NegT5: A Cross-Task Text-to-Text Framework for Negation in Question Answering

Abstract

Negation is a fundamental grammatical construct that plays a crucial role in understanding QA tasks. It has been revealed that models trained with SQuAD1 still produce original responses when presented with negated sentences. To mitigate this issue, SQuAD2.0 incorporates a plethora of unanswerable questions to enable pre-trained models to distinguish negative inquiries. In this study, we assess the performance of the model on answerable and unanswerable questions that incorporate negative words and find out that the model’s performance on unanswerable negative questions surpasses the baseline. However, the model’s performance on answerable negative questions falls short of the baseline. This outcome prompts us to surmise that SQuAD2.0 includes a substantial number of unanswerable questions, but the pattern of these questions is typically limited to the addition of negative adverbs such as “never” and “not”. As a result, the trained model tends to produce “unanswerable” responses when confronted with questions that contain negative expressions. To address this issue, we propose a novel framework, called NegT5, which adopts the text-to-text multi-task fine-tuning principle introduced in T5 for making the model able to deal with negation in QA.