AI Paper Spotlights - Geographical Erasure in Language Generation
In this talk, Pola Schwöbel, Applied Scientist at AWS, presents her EMNLP 2023 Findings paper on Geographical Erasure in Language Generation. Large language models (LLMs) encode vast amounts of world knowledge. However, since these models are trained on large swaths of internet data, they are at risk of inordinately capturing information about dominant groups. This imbalance can propagate into generated language. The paper investigates and operationalizes a form of geographical erasure wherein language models underpredict certain countries. In the talk, Pola sheds light on the consistent instances of erasure across a range of LLMs. Furthermore, we discover that erasure strongly correlates with low frequencies of country mentions in the training corpus. Lastly, the talk explores how to mitigate erasure by finetuning the LLM using a custom objective function. The talk concludes with a quick overview of other responsible AI projects by the team.