Statistical learning of predictive dependencies in the tense-aspect system of a miniature language by English and Thai first language adults
Abstract
The effects of previously learned language(s) on learning, knowledge, and use of a new language have been well documented (Jarvis & Pavlenko, 2008). It is also well known that adults extract statistical patterns of artificial grammars under both incidental and intentional conditions, by drawing on the same implicit statistical learning mechanisms used to learn their natural languages (Hamrick & Rebuschat, 2012; Newport, 2016). However, what is less clear is how statistical patterns from existing language(s) can be recruited to guide initial exposure to a novel language. One proposal is that learners form expectations for systematic patterns in a novel language based on regularities they extract from past learning (Pajak, Fine, Kleinschmidt, & Jaeger, 2016). These prior expectations reflect the cumulative experience that learners have with various patterns in their known language(s) and thus are gradient. As learners are exposed to distributional patterns and predictive dependencies of the new language, prior expectations from learners’ known language(s) are utilized to guide adaptation to the novel input. This dissertation investigated whether participants from two first language (L1) backgrounds, English (n = 30) and Thai (n = 66), utilized their prior probabilistic statistical knowledge of tense-aspect encodings to adapt to a novel language. They were presented with a miniature language that expressed the temporal meanings of completion and continuation with both English- and Thai-analogous systematic associations. To illustrate, the completion meaning ‘man read-simple past a book’ was paired in the miniature language input with an English-like encoding (i.e., associations between verbs and morphemes such as ‘harter dola katon-et’) and a Thai-like encoding (i.e., associations between aspect markers such as ‘harter dola katon lon-beng’). At four different time points during the two-day training period, participants were assessed on their accuracy on and rate of adaptation to the two tense-aspect encodings. All participants also completed an English tense-aspect cloze test, and the Thai L1 participants completed an English L2 proficiency test. Data were analyzed using Bayesian multilevel regression models. Regarding the accuracy of adaptation, results show that participants adapted more accurately to their L1-analogous encoding, when compared to the non-analogous one, and that this was true for both L1 background groups. However, the pattern of results for the rate of adaptation was unexpected: participants adapted to the two encodings at a similar rate across both L1 backgrounds, and only the subset of Thai L1 participants with lower English proficiency showed the anticipated rate of adaptation. In addition, variable sensitivity to systematic associations underpinning the English Simple Present and Past, as measured by the cloze test, showed a positive trend toward predicting English L1 and Thai L1 participants’ adaptation to the English-analogous encoding of the miniature language. These findings suggest that adaptation to distributional patterns of a novel language can be mediated by expectations for statistical regularities gleaned from past learning. Conceptualizing prior linguistic knowledge as prior expectations requires researchers to operationalize and measure prior knowledge not as a dichotomous variable such as L1 background (i.e., as when tense morphology is present or absent in the L1 vs. L2), but as a continuous variable that can be incorporated into a statistical model.
