Unanswerable Questions Recognition by Semantic Discrepancy Detection

Abstract

Machine reading comprehension (MRC) with unanswerable questions is challenging to the field of natural language processing research. Unlike previous work which ignores the mechanism of answerable and unanswerable, the semantic conflicts detection-based MRC network (SCDNet) was proposed aiming at detections of no-answer (NA) questions through semantic conflicts detection network. The basic idea is that if the given question is unanswerable, there exists semantic absence or conflicts between the question and the reference passages. Therefore, SCDNet predicts the NA probability by checking whether the passage covers the integral semantics of the question. Besides, in order to extract the exact answer from the passage, SCDNet is applied an answer length penalty in the loss function, which helps the learning objective to be more consistent with the evaluation metrics. SCDNet packs the NA question predictor and the answer extractor in a joint model and is trained in an end-to-end manner. Experiments show that SCDNet performs better than some strong baseline models, and achieve an F1 score of 72.43 and 76.96 NA accuracy on SQuAD 2.0 dataset.