Abstract
To pinpoint difficulties in the second language (L2) processing of temporarily ambiguous sentences, this study investigates first language (L1) effects and effects of verb bias, i.e. frequency information about preferential verb complements, on semantic persistence effects in L2 sentence comprehension. We tested 32 L1 German and 32 L1 Turkish intermediate-to-advanced learners of L2 English as well as 17 English natives in eye-tracking during reading and off-line comprehension. In off-line comprehension, neither natives nor L2 learners show verb bias effects, suggesting that semantic persistence of initial misinterpretations arises independently of lexical preferences. In contrast, reading times show that all groups are immediately sensitive to verb bias on-line, with L1 differences in how strongly verb bias affects garden-pathing. Unlike natives, however, L2 learners do not use verb bias to overcome semantic persistence to complete reanalysis in later segments of the sentence. We argue that, in spite of their initial sensitivity to verb bias in modulating garden paths, L2 readers suffer from processing limitations that constrain their ability to retrieve and integrate multiple sources of information for reanalysis. Our study confirms that reanalysis constitutes a key area of native versus non-native differences in L2 sentence processing.
I Introduction
When reading sentences that are temporarily syntactically ambiguous, first language (L1) and second language (L2) readers have been found to go down garden-paths, i.e. they initially adopt an incorrect syntactic analysis, and are later forced to reanalyse when the initial misanalysis is syntactically disambiguated (Frenck-Mestre and Pynte, 1997; Hopp, 2015; Jackson, 2008; Juffs, 2004; Juffs and Harrington, 1996; Rah and Adone, 2010). Garden-paths index incremental parsing in both L1 and L2 readers; yet, recent studies on L2 processing attest that reanalysis costs with temporary syntactic ambiguities may differ between L1 and L2 readers. L2 learners demonstrate larger and more extended slowdowns than native speakers (Roberts and Felser, 2011; but see Jacob and Felser, 2016), and they stay committed to the initial misanalysis for longer than native controls (Jacob and Felser, 2016; Pozzan and Trueswell, 2016). In fact, it has been argued that one major difference between L1 and L2 processing lies in the difficulty of revising initial misinterpretations of ambiguous sentences (Jacob and Felser, 2016). At present, it is unclear whether such difficulty stems from L2 learners getting garden-pathed to a larger extent than native speakers, i.e. L2 readers make a stronger commitment to the initial analysis than natives (Clahsen and Felser, 2006), or whether L2 learners experience greater reanalysis effort than natives (for discussion, Cunnings, 2017). Since reanalysis involves abandoning a previous analysis by integrating new information, L2 learners may be overtaxed by the computational demands of performing a target revision of the parse (Hopp, 2010; McDonald, 2006).
There is some evidence that adult L2 learners use various types of information in disambiguating temporary syntactic ambiguities, such as morphosyntactic information (Hopp, 2015; Jackson, 2008) and plausibility (Roberts and Felser, 2011; Williams et al., 2001). In addition, L2 readers refer to lexical information, such as transitivity (Hopp, 2015; Jegerski, 2012) and more fine-grained subcategorization preferences, such as verb bias (Dussias and Cramer-Scaltz, 2008) in incremental processing. Verb bias refers to a verb’s preference to take either a direct object or a sentential complement.
While previous studies attest that advanced L2 learners are sensitive to frequency biases for different complements, it is unclear whether the L1 may constrain L2 learners’ ability to recruit verb bias (Lee et al., 2013; Qian et al., 2019). Moreover, it has not yet been investigated systematically whether verb bias reduces garden-pathing in L2 learners, or whether it also aids in reanalysis. To explore these issues, we test whether L1 German and L1 Turkish learners of L2 English show sensitivity to verb bias information in the processing of temporary object-subject ambiguities, and we probe the extent to which verb bias facilitates syntactic reanalysis.
Using an eye-tracking during reading task and an off-line comprehension task, we employ two diagnostics for garden-pathing and reanalysis, respectively, in sentences involving temporary object-subject ambiguities like He heard the speech was quite possibly postponed until next week. First, the syntactic disambiguation region (was quite) probes the extent to which verb bias affects garden-pathing and, second, a later region (postponed until next week) taps into persistence effects, i.e. the degree to which an initial misinterpretation of the noun phrase the speech as the direct object persists, indicating incomplete reanalysis. This way, the present study disentangles how L1 and L2 readers integrate verb bias in the incremental parsing and revision of temporarily ambiguous sentences.
Our main findings suggest that native speakers can readily incorporate verb bias information which both modulates garden-pathing and aids reanalysis. In contrast, L2 learners, despite being sensitive to verb bias information, cannot effectively integrate such information for reanalysis. We conclude that L2 processing overtaxes resources, leading to failures to rely on lexical information in reanalysis.
1 Semantic persistence effects in ambiguity resolution
In the incremental processing of sentences like The professor forgot the theory revealed the underlying mechanism, readers tend to initially misanalyse the temporarily ambiguous noun phrase the theory as the direct object of the preceding verb forgot (L1 processing: Ferreira and Henderson, 1990; Frazier and Rayner, 1982; L2 processing: Juffs and Harrington, 1996; Lee et al., 2013). When they encounter the second verb revealed, they need to reanalyse the theory as the subject of an embedded clause to arrive at the target interpretation. However, many studies demonstrate that the semantic interpretation resulting from the initial direct-object analysis persists even after disambiguating information is encountered, a phenomenon that is called ‘semantic persistence’. For example, in Christianson et al. (2001), native English speakers read garden-path sentences such as While the man hunted the deer ran into the woods, in which the post-verbal noun phrase the deer can be temporarily misanalysed as the direct object of the verb of the preposed adjunct clause hunted. They subsequently answered comprehension questions such as Did the man hunt the deer?, which targeted the initially assigned direct-object interpretation. Participants made more incorrect ‘yes’ responses following temporarily ambiguous sentences than following unambiguous sentences, suggesting that native speakers did not fully reanalyse the sentence successfully.
Testing semantic persistence in on-line reading, Sturt (2007) conducted an eye-tracking reading experiment with native English speakers contrasting sentences as in (1).
(1) a. The explorers found (that) the South Pole was actually right at their feet. b. The explorers found (that) the South Pole was actually impossible to reach.
In the ambiguous (no-that) versions of (1), readers may be initially led down the garden path by virtue of a direct-object interpretation of the noun phrase the South Pole; however, this interpretation is ruled out by the following syntactic disambiguation region was actually. Additionally, the end of the sentences in (1) either matches or mismatches the direct-object interpretation of the South Pole. In the syntactic disambiguation region, the study found significant differences between the ambiguous (no-that) and the unambiguous (that) conditions, with longer reading times in the ambiguous conditions. The results also revealed robust effects of ambiguity in the final region, but only when the semantic content of this region was inconsistent with the direct-object interpretation (1b). This suggests that readers partially preserve the interpretation resulting from the initial direct-object analysis, which forces them to resolve the semantic incompatibility between the direct-object and subject interpretation of the post-verbal noun phrase even after the syntactically disambiguating copula verb has been encountered.
In a study investigating the causes of semantic persistence, Slattery et al. (2013) combined on-line reading measures and off-line interpretation questions to show that native speakers can build accurate syntactic representations during reanalysis but may fail to fully overwrite the initially assigned interpretation. As a result, semantic persistence effects seem to derive not from an inability to perform complete syntactic reanalysis, but from a failure to discard the syntactic structure built initially.
Turning to non-native sentence processing, Jacob and Felser (2016) tested whether L1 and L2 readers differed in the severity of semantic persistence effects. In an eye-tracking study, native English speakers and German-speaking L2 learners of English read sentences as in (2).
(2) While the gentleman was eating(,) the burgers were still being reheated in the microwave.
In (2), the syntactically disambiguating region (were still) was followed by a verb phrase (being reheated) expressing semantic information that was inconsistent with the direct-object interpretation of the ambiguous noun phrase. Both L1 and L2 participants showed longer reading times in the syntactic disambiguating region were still for the ambiguous sentences, with the L2 readers showing weaker garden-path effects. This was interpreted as indicating that L2 learners were less likely to revise their initial syntactic analysis. Further downstream, both groups evinced slowdowns in the semantic disambiguation region being reheated, indicative of semantic persistence. Subsequently, participants answered comprehension questions which directly targeted the direct-object misinterpretation. Both L1 and L2 participants showed semantic persistence by providing more incorrect ‘yes’ responses after ambiguous than unambiguous sentences; yet, the L2 group overall showed higher proportions of incorrect ‘yes’ answers. On the basis of these results, Jacob and Felser (2016) concluded that, despite showing sensitivity to both syntactic and semantic disambiguation cues, L2 learners were more likely to maintain the direct-object misinterpretation than L1 speakers.
In a visual world eye-tracking experiment, Pozzan and Trueswell (2016) also reported that L2 learners were more strongly committed to the initially assigned interpretation than L1 adults. In sentences like Put the frog on the napkin in the box, they continued to interpret on the napkin as the goal rather than a modifier. Oftentimes, they failed to revise an initial interpretation leading to erroneous interpretation of temporary ambiguities, spilling over into non-target-like executions of the command. These processing patterns were similar to those observed for L1 children (Trueswell et al., 1999), suggesting that the developing parser may show particular difficulties with garden-path recovery.
In sum, L2 learners appear to experience greater difficulty than native speakers in effecting syntactic reanalysis and thus exhibit increased semantic persistence effects. These findings on semantic persistence align with results from earlier studies showing that reanalysis costs are higher in L2 than L1 readers (Hopp, 2010; Juffs and Harrington, 1996), particularly during the processing of more complex garden-path sentences (e.g. Roberts and Felser, 2011).
Broadly speaking, two accounts have been proposed to explain the elevated reanalysis costs in L2 learners, which differ in whether they locate their causes in greater garden-pathing or greater reanalysis difficulty among L2 learners. According to some approaches to L2 processing like the Shallow Structure Hypothesis (SSH; Clahsen and Felser, 2006, 2018), L2 learners underuse syntactic information and instead rely more heavily on non-syntactic (lexical, semantic, plausibility) information in real-time processing. Building on this hypothesis, Jacob and Felser (2016: 922–923) claimed that L2 learners are more likely than L1 speakers to pursue the simplest syntactic analysis, as they resort more to syntactically shallow, good-enough processing. As a consequence, L2 learners are more committed to the garden-path misanalysis and find it more difficult to construct a correct and detailed syntactic analysis during reanalysis, thus showing increased semantic persistence effects. In other words, larger semantic persistence effects arise in L2 learners due to stronger initial misanalyses.
Other approaches to L2 processing argue that L2 learners are able to fully recruit syntactic information in parsing, but they sometimes fail to fully integrate multiple types of information due to cognitive resource limitations in L2 processing (Hopp, 2010; McDonald, 2006; Sorace, 2011; for illustration, see Lim and Christianson, 2013: 241). As a result, syntactic reanalysis may not be effected as efficiently during L2 processing, leading L2 learners to persist with the original analysis. On these accounts, semantic persistence effects arise due to failures in overwriting the initial misanalyses.
In the present study, we test between these approaches by manipulating information affecting the strength of garden-pathing, namely verb bias. For L1 and L2 readers, we investigate the extent to which verb bias information (1) affects initial misanalyses of temporary object–subject ambiguities and (2) modulates semantic persistence effects in on-line processing and subsequent interpretation.
2 Use of verb bias in ambiguity resolution
Verb bias refers to the use of a verb’s complementation bias, i.e. lexically specific frequency information about whether a verb is most likely to take a direct object (DO) or a sentential complement (SC) (Garnsey et al., 1997; Traxler, 2005; Trueswell et al., 1993; Wilson and Garnsey, 2009). In sentence processing research, verb bias effects have been studied in terms of whether garden-path effects vary according to an individual verb’s bias for complements, as in (3).
(3) a. The professor b. The professor
L1 processing studies using a variety of techniques show that ambiguous complement-clause sentences following a DO-bias main verb (e.g. forgot in (3a)) give rise to stronger garden-path effects than those with an SC-bias main verb (e.g. proved in (3b)) (self-paced reading: Garnsey et al., 1997; eye-tracking: Wilson and Garnsey, 2009; event-related brain potentials: Román et al., 2013; functional magnetic resonance imaging: Novais-Santos et al., 2007).
For L2 processing, Dussias and Cramer-Scaltz (2008) investigated whether L1-Spanish–L2-English learners behaved similarly to native English speakers in their use of English verb bias information. In a self-paced reading task, participants read sentences such as (4), in which DO-bias (confirm) and SC-bias (admit) verbs were followed by either a direct-object (4a&c) or a sentential-complement continuation (4b&d).
(4) a. The CIA director b. The CIA director c. The ticket agent d. The ticket agent
Both native controls and Spanish-speaking L2 learners were slower to read the disambiguating regions (italicized above) in ambiguous bias-violating (4b&c) than in ambiguous bias-matching sentences (4a&d), indicating that L2 learners show sensitivity to verb bias information during real-time sentence comprehension.
To investigate potential L1 effects, Lee et al. (2013) tested L1 Korean learners of English. Unlike English, Korean has verb-final word order, so that verb bias can never predictively guide incremental processing. In a self-paced reading task, higher and lower proficiency L1 Korean learners of L2 English read temporarily ambiguous and unambiguous complement-clause sentences that contained either DO-bias or SC-bias main verbs. Higher proficiency L1 Korean learners patterned like native speakers in that they exhibited garden-path effects (i.e. longer reading times following disambiguation in ambiguous than unambiguous sentences) only for DO-bias verbs. By contrast, lower proficiency L1 Korean learners exhibited garden-path effects in sentences with SC-bias verbs rather than in those with DO-bias verbs. These findings suggest that L2 learners whose L1 has verb-final word order can achieve native-like use of L2 verb biases at higher proficiency levels.
In a follow-up study, Qian et al. (2019) tested L1-Mandarin–L2-English learners on the same materials to check whether learners whose L1 and L2 both have SVO word order demonstrate fully native-like processing of verb bias across proficiency levels. The L1 Mandarin learners did not reveal any significant difference in the strength of garden-path effects as a function of verb bias, but still they showed verb bias effects in both ambiguous and unambiguous sentences, as reflected in longer reading times after DO-bias than after SC-bias verbs. Further, L2 proficiency level did not affect L1 Mandarin learners’ sensitivity to verb bias, although it was only the higher proficiency L1 Korean learners who displayed target-like sensitivity to verb bias by Lee et al. (2013). This contrast suggests that L1 word order and proficiency may interact in that L1 effects only play out at lower proficiency levels.
All in all, previous studies provide convergent evidence for L2 English learners’ sensitivity to verb bias information in L2 processing. Even L2 learners with a verb-final L1 can come to acquire sensitivity to selectional differences between DO-bias and SC-bias verbs to some extent. However, the scope of verb bias effects remains an open question in L2 processing. In particular, it is unclear if verb bias can also mitigate L2 learners’ difficulties with reanalysis and if the L1 modulates the use of verb bias for reanalysis. Addressing these points relates to the larger issue as to whether L1–L2 differences in the processing of temporary ambiguities are owing to differences (1) in the strength of garden-pathing or (2) in the ease of reanalysis; in other words, whether stronger semantic persistence effects observed in L2 learners result from difficulties (1) in computing fully detailed syntactic representations or (2) in fully completing the reanalysis process.
II The present study
The present study investigates potential interactions of verb bias and semantic persistence during the L2 processing of temporarily ambiguous sentences in English. Specifically, this study addresses the following research questions:
Does verb bias interact with semantic persistence effects resulting from temporary syntactic misanalysis in garden-path sentences?
Do differences in L1 word order affect L2 learners’ use of verb bias information and its interaction with semantic persistence effects?
The present study conceptually builds on Sturt (2007). As in Sturt (2007), we focus on temporary object-subject ambiguities in complement-clause sentences such as (5) in order to assess semantic persistence effects of the direct-object interpretation of post-verbal noun phrases. To this end, we manipulated the plausibility of the final segments of sentences in terms of whether they matched with the initial direct-object interpretation (factor Plausibility: semantic match (5a&c) vs. semantic mismatch (5b&d)). The semantic match and mismatch conditions differ with regard to whether the semantic content of the final segment provides a plausible or an implausible continuation for the interpretation assigned to the initial erroneous direct-object analysis.
(5) a. DO-BIAS SEMANTIC MATCH He b. DO-BIAS SEMANTIC MISMATCH He c. SC-BIAS SEMANTIC MATCH He d. SC-BIAS SEMANTIC MISMATCH He
Going beyond Sturt (2007), we manipulated the structural biases of main-clause verbs (factor Verb Bias: DO-bias (5a&b) vs. SC-bias (5c&d)), allowing us to investigate if verb bias modulates garden-path effects and thus attenuates semantic persistence effects with SC-bias verbs.
We expect that readers generate stronger expectations for a direct-object complement after DO-bias verbs (heard) and stronger expectations for an embedded sentential complement after SC-bias verbs (concluded). As a result, readers should experience increased processing difficulty when DO-bias verbs actually turn out to be followed by a sentential complement, which will manifest as longer reading times of the syntactic disambiguation region (was quite) in sentences with DO-bias verbs than in sentences with SC-bias verbs. When readers get to the final plausibility region (short and to the point/postponed until next week), longer reading times should be observed in semantic mismatch versus semantic match sentences with DO-bias verbs (5b vs. 5a) due to semantic persistence. Yet, these plausibility effects should be attenuated in sentences with SC-bias verbs (5d vs. 5c), because readers have less strongly adopted the initial direct-object interpretation than with DO-bias verbs. As a consequence, interpreting the post-verbal noun as the subject of an embedded clause should be easier for SC-bias than for DO-bias verbs, giving rise to an interaction of Verb Bias and Plausibility.
1 Hypotheses and predictions
Our first research question addresses the extent to which verb bias facilitates recovery from the incorrect direct-object analysis and thus mitigates semantic persistence effects. For native speakers, we expect to find an interaction of Verb Bias and Plausibility in later reading-time measures that index reanalysis and integration processes, once the final plausibility region has been read.
As regards L2 learners, approaches to L2 processing differ in whether they attribute heightened reanalysis effort as shown in semantic persistence effects to a stronger degree of garden-pathing or greater difficulty in revising an initial misinterpretation (Cunnings, 2017). For instance, Jacob and Felser (2016), referring to the SSH (Clahsen and Felser, 2006), posit that L2 learners overrely on non-syntactic (lexical, semantic, plausibility) information in L2 parsing. Such accounts attribute larger semantic persistence effects among L2 learners to their greater commitment to the initial misanalysis, i.e. stronger garden paths (see also Roberts and Felser, 2011), which they fail to overwrite because L2 learners allegedly underrely on syntactic information that disambiguates the sentence. However, non-syntactic information should reliably be used by L2 learners. Accordingly, lexical verb bias should modulate the strength of the garden-path effect, giving rise to main effects of verb bias in the syntactic disambiguation region. In the final region, plausibility information should also be readily used for reanalysis, which should be easier for SC-bias than for DO-bias verbs, since the initial garden-path was less strong for SC-bias verbs. Hence, like the natives, the L2 learner groups should show an interaction of Verb Bias and Plausibility.
According to other approaches, L2 processing differs from L1 parsing not in the use of a specific information type, i.e. syntax, but when multiple information types need to be integrated or recovered, e.g. in the reanalysis of temporarily ambiguous sentences. Accordingly, L2 learners likely have more difficulty recovering from an initial misinterpretation than native speakers, because L2 processing is inherently computationally more taxing (Hopp, 2010; McDonald, 2006; see also Abutalebi, 2008). In the present study, the L2 groups may show verb bias effects initially; yet, they may not demonstrate an interaction of Verb Bias and Plausibility when encountering the final region, since capacity limitations can preclude even advanced L2 speakers from retrieving and integrating additional information, e.g. verb bias, in recovering the correct interpretation.
In order to address the second research question about L1 effects, we tested L1 German and L1 Turkish learners of L2 English. We chose L1-German–L2-English and L1-Turkish–L2-English speakers, because finite verbs appear before complements in main clauses in German, while Turkish has verb-final word order in all clauses.
If L1 word order affects the use of L2 verb bias, the L1 Turkish learners, in contrast to the L1 German learners, will show a reduced ability in using verb bias information in L2 processing due to the verb-final structure of Turkish. That is, the L1 Turkish learners’ reading time patterns may reveal weak or no differences between sentences with DO-bias verbs and those with SC-bias verbs. As a result, the L1 Turkish group should not demonstrate any modulations of semantic persistence according to verb bias.
2 Participants
Sixty-four L2 learners of English participated in the study. Thirty-two participants were students of English at a German university at the time of testing. They reported German as their L1 and English as their L2. The remaining 32 participants, who were all students of English at a Turkish university at the time of testing, reported having Turkish as their L1 and English as their L2. All L2 participants had acquired English in an instructional setting.
In addition, we recruited a control group of 17 native speakers of English. All were students of English at a British university at the time of testing, and they were strongly dominant in English as per self-report. All participants had normal or corrected-to-normal vision, and they received 7.50 Euros for taking part in the study.
Table 1 summarizes all participants’ biographical information. It also presents the scores from individual differences tasks that we ran to match the groups on background factors that had been found to modulate the processing of garden-path sentences in previous studies, i.e. proficiency (e.g. Lee et al., 2013), working memory and lexical access (e.g. Hopp, 2014) as well as cognitive control (e.g. Teubner-Rhodes et al., 2016). Due to time limitations, the native speakers only took the proficiency and the lexical automaticity tasks.
Participant information and results from individual differences tasks.
a Proficiency: LexTALE
To assess participants’ proficiency in English, we administered the Lexical Test for Advanced Learners of English (LexTALE), which is a standardized placement test designed by Lemhöfer and Broersma (2012). Participants had to make a lexical (word or nonword) decision without any time limitation. Stimulus presentation and response collection were controlled by E-Prime, Version 2.0 (Schneider et al., 2002).
The L1 German participants had a mean proficiency score of 79.69 out of 100, while the mean score for the L1 Turkish participants was 72.54. These scores correspond to the B2 level (‘upper intermediate’) in the Common European Framework of Reference for Languages (CEFR). The native English participants received a mean score of 92.35, which indicates the C2 level (‘proficient user’) in the CEFR. A one-way ANOVA on the LexTALE scores showed that there was a significant difference in the mean proficiency scores between the three participant groups, F (2, 78) = 27.34, p < .001. Further independent samples t-tests demonstrated that the native speaker group performed better on the proficiency test compared to the L2 learner groups (Native vs. L1 German: t (47) = 4.94, p < .001; Native vs. L1 Turkish: t (47) = 7.47, p < .001). Additionally, the L1 German group had a statistically higher proficiency score than the L1 Turkish group (t (62) = 3.08, p = .003), even though both groups were placed in the same proficiency band according to the CEFR.
b Automaticity of lexical processing: Lexical decision task
To measure the automaticity of basic lexical processing which affects L2 sentence processing (Hopp, 2018), all participants took a lexical decision task as described in Hopp (2014), which was administered in E-Prime 2.0 (Schneider et al., 2002). A one-way ANOVA on the automaticity scores did not reveal any significant difference between the three participant groups (F (2, 78) = 0.38, p = .69), suggesting that the three groups were matched with regard to their automaticity in lexical access.
c Working memory: Reading span task
All L2 participants completed the English-language reading span task developed by Ariji et al. (2003) as described in (Hopp, 2014). An independent samples t-test did not show any statistical difference between the L1 German and L1 Turkish participants (t (62) = 0.36, p = .72). This indicated that the two participant groups were comparable in terms of their working memory capacity.
d Conflict monitoring: N-back task
To capture individual differences in cognitive control performance that have been found to affect reanalysis in garden-path sentences, a high-conflict N-back task was designed following the rationale in Teubner-Rhodes et al. (2016; see also Hussey et al, 2015; Kan et al., 2013; Novick et al., 2014). In the N-back task controlled by E-Prime 2.0 (Schneider et al., 2002), participants saw single letters presented in sequence and judged whether the current letter matched the one shown three positions back in the sequence.
The task was administered as in Teubner-Rhodes et al. (2016), and we calculated the means of proportion of correct responses to lure trials as a measure of conflict monitoring. The L1 German and L1 Turkish learners exhibited no significant differences in their conflict-monitoring scores (t (62) = 0.10, p = .92).
In all, this means that any potential differences in syntactic ambiguity resolution between the two learner groups cannot be obviously attributed to these background factors. To check for effects of proficiency, we added proficiency as a fixed effect in all analyses of the L2 learner groups.
III Materials and procedure
1 Eye-tracking study
a Materials
For the reading experiment, we constructed 24 quadruplets of critical sentences by crossing the factors Verb Bias (DO-bias vs. SC-bias) and Plausibility (semantic match vs. semantic mismatch) as in (6a-d) below (for a full list of the experimental items, see Appendix 1 in supplemental material). All critical sentences were preceded by a lead-in sentence and were followed by a final sentence.
(6) The account executive chaired the staff meeting in the conference hall. a. DO-BIAS SEMANTIC MATCH He heard the speech was quite possibly short and to the point. b. DO-BIAS SEMANTIC MISMATCH He heard the speech was quite possibly postponed until next week. c. SC-BIAS SEMANTIC MATCH He concluded the speech was quite possibly short and to the point. d. SC-BIAS SEMANTIC MISMATCH He concluded the speech was quite possibly postponed until next week. The company has over 500 employees, which makes staff meetings big events.
The critical sentences started with a subject pronoun (he), followed by a main-clause verb that has either DO- or SC-bias (heard vs. concluded). All verbs, including 24 DO- and 24 SC-bias verbs, were taken from previous studies on verb bias (Clifton and Frazier, 2012; Garnsey et al., 1997; Trueswell et al., 1993; Wilson and Garnsey, 2009). 1 The main-clause verb was followed by a noun phrase (the speech), which could be temporarily interpreted as the direct object of the preceding main-clause verb. This was followed by an auxiliary (was), which syntactically disambiguated the preceding noun phrase towards the subject of the embedded clause. In the final segment of the critical sentences, the content was either semantically matched or semantically mismatched with the direct-object interpretation of the noun phrase (short and to the point vs. postponed until next week). Therefore, the mismatch conditions additionally disambiguated the initial direct-object interpretation semantically. In order to increase the distance between the regions of syntactic and semantic disambiguation, adverbs and/or a past participle were inserted between the auxiliary and the final segment (quite possibly).
To ensure that any effects of verb bias were not confounded with the acceptability of the post-verbal noun phrases as direct objects of the main-clause (DO- or SC-bias) verbs, we administered an acceptability judgment task to 20 native-speaker English controls as well as 20 L1 German and 20 L1 Turkish learners of L2 English. These participants were recruited from the same populations as those tested in the eye-tracking reading study, yet they did not participate in the study. In the acceptability judgment task, participants were presented with simple sentences containing the post-verbal noun phrase as the direct object (e.g. The account executive heard/concluded the speech.), and they were asked to rate the acceptability of the given sentences on a seven-point scale, ranging from 1 (completely unacceptable) to 7 (completely acceptable). All 48 verbs were presented across two lists, containing 24 sentences each, along with 12 implausible objects (e.g. The band played the beer.) and 12 ungrammatical sentences (e.g. The mother slept the baby.). The results from English natives and L1 German and L1 Turkish learners of L2 English confirmed that the post-verbal noun phrases used in the eye-tracking reading study were acceptable as direct objects, irrespective of whether they followed a DO-bias or an SC-bias verb (Native: mean acceptability rating = 6.74 (SD = 0.73) vs. 5.27 (SD = 2.23); L1 German: mean acceptability rating = 6.19 (SD = 1.50) vs. 4.91 (SD = 2.27); L1 Turkish: mean acceptability rating = 6.38 (SD = 1.22) vs. 4.92 (SD = 2.12)). 2
b Procedure
The 24 experimental items were distributed across four lists according to a Latin Square design. These experimental items were combined with 56 filler sentences containing different structures, which were also preceded by a lead-in sentence and followed by a final sentence. To ensure that participants read the sentences attentively, half of the experimental and filler items were followed by a yes/no comprehension question that asked about the content of various parts of the lead-in, critical and final sentences (e.g. Were the staff meetings held in the dining room?). Twenty of the comprehension questions required a ‘yes’ response, while the remaining 20 required a ‘no’ response.
All experimental and filler items were presented in a pseudo-randomized order with a new randomization for each participant. The main experiment was preceded by instructions and a practice block of three items.
The recording of the reading-times and responses was controlled by an SMI RED High Speed eye-tracker with a spatial resolution below 0.4 degrees. Tracking speed was 500 Hz. The sentences were presented on a single line of text in Arial 24 point with white letters on a dark background on a computer screen with a resolution of 1680 × 1050 pixels. Participants were seated in front of the screen at a distance of approximately 700 mm. An adjustable chin rest was used to minimize head movements. Participants were asked to read the sentences for comprehension at their normal reading speed. They controlled the progress of the experiment using a keypad with three active buttons: a ‘Go’ button and two response buttons for answering the comprehension questions. Prior to participants reading the first item in the main experiment, the eye-tracker was calibrated with a 9-point calibration grid for both eyes. The calibration was repeated when visual acuity was below 0.5 degrees. The entire experiment lasted 25–30 minutes.
For the reading-time analysis, we defined two regions of interest in the experimental sentences: the syntactic disambiguation region which consisted of two words, including the auxiliary and the word immediately following it (was quite), and the final region containing the words which manipulated the semantic (mis)match information (short and to the point or postponed until next week). For each region of interest, we analysed first-pass reading time, second-pass reading time, total reading time, and number of regressions. First-pass reading time is the sum of all fixations in a particular region until this region is first exited to either the right or left. First-pass reading time has been taken to index first-pass syntactic processing (Rayner, 1998). Second-pass reading time is the sum of all fixations in a region after this region has already been exited to either the right or left, including zero times when the region is not refixated. Total reading time refers to the sum total of all fixations in a region. The number of regressions expresses the sum of backward eye-movements out of a region. The latter three measures are typically associated with reanalysis and integration processes (Clifton et al., 2007).
2 Off-line comprehension task
To further examine reanalysis and semantic persistence, an off-line comprehension task was designed as a follow-up to the main reading experiment. This off-line task contained the 24 quadruplets of critical sentences used in the reading experiment. Unlike in the reading experiment, the critical sentences were not followed by a final sentence. Instead, a yes/no comprehension question was given below each critical sentence as shown in (7). The critical sentences in the semantic match conditions were followed by a comprehension question that targeted the content of the embedded clause (7a), while the sentences in the semantic mismatch conditions were followed by a comprehension question that probed the direct-object interpretation (7b). Following Slattery et al. (2013), the rationale here was to test both whether participants achieved the correct syntactic assignment of the temporarily ambiguous noun phrase as the subject of an embedded clause for (6a&c) and whether they exhibited persistence in the initially assigned direct-object misinterpretation for (6b&d).
(7) a. Was the speech short and to the point? b. Did the account executive hear/conclude the speech?
Four counterbalanced presentation lists were created, with one of the four conditions of each critical sentence occurring in each list. The participants saw each critical sentence in a different version in terms of both Verb Bias and Plausibility from the one presented in the reading experiment. This way we controlled for potential priming effects from the reading experiment.
The off-line comprehension task was programmed in E-Prime 2.0 (Schneider et al., 2002). Each of the 24 experimental items was presented on a single line of text in Arial 12 point with white letters on a black background on a 15.6 inch screen. Participants were instructed to read the sentences carefully and to answer the comprehension questions as accurately as possible. Participants were allowed to take as much time as they needed to complete the task.
The correct answer for the comprehension questions in the match conditions was ‘yes’, while for the questions in the mismatch conditions it was ‘no’. The percentages of accuracy were calculated for each of the four conditions per participant.
IV Results
1 Off-line comprehension accuracy
To analyse the off-line comprehension data, we conducted logistic mixed-effects models (see Jaeger, 2008), including Question Type (embedded-content vs. direct-object), Group, and their interaction as fixed effects, random intercepts for participants and items, and random slopes for Question Type (both for participants and items) and Group (for items). All fixed effects were coded using sum coding, with the exception of the factor Group, for which we used treatment coding in order to contrast both L1 German and L1 Turkish learners of English with native controls. The model analyses were performed separately for the DO-bias conditions (10a&b) and the SC-bias conditions (10c&d). 3
a DO-bias semantic (mis)match conditions
For off-line comprehension questions following the DO-bias conditions, the overall response accuracy was 79% (SD = 13%) for the native English speakers, 79% (SD = 13%) for the L1 German learners, and 69% (SD = 15%) for the L1 Turkish learners. The model results showed that the L1 Turkish learners exhibited a trend towards being less accurate at responding to off-line comprehension questions compared to the native speakers (estimate = −0.79, SE = 0.42, z = −1.87, p = .061), while the L1 German learners’ overall accuracy did not differ from the native speakers’ accuracy (estimate = 0.43, SE = 0.49, z = 0.89, p = .373).
The results also revealed a significant main effect of Question Type (estimate = 4.03, SE = 1.59, z = 2.54, p = .011). For all three participant groups, the accuracy for the direct-object questions was found to be considerably lower than for the embedded-content questions (Native: 58% (SD = 26%) vs. 100% (SD = 0%); L1 German: 63% (SD = 21%) vs. 95% (SD = 11%); L1 Turkish: 44% (SD = 29%) vs. 94% (SD = 13%); see Figure 1). When checking for group differences, the models did not return any significant interaction of Question Type and Group in comparisons of the L2 learner groups against the natives (L1 German vs. Native: estimate = −0.22, SE = 0.68, z = −0.33, p = .743; L1 Turkish vs. Native: estimate = 0.18, SE = 0.69, z = 0.26, p = .798).
Mean percentages of accuracy by verb bias, participant group, and question type.
b SC-bias semantic (mis)match conditions
For off-line comprehension questions following the SC-bias conditions, the overall accuracy was 78% (SD = 12%) for the native English speakers, 74% (SD = 14%) for the L1 German learners, and 70% (SD = 12%) for the L1 Turkish learners. Both the L1 German and the L1 Turkish learners’ overall accuracy rates were indistinguishable from the native speakers’ accuracy rates (L1 German vs. Native: estimate = −0.21, SE = 0.42, z = −0.51, p = .608; L1 Turkish vs. Native: estimate = −0.33, SE = 0.47, z = −0.71, p = .475).
For the SC-bias conditions, the analyses revealed a significant main effect of Question Type (estimate = 1.23, SE = 0.50, z = 2.48, p = .013), yet no interaction of Question Type and Group (L1 German vs. Native: estimate = 0.08, SE = 0.48, z = 0.17, p = .863; L1 Turkish vs. Native: estimate = 0.75, SE = 0.49, z = 1.52, p = .128), reflecting that all three participant groups showed considerably lower accuracy for the direct-object questions than for the embedded-content questions (Native: 67% (SD = 22%) vs. 90% (SD = 12%); L1 German: 63% (SD = 23%) vs. 86% (SD = 18%); L1 Turkish: 51% (SD = 26%) vs. 90% (SD = 13%); see Figure 1).
In sum, all groups showed evidence of syntactic reanalysis effects, i.e. correct answers to the embedded-content questions, irrespective of whether the critical sentences contained DO-bias or SC-bias verbs. At the same time, all groups tended to answer ‘yes’ to the direct-object questions, suggesting that they could not entirely abandon the initial direct-object interpretation. 4
2 Reading times
For comprehension questions following half of the experimental and filler items, the overall accuracy was 87% (range: 78–97%) for the native control group, 88% (range: 75–100%) for the L1 German group and 85% (range: 75–95%) for the L1 Turkish group. This confirmed that all participants read the sentences carefully for comprehension.
For all reading-time measures, the data for a particular region were excluded if the first-pass reading time measure for that region was zero (Hopp and León-Arriaga, 2016). In addition, fixations shorter than 80 ms were removed because readers cannot extract sufficient information in such short fixations (Rayner, 1998). Fixations longer than 2,000 ms were also deleted because such long fixations typically indicate tracker loss (Rayner, 1998). In total, less than 9% of all data points were excluded due to these criteria.
We analysed the reading-time data for two regions of interest, namely the syntactic disambiguation region comprising the auxiliary and the word immediately following it, and the final region comprising the words which manipulated the semantic (mis)match information. To adjust for differences in region length and to remove individual differences in reading speed, we performed all statistical analyses on length-corrected residual reading times (see Trueswell et al., 1994). Residual reading times were computed by fitting a separate linear mixed-effects model for each reading time measure in all sentences (including fillers), with region length as the fixed effect and with random intercepts and slopes for each participant. Table 2 lists both raw (unadjusted) and residual reading times per condition and region of interest for all three participant groups.
Raw and residual mean reading times (in ms) and number of regressions by region, participant group, and condition.
Notes. Standard deviations are given in parentheses. All participants, Native English speakers: n = 17, L1 German learners: n = 32, L1 Turkish learners: n = 32. DO = direct object. SC = sentential complement.
All continuous reading time data were analysed with linear mixed-effects models (Baayen et al., 2008) using ‘R’, version 3.4.2 (R Core Team, 2017) and the lmer() function of R’s lme4 package (Bates et al., 2012). For all analyses, the initial model contained a maximal random-effects structure that included all fixed effects and interactions among them, random intercepts for participants and items as well as random slopes for all fixed effects and their interactions (Barr et al., 2013). If the maximal model failed to converge, the random-effects structure was simplified by removing the random slopes of the fixed effects, one at a time, on the basis of the values in the Hessian matrix. All categorical fixed effects were coded using sum coding. All continuous fixed effects were centered to avoid collinearity, and p-values were calculated following the procedure described in Mirman (2014).
In the first step of the reading-time analysis, we performed linear mixed-effects models with Group, Verb Bias, Plausibility, and their interactions as fixed effects to directly compare the processing patterns of the L2 learner groups with those of the native control group. The results of the omnibus analysis are shown in Table 3.
Linear mixed-effects models by region and reading-time measure.
Notes. DO = direct object. SC = sentential complement. DV = dependent variable (i.e. reading-time measure). SE = standard error. All participants, L1 German learners: n = 32, L1 Turkish learners: n = 32, Native English speakers: n = 17. *p < 0.1. **p < 0.05.
In addition to a three-way interaction between Group, Verb Bias and Plausibility in total reading times in the syntactic disambiguation region, the analysis yielded two-way interactions of Group and Plausibility in second-pass reading times and the number of regressions in the final region. To fully capture each participant group’s processing patterns, we analysed the reading times of each group separately for the two relevant regions. To determine the size of any observed effects of Verb Bias and Plausibility, we also calculated Cohen’s d effect sizes using R’s effectsize package (Ben-Shachar et al., 2020). In the following, we report the results of the maximal models for each group that reached convergence for all reading-time measures in each region of interest (see Table 4).
Linear mixed-effects models by participant group, region, and reading-time measure.
Notes. DO = direct object. SC = sentential complement. DV = dependent variable (i.e. reading-time measure). SE = standard error. All participants, Native English speakers: n = 17, L1 German learners: n = 32, L1 Turkish learners: n = 32. *p < 0.1. **p < 0.05.
3 Native control group
a The syntactic disambiguation region
There was a trend towards an interaction of Verb Bias and Plausibility in first-pass reading times. There were no main effects or interactions in second-pass reading times. Importantly, the native speakers demonstrated a significant interaction of Verb Bias and Plausibility in total reading times. To investigate the nature of this interaction, we computed another linear mixed-effects model for the DO-bias conditions and for the SC-bias conditions separately. The model contained the factor Plausibility as fixed effect, random intercepts for participants and items, and random slopes of the factor Plausibility for participants. This analysis demonstrated that the syntactic disambiguation region was read more slowly in the semantic mismatch conditions (143 ms) than in the semantic match conditions (29 ms) when the main-clause verbs had DO biases (estimate = −62.01, SE = 30.03, t = −2.06, p = .039, d = −0.28). When the main-clause verbs had SC biases, however, the native control group made no significant difference between the semantic match and mismatch conditions (80 ms vs. 46 ms, estimate = 16.65, SE = 30.26, t = 0.55, p = .582, d = 0.08).
Additionally, the main effect of Verb Bias was significant in the number of regressions, reflecting that there were more regressions for DO-bias verbs than for SC-bias verbs (67 vs. 48, estimate = 0.05, SE = 0.03, t = 2.01, p = .044, d = 0.20).
b The final region
No significant effects or interactions arose in any reading-time measure.
4 L1 German group
a The syntactic disambiguation region
The L1 German learners showed a trend towards a main effect of Plausibility in first-pass reading times (estimate = −16.41, SE = 8.85, t = −1.85, p = .064, d = −0.14). Further, there was a main effect of Verb Bias in second-pass reading times (estimate = 29.60, SE = 10.86, t = 2.72, p = .006, d = 0.21), indexing that the L1 German learners took significantly more time to read the syntactic disambiguation region if the main-clause verb had a DO bias (36 ms) than if it had an SC bias (–23 ms). There was also a tendency towards an interaction of Plausibility and Proficiency in second-pass reading times, with the difference between the semantic match and mismatch conditions being bigger with increasing proficiency. Total reading times also showed a significant main effect of Verb Bias (estimate = 24.31, SE = 12.06, t = 2.01, p = .044, d = 0.15), with longer reading times for DO-bias verbs (56 ms) than for SC-bias verbs (5 ms), and a trend towards an interaction of Plausibility and Proficiency. In the number of regressions, we found only a numerical effect of Verb Bias (DO-bias: 107, SC-bias: 85, estimate = 0.03, SE = 0.02, t = 1.68, p = .093, d = 0.12).
b The final region
The L1 German learners showed a main effect of Verb Bias in first-pass reading time (estimate = 45.68, SE = 18.36, t = 2.49, p = .013, d = 0.19), with significantly longer first-pass reading times in DO-bias conditions (113 ms) than in SC-bias conditions (20 ms). The effect of Verb Bias numerically reversed in second-pass reading times (estimate = −33.50, SE = 19.33, t = −1.73, p = .083, d = −0.14). Moreover, there was a significant main effect of Plausibility in second-pass reading times (estimate = −42.55, SE = 20.73, t = −2.64, p = .011, d = −0.22) such that the L1 German group took more time reading the semantic mismatch conditions (–45 ms) than the semantic match conditions (–128 ms), regardless of the biases of the main-clause verbs. In total reading times, the L1 German learners did not show any effects or interactions, but they exhibited a significant effect of Proficiency.
5 L1 Turkish group
a The syntactic disambiguation region
There were no significant effects or interactions in first-pass reading times. The group exhibited a trend towards effects of Verb Bias in second-pass (DO-bias: 54 ms vs. SC-bias: –18 ms, estimate = 33.67, SE = 18.16, t = 1.85, p = .064, d = 0.18) and total reading times (DO-bias: 102 ms vs. SC-bias: 35 ms, estimate = 32.22, SE = 17.12, t = 1.88, p = .060, d = 0.16). No other effects or interactions approached significance in these reading-time measures. The number of regressions showed a numerical effect of Verb Bias (estimate = 0.05, SE = 0.03, t = 1.87, p = .062, d = 0.16) and a significant effect of Proficiency.
b The final region
No significant effects or interactions were found in first-pass reading times. In second-pass reading times, however, the L1 Turkish learners showed a main effect of Plausibility (estimate = −74.62, SE = 34.12, t = −2.19, p = .029, d = −0.25), i.e. significantly longer second-pass reading times in the semantic mismatch conditions (–24 ms) than in the semantic match conditions (–176 ms); yet, this effect did not surface in total reading times (–41 ms vs. –118 ms). In the number of regressions, again, the L1 Turkish learners evinced a significant effect of Plausibility (estimate = −0.14, SE = 0.04, t = −3.07, p = .002, d = −0.28), with more regressions for semantic mismatch sentences (n = 359) than for semantic match sentences (n = 275), as well as a main effect of Proficiency, yet no interactions.
In sum, the native group showed both effects of Verb Bias and an interaction of Verb Bias and Plausibility in later measures in the syntactic disambiguation region. The L1 German group showed robust effects of Verb Bias in the syntactic disambiguation region as well as in the final region. The L1 Turkish group displayed the same pattern as the L1 German group, although the effects of Verb Bias in the syntactic disambiguation region did not reach statistical significance and were associated with very weak effect sizes. Critically, for neither L2 group were there any interactions of Verb Bias and Plausibility. 5
V Discussion
The aim of the present study was to investigate (1) whether verb bias modulates reanalysis in the processing of temporary object–subject ambiguities and, in turn, semantic persistence effects resulting from initial syntactic misanalysis, and (2) whether differences in L1 word order affect L2 learners’ use of verb bias information and semantic persistence effects.
Both the native English speakers and the L2 English learners exhibited effects of verb bias in the syntactic disambiguation region. The greater reanalysis difficulty observed in complement-clause sentences with DO-bias compared to SC-bias verbs indicates that, like the native speakers, the L1 German and L1 Turkish learners of L2 English had a stronger expectation for a direct object following DO-bias verbs and a stronger expectation for a sentential complement following SC-bias verbs. As a consequence, they experienced increased reanalysis difficulty in sentences with DO-bias verbs in the syntactic disambiguation region than with SC-bias verbs. This finding is in line with earlier studies suggesting that verb bias guides L1 and L2 comprehenders’ predictions about upcoming complement structure and thus affects garden-pathing in temporary object-subject ambiguities (Dussias and Cramer-Scaltz, 2008; Lee et al., 2013). 6
As for L1 effects on the use of L2 verb bias, the L1 German group demonstrated significant effects of verb bias in second-pass and total reading times in the syntactic disambiguation region, while the L1 Turkish group only approached significant effects of verb bias in the same reading-time measures. These differences suggest a certain, but limited role for L1 effects (see also Qian et al., 2019). Given that the verb biases were largely similar between English and the respective L1s, the finding that verb bias effects differ across the L2 groups may be attributable to the difference in word order between L1 and L2 (SOV vs. SVO). Compared to the L1-German–L2-English learners whose L1 and L2 both require finite verbs to appear before complements in main clauses, the L1-Turkish–L2-English learners might have shown weaker sensitivity to verb bias due to the Turkish SOV word order. Unlike a previous study on Korean learners of English by Lee et al. (2013), we did not find any effects of proficiency on the L1 Turkish learners’ sensitivity to verb bias. Since both Korean and Turkish share SOV word order, we suggest that our L1 Turkish learners were overall more proficient than Lee et al.’s (2013) lower proficiency L1 Korean learners, so that we cannot rule out that we would have found proficiency effects had we included less proficient L1 Turkish learners.
Although the L2 English learners were able to use verb bias information initially in the same way as the native English speakers, the interactions between Group, Verb Bias and Plausibility obtained in later reading-time measures associated with reanalysis suggest group differences in the way verb bias is used in reanalysis. The native speakers showed an interaction of Verb Bias and Plausibility. For DO-bias verbs, the natives demonstrated semantic persistence effects, as in Sturt (2007), while for SC-bias verbs, there was no significant difference between the semantic match and mismatch conditions. In other words, verb bias reduced the extent to which native readers continued to pursue a direct-object interpretation. We found the interaction of Verb Bias and Plausibility in late reading measures in the syntactic disambiguation region and not in the final region (as in Sturt, 2007), but we submit that the effect emerged most clearly in the re-reading of earlier material expressing the verb bias.
In contrast, the L1 German and L1 Turkish learners of L2 English did not show any interactions of Verb Bias and Plausibility in any region. As indexed by the Group by Plausibility interaction and the subsequent by-group analyses, they only showed significant main effects of Plausibility in the final region. In other words, the L2 learners remained committed to the direct-object misinterpretation to similar degrees across DO and SC-bias conditions, regardless of their earlier sensitivity to verb bias. Hence, they could not recruit verb bias for attenuating semantic persistence effects during on-line comprehension.
This pattern of findings raises the question of why semantic persistence effects get modulated by verb bias in L1, but not in L2 parsing. Considering that the L2 English learners showed target-like initial sensitivity to verb bias in the syntactic disambiguation region, the result that verb bias does not affect persistence of the incorrect direct-object interpretation in L2 parsing cannot be explained in terms of L2 learners’ inability to utilize verb bias information. Hence, it is unlikely that L2 speakers overrely on non-syntactic (lexical and plausibility) information per se, as e.g. suggested by the SSH (Clahsen and Felser, 2006). If lexical information were privileged in guiding on-line L2 sentence comprehension, verb bias should have continued to mitigate later semantic persistence effects. Instead, the effects of verb bias observed in the syntactic disambiguation region appeared to fade too quickly to attenuate the original interpretation of the post-verbal noun as an object which could then reduce the effort of the reanalysis occasioned by the semantic disambiguation in the final region.
Our findings therefore suggest that the causes for differences between L1 and L2 speakers reside in the reanalysis effort itself when readers need to integrate different types of information, namely, lexical, syntactic and plausibility information, to build the target sentence interpretation. For verb bias to attenuate semantic persistence effects, the reader needs to retrieve the lexical information from the first verb in order to construct the target interpretation when getting to the sentence-final segment. While such retrieval is effected in native speakers, as demonstrated by their elevated re-reading times of the syntactic disambiguation for DO-bias verbs, L2 learners do not seem to retrieve and integrate verb bias information later downstream in the parse. Such a failure likely reflects the higher cognitive demands of processing a non-native language which constrains L2 speakers in their ability to recover from the initial direct-object interpretation depending on earlier-encountered lexical information. Consequently, later downstream, temporarily ambiguous sentences continue to give rise to semantic persistence effects for both DO-bias and SC-bias verbs. This interpretation is consistent with approaches to L2 sentence processing which argue that many L1/L2 processing differences derive from limitations in L2 processing efficiency and information integration ability (Hopp, 2010; McDonald, 2006; see also Lim and Christianson, 2013).
To further tap into the causes of semantic persistence effects, we employed an additional off-line comprehension task. All groups had low accuracy rates for the direct-object questions, indicating that L1 and L2 speakers of English experienced semantic persistence effects of the incorrect direct-object interpretation, whereas the high accuracy rates for the embedded-content questions show that L1 and L2 speakers of English were able to assign the globally correct syntactic structure at the same time (see also Hopp, 2015; Roberts and Felser, 2011). In conjunction, these findings index that L1 and L2 comprehenders have difficulty abandoning the initial erroneous interpretation, even though they compute sufficiently detailed syntactic representations to arrive at the target interpretation (Cunnings, 2017; Slattery et al., 2013; see also Fujita and Cunnings, 2018).
At the same time, both the L2 English learners and the native controls tended to answer the direct-object questions incorrectly, and we did not observe any statistical difference in the accuracy rates between the L1 and L2 groups. These findings revealed that our L2 learner groups were not more strongly committed to the initial direct-object misinterpretation and therefore did not exhibit greater semantic persistence than the native control group. This finding runs counter to Jacob and Felser (2016), who reported that their L1-German–L2-English participants exhibited significantly lower accuracy for the direct-object questions than did native speakers. We submit that the discrepancy is likely due to the fact that different types of garden-path sentences were used in the two studies. In contrast to Jacob and Felser (2016), we focused on temporary object-subject ambiguities in complement-clause sentences rather than in sentences containing preposed adjunct clauses. As argued by Roberts and Felser (2011), it appears that L2 learners are more likely to complete reanalysis in complement-clause sentences (‘weak’ garden paths) than in preposed adjunct-clause sentences (‘strong’ garden paths), since garden-path recovery involves relatively less processing effort in complement-clause sentences. Overall, the comparison of the results reported here with those of Jacob and Felser (2016) indicates that semantic persistence effects are not necessarily more pronounced in L2 than in L1 comprehension, and that the strength of semantic persistence effects may vary depending on the types of garden-path constructions.
To conclude, the reading data showed that both L1 German and L1 Turkish learners of L2 English were similar to native speakers in their sensitivity to verb bias on-line, with L1 differences affecting the strength of verb bias effects; yet, unlike natives, L2 learners failed to demonstrate modulations of semantic persistence according to verb bias. We argue that L2 learners’ processing resource limitations constrain the L1 German and L1 Turkish learners in retrieving verb bias to overcome semantic persistence effects. In future research, it would be interesting to investigate whether near-native L2 learners show fully native-like processing. In all, the study adds to the growing body of studies documenting that L2 speakers suffer greater difficulty than native speakers in recovering from garden-path misanalysis. It extends previous research by showing that L2 speakers use verb bias information to modulate garden-pathing, yet that they cannot integrate it in reanalysis to attenuate semantic persistence effects resulting from garden-paths in parsing or interpretation.
Supplemental Material
sj-pdf-1-slr-10.1177_0267658321997904 – Supplemental material for Verb bias and semantic persistence effects in L2 ambiguity resolution
Supplemental material, sj-pdf-1-slr-10.1177_0267658321997904 for Verb bias and semantic persistence effects in L2 ambiguity resolution by Duygu Fatma Şafak and Holger Hopp in Second Language Research
Footnotes
Acknowledgements
We would like to thank Bilal Kırkıcı and Christina Schelletter for their dedicated support during our data collection in their respective universities. We also extend our thanks to Kiel Christianson and two anonymous reviewers for their constructive comments, which helped us to improve the manuscript. All remaining shortcomings are our responsibility.
Declaration of Conflicting Interests
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 no financial support for the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
