Geographical Erasure in Language Generation
This addresses a fairness issue in AI for users affected by biased language generation, but it is incremental as it builds on existing work about biases in LLMs.
The paper tackles the problem of geographical erasure in large language models (LLMs), where certain countries are underpredicting in generated language, and finds that this erasure correlates with low frequencies of country mentions in training data, with mitigation achieved through finetuning using a custom objective.
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. In this work, we study and operationalise a form of geographical erasure, wherein language models underpredict certain countries. We demonstrate consistent instances of erasure across a range of LLMs. We discover that erasure strongly correlates with low frequencies of country mentions in the training corpus. Lastly, we mitigate erasure by finetuning using a custom objective.