Poster C57, Friday, August 17, 10:30 am – 12:15 pm, Room 2000AB
The Impact of Second Language Age of Acquisition and Language Usage Entropy on Reinforcement Learning among Bilingual Adults in a Non-Verbal Decision Making Task
Mehrgol Tiv1, Jason Gullifer1, A. Ross Otto1, Debra Titone1;1McGill University
We investigated whether bilingual language experience modulates the interplay of two behaviorally and neurally dissociable cognitive systems that contribute to reinforcement learning in a non-verbal decision making task. One system draws on simple reward-driven associations ("model-free") to make choices, and the other leverages information about the environment’s structure and the decision-maker's current context and goals ("model-based") to prospectively plan choices. Of relevance, neural networks related to bilingual language control (e.g., Green & Abutalebi, 2012) overlap with those involved in reward valuation, and could thus manifest as behavioral differences in a decision task that has dynamically shifting reward probabilities. We specifically examined individual differences in second language (L2) age of acquisition (AoA) and bilingual language entropy (entropy = 0 indicates compartmentalized use whereas entropy = 1 indicates an even integration of languages; Gullifer et al., in revision), which we hypothesized could modulate bilingual choice behavior (i.e. model-free or model-based) on a non-verbal decision making task. A total of 45 sequential, bilingual French-English adults performed a two-step decision task (Daw et al., 2011; Otto et al., 2014) where both the transition type between stages and the probability of receiving a reward dynamically varied. We first compared model-free vs. model-based behavior using a factorial mixed-effects logistic regression on “stay” probability as a function of the prior trial’s reward (reward vs. no reward) and transition type (common vs. rare). We then examined the interaction of L2 AoA and language entropy on choice behavior in conjunction with model-free and model-based predictors. Consistent with past work, we found a significant interaction between reward and transition type, indicating that bilingual choice behavior was both model-free and model-based to differing degrees. Crucially, L2 AoA and language entropy jointly modulated choice behavior. Bilinguals having more integrated L1/L2 capacities (earlier L2 AoA, greater language entropy), relied more on simple reinforcement to guide their decisions, whereas bilinguals having more compartmentalized L1/L2 capacities (later L2 AoA, lower language entropy) demonstrated greater sensitivity to model-based reward frequencies. These preliminary results suggest that bilingual adults successfully integrate contextual information with reward in a two-step decision task, but that bilinguals who acquired an L2 later in life and compartmentalize their language usage rely upon model-based learning to a greater degree. There are minimally two possible explanations for this experiential effect. First, it is possible that late compartmentalized bilinguals are more reliant upon context-based cues or recruit additional control capacities when communicating to suppress a highly entrenched L1 during L2 processing. In contrast, early integrated bilinguals may be substantially more practiced at managing language co-activation because L1 processing is less entrenched and L2 processing is more proficient. Second, it is possible that differences in bilingual language experience relate to other cognitive propensities that would impact task performance, such as tolerance of ambiguity or reward uncertainty, given that integrated bilinguals explore unrewarded choices more than compartmentalized bilinguals. In our ongoing work, we are pursuing these options, as well as examining how bilingual experience and reinforcement learning both relate to individual differences in cognitive control.
Topic Area: Multilingualism