"You are grounded!": Latent Name Artifacts in Pre-trained Language Models
This addresses biases in language models that can perpetuate stereotypes and affect fairness in NLP applications, representing an incremental improvement in bias detection and mitigation.
The paper tackles the problem of latent name artifacts in pre-trained language models, where names like 'Donald' are associated with specific entities or negative sentiment, and demonstrates that these biases can affect downstream tasks such as reading comprehension, but may be mitigated by additional pre-training on different corpora.
Pre-trained language models (LMs) may perpetuate biases originating in their training corpus to downstream models. We focus on artifacts associated with the representation of given names (e.g., Donald), which, depending on the corpus, may be associated with specific entities, as indicated by next token prediction (e.g., Trump). While helpful in some contexts, grounding happens also in under-specified or inappropriate contexts. For example, endings generated for `Donald is a' substantially differ from those of other names, and often have more-than-average negative sentiment. We demonstrate the potential effect on downstream tasks with reading comprehension probes where name perturbation changes the model answers. As a silver lining, our experiments suggest that additional pre-training on different corpora may mitigate this bias.