"You are grounded!": Latent Name Artifacts in Pre-trained Language Models

Pre-trained language models (LMs) may perpetuate biases originating in their\ntraining corpus to downstream models. We focus on artifacts associated with the\nrepresentation of given names (e.g., Donald), which, depending on the corpus,\nmay be associated with specific entities, as indicated by next token prediction\n(e.g., Trump). While helpful in some contexts, grounding happens also in\nunder-specified or inappropriate contexts. For example, endings generated for\n`Donald is a' substantially differ from those of other names, and often have\nmore-than-average negative sentiment. We demonstrate the potential effect on\ndownstream tasks with reading comprehension probes where name perturbation\nchanges the model answers. As a silver lining, our experiments suggest that\nadditional pre-training on different corpora may mitigate this bias.\n