Simulating code-switched sentence production using artificial neural networks
- Chara Tsoukala
- Thursday 7 March 2019
2311 BD Leiden
People who speak several languages are able to switch from one to the other, a process called code-switching, between or even within sentences. The underlying mechanisms, however, are still not well understood; using computational modeling one can simulate code-switching behavior in multilinguals with the goal to explain the process. One way to model human cognition is by employing artificial neural networks (ANN), a.k.a. connectionist models. In this talk, I will give a brief introduction to ANNs and more particularly to the Dual-path model (Chang, 2002), a model of sentence production that is trained on message-sentence pairs and learns to produce a sentence, word by word, given its semantic representation. For instance, the simple message “AGENT=DEF, WAITER; ACTION=EAT;” is expressed in English as “The waiter is eating”. We have extended the Dual-path model to handle two or more languages (Tsoukala et al., 2017) and we have simulated sentence production in early Spanish-English bilinguals and late speakers of English who have Spanish as a native language. We then manipulated language control to allow the model to produce sentences in either language or to code-switch. I will show how the code-switched patterns that the model produced are similar to the ones observed in bilinguals, for instance in the Puerto-Rican community in the US (Poplack, 1980), and how the model can be used to test specific hypotheses regarding code-switched production.