Language Models as Measurement Apparatus for Culture

Imagine you're trying to use language models to analyze cultural trends, like how people talk in movies or TV shows. The challenge is that these models don't just passively reflect culture; they actively shape it based on how they're built and trained. For instance, if a model is trained on biased data, it might overlook important cultural nuances or reinforce stereotypes. This is what's called the 'erasure of cultural markers.' When you rely on these models without understanding their limitations, you risk misrepresenting the very culture you're trying to study.

This paper argues that we need to be more aware of how our tools influence our understanding of culture. It introduces the idea of the 'agential cut,' which is about recognizing the boundaries we create between the data we collect and the models we use. By examining how these boundaries are drawn, the author highlights that our models often carry the cultural biases of the data they were trained on. The paper includes case studies that show how these issues manifest in real-world applications, like analyzing dialogue in films.

What changes with this approach is that it encourages a more thoughtful design of language models, one that takes into account the cultural implications of their use. For builders, this means being more intentional about the data and methods we choose, ensuring that our models not only perform well but also respect and accurately represent the cultures they engage with.

This work introduces a new theoretical framing for understanding cultural phenomena in NLP.

The claims are supported by case studies, though broader empirical validation is needed.