Session 5G

Variational learning (VL) is a popular framework for modelling L1 acquisition and language change [1]. Its use in elucidating SLA data has so far remained limited, however [2]. In this contribution, we describe how VL models can be fit to longitudinal learning data, and report that these models sometimes outperform more flexible but less theoretically motivated statistical models in explaining the L2 acquisition of Spanish tense and aspect.
The acquisition of grammatical aspect has been conceptualised in generative approaches as the acquisition of a [±perfective] feature. In Spanish, past tense morphology encodes both tense and aspect, for which two past tenses may be used: while the Preterite marks the completeness of an action performed in the past, the Imperfect denotes imperfective aspect, which may encompass habitual, continuous or progressive meanings. On the other hand, English has an inherent [+perfective] feature [3]. Thus, the L1 English/L2 Spanish speaker's learning task is to acquire a consistent mapping between Preterite/Imperfect morphology and perfective/imperfective semantics.
We employ the LANGSNAP corpus [4], which consists of production data from 27 L2 learners over a period of 20 months. Two native speakers of Spanish analysed the corpus data, tagging the learners' productions of Preterite and Imperfect as either correct or incorrect depending on the semantic context in question. We fit four models to these data using maximum likelihood estimation and Akaike's Information Criterion: NL, a model of no learning that assumes a constant probability of employing the two tenses; OVL, an ordinary VL model in which the learner's accuracy increases monotonically over time towards full accuracy; MVL, a modified VL model which imposes a ceiling on eventual attainment; and GAM, a Generalised Additive Model with cubic splines that can account for nonmonotonic (e.g. U-shaped) learning trajectories.
After pruning 10 learners who attest no change (best-fitting model is NL), we examine the remaining 17 speakers who show evidence of learning over the duration of the corpus. Seven learners are best described by the GAM, as their learning trajectories exhibit nonmonotonic reversals. For the remaining 10 learners, however, the OVL or MVL model produces a fit that is at least as good as, and in some cases better than, the GAM's. Since the OVL and MVL models have a foundation in learning theory - unlike the GAM - we regard this as substantial support for the VL framework. We suggest that future work should, however, look for ways of extending VL models so that nonmonotonic learning trajectories can be derived as explicit mathematical predictions.


[1] Yang, C. D. (2002) Knowledge and learning in natural language. OUP.
[2] Rankin, T. (2014) Variational learning in L2. Linguistic Approaches to Bilingualism, 4(4), 432-461.
[3] Montrul, S. & Slabakova, R. (2002). The L2 acquisition of morphosyntactic and semantic properties of the aspectual tenses Preterite and Imperfect. In Pérez-Leroux A. & Liceras, J. (Eds.), The acquisition of Spanish morphosyntax, 115-151. Kluwer.
[4] Tracy-Ventura, N., Mitchell, R. & McManus, K. (2016) The LANGSNAP longitudinal learner corpus. In Alonso-Ramos, M. (Ed.), Spanish learner corpus research, 117-142. Benjamins.

Previous studies on L1 morphosyntactic attrition have primarily investigated immersed late bilinguals with extensive L2 exposure (e.g., over 5 years in Chamorro et al., 2016; or 6 in Tsimpli et al., 2004). However, little is known about the potential changes experienced in bilinguals' L1 during the first five years of L2 exposure in an immersion setting. Additionally, the (in)stability of an L1 in instructed late bilinguals has also been overlooked. Thus, our study will fill these gaps by exploring how the L1 of L1Spanish-L2English immersed and instructed bilinguals might change in the production and comprehension of subject anaphoric expressions, a domain which has shown to be vulnerable for attriters (Chamorro & Sorace, 2019; Gürel, 2019). 
207 L1Spanish-L2English participants (33 beginners, 81 advanced instructed learners in Spain, and 93 immersed learners in the UK) completed two oral video-retelling tasks and a picture selection task (PST) which tested the predictions from the Position of Antecedent Strategy (Carminati, 2002), i.e., whether null pronouns select subject antecedents and overt pronouns bias towards object antecedents. Data were also collected from a background questionnaire (Bilingual Language Profile, Birdsong et al., 2012), which provided a continuous dominance score. Firstly, the production tasks triggered the use of 3rd person singular subject anaphora in topic continuity (TC), where null pronouns are expected in native Spanish (Lozano, 2016; Martín-Villena & Lozano, 2020). All 3rd person subject anaphoric expressions (N=9225) in TC were analysed using a fine-grained tagset implemented in the UAMCorpusTool (O'Donnell, 2009), which provides χ2 statistics. Secondly, the PST (adapted from Tsimpli et al., 2004) contained 20 experimental items where a main clause, with two potential same-gender antecedents in subject and object position respectively, was followed by a subordinate clause introduced by either cuando('when')/mientras('while') with either a null or an overt pronoun. Participants selected which picture (subject vs. object bias) best matched their interpretation of the sentence. The data from the PST were analysed using linear mixed-effect models in R (Bates et al., 2015), and the final model included Pronoun (overt/null), Conjunction (while/when), the BLP score and Group (beginners/instructed-advanced/immersed-advanced) and their interactions in the fixed-effect structure, and the maximal converging random-effect structure allowed by the design. 
The results from the production tasks show that advanced L1Spanish-L2English bilinguals significantly produce more unexpected overt anaphoric expressions in TC than beginners overall, with significant differences between the two advanced bilingual groups. Additionally, the results from the PST indicate that 1)the two advanced bilingual groups differed from beginners only in the overt pronoun condition, and 2)that the more dominant instructed and immersed participants were in English as measured by the BLP, the more flexible their interpretation of the overt pronoun was, biasing more towards the subject (Chamorro et al., 2016). These results shed light on the unexplored initial stages of L1 morphosyntactic attrition in immersed bilinguals, provide new evidence of variability in the L1 in advanced instructed bilinguals and call for the use of continuous measures to investigate gradience in bilingualism. 

Our ability to engage in fast, efficient language use is facilitated by predictive sentence processing, referring to the rapid integration of various kinds of incoming linguistic and contextual information to pre-activate expected upcoming linguistic information before it is actually encountered in the input. While first-language (L1) speakers have been found to use a wide range of linguistic cues predictively (Huettig, 2015; Pickering & Gambi, 2018), less is known about second-language (L2) predictive processing. L2 speakers appear to exploit semantic information predictively, but studies using morphosyntactic cues have produced mixed outcomes (Godfroid, 2020). Furthermore, much is left to explore regarding the factors that modulate L1 and L2 prediction.
Our study further explores morphosyntax-based prediction by testing whether adult L1 (n = 31) and advanced L2 (n = 30) speakers of German can use the grammatical number information encoded in German verb morphology to generate predictions during listening, and the influence of working memory, awareness of the predictive cue, and several L2-related individual differences thereon. We conducted two visual-world eye-tracking experiments, differing in their target structure: while experiment 1 used productive German verb inflection as a predictive cue (regular ,'weak' conjugation through affixation), experiment 2 focused on unproductive inflection (irregular, 'strong' conjugation through stem allomorphy). Participants heard German sentences (in VSO word order) and had to match each sentence with one of two pictures, displaying identical action scenes but varying in the number of agents. The number information provided by the verb allowed participants to predict whether the upcoming subject would be singular or plural.
Both target structures yielded significant prediction effects – measured as predictive looks towards the target picture and predictive button-presses – in both groups. Prediction started somewhat later in the reaction-time data of the L2 group than in the L1 group, and emerged also later when it was based on stem allomorphy than when it was based on affixes. Higher working memory scores were linked to faster predictive presses. In the L2 group, prediction overall was facilitated when general proficiency was high; moreover, a lower age of onset of learning German had a beneficial effect on affix-based predictive presses, and a higher general frequency of using German in daily life facilitated allomorphy-based prediction. Post-experiment debriefings revealed that predictive processing was of an aware and strategic kind in experiment 1; in experiment 2, only half of the participants had become aware of the stem allomorphy as a predictive cue, yet awareness did not appear to modulate prediction. These findings suggest a clear reliance on subtle morphosyntactic cues during online processing in both L1 and L2 German.


References:
Godfroid, A. (2020). Eye tracking in second language acquisition and bilingualism: A research synthesis and methodological guide. Routledge. https://www.routledge.com/Eye-Tracking-in-Second-Language-Acquisition-and-Bilingualism-A-Research/Godfroid/p/book/9781138024670 
Huettig, F. (2015). Four central questions about prediction in language processing. Brain Research, 1626, 118–135. https://doi.org/10.1016/j.brainres.2015.02.014 
Pickering, M. J., & Gambi, C. (2018). Predicting while comprehending language: A theory and review. Psychological Bulletin, 144(10), 1002–1044. https://doi.org/10.1037/bul0000158