among focusing, thematic roles and coherence relations. Language and Cognitive Process 15: 225–62.

84.

Stewart

Pickering

Sanford

(2000) The time course of the influence of implicit causality information: Focusing versus integration accounts. Journal of Memory and Language 42: 423–43.

85.

Van Berkum

JJA

Brown

Hagoort

(1999) Early referential context effects in sentence processing: Evidence from event-related brain potentials. Journal of Memory and Language 47: 147–82.

86.

Van Berkum

JJA

Koornneef

Otten

Nieuwland

(2007) Establishing reference in language comprehension: An electrophysiological perspective. Brain research 1146: 158–71.

87.

Van Berkum

JJA

Zwitserlood

Bastiaansen

MCM

Brown

Hagoort

(2004) So who’s ‘he’ anyway? Differential ERP and ERSP effects of referential success, ambiguity and failure during spoken language comprehension. Paper presented at the Annual meeting of the Cognitive Neuroscience Society (CNS-2004), San Francisco, April 18–20.

88.

Van Silfhout

Evers-Vermeul

Sanders

(2015) Connectives as processing signals: How students benefit in processing narrative and expository texts. Discourse Processes 52: 47–76.

89.

Verhagen

(2000) Concession implies causality, though in some other space. In: Couper-Kuhlen

Kortmann

(eds) Cause – condition – concession – contrast (pp. 361–80). Berlin: Mouton de Gruyter.

90.

Wolf

Gibson

Desmet

(2004) Discourse coherence and pronoun resolution. Language and Cognitive Processes 19: 665–75.

91.

(2015) The influence of information status on pronoun resolution in Mandarin Chinese: Evidence from ERPs. Frontiers in Psychology 6: 1–17.

92.

Zhou

(2016) Topic shift impairs pronoun resolution during sentence comprehension: Evidence from event-related potentials. Psychophysiology 53: 129–42.

93.

Chen

(2017) How pronoun resolution was influenced by topic structure and verb-based implicit causality? Evidence from event-related potentials. Foreign Language Teaching and Research [in Chinese] 49: 323–34.

94.

Jiang

Zhou

(2015) When a causal assumption is not satisfied by reality: Differential brain responses to concessive and causal relations during sentence comprehension. Language, Cognition and Neuroscience 6: 704–15.

95.

Chen

Panther

(2018) Influence of concessive and causal conjunctions on pragmatic processing: Online measures from eye movements and self-paced reading. Discourse Processes 55: 387–409.

96.

Zorana

(2013) Assessing learners’ comprehension of logical connectives in L2 texts. Theory and Practice in Language Studies 3: 7–16.

97.

Zufferey

Mak

Degand

Sanders

(2015) Advanced learners’ comprehension of discourse connectives: The role of L1 transfer across on-line and off-line tasks. Second Language Research 31: 389–411.