关键词: 机器阅读理解, 问题边界感知, 答案边界回归, 片段抽取, SQuAD1.1

Abstract: In span extraction machine reading comprehension tasks, methods based on fine-tuning of pre-trained language models have excellent performance. However, such methods commonly use spliced questions and contexts as input and predict start and end boundaries by fine-tuning to obtain answers directly, leading to the problem that the models do not predict boundaries accurately enough. To alleviate this problem, a boundary-aware approach is proposed in this paper. The method emphasises the question boundaries and incorporates the perceived semantic information of the question boundaries into the answer boundary regressor, enabling further adjustment of the prediction boundaries for deviations. The method effectively enhances the model’s perception of question boundary information and enables the calibration of predicted answer boundaries. Experiments on the public datasets SQuAD1.1、HotpotQA and NewsQA datasets confirm the effectiveness of the proposed method and validate the need for each module of the method through ablation experiments.

Key words: machine reading comprehension, question boundary-aware, answer boundary regression, span extraction, SQuAD1.1