基于问题分解的多跳机器阅读理解模型  

A multi-hop reading comprehension method based on question decomposition

作　　者：周展朝 刘茂福[1,2] 胡慧君 ZHOU Zhan-zhao;LIU Mao-fu;HU Hui-jun(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan 430065,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065

出　　处：《计算机工程与科学》2022年第8期1506-1513,共8页Computer Engineering & Science

基　　金：国家社会科学基金(11&ZD189);湖北省教育厅人文社会科学研究项目(17Y018)。

摘　　要：多跳机器阅读理解是自然语言处理领域最困难的任务之一,需要在多个段落之间进行推理。多跳机器阅读理解任务中的复杂问题一般由多个简单问题融合而成,可以通过分解复杂问题使模型更好地理解问题。因此,针对复杂多跳问题,提出了一种基于问题分解的多跳阅读理解模型。该模型首先将多跳问题分解为多个单跳问题,然后利用单跳阅读理解模型对其进行求解。将问题分解视作一个阅读理解任务:多跳问题是问题分解的上下文,而包含问题答案的证据段落则是问题。阅读理解任务捕捉了多跳问题和证据段落之间的交互语义信息,可以指导多跳问题中单跳问题的抽取。所提模型的BLEU值和Rouge-L值分别为71.48%和79.29%。实验结果表明,该模型对多跳机器阅读理解是有效的。Multi-hop machine reading comprehension is one of the most difficult tasks in the field of natural language processing,which requires reasoning between multiple paragraphs.The complex question in multi-hop machine reading comprehension task is usually composed of several simple questions and decomposing the complex question can make the model better understand the question itself.Therefore,for the complex multi-hop question,a multi-hop reading comprehension model based on question decomposition is proposed.The multi-hop question is first decomposed into several single-hop questions,and then the single-hop reading comprehension model is used to solve them.The question decomposition is regarded as a reading comprehension task:the multi-hop question is the context of the question decomposition,while the evidence paragraph containing the answer to the question is the question.Machine reading comprehension task captures the interactive semantic information between multi-hop question and evidence paragraph,which can guide the extraction of single-hop questions in multi-hop questions.The values of the proposed model on BLEU and Rouge-L are 71.48%and 79.29%,respectively.Experimental results show that this model is effective for multi-hop machine reading comprehension.

关 键 词：机器阅读理解 推理 多跳问题 问题分解 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]