“You are grounded!”: Latent Name Artifacts in Pre-trained Language Models
Vered Shwartz, Rachel Rudinger, Oyvind Tafjord · 2020
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 underspecified 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. ModelMain Corpus Type Gen. Cls.Named Entities from News Named Entities from History Model Minimal News History Infrml Avg Minimal News History Infrml Avg GPT 0.0 7.0 12.7 1.4 5.3 0.0 21.9 39.1 7.8 17.2 GPT2-small 22.5 63.4 50.7 15.5 38.0 12.5 29.7 56.2 12.5 27.7 GPT2-medium 33.8 64.8 49.3 12.7 40.2 21.9 32.8 62.5 4.7 30.5 GPT2-large 43.7 66.2 47.9 16.9 43.7 29.7 29.7 56.2 12.5 32.0 GPT2-XL 50.7 62.0 45.1 21.1 44.7 28.1 31.2 60.9 14.1 33.6 TransformerXL 14.1 18.3 15.5 12.7 15.2 35.9 43.8 51.6 37.5 42.2 XLNet-base 4.2 33.8 12.7 4.2 13.7 0.0 34.4 23.4 3.1 15.2 XLNet-large 11.3 40.8 23.9 9.9 21.5 6.2 29.7 31.2 7.8 18.7 Average 22.5 44.5 32.2 11.8 27.7 16.8 31.7 47.6 12.5 27.1