PREDICTIVE PROCESSING AND L2 LEARNERS

Adult psycholinguistic research has shown that native speakers not only integrate each word into a syntactic structure incrementally but also activate structural representations before they appear in the linguistic input (e.g., Altmann & Kamide, Reference Altmann and Kamide1999). This predictive behavior has been shown to be an integral component of language processing that makes comprehension fast and efficient (Federmeier, Reference Federmeier2007; Lau et al., Reference Lau, Stroud, Plesch and Phillips2006; Staub & Clifton, Reference Staub and Clifton2006). A critical issue in L2 processing research is whether the same processing strategies that have been identified in native speakers of a target language are also found in L2 learners’ processing (Grüter, Rohde, & Schafer, Reference Grüter, Rohde, Schafer, Orman and Valleau2014, Reference Grüter, Rohde and Schafer2016; Kaan, Reference Kaan2014).

One of the areas that have been extensively studied in L2 processing research is gender agreement during lexical and syntactic processing. In languages such as French and Spanish, grammatical gender is marked on the determiner, which precedes the noun, and the determiner serves as a predictive cue for the upcoming noun. For instance, Dahan, Swingley, Tanenhaus, and Magnuson (Reference Dahan, Swingley, Tanenhaus and Magnuson2000) showed that, in the absence of grammatical gender, French native speakers looked at pictures with names that shared an initial sound with the target (e.g., vase “vase” and vache “cow”) more than at pictures with phonologically unrelated names, demonstrating cohort effects. However, such effects disappeared when the nouns were preceded by gender-informative determiners. Several L2 studies suggest that L2 learners are less efficient in using gender-marked determiners for generating predictions (Guillelmon & Grosjean, Reference Guillelmon and Grosjean2001; Grüter et al., Reference Grüter, Lew-Williams and Fernald2012; Martin et al., Reference Martin, Thierry, Kuipers, Boutonnet, Foucart and Costa2013).

In response to L2 learners’ nonnativelike behavior in generating predictions, Grüter et al. (Reference Grüter, Rohde, Schafer, Orman and Valleau2014) proposed that learners have a reduced ability to generate expectations, which they call the RAGE hypothesis (see also Grüter, Rohde, & Schafer, Reference Grüter, Rohde and Schafer2016). The authors claim that the reduced ability of anticipatory processing is attributed to a limited capacity of the L2 processor, which does not have enough resources for updating predictions after dealing with the immediately required processes. Exceptions to this pattern have been observed, such that L2 learners demonstrated nativelike predictive processing (Dowens et al., Reference Dowens, Vergara, Barber and Carreiras2010; Dussias et al., Reference Dussias, Valdés Kroff, Guzzardo Tamargo and Gerfen2013; Foote, Reference Foote2010; Hopp, Reference Hopp2013; Keating, Reference Keating2009). To resolve the discrepancy, Kaan (Reference Kaan2014) proposed an individual difference account, in which the potential differences between native and nonnative predictive processing lie in the same factors that drive individual differences in native speakers; in particular, Kaan considered factors such as stored differences in frequency information and the accuracy and consistency of lexical representation. Frequency information bias moderates the outcomes of predictions, because predictions are generated on the basis of the likelihood of a certain word and syntactic frame occurring in the context (MacDonald, Pearlmutter, & Seidenberg, Reference MacDonald, Pearlmutter and Seidenberg1994).

The effect of frequency information can help us account for positive evidence offered by some studies that demonstrate nativelike predictive processing in only advanced L2 learners. Nevertheless, typical L2 learners receive language input that differs from that of native speakers in its nature and quantity. The logic is that some L2 learners accumulate the relevant frequency information, and therefore, their manner of predictive processing approximates that of native speakers. One convincing case comes from studies on verb subcategorization bias (Dussias & Cramer Scaltz, Reference Dussias and Cramer Scaltz2008; Dussias, Marful, Gerfen, & Molina, Reference Dussias, Marful, Gerfen and Molina2010; Frenck-Mestre & Pynte, Reference Frenck-Mestre and Pynte1997). Lee, Lu, and Garnsey (Reference Lee, Lu and Garnsey2013) examined how learners of English with an L1 of Korean used verb information when they read sentences such as (3) and (4). These sentences were presented with or without the complementizer that.

1. (3) The ticket agent admitted (that) the mistake might be hard to correct.

2. (4) The club members understood (that) the bylaws would be applied to everyone.

It was hypothesized that L2 learners experience garden-path effects at the auxiliary would in (4) with understood, but there should not be garden-path effects at might in (3) with admit. This is because the verb admit has a strong bias for a sentential complement, whereas the verb understand takes a direct object more frequently than a sentential complement, which leads speakers to anticipate the direct-object frame (Gahl, Jurafsky, & Roland, Reference Gahl, Jurafsky and Roland2004; Garnsey, Pearlmutter, Myers, & Lotocky, Reference Garnsey, Pearlmutter, Myers and Lotocky1997). Lee et al. (Reference Lee, Lu and Garnsey2013) found that when reading (4), both native speakers and L2 learners with advanced proficiency demonstrated a slowdown in reading time at the auxiliary, regardless of the presence of that. However, lower proficiency learners processed the auxiliary faster for verbs with sentential complement bias than for optimally transitive verbs but only with the presence of that. This pattern of results confirms that the L2 participants used the likelihood of a verb occurring with a direct object or with a sentence complement for processing complements. However, the lower proficiency learners needed an additional cue (i.e., the presence of that) to use verb frequency information for real-time processing. Limited exposure due to the learning experience leads to an underrepresentation of words and weaker associative connections (Chen, Reference Chen1990; Dufour & Kroll, Reference Dufour and Kroll1995), and therefore, these results support incremental processing being a developmental phenomenon.

In verb-final languages, verb subcategorization information is not as helpful as it is in English. Instead, case markers are known to guide incremental processing, allowing native speakers of verb-final languages to predict upcoming arguments in a sentence in Dutch and German (Bader & Meng, Reference Bader and Meng1999; Fanselow, Kliegl, & Schlesewsky, Reference Fanselow, Kliegl, Schlesewsky, Kemper and Kliegl1999). This difference in the types of cues used in predictive processing naturally leads us to question whether L2 learners are able to use case-marker cues for processing. Evidence from L2 processing studies in German and Dutch demonstrated that learners may have difficulty integrating case and gender assignment information (Havik, Roberts, Van Hout, Schreuder, & Haverkort, Reference Havik, Roberts, Van Hout, Schreuder and Haverkort2009; Hopp, Reference Hopp2006; Jackson, Reference Jackson2008). When processing temporarily ambiguous wh-questions, such as in Examples (5)–(7), L2 learners do not demonstrate a preference for a subject-first order to the same extent as native speakers do; the subject preference was only observed for simple past-tense sentences, such as (5), in which the verb is present early in the sentence, and not in those such as (7). However, German native speakers exhibited a processing preference for subject-first sentences, regardless of the location of the lexical verb (Jackson, Reference Jackson2008).

1. (5)

   

2. (6)

   

3. (7)

   

Subsequent studies have shown that highly advanced L2 learners can demonstrate nativelike strategies when they process sentences in a larger discourse context (Hopp, Reference Hopp2009) or sentences in which the initial wh-word is unambiguously marked with wer “who” or wen “whom” (Jackson & Dissias, Reference Jackson and Dussias2009). These results collectively suggest that highly proficient L2 learners may be able to acquire nativelike predictive processing in head-final languages too, but such processing strategies may be susceptible to a degree of ambiguity in the predictive cues and to the availability of lexical cues.

The Japanese case system does not have dichotomous rules, as in the assignment of either nominative or accusative case to Welche in (5). The Japanese nominative case, –ga, may correspond, before the verbal disambiguation, to the agent, the logical object (a theme or patient), or the logical object of a predicate adjective. Japanese therefore presents even greater ambiguity of case-marker cues than Germanic languages do. Despite this ambiguity, previous psycholinguistic research has shown that Japanese native speakers begin projecting the structure of an upcoming sentence prior to the verb and that case markers are one of the major determinants of that process (Aoshima et al., Reference Aoshima, Phillips and Weinberg2004; Kamide et al., Reference Kamide, Altmann and Haywood2003; Miyamoto, Reference Miyamoto2002; Nakano, Felser, & Clahsen, Reference Nakano, Felser and Clahsen2002; Yamashita, Reference Yamashita1997). Adapting case-marker-driven processing may be a challenging task for L2 learners. The Japanese case system is known to be one structure that is difficult to master by L2 learners (Iwasaki, Reference Iwasaki and Karimi2008). Studies in the competition model have shown that L2 learners of Japanese demonstrate a reliance on word-order cues, failing to employ information from case markers for comprehension (Kilborn & Ito, Reference Kilborn, Ito, MacWhinney and Bates1989; Rounds & Kanagy, Reference Rounds and Kanagy1998; Sasaki, Reference Sasaki1991, Reference Sasaki1994; for a review, see Sasaki & MacWhinney, Reference Sasaki, MacWhinney, Nakayama, Mazuka and Shirai2006).

To directly examine predictive processing by L2 learners of Japanese, Mitsugi and MacWhinney (Reference Mitsugi and MacWhinney2016) replicated Experiment 3 of Kamide et al. (Reference Kamide, Altmann and Haywood2003) with L2 learners with L1 of English. In line with the results obtained in the original study (Kamide et al., Reference Kamide, Altmann and Haywood2003), Mitsugi and MacWhinney (Reference Mitsugi and MacWhinney2016) demonstrated that Japanese native speakers showed anticipatory eye movements, looking at the theme object in the visual scene after hearing the sequence of a nominative-marked noun and a dative-marked noun in the ditransitive condition, as in (8), more than in the monotransitive condition, in which the sequence of a nominative-marked noun and an accusative-marked noun signals no theme object to follow, as in (9), before the point at which the theme was mentioned in the input.

1. (8)

   

2. (9)

   

In contrast, the L2 learners did not generate predictions. L2 learners regulated their attention to the theme object more in the ditransive condition than in the monotransitive condition only after the target words were mentioned. Mitsugi and MacWhinney (Reference Mitsugi and MacWhinney2016) accounted for the absence of predictions with the complexity of cues. That is, the process involved anticipation of the thematic role and lexical semantics based on the combined information of two preverbal nouns, which is more complicated than assigning a thematic role to either a nominative or an accusative noun by a single unambiguous cue, as in Germanic languages. Clearly, this interpretation remains speculative, and further testing of different linguistic structures for predictive processing is necessary. Building on Mitsugi and MacWhinney (Reference Mitsugi and MacWhinney2016), this study examined whether case markers are efficiently used by L2 learners of Japanese by looking at a different type of linguistic outcome prediction: active and passive voice. In the next section, we lay out the key features of the structure in question.

THE STRUCTURE AND PROCESSING OF JAPANESE

Japanese actives and passives have been considered a binary opposition in grammatical voice (Shibatani, Reference Shibatani and Shibatani1988). The shift to the passive voice from the active is expressed as a part of the inflection on the verb, which appears in sentence-final position. Examples (10) and (11) illustrate an active sentence and its passive counterpart, respectively. Generative approaches see the passive formation as a process of syntactic derivation from the active counterpart (Chomsky Reference Chomsky1995), and the process of the derivation of a passive sentence from an active sentence is as follows: the direct objet noun, which is originally marked with the accusative case –o, is promoted to the sentence-initial subject position. It is now marked with the nominative case affix –ga in the passive sentence shown in (11). The initial subject, in turn, is demoted, and it is marked with the dative case marker –ni. Finally, the passive morpheme –(r)are attaches to the verb root (Tsujimura, Reference Tsujimura1996).Footnote ²

1. (10)

   

2. (11)

   

In terms of the acquisition, it has been reported that Japanese-speaking children have difficulty comprehending passive sentences. Children exhibit non-adult-like interpretations until around age 5 or 6 (Sano, Endo, & Yamakoshi, Reference Sano, Endo, Yamakoshi, Do, Domínguez and Johansen2001). Possible sources of this difficulty have been explored; some researchers have argued that the structural complexities of passives (i.e., the A-chain maturation hypothesis; Borer & Wexler, Reference Borer and Wexler1987) contribute to this comprehension difficulty (Minai, Reference Minai, Nakayama and Quinn2000; Sugisaki, Reference Sugisaki1999). Others posit that the difficulty resides in the assignment of thematic roles in an atypical order (i.e., patient before agent). Hakuta (Reference Hakuta1982) showed that Japanese children misinterpret a first noun phrase of the sentence as the thematic agent when it is marked in the nominative case, drawing a parallel with English-speaking children, who use the canonical-order strategy (Slobin & Bever, 1982). Similarly, Clancy (Reference Clancy and Slobin1985) demonstrated that Japanese children acquired scrambled object–subject–verb passives earlier than they did subject–object–verb passives. In recent eye-tracking studies, Choi and Trueswell (Reference Choi and Trueswell2010) attributed the children's disproportionate reliance on cues to the order in which those cues become available (see Huang, Zheng, Meng, & Snedeker, Reference Huang, Zheng, Meng and Snedeker2013, for a study on children processing Chinese passives). Children were able to, like adults, incrementally build structural representations as soon as they had access to the relevant evidence, such as case markers in Korean (which has a case system similar to that of Japanese). However, when this initial structural commitment turned out to be incorrect, the children did not appear to revise their initial interpretation and made errors that were inconsistent with late-arriving grammatical cues.

Although L1 acquisition studies have focused on children's comprehension difficulty with passives experimentally using, for instance, act-out comprehension tasks, prior L2 studies on Japanese passives have been rather descriptive in nature, focused on error analysis and narrative analysis (Mizutani, Reference Mizutani1985; Otsuka, 1995; Reference SatoSato, 1997). These studies showed that the passive construction poses a great challenge for L2 learners of Japanese too, which is indexed by the delayed and scarce use of the structure by those speakers. For instance, Mizutani (Reference Mizutani1985) pointed out that learners of Japanese with L1 English had a tendency to use the active voice in contexts in which the passive voice would be more appropriate. Similarly, examining oral narrative samples, Watanabe (Reference Watanabe1996) found evidence that learners did not effectively handle perspectives, frequently and inconsistently changing the point of view in successive clauses, whereas native speakers told a story consistently from the main character's viewpoint. On the basis of such production patterns, Tanaka (Reference Tanaka and Tanaka1998) argued that difficulties with the use of Japanese passives stem from the relationship between structures and perspectives, in which Japanese speakers are required to use different syntactic structures, such as actives, direct passives, indirect passives, and benefactives, depending on the point of view in compound sentences.

Although these L2 studies can inform us about the pattern of learners’ errors and their relationship with surrounding structures, it eschews the significance of how, exactly, L2 learners process Japanese passives in real time. With this study, we take steps toward closing this gap. Taking a psycholinguistic approach, we systematically compared the time course of processing of Japanese passives by L2 learners with that of native Japanese speakers.

THE PRESENT STUDY

The present study examined the extent to which L2 learners incrementally assign thematic roles to preverbal nouns and predictively activate a structural representation of the upcoming verb voice before encountering the verb. This study further considered the process in which verb morphology information feeds into L2 learners’ comprehension. The processing patterns of active and passive sentences, as in Examples (12) and (13), were compared. We used the visual-world paradigm, with the two-choice picture-identification method, in which one picture matched the active reading, as in (12), and the other matched the passive reading, as in (13). Figure 1 demonstrates the setup for a visual scene.

Figure 1.

An example of a visual scene used in the experiment.

1. (12)

   

2. (13)

   

The first research question posed in the current study is whether L2 learners of Japanese incrementally assign thematic roles when hearing preverbal nouns. Active sentences involve a preverbal sequence of a noun in the nominative case and a noun in the accusative case, as in (12), and passive sentences have a sequence of a noun in the nominative case and a noun in the dative case, as in (13). We hypothesized that if the participants effectively use the information provided by these case-marked nouns, they should predict a verb with active voice after hearing a sequence of a noun in the nominative case and a noun in the accusative case, and the participants should regulate their anticipatory eye movements toward the image depicting the active reading (see left image in Figure 1). After hearing a sequence of a noun in the nominative case and a noun in the dative case, they should anticipate that the upcoming verb would involve passive morphology and should regulate their anticipatory eye movements toward the image depicting the passive reading (see right image in Figure 1). Notice that the visual scene provides the meaning of the verb, and therefore, the participants should not anticipate the verb turning out to be one of the rare monotransitive verbs that take a noun marked in the dative case as its theme (e.g., au “meet”) or to be a construction in which the dative noun is interpreted as a goal after a ditransitive verb (e.g., watasu “pass”).Footnote ³ If, however, L2 learners do not incrementally assign thematic roles, they will initially extract identical information from the preverbal nouns. Consequently, there should be no difference in their attention to the pictures in the active and passive conditions.

The second question is how the information from the disambiguating verb feeds into comprehension. Evidence from previous psycholinguistic studies with native Japanese speakers suggests that when the initial assignment of thematic roles turns out to be wrong at a later point, native speakers revise the interpretation (Hirotani, Makuuchi, Rüschemeyer, & Friederici, Reference Hirotani, Makuuchi, Rüschemeyer and Friederici2011; Mazuka & Itoh, Reference Mazuka, Itoh, Mazuka and Nagai1995). In the passive condition, L2 learners’ eye movements to the picture describing the passive action are to increase over the course of processing the disambiguating verb when they correctly predict verb voice. However, if the participants analyze the preverbal fragment as a part of an active sentence in the passive condition, upon arriving at the passive morphology, –rare, their initial assignment will need revision for the sentence to be correctly interpreted as having passive voice.

Participants

Twenty native speakers of Japanese and 29 L2 learners of Japanese with an L1-English background took part in this study. They all received US$10 for their participation. The L2-learner participants were undergraduate students at a North American university, and the native-speaker participants were also recruited from this university's community. The learners were enrolled in forth- to eighth-semester Japanese courses at the time of data collection. The L2 learners’ exposure to the Japanese language was primarily through formal classroom instruction, but several L2 participants had studied abroad. Participants’ biographical information was assessed by a language learning history questionnaire, which was administered after the eye-tracking task.

Because the L2 learners were recruited from various course levels, it was necessary to assess their proficiency levels and to take them into account when examining the learners’ processing. We used Part I of the Japanese Skill Test (Itomitsu, Reference Itomitsu1996). This test is a criterion-reference test and is reported to be effective and reliable in determining Japanese learners’ proficiency (Eda, Itomitsu, & Noda, Reference Eda, Itomitsu and Noda2008). The test consists of 45 multiple-choice items assessing a variety of grammatical structures in Japanese, including case-marking, conjugation, and sentence structure rules. The L2 participants took this proficiency test on a different day, after completing the eye-tracking session. Table 1 provides the biographic information and the proficiency test results obtained from the 20 L2 participants whose data points were included in the eye-tracking analysis.

Table 1.

Participant biographical information and results of offline measures

The L2 learners also completed a translation task in which they translated Japanese sentences to English. The translation task was administered after the eye-tracking task so as not to influence the results of the main part of the study. This task was used to make sure that the L2 learners had the relevant grammatical knowledge. The task covered but was not limited to active and passive structures; it also included sentences with causatives, causative-passives, and relative clauses, all of which required the participants to pay attention to thematic role assignments. There were 16 items that directly tested the translation of active and passive structures. A dichotomous scoring method was used, giving 1 point for each correct thematic role pair (i.e., agent–patient) and verb voice. The results of the translation task showed that the L2 learners did have the relevant grammatical knowledge (M = 31.2 points, 97%, SD = 0.98). For the translation task, the cutoff point was set at 91%, determined by selecting the point that was 2 SD away from the mean. All of the participants were carried forward for analysis.Footnote ⁴

Materials

Twelve experimental sentence pairs were constructed on the basis of the two patterns illustrated in (12) and (13). The active sentences involved a preverbal sequence of a nominative-marked noun and an accusative-marked noun, followed by a transitive verb with active voice, and the passive sentences had a sequence of a nominative-marked noun and a dative-marked noun, followed by a transitive verb with the passive morphology –rare. A modal adverbial phrase (e.g., “badly,” “immediately,” and “seriously”) was added between the second noun and the sentence-final verb. The hearsay morpheme (i.e., the reported evidential) was added in order to manipulate the experimental stimuli so that the sentences would not end with a verb root. The vocabulary items were drawn from the textbooks used in the first- to third-semester Japanese courses at the institution where the study was conducted in order to ensure lexical familiarity. Figure 2 provides a visual demonstration of the regions of interest.

Figure 2.

The periods of interest tagged in the audio stimuli.

Two female native speakers recorded experimental sentences at a natural speaking rate. The sentences were produced using standard intonation. The recording was done with a sampling rate of 44,100 Hz. To analyze eye movements with reference to the corresponding spoken sentences, we placed markers on the recorded stimuli. We marked the onset of the first noun and the offset of the second noun (in the active condition, onnnanohito-ga otokonohito-o “woman-NOM man-ACC”; in the passive condition, onnnanohito-ga otokonohito-ni “woman-NOM man-DAT”) and the onset and offset of the adverbial phrase (hidoku “badly”) and the sentence-final verb phrase (in the active condition, tataitasoudesu “hit-PAST-HSY-COP”; in the passive condition, tatakaretasoudesu “hit-PASS-PAST-HSY-COP” in each target sentence. A 200-ms slice of silence was added between the offset of the second noun and the onset of the adverbial phrase, and a 500-ms slice was added between the offset of the adverbial phrase and the onset of the verb phrase to control the duration of the critical region.

To examine the effect of predictive processing, analyses were conducted on the participants’ looks to the active picture scenes during the audio presentation of the adverbial phrase. Hereafter, this period will be called the critical region. The duration of the critical region was 1200 ms, which approximately corresponds to the preadverbial slice of silence (200 ms), the mean duration of the adverbial phrase period (500 ms), and the postadverbial silence (500 ms). In order to take into account the time spent on planning and regulating eye movements, the period of interest was offset by 200 ms (Matin, Shao, & Boff, Reference Matin, Shao and Boff1993). The statistical analysis was performed on the critical region, which was from 200 ms after the offset of the second noun and lasted 1200 ms. Table 2 reports the mean durations of these sentence regions.

Table 2.

Mean duration of sentence regions by condition

Note: A 200-ms slice of silence was added between the offset of second noun and the onset of adverbial phrase, and a 500-ms slice was added between the offset of the adverbial phrase and the onset of the verb phrase.

For each verb, two line drawings were created (see Figure 1). These images were drawn by hand, then scanned and edited for stimulus presentation. One drawing depicted the active reading (i.e., “the woman hit the man”), and the other drawing reversed the agent and patient roles (i.e., “the woman was hit by the man”). For each condition, 12 visual scenes were created. In order to ensure the reliability of these drawings, two native speakers of Japanese and one native speaker of English with advanced Japanese proficiency who did not participate in the main experiments, were asked to match the experimental sentences with the corresponding drawings. There was no disagreement or confusion in the matching procedure. These images were presented on the eye-tracking monitor at a resolution of 1680 × 1050 pixels.

The prerecorded spoken sentences were randomly assigned to the two experimental lists, such that the active and passive readings of the sentences each appeared on one of the lists, and each list contained an equal number of active and passive sentences. Each experimental list comprised 12 items, with 6 items per experimental condition, and these experimental sentences were combined with 36 filler sentences. The filler sentences had lengths and complexity similar to those of the experimental sentences.Footnote ⁵ The picture scenes and recorded spoken sentences were presented in a random order determined by the stimulus presentation system associated with the eye-tracking system. The positions of the active and passive picture scenes were counterbalanced.

The participants’ eye movements were recorded with an EyeLink 1000 desktop-mounted tracking device manufactured by SR Research, with a monocular sampling rate of 1000 Hz. The tracking device was connected to a PC that controlled the stimulus display and data storage. The visual stimuli were presented on a color 22-in. ViewSonic monitor. The participants were seated approximately 65 cm from the monitor and rested their chins comfortably on a chin rest. The participants’ eye positions were calibrated at the beginning of the experimental session and whenever it was necessary thereafter, using a 9-point calibration procedure for visual acuity below 0.5 degrees.

Analysis

Data treatments

We calculated response accuracy on the eye-tracking task (i.e., mouse clicks on the correct picture scenes). We excluded data from any participants whose responses were incorrect on more than one-third of the items in the eye-tracking study in order to ensure the reliability of their eye movement patterns. The data from all of the native Japanese participants were carried forward for analysis (response accuracy for actives, M = 1.00, SD = 0; for passives, M = 0.96, SD = 0.07). The data from 9 L2-learner participants were removed because of their low response accuracy; 7 of them were enrolled in fourth-semester, 1 in sixth-semester, and 1 in eighth-semester Japanese courses. The remaining data, from 20 native speakers of Japanese (8 males, 12 females) and 20 L2 learners of Japanese (10 males, 10 females), were submitted to analysis (response accuracy for actives, M = 0.91, SD = 0.10; for passives, M = 0.76, SD = 0.21).

The accuracy data from the native-speaker and the L2-learner groups were analyzed using logit mixed models (Jaeger, Reference Jaeger2008). We first compared the native-speaker and the L2-learner groups in a single model. The model included the fixed effects condition and group.Footnote ⁶ The random-effect structure included random slopes for participant and for item, together with a by-participant random slope for condition. We used treatment coding for both fixed effects: for condition, with the active condition as the baseline, and for group, with the native-speaker group as the baseline. There was a significant effect of condition (estimate = –1.43, Wald Z = –2.52, p = .011), indicating that the native-speaker group performed less accurately on passive sentences than on active sentences. There was also a significant effect of group (estimate = –2.50, Wald Z = –4.43, p < .001), suggesting that the L2 learners were less accurate than the native speakers in the active condition. In addition, we separately analyzed the L2-learner group using a model that includes fixed effects of condition and proficiency and the interaction of these two effects. The random-effect structure of the model was the same as that of the joint model described earlier. The L2 learners performed less accurately on passive sentences than on active sentences (estimate = –1.29, Wald Z = –2.10, p = .026). There was no effect of proficiency. Table 3 shows the results of the logit mixed models.

Table 3.

Fixed effects in the logit mixed model of response accuracy in native speakers and in L2 learners

Note: L2, Second language.

For each participant, the eye-tracking data was analyzed only for the sentences for which they had clicked the picture object correctly, following the most often employed procedure for selecting the measure of performance in behavioral tasks. This process led to the exclusion of 2% of the native-speaker data and 16% of the L2-learner data. The average track loss was 14% for the native-speaker group and 13% for the L2-learner group. There was no trial with more than 40% track loss, and therefore, all data points were submitted to analysis. For the analysis of the verb disambiguation region, we included a total of 107 data points from the native speakers (59 target-initial trials and 47 distractor-initial trials) and 88 data points from the L2 learners (38 target-initial trials and 36 distractor-initial trials).

Analysis of the critical region

To assess the time course of processing, we used growth curve analysis, adapted for visual-world data (Mirman, Dixon, & Magnuson, Reference Mirman, Dixon and Magnuson2008). Growth curve analysis is a multilevel regression technique that allows for the estimation of inter- and intraindividual patterns that change over time (Raudenbush & Bryk, 2002). Although there is no consensus on the manner in which time course data from the visual-world method should be analyzed (for topics on eye-tracking data analysis, see the vol. 59 special issue of Journal of Memory and Language in 2008), our data characteristics are in line with the assumptions of growth curve modeling. With growth curve modeling, it is possible to analyze processing trajectories that change nonlinearly over time, observed in visual-world paradigm data when the proportion of fixations at the beginning is low and the curve levels off at the tails. The time effect was represented using orthogonal polynomials. This representation allows us to independently assess the various characteristics of the data (Mirman et al., Reference Mirman, Dixon and Magnuson2008). For example, the constant term captures the overall effect of condition for the entire analysis window; this is important because the critical region in the present experiment was in the middle of the sentence, and there may have been some preexisting effects at the onset of the analysis window. Similarly, the linear term captures the overall rate of increase or decrease of the effect over time. Including a higher order term does not change the value or the interpretation of the estimated lower order terms (Bollen, Reference Bollen, Cudeck and MacCallum2007), which makes the results more interpretable.

The analysis was conducted on fixations to the active picture scenes in 100-ms bins for the critical-region duration of 1200 ms. First, we analyzed all participants in a single model, directly comparing the effect between the native speakers and the L2 learners. We used treatment coding for both fixed effects: for condition, with the active condition as the baseline, and for group, with the native-speaker group as the baseline. This choice of coding is suitable because we are interested in examining the effect of condition, controlling for group, and the effect of group, controlling for condition. Having the active condition and the native-speaker group as baselines helps us determine how native speakers of Japanese process active and passive sentences and whether L2 learners’ processing pattern is different from that of native speakers when comprehending active and passive sentences. Our model included fixed effects of group and condition and the interaction of these two effects, a second-order orthogonal polynomial (the linear and quadratic terms), crossed with the fixed effects and their interaction. The model also included a random slope for participant, together with by-participant random slopes for condition, crossed with the linear and quadratic terms. By plotting the fixation trajectories, we judged that the second-order orthogonal polynomial sufficiently characterizes the data, and therefore, the cubic term was excluded (for a discussion of interpreting cubic terms in cognitive psychology data, see Mirman et al., Reference Mirman, Dixon and Magnuson2008). We analyzed the L2-learner data separately, using a model with the fixed effects of condition and proficiency and their interaction, crossed with the linear and quadratic terms. The random-effect structure was the same as in the joint model. In modeling the effect structure, we followed the “keep it maximal” approach (Barr, Levy, Scheepers, & Tily, Reference Barr, Levy, Scheepers and Tily2013). All growth curve analyses in the current study were performed using the lme4 package in the statistical software environment R (Bates, Maechler, Bolker, & Walker, Reference Bates, Maechler, Bolker and Walker2014).

Analysis of the verb disambiguation region

To assess whether the native Japanese speakers and the L2 learners were sensitive to morphological information from the passive verb, we conducted an onset-contingent analysis to track the participants’ eye movements separately for target-initial trials (i.e., those in which the participants started by looking at the passive picture scene) and distractor-initial trials (i.e., those in which the participants began by looking at the active picture scene; Fernald, Pinto, Swingley, Weinberg, & McRoberts, Reference Fernald, Pinto, Swingley, Weinberg and McRoberts1998; Fernald, Zangl, Portillo, & Marchman, Reference Fernald, Zangl, Portillo, Sekerina, Fernández and Clahsen2008). These trials were divided based on the fixation location at the onset of the verb. In successful comprehension, the participants who start with looks at the target picture scene should keep looking at the target, whereas the participants who initially look at the distractor should shift quickly to the target picture (Fernald et al., Reference Fernald, Pinto, Swingley, Weinberg and McRoberts1998, Reference Fernald, Zangl, Portillo, Sekerina, Fernández and Clahsen2008).

We used treatment coding for the location of initial gaze (i.e., initial gaze), with the target-initial trials as the baseline, and also used treatment coding for group, with the native-speaker group as the baseline. We used growth curve analysis of the verb disambiguation region with the fixed effects of group and trial type and interaction of group by trial type, crossed with a second-order orthogonal polynomial. The random-effect structure included a random slope for participant, by-participant random slopes for trial type, crossed with the linear and quadratic terms.