Whether the sentence terminal word meets participants’ prediction is reflected by its cloze probability, the percentage of a specific word that a normative group use to complete a sentence. Take the discussed sentence as an example, the sentence
“Tom has washed all the dishes on the ______” could be completed by “table,”
“countertop,” and other reasonable alternatives. If “table” is offered by 90% and
“countertop” by 5% of the normative group, “table” would be categorized as a high-cloze word while “countertop” as a low-cloze word. Behavioral studies had found a facilitating effect in processing high-cloze words (i.e., shorter response time) compared with processing low-cloze words (Schwanenflugel & Shoben, 1985). As for the ERP studies, the results showed that the amplitude of N400 component was graded with the cloze probability of a terminal word. Specifically, low-cloze words elicited larger N400 amplitude than high-cloze words (Federmeier et al., 2007; Kutas
& Hillyard, 1984; Van Petten & Luka, 2012). Some researchers believed that N400 indicated the degree of easiness in contextual integration (e.g., Kutas & Federmeier, 2000).
In addition to the amplitude of N400, the ERP components of high-cloze words and low-cloze words also differ in the late time window (between 500 to 900ms). In a
systematic survey, Van Petten and Luka (2012) detected that when sentences were completed by unexpected endings (low-cloze words), a late positivity was found at the frontal site. The scalp topography was different from a canonical P600, which was found to be larger at the central-parietal site, and was sensitive to syntactic violations (Osterhout & Holcomb, 1992). Researches in the past seemed not notice the late frontal positivity effect, so they left this time window unanalyzed. According to Van Petten and Luka (2012), among 18 studies that had analyzed this time window, 17 of them reported a frontal rather than parietal positivity. However, these studies had different interpretations toward this positivity. Federmeier et al. (2007) argued that the late frontal positivity indicated a need to reallocate resources to revise the prediction.
As for Thornhill and Van Petten (2012), they considered the late frontal positivity an indication of a lexical form mismatch. Another perspective is proposed by Kutas (1993), who speculated that the frontal positivity reflected inhibition of words that were expected but not presented. Nevertheless, as Van Petten and Luka (2012) pointed out, Kutas’ proposal was a fairly strong claim and should be further tested.
Admittedly, the exact cognitive function indexed by this frontal positivity remained unclear. The consensus so far was that the late frontal positivity indicated a conceptual or lexical mismatch between the expected word and the presented word.
While low-cloze words were less expected candidates in a sentence, the least expected case was created by sentences ended with a semantically inappropriate word.
This line of research examined the differences between semantic congruous and incongruous sentences. Kutas and Hillyard (1980a) were researchers who initiated this line of research.3 Their congruous and incongruous materials were sentences like
“It was his first day at work” and “He spread the warm bread with socks.” The ERP
3 In fact, it is in Kutas and Hillyard (1980a) that the effect of N400 is discovered.
recordings revealed that compared to congruous sentences, a larger N400 was elicited from the onset of the problematic word in incongruous sentences. A considerable amount of researches had been devoted to this line of research, and they all achieved similar results (Andrews et al., 1993; Revonsuo, Portin, Juottonen, & Rinne, 1998).
Recently, Van Petten and Luka (2012) added that while processing incongruous sentences, N400 was usually followed by a large positivity. The positivity was larger at the parietal site, which was fairly similar to the scalp distribution of a canonical syntactic P600/LPC. In fact, researches of complex sentences processing also found such a late parietal positivity. Specifically, their experiment stimuli were semantically incongruous and did not violate syntactic regulations (Kim & Osterhout, 2005;
Kuperberg, Sitnikova, Caplan, & Holcomb, 2003). Based on these findings, P600/LPC was interpreted as an attempt to revise/reanalyze the interpretation of the relatively complex sentence. Van Petten and Luka (2012) contended that the dichotomy of “syntax P600 and semantic N400” could be reconsidered. The late parietal positivity could be elicited by incongruous sentences, and it reflected an attempt to reanalyze the problematic sentences.
In sum, compared with high-cloze words, less expected low-cloze words elicited not only larger N400 but also larger late positivity at the frontal site. As for the least expected incongruous sentence completions, they elicited a larger N400 and a larger late positivity at the parietal site compared with congruous sentence completions. It could be possible that since the posterior positivity was too strong, existing studies did not examine frontal positivity in incongruous sentences. Van Petten and Luka (2012) argued that the two signals, late frontal positivity and late parietal positivity, should arise from different brain regions and should be attributed to different functional processes. In the current study, we will include both high-cloze
and low-cloze words and we predict that a late frontal positivity will be observed in low-cloze word conditions.
2.3 Expectedness of Sentence Code
2.3.1 Code-mixing
What we have discussed so far are studies on expectedness of sentence processing within one language. In a bilingual speech community, people tend to switch languages back and forth in conversations. The mixing use of two languages within one sentence was defined as code-mixing (Bokamba, 1989). Taken language to be used as an unpredictable factor, researchers initiated several lines of studies and each line had its own interest of focus. Some studies examined code-mixed word recognition in contexts (Li, 1996; Chien, 2000), some explored the effect of ambiguous cognates/interlingual homographs (Altarriba, Carlo, & Kroll, 1992; Duyck, Van Assche, Drieghe, & Hartsuiker, 2007; Schwartz & Kroll, 2006), and still some investigated code-mixed sentence processing (Chen, 2004; Moreno, Federmeier, &
Kutas, 2002). Since the former two lines of studies are beyond the scope of the current study, our discussion will be focused on code-mixed sentence processing.
In early off-line studies, Kolers (1966b) and Macnamara and Kushnir (1971) found that as the number of code-mixed words increased in a text, participants’
reading time also increased. However, the methods of both studies were criticized. It was argued that other factors, such as the grammaticality of experiment stimuli and the requirement of the task, could all confound with the results (Grosjean, 1982;
Paradis, 1980c). Recently, as computers and other new techniques were applied into
research, on-line experiments were conducted to investigate this inquiry. Under such a paradigm, participants were presented with code-mixed/non-mixed sentences and were requested to make a response by button-pressing. Participants’ response times were recorded and analyzed. The results of a majority of studies had shown that response times to non-mixed sentences were significantly shorter than mixed ones (Liao & Chan, 2011, 2012; Proverbio, Leoni, & Zani, 2004). Notice that such a processing cost was absent in Chen (2004). However, it was possible that the insignificant effect in Chen was due to the limited number of participants (n = 10).
In neurolinguistic researches, both fMRI and ERP techniques were adopted to investigate language switching. fMRI studies revealed that processing code-mixed sentence was a complex task. Code-mixed sentence processing was accomplished by an extensive neurological network, and there was no specific brain region specialized for language switching. Compared with non-mixed conditions, Wang, Xue, Chen, Xue and Dong (2007) reported that code-mixed conditions activate brain regions in the right superior prefrontal cortex, left middle and superior frontal cortex, and right middle cingulum and caudate. Abutalebi et al. (2007) presented similar results. That is, language switching involved bilateral prefrontal and temporal associative regions. In addition, Abutalebi et al. also investigated differences between code-switching (i.e., switching languages between sentences) and code-mixing (i.e., switching languages within a sentence frame). They discovered that while processing code-switched sentences activated brain regions related to lexical processing, processing code-mixed sentences entailed brain structures related to syntactic and phonological processing.
As for ERP studies, Moreno et al. (2002) was the first study that adopted the ERP technique to examine language switching. The experiment materials were English sentences completed by an English expected word, by a lexical switch (an
English synonym of the expected word), and by a code switch (the Spanish equivalent of the expected word). The ERP recordings revealed that while lexical switch condition elicited an N400 effect, code switch condition elicited a left anterior negativity (LAN) between 250 to 450 ms, which was more sensitive to syntactic processing, especially those that taxed working memory (King & Kutas, 1995).
Moreno et al. proposed that if this effect was a LAN, then it might be resulted from different morphological agreement systems in Spanish and English. Still, other language switching studies (Alvarez, Holcomb, & Grainger 2003; Chen 2004;
Proverbio et al., 2004) consistently reported an N400 effect rather than a LAN in this time window. Compared with an expected ending, language switching seemed to cause problems in lexical integration.
Back to the study of Moreno et al. (2002), another major difference between the lexical switch condition and the code switch one was observed in a late time window. Only the code switch condition elicited a large positive complex (LPC, also known as P600), especially at the parietal and occipital sites. Moreno et al. argued that the P600/LPC component revealed that participants required more resources for stimulus evaluation. In fact, although P600/LPC component was originally proposed as an indication of syntactic violations (Osterhout & Holcomb, 1992), as has been discussed in Session 2.2, recent studies suggested that P600/LPC could reflect a process of sentence reanalysis, such that complicated sentences (e.g., semantically incongruous sentences, garden-path sentences, etc.) would elicit this effect (Kim &
Osterhout, 2005; Osterhout, Holcomb, & Swinney, 1994). In addition, it was suggested that P600/LPC could be related to executive control in sentence level (Kolk
& Chwilla, 2007): schizophrenic patients, who may have difficulties in monitoring speech perception, showed a reduced P600/LPC effect (Kuperberg, Sitnikova, Goff, &
Holcomb, 2006). While it seemed that P600/LPC could be associated with various cognitive functions, in the studies of language switching, P600/LPC was usually considered an indication of sentence reanalysis, with its amplitude correlated to participants’ language proficiency (Moreno et al., 2002; Proverbio et al., 2004).
Evidence from both psycholinguistic and neurolinguistic studies have shown that language switching requires cognitive resources. However, when switching directionality (i.e., switching from Language A into Language B versus switching from Language B to Language A) is taken into consideration, the results may not be so consistent. Specifically, a considerable amount of studies had found that the processing cost in two switching directionalities was asymmetric (Alvarez et al., 2003;
Liao & Chan, 2011, 2012; Meuter & Allport, 1999). To further look into this phenomenon, switching directionality is discussed in more detail in the following session.
2.3.2 Switching Directionality
2.3.2.1 Theoretical Models
Psycholinguists were amazed at bilinguals’ capability of switching languages back and forth. Many models were proposed to account for bilingual language processing, and each model had its own focus of interest. Some of the models emphasized on whether bilinguals have separated mental lexicons (e.g., Distributed Feature Model, De Groot, 1992a); some aimed to address the developmental stages of acquiring a new language (e.g., Revised Hierarchical Model, Kroll, & Stewart, 1994;
Word Association Model and Concept Mediation Model, Potter, So, Van Eckardt, &
Feldman, 1984); still some concerned levels of control/activation in bilinguals’ two languages and its influence in language switching (e.g., Inhibitory Control Model, Green, 1998; Bilingual Interactive Activation Plus model; Dijkstra & van Heuven, 2002). For the purpose of the current study, we will focus on the control/activation in bilinguals and discuss the Inhibitory Control model (ICM henceforth) and the Bilingual Interactive Activation Plus model (BIA+ henceforth).
Green’s ICM (1998) is a model specific to language production. In this model, a planned concept activated both the bilingual lexico-semantic system and the supervisory attentional system (SAS) (see Figure 1). While in the bilingual lexico-semantic system all the lexical words were tagged for their language information, the SAS regulated the language task schema, which inhibited the lemmas of an unintended language and activated lemmas of an intended language in the bilingual lexico-semantic system. As lemmas were selectively activated/inhibited based on the requirements of the task, it was implicated that the language as a whole would be affected. In addition, since lexical nodes in L1 were usually highly activated, a stronger inhibition of L1 was needed to perform a task in L2. If the required task was to switch language from L2 back into L1, more efforts were required to reactivate lemmas in L1. In other words, switching from L2 to L1 would be more difficult than the other way round. Besides, this model distinguished word identification system from decision required by a particular task.
Figure 1: The Inhibitory Control Model (ICM), proposed by Green (1998)
As for BIA+ (Dijkstran & van Heuven, 2002), it is a language comprehension model. BIA+ presupposed that there was an integrated lexicon of L1 and L2, and the access to words was nonselective. The processing of input information was initiated in a bottom-up manner. In fact, BIA+ contained two separated systems, respectively a word identification system and a task/decision system (see Figure 2). In the word identification system, the orthographic, phonological, and semantic representations formed a highly interactive network with language nodes. As for the task/decision system, it included all the non-linguistic information that could affect word processing (e.g., experiment instruction and participants’ strategy) and reset the parameters for attention allocation. Moreover, this model suggested that the resting-level activation of words in different languages reflected their frequency of usage. In most of the cases, a bilingual’s L1 was more frequently used than his L2.
Therefore, the resting-level activation of words in L1 was higher. Since the representations of L1 were more active in a bilingual’s mind, the cross-linguistic influence from L1 to L2 would be larger. In other words, sentences switched from L1 to L2 would be more difficult to process than the other way round.
Figure 2: The Bilingual Interactive Activation Plus model (BIA+), proposed by Dijkstra and van Heuven (2002)
While both ICM (Green, 1998) and BIA+ (Dijkstra & van Heuven, 2002) shared the idea that the word processing system should be separated from the task/decision system, their predictions varied greatly since production and perception were different in nature. To begin with, in a production model (such as ICM), language selection was presumed to happen at an early stage. Cognitive control was needed to inhibit the language that was not currently used (Ye & Zhou, 2009). However, in a language comprehension model (such as BIA+), language selection happened at a fairly late stage. As hypothesized by BIA+, the language nodes were not activated until relatively late in the processing sequence. The role of cognitive control was to help bilinguals resolving sentence ambiguities and arriving at an appropriate interpretation (Ye & Zhou, 2009). The last difference between the ICM and BIA+ was the asymmetric processing cost they predict. ICM argued for the effect of inhibition. To perform a task in L2, the dominant L1 should be strongly inhibited. Since the inhibition of L1 was stronger than that of L2, more efforts were needed to reactivate it.
Thus, switching from L2 into L1 would be more difficult. In contrast, BIA+
contended that words in L1 and L2 might have different resting-level of activations.
Since L1 tended to be more frequently used than L2, the resting-level activation of words in L1 should be higher. In addition, because L1 was more dominant in the participants’ mind, the cross-linguistic influence from L1 to L2 would be larger. Thus, switching from L1 into L2 would be more difficult. Since the current study investigates bilingual language switching in a comprehension mode, in the following session, we review studies that test BIA+ and studies concern switching directionality in language comprehension.
2.3.2.2 Empirical Studies
As proposed by BIA+ (Dijkstra & van Heuven, 2002), the frequency of using L1 and L2 would result in different resting-level of activations in the two languages. In most cases, L2 is not as frequently used as L1, and thus the representations of L2 are not as activated as L1. Temporal delay of processing L2 had been attested by a series of studies (Ardal, Donald, Meuter, Muldrew, & Luce, 1990; Moreno & Kutas, 2005).
In fact, BIA+ further suggested that switching from L1 to L2 would be more difficult than the other way round. To examine this effect, researchers examined language switching between decontextualized lexical items (Alvarez et al., 2003; Wang, 2008).
The results confirmed the prediction of BIA+, showing that switching from L1 to L2 should tax more efforts than the other way round. Since processing cost of language switching in sentence context was not presupposed by BIA+, researchers further inserted the lexical items into a sentence context (so that the experiment stimuli were
code-mixed sentences), hoping to uncover whether such a cost exists in sentence level.
Inconsistent results arose among different studies. Chen (2004) reported that there was no cost in processing code-mixed sentences, and that there was no difference between two switching directionalities. As has been discussed in Session 2.3.1, the insignificant results between Chen’s non-mixed and code-mixed conditions could be resulted from her limited number of participants. In addition, some potential flaws in Chen’s design made her conclusion less persuasive. For instance, Chen’s two experiments respectively looked into one of the two switching directionalities.
However, the complexity of stimuli in her two experiments seemed to be different, and it would thus be problematic to directly compare how participants responded to them.
Different from Chen’s results (2004), the behavioral data in Proverbio at al.
(2004) revealed that there was a processing cost in code-mixed conditions.
Specifically, the response times to sentences switched from L1 to L2 were longer than the other way round. Similar results that switching from a more dominant language to a less dominant one was more difficult could be found in Liao and Chan (2011, 2012).
In these studies, we examined how bilinguals in Taiwan processed Mandarin-Taiwanese code-mixed sentences. Response times to the target words were measured when the participants performed a sentence comprehension task. The recruited participants were all simultaneous bilinguals, who acquired both languages before the age of six, and were proficient in both languages. In addition, the participants used Mandarin more frequently than Taiwanese (as indicated from the questionnaire). Thus, the profile of bilinguals’ language background in our studies was different from that in existing literatures, in which the participants’ L2 was
usually acquired after critical age with various proficiency levels. Still, our results revealed that while comprehending sentences switched from Mandarin to Taiwanese required extra time, comprehending sentences switched from Taiwanese to Mandarin did not. We suggested that Mandarin was more dominantly used, so it was more active in the participants’ mind and thus was easier to retrieve. The results seemed to be in line with the prediction of BIA+.
Results from behavioral studies showed that switching from L1 to L2 (or from a more dominant to a less dominant language) seemed to be more difficult, fMRI studies reported a similar pattern. Both Abutalebi et al. (2007) and Wang et al. (2007) agreed that switching from a more dominant language into a less dominant one required more cognitive efforts. Specifically, brain regions such as bilateral frontal cortices and anterior cingulate cortex would be activated. These brain regions were related to executive functions. In other words, it was suggested that switching into a less dominant language required cognitive control.
As for ERP studies, N400 and LPC/P600 were the components usually found among language switching studies (Alvarez et al., 2003; Brenders, 2004; Chen, 2004;
Moreno et al., 2002; Proverbio et al., 2004). N400 was reported to be larger in switching from L1 to L2, and was interpreted as encountering difficulties in lexical interpretation. However, the scalp topography of N400 was slightly different among these studies. For instance, N400 was found to be larger in the midline in Alvarez et al., but larger in the bilateral electrodes in Proverbio at al. As for LPC/P600, it seemed that this component was relatively insensitive to switching directionality. Some of the
Moreno et al., 2002; Proverbio et al., 2004). N400 was reported to be larger in switching from L1 to L2, and was interpreted as encountering difficulties in lexical interpretation. However, the scalp topography of N400 was slightly different among these studies. For instance, N400 was found to be larger in the midline in Alvarez et al., but larger in the bilateral electrodes in Proverbio at al. As for LPC/P600, it seemed that this component was relatively insensitive to switching directionality. Some of the