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

发布时间：2018-05-28 11:49

本文选题：深度神经网络 + 深度学习　；参考：《小型微型计算机系统》2017年06期

【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.
[Abstract]:Answer extraction is the core content in the question answering system. The key to find the answer is to fully understand the syntax and semantics of the question and answer sentence. Because of the complexity of syntactic and semantic analysis of Chinese natural language, the difficulty of artificial feature extraction and the strong subjectivity, it has become the focus and difficulty of the Chinese question answering system. In this paper, we use the idea of deep learning to actively learn candidate depth features, and transform the problem of answer extraction into feature learning and classification. That is, word vector representation question sentence and answer sentence, through long and short memory neural network to actively learn its deep semantic correlation, with the help of dependency syntax tree analysis of syntactic structure characteristics, construct depth neural network learning question sentence. The internal correlation between the answer sentence and the candidate answer. The experimental results show that the proposed method still has a good performance of answer extraction and the MRR value is 0.71 without the need to make the complicated syntactic and semantic features.
【作者单位】：郑州大学信息工程学院;
【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
【分类号】：TP183;TP391.3

【相似文献】