On Commonsense Cues in BERT for Solving Commonsense Tasks
This addresses the problem of understanding model reliance on spurious associations versus genuine cues in commonsense reasoning for AI researchers, though it is incremental as it builds on prior work on BERT's commonsense capabilities.
The study investigated whether BERT uses structural commonsense cues for solving commonsense tasks, finding that it does rely on relevant knowledge and that this presence correlates positively with model accuracy.
BERT has been used for solving commonsense tasks such as CommonsenseQA. While prior research has found that BERT does contain commonsense information to some extent, there has been work showing that pre-trained models can rely on spurious associations (e.g., data bias) rather than key cues in solving sentiment classification and other problems. We quantitatively investigate the presence of structural commonsense cues in BERT when solving commonsense tasks, and the importance of such cues for the model prediction. Using two different measures, we find that BERT does use relevant knowledge for solving the task, and the presence of commonsense knowledge is positively correlated to the model accuracy.