Generative Language Models and African American Language:
The What, the How, and the Why
What do generative language models know about African American Language (AAL)? Where are they currently falling short, what do they tell us about the transparency of grammatical patterning in African American Language, and what are the implications of their deficiencies for further model development? This talk explores ongoing work by a cross-disciplinary team on the nature of the relationship between GLMs and the language of African Americans. I’ll report on three interconnected studies: one exploring the nature of the models’ ability to create humanlike language from AAL across four models in a translation task; another exploring the specific linguistic features of AAL which seem to give the models the most difficulties, and a third which shows a probable cause rooted in the AAL representation in the underlying data. I’ll conclude with some observations about what GenAI can tell us about AAL linguistics, and some possible next steps for the future.