面向机器阅读理解的候选句抽取算法  

Candidate Sentences Extraction for Machine Reading Comprehension

作　　者：郭鑫[1] 张庚 陈千[1,2] 王素格[1,2] GUO Xin;ZHANG Geng;CHEN Qian;WANG Su-ge(School of Computer&Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing,Ministry of Education,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]计算智能与中文信息处理教育部重点实验室,太原030006

出　　处：《计算机科学》2020年第5期198-203,共6页Computer Science

基　　金：山西省应用基础研究计划项目(201701D221101,201901D111032);国家自然科学基金项目(61502288,61403238,61673248);山西省重点研发计划项目(201803D421024)。

摘　　要：使机器理解人类自然语言是人工智能在认知领域的终极目标,机器阅读理解是自然语言处理技术中继语音识别、语义理解之后的一大挑战,要求计算机具有一定的背景常识,全面理解给定文本材料,并根据材料内容对相应的问题作答。随着深度学习的快速发展,阅读理解成为当前人工智能的热点研究方向,涉及机器学习、信息检索、语义计算等核心技术,在聊天机器人、问答系统、智能化教育等多个领域具有广泛的应用前景。文中聚焦微阅读模式,根据问题或选项从给定文本材料中抽取包含答案的候选句,缩小推理范围,为进一步实现机器阅读理解提供技术支持。传统基于特征的方法耗费大量人力,文中将答案候选句抽取看成一种语义相关度计算问题,提出了一种答案候选句排序方法,即Att-BiGRU/BiLSTM模型。首先,利用双向长短期记忆和门控循环单元来编码句子中表达的语义信息;其次,设计Atten结构,结合相异性和相似性对语义相关度进行建模;最后,采用Adam算法来学习模型的参数。在SemEval-SICK数据集上的实验结果显示,该模型在测试集上的pearson指标超过了基线方法BiGRU将近0.67,在MSE指标上超过BiGRU方法16.83%,收敛速度更快,表明双向和Atten结构能大大提高候选句抽取的精度。The ultimate goal of artificial intelligence is to let machine understand human natural language in cognitive field.Machine reading comprehension raises great challenge in natural language processing which requires computer to have certain common knowledge,comprehensively understand text material,and correctly answer the corresponding questions according to that text material.With the rapid development of deep learning,machine reading comprehension becomes the current hotspot research direction in artificial intelligence,involving core technologies such as machine learning,information retrieval,semantic computing and has been widely used in chat robots,question answering systems and intelligent education.This paper focuses on micro-rea-ding mode,and answer candidate sentences containing answers are extracted from given text,which provide technology support for machine reading comprehension.Traditional feature-based methods consumes lots of manpower.This paper regards candidate sentences extracting as a semantic relevance calculation problem,and proposes an Att-BiGRU/LSTM model.First,LSTM and GRU are used to encode the semantic expressed in a sentence.Then,the dissimilarity and similarity are captured with an Atten structure for semantic correlation.Last,adam optimizer is used to learn the model parameters.Experiment results show that Att-BiGRU model exceeds the baseline method of nearly 0.67 in terms of pearson,16.8%in terms of MSE on SemEval-SICK test dataset,which proves that the combination of the bidirectional and Atten structure can greatly improve the accuracy of the candidate sentences extraction,as well as the convergence rate.

关 键 词：长短期记忆模型 门控循环单元 候选句抽取 语义相关度计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]