To enable naturalistic conversational agents for multilingual users, dialogue systems need to be extended to converse with bilinguals, potentially using multiple languages in an utterance (i.e. code-mixing). Yet little is known about human preferences for code-mixing in the context of a dialogue. To fill this gap and to study preferred code-mixing styles, we incorporate linguistically-motivated strategies of code-mixing into a rule-based goal-oriented dialogue system.
We collect a corpus of 587 human–computer text conversations between our dialogue system and fluent Spanish–English bilinguals. From this new corpus, we analyze the amount of elicited code-mixing, types of code-mixing strategies people use, and whether they entrain to the system’s code-mixing. Based on these exploratory findings, we give recommendations for future code-mixing dialogue systems.