Modeling the effects of entrenchment and memory development on second language acquisition

Thumbnail Image
Publication or External Link
Osthus, Peter Daniel
DeKeyser, Robert
The observation that language learning outcomes are less consistent the older one becomes has motivated a large portion of second language acquisition research (e.g., Hartshorne, Tenenbaum, & Pinker, 2018; DeKeyser, 2012). Hypotheses about the underlying mechanisms which lead to age-related declines are traditionally tested with human subjects; however, many hypotheses cannot be fully evaluated in the natural world due to maturational and environmental constraints. In these scenarios, computational simulations provide a convenient way to test these hypotheses. In the present work, recurrent neural networks are used to study the effects of linguistic entrenchment and memory development on second language acquisition. Previous computational studies have found mixed results regarding these factors. Three computational experiments using a range of languages were conducted to understand better the role of entrenchment and memory development in learning several linguistic sub-tasks: grammatical gender assignment, grammatical gender agreement, and word boundary identification. Linguistic entrenchment consistently had a negative, but marginal, influence on second language learning outcomes in the gender assignment experiment. In the gender agreement and word boundary experiments, entrenchment rarely affected learning outcomes. Starting with fewer memory resources consistently led to poorer outcomes across learning tasks and languages. The complexity of the learning task and the regularity of the formal cues present in the linguistic input affected outcomes. In the gender assignment experiment, the first language influenced second language outcomes, especially when the second language had fewer gender classes than the first language. These results suggest that the effects of entrenchment and memory development on second language learning may be dependent upon the language pairs and the difficulty of the modeling task.