The sociolinguistic foundations of language modeling

Maya: Imagine asking a language model a question — something practical — but you write it in African American English. It's the same question, just a different variety of language. The model stumbles. Is that a bug someone can fix, or is it something deeper about what language models fundamentally are? Adam: Grieve and colleagues argue it is the second thing. Their claim is that language models do not model "language." They model specific varieties of language. Whatever texts went into training, that is the variety the model knows. Maya: I would have expected the answer to be yes, training data has gaps, and you fix those gaps by adding more data. That is an engineering problem. Adam: The paper's move is to argue that the framing itself is incorrect. There is no such thing as language in the abstract — only varieties exist. They define a variety as a population of texts bounded by three factors: dialect, register, and period. Maya: Unpack those. Adam: Dialect refers to who produced the texts, including factors like region, class, and ethnicity. Register indicates the context, such as a legal brief compared to a text message; both are in English but are radically different. Period refers to when language changes over time. A specific variety combines all three factors. For example, contemporary conversational Canadian French is a variety, as is standard written American English from the 2010s. Maya: Every corpus is a mix of varieties, though you cannot avoid that. Adam: Right, and Grieve and colleagues say the smallest common variety that encompasses all your texts is what your model learns. Most large corpora are convenience samples — whatever was easy to scrape, which means predominantly standard written English from the internet. Maya: So the bias story isn't a values problem added to an otherwise neutral system. Adam: That's exactly the reframe. The paper distinguishes between two kinds of harm. Quality-of-service harms, where the model performs worse for speakers of underrepresented varieties, and stereotyping harms, where the model reproduces negative associations embedded in the training texts. This tracing back to the work of Bender and colleagues in 2021. Both harms result from which dialects and registers the corpus represents. Maya: What about fine-tuning? Adapting a model to a specific domain — doesn’t that solve the issues? Adam: It helps, but domain adaptation can mitigate the problem. It cannot substitute for representative pretraining data. You are adjusting a model that already has a baked-in variety. Alignment has the same issue — the humans giving feedback speak particular varieties. Therefore, alignment is also variety specific, not universal. Maya: That's the step I hadn't worked through. I had been treating alignment as the place where you impose values on a neutral substrate. But if the substrate already encodes a variety, the feedback layers one variety's values on top of another variety's language. Adam: The paper makes the same point about scale. They cite the Chinchilla finding, which identifies an optimal ratio of approximately twenty tokens per parameter. However, scaling up a biased sample results in a more capable and confident model of the wrong variety. More data amplifies whatever was already present. Maya: So the solution is the corpus itself. Adam: The research agenda they outline involves representative sampling across dialects, registers, and periods. This approach is deliberate, similar to how survey researchers stratify a sample, rather than simply selecting the most accessible texts. They urge natural language processing researchers and sociolinguists to collaborate on this. Maya: What I’m landing on is that this reframes the whole set of problems. Bias, domain adaptation, alignment, and scale — these look like separate issues until you see them through the variety lens. Then they are all symptoms of the same upstream decision about corpus composition. Adam: The practical implication is direct. When a language model fails you — gets your dialect wrong, handles your register poorly, or misses recent usage — there's a corpus decision somewhere upstream. The model is its data, and data is a sociolinguistic choice. Maya: This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.