基于多尺度卷积的阅读理解候选句抽取

Evidence Sentence Extraction for Reading ComprehensionBased on Multi-scale Convolution

机器阅读理解作为检验机器是否具有理解人类自然语言能力的重要任务之一,受到了越来越广泛的关注。该文针对选择型阅读理解任务中特征提取不全面和交互不充分的问题,提出一种基于多尺度卷积的候选句抽取模型。首先,使用预训练模型编码句子语义信息,并利用多种特征辅助编码提升模型性能。其次,为了充分利用文本信息,采用多尺度卷积捕捉不同尺度的文本特征。再次,使用Focal Loss解决阅读理解中正负样本不均衡的问题,最后,选取top-20作为候选句。该文的方法在两个阅读理解选择题数据集上进行测试,实验结果表明,多尺度卷积模型效果优于基线模型,F1值较最优基线模型结果分别提升3.66%和4.82%,验证了方法的有效性。

Machine reading comprehension is a popular task to test whether a machine can understand natural language.Aiming at the choice reading comprehension items,we propose a multi-scale convolution based evidence sentence extraction model to extract more comprehensive features.Firstly,we utilize the pre-trained model to encode the semantic information for sentences,and use various features to assist the encoding to improve the performance of the model.Then,the multi-scale convolution is adopted to capture the text features at different scales,with the focal loss to alleviated the unbalanced sample issue.Finally,top-20 sentences are selected as the evidence sentences.Experimented on two datasets of reading comprehension,the proposed method improves the F 1 values by 3.66%and 4.82%,respectively,compared with the optimal baseline models.