THEORIES OF L2 MORPHOLOGICAL PROCESSING

Ullman (Reference Ullman2001, Reference Ullman2004; see also Pinker, Reference Pinker1999; Pinker & Ullman, Reference Pinker and Ullman2002) proposed that native speakers’ lexical and grammatical processing depend on two different memory systems: respectively, declarative memory and procedural memory. Declarative memory is used for the learning and knowledge of facts and events, and it can be recollected consciously; for language, it is the memory system that allows native speakers to associate meanings to word forms. In contrast, procedural memory is used for the learning and control of skills and habits, and it cannot be recollected consciously; for language, it is the memory system that enables native speakers to compute grammatical structures (e.g., phonology, morphology, syntax, and semantics). To explain L2 learners’ poorer ability to compute grammatical structures (compared to native speakers) but their relatively intact ability to learn lexical words, Ullman (Reference Ullman2001) proposed that L2 learners increasingly rely on declarative memory as their age of first exposure to the target language increases, and that their reliance on declarative memory should decrease as their exposure to and use of the target language increases (Ullman, Reference Ullman2004). This does not imply that multimorphemic words in the lexicon are not analyzed but that the lexical analyses are not procedural computations: they are instead derived from pattern recognition or from rules learned during pedagogical instruction. Thus, according to Ullman's model, the processing of grammatical structure, including agreement morphology, is qualitatively different for native and nonnative speakers, at least nonnative speakers at lower levels of proficiency.Footnote ¹

One prediction that the declarative-procedural (DP) model makes is that L2 learners should process regularly and irregularly inflected words somewhat similarly. That is, they should store morphologically simple and complex words in declarative memory as single units, whether or not morphological, nonprocedural generalizations have taken place within the lexicon. In the psycholinguistic research on this topic, three main methods have been used to determine how L2 learners (and native speakers) process regularly and irregularly inflected words. The first one examines whether the processing of inflected words is subject to frequency effects; because frequency effects have been associated with lexical storage (i.e., the more frequently a word is encountered, the stronger the memory traces), the DP model predicts that L2 learners should show token frequency effects for both regularly and irregularly inflected words, whereas native speakers should show such effects only for irregularly inflected words. The second method aims to determine whether regularly inflected words prime their stem; if inflected words and their stem have separate lexical entries, as would be predicted by the DP model, then L2 learners’ processing of stems should not be primed by their inflected form, unlike that of native speakers. Finally, the third method examines the electrophysiological brain responses evoked by regular words that were irregularly inflected and irregular words that were regularly inflected; the DP model predicts that both word forms should elicit similar electrophysiological responses for L2 learners but different electrophysiological responses for native speakers.

Using these methods, the existing L2 research on this topic has revealed mixed findings. On the one hand, in a multiple-choice task where participants chose the correctly inflected verb form, Birdsong and Flege (Reference Birdsong, Flege, Do, Domínguez and Johansen2001) found that advanced Korean and Spanish L2 learners of English show stronger frequency effects for irregular verbs (e.g., swam) than for regular ones (e.g., helped). Similarly, in an event-related potential (ERP) study, Hahne, Mueller, and Clahsen (Reference Hahne, Mueller and Clahsen2006) report that both native German speakers and Russian L2 learners of German showed a P600 effect (typically associated with grammatical violations) for regular verbs that were incorrectly inflected (e.g., getanzen “danced”; correct form: getanzt) and an N400 effect (typically associated with lexical/semantic violations) for irregular verbs that were regularly inflected (e.g., gelauft “run”; correct form: gelaufen), consistent with a dual route for processing regularly and irregularly inflected words and unlike what the DP model would predict for L2 learners.

In contrast, in a masked priming study, Silva and Clahsen (Reference Silva and Clahsen2008) found that regularly inflected verbs (e.g., prayed) do not prime their stem (e.g., pray) for advanced Chinese, German, and Japanese L2 learners of English (unlike native speakers). In line with these findings, using lexical decision and masked priming experiments, Neubauer and Clahsen (Reference Neubauer and Clahsen2009) found that advanced Polish L2 learners of German showed frequency effects for both regularly (e.g., gespielt “played”) and irregularly (e.g., gelaufen “run”) inflected words, whereas native German speakers showed frequency effects only for irregularly inflected ones. Similarly, Bowden, Gelfand, Sanz, and Ullman (Reference Bowden, Gelfand, Sanz and Ullman2010) found frequency effects for both regular (e.g., pesco “(I) fish”) and irregular (e.g., pienso “I think”) present-tense verbs in the production latencies of intermediate to advanced English L2 learners of Spanish, but not for regular verbs in the production latencies of native Spanish speakers. These results suggest that L2 learners store morphologically complex words as single units in declarative memory, consistent with the predictions of the DP model. One possibility, which is also consistent with the DP model, is that L2 learners rely more on declarative memory in earlier stages of development. The findings of Osterhout, McLaughlin, Pitkamen, Frenck-Mesitre, and Molinaro (Reference Osterhout, McLaughlin, Pitkanen, Frenck-Mestre and Molinaro2006) and Osterhout et al.'s (Reference Osterhout, Poliakov, Inoue, McLaughlin, Valentine and Pitkanen2008) ERP study are consistent with this hypothesis: for novice English L2 learners of French, morphological anomalies that yield an N400 effect shortly after the onset of instruction eventually yield a P600 effect after several months of instruction (for a proposed development, see Steinhauer, White, & Drury, Reference Steinhauer, White and Drury2009).

A different approach to L2 learners’ processing of agreement morphology is Clahsen and Felser's (Reference Clahsen and Felser2006a) shallow structure hypothesis (SSH), according to which late L2 learners compute shallower and less detailed syntactic structures, and rely more on lexical, semantic, and pragmatic information than native speakers. SSH claims that the processing of grammatical structure is qualitatively different for native and nonnative speakers, regardless of proficiency in the target language. This hypothesis was proposed to explain reduced or absent structure-based effects in L2 learners’ processing of temporary syntactic ambiguities and filler-gap dependencies (e.g., Felser & Roberts, 2007; Felser, Roberts, Marinis, & Gross, Reference Felser, Roberts, Marinis and Gross2003; Marinis, Roberts, Felser, & Clahsen, Reference Marinis, Roberts, Felser and Clahsen2005; Papadopoulou & Clahsen, Reference Papadopoulou and Clahsen2003). Whereas Clahsen and Felser (Reference Clahsen and Felser2006a) originally claimed that highly proficient L2 learners may process agreement morphology in a similar manner as native speakers (unlike syntax), in more recent studies, Clahsen and colleagues suggested that L2 learners may be less affected by internal morphological structure than native speakers (for a discussion, see Clahsen, Felser, Neubauer, Sato, & Silva, Reference Clahsen, Felser, Neubauer, Sato and Silva2010). Clahsen and Felser (Reference Clahsen and Felser2006b) also suggest that L2 learners may have difficulty computing agreement dependencies, with nativelike processing being restricted to local domains such as “morphosyntactic agreement between closely adjacent constituents” (p. 111). This then makes the prediction that although L2 learners may be able to decompose morphologically complex words, they should have difficulty computing agreement when it is not local.

Keating (Reference Keating2009) tested this prediction in an eye-tracking study on gender agreement with beginner, intermediate, and advanced English L2 learners of Spanish. He found that advanced L2 learners, but not beginner or intermediate L2 learners, were sensitive to gender agreement violations, and, unlike native speakers, they showed this sensitivity only when the agreement dependency was between two adjacent words (e.g., *una fiesta pequeño “a small-masc party-fem”). Keating concluded that his results support SSH, but he admitted that linear and structural distances were confounded in his study. Using self-paced reading experiments, Sagarra (Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007) and Sagarra and Herschensohn (Reference Sagarra and Herschensohn2010) similarly found that intermediate English L2 learners of Spanish, but not beginner L2 learners, were sensitive to gender and number violations that occurred between two adjacent words (e.g., *el prototipo famosa “the prototype-masc famous-fem”), whereas Jiang (Reference Jiang2004, Reference Jiang2007) reported that, unlike native speakers, intermediate to advanced Chinese L2 learners of English were not sensitive to number agreement violations in sentences where the two words in the agreement dependency were not adjacent (e.g., *The bridge to the island were about ten miles away). The distance between two words in an agreement dependency may thus contribute to L2 learners’ lack of sensitivity to number agreement violations.

In contrast, Foote (Reference Foote2011) found that both advanced English L2 learners of Spanish and native Spanish speakers slowed down when parsing gender and number agreement violations, even when the two words in the agreement dependency were not adjacent (e.g., *El pollo del taco está rica pero picante “The chicken-masc of the taco is tasty-fem but spicy”). Her findings perhaps revealed proficiency differences between her group of L2 learners and those in Keating (Reference Keating2009), suggesting that nativelike attainment is not impossible, even in longer agreement dependencies. Other behavioral studies have indeed shown that high-proficiency L2 learners can perform similarly to native speakers in experiments that require them to use agreement morphology in sentence processing (e.g., Havik, Roberts, Van Hout, Schreuder, & Haverkort, Reference Havik, Roberts, Van Hout, Schreuder and Haverkort2009; Hopp, Reference Hopp2006, Reference Hopp2010), and recent ERP studies have even reported that very advanced L2 learners can detect agreement violations in a qualitatively similar way (i.e., eliciting a LAN effect and a P600 effect) as native speakers (Ojima, Nakata, & Kakigi, Reference Ojima, Nakata and Kakigi2005; Rossi, Gugler, Friederici, & Hahne, Reference Rossi, Gugler, Friederici and Hahne2006); the ease with which L2 learners detect such violations, however, is at least partially dependent on whether the native language has a similar agreement rule (e.g., Foucart & Frenck-Mestre, Reference Foucart and Frenck-Mestre2011; Tokowicz & MacWhinney, Reference Tokowicz and MacWhinney2005).

The results of Jiang (Reference Jiang2004) and Keating (Reference Keating2009) suggest that linear distance between two elements in an agreement dependency can potentially increase the difficulty with which L2 learners process such dependencies. Such a difficulty should arise if long-distance agreement dependencies impose a higher WM load on L2 learners than short-distance ones. We might thus expect a relationship between L2 learners’ ability to process such dependencies and their WM capacity.

A third type of account that predicts such a relationship is McDonald's (2006) cognitive processing account. McDonald (Reference McDonald2006) proposed that L2 learners’ variable use of and sensitivity to grammatical structures is due to difficulties in basic cognitive processes that depend in part on grammatical knowledge (i.e., these processes should improve with increasing proficiency) but that are not grammar specific: “(1) low L2 [working-]memory capacity, (2) poor L2 decoding ability, and (3) slow L2 processing speed” (p. 382). Previous research has indeed shown that learners tend to have a lower WM capacity in the L2 than in the native language (e.g., Harrington & Sawyer, Reference Harrington and Sawyer1992; Osaka & Osaka, Reference Osaka and Osaka1992; Service, Simola, Mesanheimo, & Maury, Reference Service, Simola, Mesanheimo and Maury2002; Van den Noort, Bosch, & Hugdahl, Reference Van den Noort, Bosch and Hugdahl2006), suggesting that potential limitations in proficiency may have a negative outcome on L2 learners’ performance in WM tasks. Furthermore, L2 learners have been found to have more difficulty identifying words in noise (e.g., Mayo, Florentine, & Buus, Reference Mayo, Florentine and Buus1997; Meador, Flege, & MacKay, Reference Meador, Flege and MacKay2000; Van Wijngaarden, Steeneken, & Houtgast, Reference Van Wijngaarden, Steeneken and Houtgast2002) and to have slower processing times than native speakers (for discussion, see Frenck-Mestre, Reference Frenck-Mestre, Heredia and Altarriba2002). McDonald (Reference McDonald2006) proposed that these cognitive processes interact with grammatical processing such that L2 learners’ performance on corresponding cognitive tasks should correlate with their performance on grammatical tasks, and the impact of these cognitive processes should increase as the complexity of grammatical structure increases. Given the research that shows strong ties between L2 WM capacity and L2 comprehension (e.g., Altekin & Erçetin, Reference Alptetin and Erçetin2010; Harrington & Sawyer, 1992; Miyake & Friedman, Reference Miyake, Friedman, Healy and Bourne1998; Service et al., Reference Service, Simola, Mesanheimo and Maury2002; Walter, Reference Walter2004), she predicted L2 grammatical performance to be more strongly dependent on WM capacity in the L2 than on WM capacity in the native language.

To test these predictions, McDonald (Reference McDonald2006) conducted a first set of experiments with L2 learners of English from various language backgrounds and native English speakers. The participants’ grammatical performance was assessed using an auditory grammaticality judgment task that included agreement violations (regular past tense, third person singular, regular plural, present progressive, irregular past tense, irregular plural). WM capacity was assessed by asking the participants to report a list of English words they had heard in order of sizes (i.e., from smallest to largest; for details, see Montgomery, Reference Montgomery2000). Decoding ability was assessed with a gating task in which the participants identified words after hearing segments of the words. Processing speed was assessed by asking the participants to detect words in sentences. Note that L2 learners’ performance on all the cognitive tasks depended in part on their proficiency in English, at least their knowledge of words and phonological processes. The L2 learners’ results on each of the tasks were significantly different from those of the native speakers: L2 learners evidenced poorer grammaticality judgments, smaller WM capacity, poorer decoding abilities (i.e., more gates necessary), and slower word detection times than native speakers. Significant correlations (in the predicted directions) were found between L2 learners’ judgments (accuracy and/or response times) and their WM capacity and decoding ability. Because the word detection task involved the recognition of words in complete sentences, McDonald (Reference McDonald2006) suggested that it may have involved too much syntactic knowledge to be a good measure of processing speed alone.

In a second set of experiments, McDonald (Reference McDonald2006) asked groups of native English speakers to complete the same grammaticality judgment task under processing loads related to either WM capacity (low or high digit loads), decoding ability (listening through white noise), or processing speed (eliciting responses under time pressure or presenting compressed speech). The participants also completed the same WM and decoding (i.e., gating) tasks as in the first set of experiments. The results showed poorer performance under a load (in the second set of experiments) than under no load (in the first set of experiments) for all the groups except for the group under a low memory load. For the grammaticality judgment task, significant correlations were found between the performances of (a) the participants under a memory load and their WM capacity, (b) the participants under a decoding load and their decoding ability (as measured by the gating task), and (c) the participants under time pressure and their response times in the grammaticality judgment task. The performances of the native speakers who had a high memory load or who listened to speech through white noise correlated most with L2 learners’ performance.

The findings of this study, specifically those for native speakers, suggest relationships between the above cognitive processes (i.e., WM capacity, decoding ability, and processing speed) and grammatical accuracy. For L2 learners, because their performance on the cognitive tasks (at least those testing for decoding ability and processing speed) should have been correlated with their general proficiency in English (including lexical and phonological knowledge), it is unclear whether it was these cognitive abilities rather than their general proficiency that were significant predictors of grammatical accuracy. Decoding ability and processing speed, in particular, may be strongly tied to L2 learners’ automaticity in the target language, which increases with proficiency (for discussion, see Segalowitz & Hulstijn, Reference Segalowitz, Hulstijn, Kroll and De Groot2005). Although WM capacity should improve with increasing proficiency (e.g., Service et al., Reference Service, Simola, Mesanheimo and Maury2002; Van den Noort et al., Reference Van den Noort, Bosch and Hugdahl2006), it is likely to be less dependent on proficiency than decoding ability and processing speed, as native speakers show a great deal of variability in their WM capacity (e.g., Baddeley & Wilson, Reference Baddeley and Wilson1985; Daneman & Carpenter, 1980; Roberts & Gibson, Reference Roberts and Gibson2002; Salthouse, Reference Salthouse1990; Turner & Engle, Reference Turner and Engle1989; Waters & Caplan, Reference Waters and Caplan1996a, Reference Waters and Caplan1996b). Hence, by testing for both general proficiency and WM capacity and by disentangling them, researchers may capture a great deal of variation among L2 learners in their sensitivity to agreement morphology. Different models of WM capacity make different predictions for L2 learners’ processing of agreement morphology, however. Let us consider some of these models before we discuss the role of WM capacity in L2 morphosyntactic processing.

WM CAPACITY AND ITS ROLE IN L2 MORPHOSYNTACTIC PROCESSING

The first model of WM, put forward by Baddeley and Hitch (Reference Baddeley, Hitch and Bower1974), consists of a phonological loop, a visuospatial sketchpad, and a central executive component responsible for attention control. The first two systems were proposed to hold speech-based and visuospatial information in temporary stores, whereas the third system was responsible for allocating attention and controlling the information in these stores. More recently, Baddeley (Reference Baddeley2000) proposed a fourth component, the episodic buffer, which represents an interface between the three WM systems and long-term memory (for a discussion, see Baddeley, Reference Baddeley2007).

Whereas much research on auditory language comprehension has focused on the role of the phonological loop (e.g., Gathercole & Baddeley, Reference Gathercole and Baddeley1993), in the area of sentence processing, it is the ability of speakers to process sentences while holding information in memory that has received more attention. Two types of WM theories have been proposed: the single-resource theory, according to which humans have a set of verbal processing resources dedicated to all verbal tasks (e.g., Just & Carpenter, Reference Just and Carpenter1992; King & Just, Reference King and Just1991; MacDonald, Just, & Carpenter, Reference MacDonald, Just and Carpenter1992; Miyake, Carpenter, & Just, Reference Miyake, Carpenter and Just1994); and a separate sentence-interpretation resource theory, according to which part of verbal WM is dedicated to “interpretive” processes, such as assigning syntactic structure to a sentence and using it to understand its meaning, and part of it is dedicated to “postinterpretive” processes, such as using the meaning of the sentence to perform other operations (e.g., Caplan & Waters, Reference Caplan and Waters1995, Reference Caplan and Waters1996; Waters & Caplan, Reference Waters and Caplan1996a, Reference Waters and Caplan1996b). Caplan and Waters (1999) propose that although both interpretive and postinterpretive processes degrade as WM capacity decreases or as WM load increases, interpretive processes are not differentially affected by the size of WM capacity and of WM load. They reviewed several studies showing that speakers with lower reading spans or under a larger WM load do not show a greater effect of syntactic complexity than speakers with higher reading spans or under a smaller WM load; by contrast, speakers with lower reading spans or under a large WM load have more difficulty processing sentences that contain more propositions than speakers with higher reading spans or under a small WM load, suggesting that postinterpretive processes are differentially affected by WM capacity (for a discussion, see Caplan & Waters, 1999).

WM capacity in sentence processing studies is typically measured with reading span tasks, which assess the ability of speakers to process sentences while holding words or digits in memory. One of the well-established reading span tasks is that of Daneman and Carpenter (Reference Daneman and Carpenter1980), in which speakers read aloud sentences and recall the last word of the set of sentences they read, with each set containing from two to six sentences. The participants’ reading span score is the total number of words that were correctly recalled. Waters and Caplan (Reference Waters and Caplan1996a) identify a number of limitations with this task, including the possibility that the participants may process sentences only at a superficial level when reading them aloud and that there might be a trade-off between their focus on the sentence meanings and their word recall. The test may thus reflect the storage component of WM more than the processing component. Instead, Waters and Caplan (Reference Waters and Caplan1996a) developed a reading span task in which the participants judged the acceptability of sentences based on thematic role assignments (from the animacy requirements of verbs), and then recalled the last word of sets of sentences (for other versions requiring the participants to focus on the meaning or acceptability of sentences, see Tirre & Pena, Reference Tirre and Peña1992; Turner & Engle, Reference Turner and Engle1989). The WM score is a composite score that reflects both the processing component (i.e., accuracy and reaction times when judging the acceptability of sentences) and the storage component (i.e., accuracy of the word recall) of WM, with the latter being computed as a function of set sizes (i.e., the highest set size to yield an accurate word recall), thus better capturing the upper limit of the storage component. In L2 studies, however, the dependent variables associated with the processing component of L2 reading-span tests are likely to be correlated with L2 proficiency. Solutions to this problem include to test for WM capacity in the native language (assuming that WM is language independent; e.g., Osaka & Osaka, Reference Osaka and Osaka1992) or to use only the L2 word recall scores as a measure of WM capacity, as it is less likely to covary with proficiency.

A number of studies have investigated the role of WM capacity in L2 sentence processing (for a review of studies that focus on L2 learning in general, see Ardila, Reference Ardila2003; Williams, Reference Williams, Gass and Mackey2011). Some of these morphosyntactic studies suggest that reading span is not a determining factor in L2 learners’ ability to process complex syntactic structures. For example, Juffs (Reference Juffs2004, Reference Juffs2005) report that the ability of Chinese, Japanese, and Spanish L2 learners of English to process sentences that contain temporary syntactic ambiguities (e.g., After the children cleaned the house looked very neat and tidy) or wh- movements (e.g., Who did the woman suggest the manager liked at the office?) is not differentially affected by WM capacity in the native or target language, as measured by Harrington and Sawyer's (1992) reading-span test, which is based on Daneman and Carpenter's (1980) test. In a cross-modal priming study, Felser and Roberts (Reference Felser and Roberts2007) found that high-span native English speakers, but not low span ones (as assessed by Harrington & Sawyer's [1992] reading-span test), showed evidence of antecedent reactivation in indirect object (experimental) gap positions but not in pregap (control) positions of wh- sentences, indicating that they posited a syntactic gap in indirect object position; by contrast, Greek L2 learners of English showed evidence of antecedent reactivation in both experimental and control positions, and their performance was not differentially affected by WM capacity.

In contrast, studies on the processing of agreement morphology suggest that reading span can be good predictor of L2 learners’ ability to process agreement dependencies. Sagarra (Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007) examined the relationship between low-proficiency Spanish L2 learners’ WM capacity and their sensitivity to gender agreement violations. Because L2 learners’ WM capacity has been suggested to be language independent (Osaka & Osaka, Reference Osaka and Osaka1992), and because WM capacity in the target language can be correlated with L2 proficiency, Sagarra (Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007) measured WM capacity in L2 learners’ native language, English. The L2 learners completed a reading-span test adapted from Waters and Caplan (Reference Waters and Caplan1996a), with the scores including both the L2 learners’ word recall and accuracy rates. She found that the L2 learners with higher WM scores processed sentences with gender agreement violations more accurately than L2 learners with lower WM scores. Similarly, Sagarra and Herschensohn (Reference Sagarra and Herschensohn2010) found that the WM scores of their intermediate Spanish L2 learners (assessed in the same manner as in Sagarra, Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007) correlated positively with these L2 learners’ reading times at the critical region of sentences with gender agreement violations and with their grammaticality judgments on the same sentences, suggesting that these L2 learners were more sensitive to gender agreement violations as their WM capacity increased.

Havik et al. (2009) also report WM effects in their study on the processing of subject–object relative clause ambiguities by German L2 learners of English. Their study focused on sentences where agreement on the verb disambiguated between the two different syntactic structures (e.g., Daar is de machinist die de conducteurs heeft/hebben bevrijd uit het brandende treinstel “That is the engine driver who has saved the guards/who the guards have saved from the burning train-carriage”). They manipulated the length of the agreement dependency, such that an adverbial phrase intervened between the relative clause subject and verb (e.g., Daar is de machinist die de conducteurs na het ongeluk met de trein heeft/hebben bevrijd uit het brandende treinstel “That is the engine driver who after the accident with the train has saved the guards/who the guards after the accident with the train have saved from the burning train-carriage”). Their results revealed that only the high-span L2 learners (as measured by Dutch and German versions of Daneman & Carpenter's [1980] reading-span task) had more difficulty in the processing of object relative clauses than in the processing of subject ones, and they did so only when the agreement dependency was short. This suggests that WM capacity influenced their use of agreement morphology for the processing of syntactic structure, but that the long agreement dependencies they tested might have been too taxing for WM capacity to be a good predictor of L2 learners’ performance. Similar results were found for the low-span native speakers, whereas the high-span native speakers showed an effect of relative clause type in both the sentences containing short and long agreement dependencies.

By contrast, Foote (Reference Foote2011) did not find a relationship between her advanced Spanish L2 learners’ sensitivity to gender and number agreement violations as a function of length of the agreement dependency and L2 learners’ WM capacity in Spanish (as assessed by a reading-span test adapted from Waters & Caplan, Reference Waters and Caplan1996a). To compute sensitivity to agreement as a function of distance, she first subtracted the participants’ reaction times on grammatical items from their reaction times on ungrammatical items separately for the two conditions where the agreeing words were and were not adjacent. The latencies that these subtractions yielded in the nonadjacent condition were then subtracted from the corresponding latencies in the adjacent condition. The WM scores were the total number of words that each participant had recalled in the task. This absence of relationship, however, could be due to the way in which she computed the effect of grammaticality as a function of distance or to the way in which she computed WM scores.Footnote ²

In summary, the findings of Havik et al. (2009), McDonald (Reference McDonald2006), Sagarra (Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007), and Sagarra and Herschensohn (Reference Sagarra and Herschensohn2010) indicate that WM capacity can be a good predictor of L2 learners’ ability to process agreement morphology. The present study further investigates the roles WM capacity, as well as the role of proficiency, in the processing of number agreement dependencies as a function of the length of the agreement dependency. It adds to the previous studies on the processing of agreement morphology in that (a) it focuses on a different target language, French; (b) unlike Keating (Reference Keating2009) and Foote (Reference Foote2011), it carefully controls the structure of the sentences in which the agreement dependency is short versus long; and (c) it examines the relationship between L2 learners’ processing of agreement morphology and their WM capacity in both the native and target languages, as well as the relationship between their WM capacity in both languages.

This study focuses specifically on L2 learners’ sensitivity to number agreement violations in third-person direct object clitics that are close to or distant from their antecedents. Third-person direct object clitics in French include le “him/it-masc,” la “her/it-fem,” and les “them.” They agree in gender (when singular) and number with their antecedent. Unlike in English, these pronouns occur in preverbal position (Marie le mangera, “Marie it will-eat”) and they cannot be stressed (*Marie LE mangera; Marie le mangera LUI, “Mary it will-eat IT”). L2 learners’ sensitivity to number agreement is examined in sentences where the object clitic agrees or does not agree with its left-dislocated antecedent (e.g., Ce fruit/Ces fruits Marie le (*les)/les (*le) mangera pour sa collation avant l'entretien, “This fruit/These fruits Marie it (*them)/them (*it) will-eat for her snack before the interview”), and where the object clitic is close to or distant from its left-dislocated antecedent (e.g., Ce fruit/Ces fruits avant l'entretien Marie le (*les)/les (*le) mangera pour sa collation, “This fruit/These fruits before the interview Marie it (*them)/them (*it) will-eat for her snack”).Footnote ³ Left-dislocated phrases are common in spoken French, and they make it possible to test for the participants’ sensitivity to short and long number agreement dependencies without introducing a context prior to the critical sentences.

Because English pronouns also agree in number with their antecedents, the goal of the present study is not to determine whether nonnative speakers can learn to use morphosyntactic information not instantiated in the native language, but rather to investigate the effects of proficiency and WM capacity on L2 learners’ sensitivity to number agreement violations as a function of distance between object clitics and their antecedents. Because the intervening material between the clitic and its antecedent adds to the propositional content of the sentence processed immediately before the clitic, both the single-resource theory and the separate sentence-interpretation-resource theory of WM capacity would predict that L2 learners’ sensitivity to agreement violations in short- and long-distance dependencies will be differentially affected by their WM capacity. This would also be consistent with the findings of the previous studies on L2 learners’ processing of agreement morphology (e.g., Havik et al., 2009; McDonald, Reference McDonald2006; Sagarra, Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007; Sagarra & Herschensohn, 2010). Furthermore, we predict that less proficient L2 learners will be less sensitive to agreement violations than more proficient L2 learners, and L2 learners will be less sensitive to agreement violations when the clitic is distant from its antecedent than when it is close to it.

The participants completed a region by region self-paced reading task that tested these predictions. They also completed an acceptability judgment task that assessed whether they had explicit knowledge of number agreement between object clitics and left-dislocated antecedents. Finally, all the participants completed a reading-span task in French, and the L2 learners also completed a reading-span task in English, both of which were adapted from Waters and Caplan (Reference Waters and Caplan1996a).

GENERAL DISCUSSION

The results of the present study showed that whereas both mid- and high-level L2 learners were able to identify (and correct) number agreement violations in object clitics in an offline acceptability judgment task, only the high-level L2 learners showed sensitivity to such violations in an online self-paced reading task. Of importance, the high-level L2 learners’ sensitivity to number agreement violations did not appear to vary (other than numerically) as a function of the distance between the clitic and its antecedent. Furthermore, despite the tendency for these L2 learners to show more sensitivity to agreement violations in the long-distance condition when the clitic is plural than when it is singular, the results did not show a significant interaction between grammaticality and clitic number, suggesting that these L2 learners were able to detect agreement violations, irrespective of clitic number.

These results are in line with those of Foote (Reference Foote2011), who also found that L2 learners were sensitive to agreement violations irrespective of the length of the agreement dependency, but they differ from those of Keating (Reference Keating2009), who report sensitivity to agreement only when the two words in the agreement dependency were adjacent. In his study, Keating manipulated his sentences by increasing both the linear and structural distance between the two words in the agreement dependency. In our study, other than for the intervening adverbial phrase, which increased the linear distance between the clitic and its antecedent, the sentences in the long-distance conditions had identical syntactic structures to those in the short-distance conditions. One possibility is that the structural distance between two elements in an agreement dependency increases L2 learners’ processing load more than does linear distance. Alternatively, the distance between the clitic and its antecedent in the long-distance condition may not have been sufficiently large to affect the high-level L2 learners’ sensitivity. L2 research that teases linear and structural distances apart and that manipulates the length of agreement dependencies is necessary to determine which of these explanations is likely to be correct.

One question that arises from these results, then, is why the midlevel L2 learners were able to identify (and correct) number agreement violations in the acceptability judgment task and yet failed to show a significant effect of grammaticality in the self-paced reading task. Because these L2 learners showed an effect of clitic number in their results, it is unlikely that they simply failed to parse the plural marking on the object clitic. A more likely explanation is that they failed to establish an agreement dependency between the clitic and its antecedent, possibly due to two factors: the left-dislocated syntactic structure that was used to express the antecedent, and the preverbal position of the object clitic in French, unlike that of object pronouns in English. The midlevel L2 learners may not have yet developed the processing routines necessary to link dislocated phrases to preverbal object clitics, which might then make it more difficult for them to establish an agreement dependency between these phrases and object clitics.

The results of this study also showed that L2 learners had lower reading-span scores in French than native French speakers, consistent with the findings of previous studies (e.g., Harrington & Sawyer, 1992; Osaka & Osaka, Reference Osaka and Osaka1992; Service et al., Reference Service, Simola, Mesanheimo and Maury2002; Van den Noort et al., Reference Van den Noort, Bosch and Hugdahl2006). L2 learners’ acceptability judgments on the WM task in French were found to correlate with their proficiency scores, and their response times and words recalled on the WM task in French were found to correlate with, respectively, their response times and words recalled on the WM task in English. This indicates that WM capacity is at least partially dependent on proficiency, and it covaries with WM capacity in the native language, as has been found in previous work (e.g., Altekin & Erçetin, Reference Alptetin and Erçetin2010; Harrington & Sawyer, 1992; Service et al., Reference Service, Simola, Mesanheimo and Maury2002; Van den Noort et al., Reference Van den Noort, Bosch and Hugdahl2006). These findings confirm that, as a construct, WM capacity is likely to be language independent, but potential limitations in proficiency can have a negative outcome on L2 learners’ performance in such tasks.

WM capacity in French, as computed from the word-recall scores (which did not correlate with proficiency), was also found to be a weak predictor of L2 learners’ sensitivity to agreement violations in sentences containing a plural clitic in the short-distance condition and in sentences containing a singular clitic in the long-distance condition. These results are consistent with those of Havik et al. (2009), who found that only high-span L2 learners could use verbal agreement to extract the syntactic structure of relative clauses in Dutch. In the present study, it is unclear why the relationship between reading span and sensitivity to agreement morphology was not found in the four distance and number conditions. Havik et al. (2009) report that high-span L2 learners did not outperform low-span ones when the distance between the subject and the agreeing verb was too large, potentially because this distance was simply too taxing for the L2 learners. In our study, one might suspect that the sentences with singular clitics in the short-distance condition were the easiest and that those with plural clitics in the long-distance condition were the most difficult, thus perhaps resulting in insufficient variability in L2 learners’ sensitivity to agreement violations for a relationship between it and WM capacity to be found. Because clitic number was not found to interact with grammaticality, this hypothesis is of course speculative. Nonetheless, the results suggest that WM capacity can be a factor in determining whether L2 learners show sensitivity to agreement violations in online sentence processing.

Contrary to predictions, WM capacity did not differentially affect L2 learners’ performance on sentences in the short- and long-distance conditions. Because the intervening material between the clitic and its antecedent adds to the propositional content of the sentence processed immediately before the clitic, both the single-resource theory and the separate sentence-interpretation-resource theory of WM capacity predicted that L2 learners’ sensitivity to agreement violations in short- and long-distance dependencies would be differentially affected by their WM capacity, but this was not the case. As mentioned above, the linear distance imposed between the clitic and its antecedent may not have been sufficiently large to elicit a difference between the short- and long-distance conditions. Further research should further investigate the effect of the size of agreement dependencies on L2 learners’ processing of agreement morphology.

The results also showed that despite the significant correlation between L2 learners’ WM scores in French and English, WM capacity in English was not a significant predictor of L2 learners’ sensitivity to agreement violations. The fact that the word-recall scores in the French task could predict L2 learners’ sensitivity to agreement perhaps suggest that a similar process underlies their storing of French words and their detecting agreement violations in French. These findings, together with the absence of clear distance effect in L2 learners’ processing of object clitics, are compatible with Ullman's (2001, 2004) DP model. The DP model would predict that L2 learners would store singular and plural object clitics as separate lexical entries, and their establishing an agreement dependency between the clitic and its antecedent would not necessarily be more difficult when the clitic is close to its antecedent than when it is distant from it. Although this model does not make predictions that are specific to WM, it might predict that L2 learners’ ability to store words in WM would be related to their ability to process inflected words that are stored as single units in declarative memory. In the case of French clitics, although the plural clitic is formed by adding the plural marker “s” to the masculine singular clitic “le,” the phonology of these forms is suppletive rather than agglutinative, with the plural “s” not being pronounced and with the vowel changing from /ə/ or /a/ to /e/. It is therefore possible that L2 learners store these words as separate lexical entries, although whether or not native speakers do so cannot be established on the basis of our results. Further research should try to tease these two scenarios apart.

Even though the present results are compatible with both views of L2 morphological processing, namely, that L2 learners and native speakers process agreement morphology in a quantitatively or qualitatively different manner, the findings of this study are quite revealing for understanding the factors that regulate L2 learners’ processing of agreement morphology. They indicate that L2 learners become more sensitive to agreement morphology in sentence processing as their proficiency in French increases, and they are more likely to show sensitivity to agreement morphology if they have a high WM capacity in the target language. Models that incorporate both proficiency and WM capacity can thus explain a great deal of inter-L2-learner variability in the processing of number agreement. Theories such McDonald's (2006) cognitive processing account assume that WM capacity is a determining factor in L2 learners’ sensitivity to agreement violations. The DP model assumes that declarative memory plays an important role in L2 learners’ learning of agreement morphology. By integrating the roles of general proficiency, WM capacity, and declarative memory, these two theories can potentially explain a great deal of inter-L2-learner variability, arguably more than Clahsen and Felser's (2006) SSH.

The main limitation of this study perhaps lies in its use of number agreement violations as a diagnosis for L2 learners’ and native speakers’ sensitivity to morphological information in sentence processing. This method, which was also used by Foote (Reference Foote2011), Jiang (Reference Jiang2004, Reference Jiang2007), Keating (Reference Keating2009), Sagarra (Reference Sagarra, Cabrera, Camacho, Déprez, Flores and Sánchez2007), Sagarra and Herschensohn (Reference Sagarra and Herschensohn2010), and can lead participants to alter their normal processing routines either by drawing their attention to the agreement violations or by leading them to develop processing strategies for ignoring these violations. The relationships found between L2 learners’ sensitivity to number agreement violations and their WM capacity and proficiency suggest that the task successfully tapped into L2 learners’ processing of agreement morphology, but further research should examine their processing of agreement morphology in grammatical contexts where it has interpretive consequences (e.g., in cases where it would influence the syntactic structure assigned to the sentence, as in Havik et al., 2009; Hopp, Reference Hopp2006, Reference Hopp2010).