句法分析和深度神经网络在中文问答系统答案抽取中的研究  

Research on Syntactic Analysis and Deep Neural Network in the Answer Extraction for Chinese Question Answering System

作　　者：李超[1] 柴玉梅[1] 高明磊[1] 昝红英[1] 

机构地区：[1]郑州大学信息工程学院,郑州450001

出　　处：《小型微型计算机系统》2017年第6期1341-1346,共6页Journal of Chinese Computer Systems

基　　金：国家社会科学基金项目(14BYY096)资助;国家自然科学基金项目(61402419;61272221)资助;国家"八六三"高技术研究发展计划项目(2012AA011101)资助

摘　　要：答案抽取是问答系统中的核心内容,问题及答案句的句法和语义充分理解是找出答案的关键.由于中文自然语言句法语义分析复杂,人工提取特征难度较大、主观性较强,使之成为中文问答系统的研究重点和难点.为此本文提出利用深度学习的思想主动学习候选答案深度特征,将答案抽取问题转化为特征学习与分类问题.即用词向量表征问题句和答案句,通过长短时记忆神经网络主动学习其深层语义相关,借助依存句法树分析句法结构特征,构造深度神经网络学习问题句、答案句和候选答案之间的内在关联信息.实验表明,该方法在不需要制定繁琐句法语义特征的条件下,仍具有较好的答案抽取性能,MRR值达到0.71.The answer extraction is the core content in the question answering system. The syntactic and semantic understanding of the question and answer sentence is the key to find out the answer. Because of the complexity of the syntactic and semantic analysis of Chinese natural language, manual extraction of features is not only difficult but also subjective ,so it becomes the focus and difficulty of Chinese question answering system research. To this end, we proposes to use the theory of deep learning to actively study the candi- date answer feature, and the answer extraction problem is transformed into problem of feature learning and classification in this paper. That is, we use word vector to represent of questions and answer sentences, then learn the deep semantic correlation with the Long- Short Term Memory（LSTM） model,and analyze the syntactic structure characteristics by using the dependency syntax tree. At last, we construct deep neural network to learn the intrinsic link between the question sentence, the answer sentence and the candidate an- swers. Experimental results show that we still have a better answer extraction performance and MRR of the proposed method is 0.71 under the condition of no need to formulate the characteristic rules.

关 键 词：深度神经网络 深度学习 句法分析 问答系统 答案抽取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]