基于深度学习的短文本分类及信息抽取研究

发布时间：2018-03-14 03:34

本文选题：深度学习　切入点：信息抽取　出处：《郑州大学》2017年硕士论文　论文类型：学位论文

【摘要】：互联网的发展和网络信息的爆炸式增长,给人们带来更全面、及时信息的同时,也使得用户快速、精准地找到所需要的信息变得更加困难。信息抽取可从海量的数据中检索并返回给用户更加准确、简洁的信息,可以更好地满足用户的需求。文本分类可减少信息抽取的选择空间,并可针对不同的信息类型,制定不同的策略,是信息抽取不可或缺的前提步骤。现阶段,自然语言的句法和语义充分理解是文本分类和信息抽取的关键。人工提取自然语言的句法语义特征难度大、主观性较强,深度学习可自我学习特征,对自然语言理解具有可行性。利用深度学习的思想,可以主动学习文本的句法语义特征,进而学习到所抽取信息的深度特征,降低人工特征的制定的难度,并有较好的客观性。本文在文本分类和信息抽取问题上,借助CNN模型、LSTM模型以及传统句法树的优势,构造深度神经网络模型,挖掘文本的深度特征。本文主要工作有:在文本分类上,改进传统卷积神经网络模型(CNN),提出多粒度卷积核的方法,并联合长短时记忆人工神经网络模型(LSTM),借助两个模型的优势,提出一种新的学习模型(L-MFCNN),较好的对词序语义学习和深度特征挖掘。实验结果表明,该方法在没有制定繁琐的人工特征规则的条件下,仍然有较好的性能。在信息抽取上,本文用词向量表示问题句和候选信息句,并利用长短时记忆神经网络(LSTM)学习问题句和候选信息句的语义相关特征,再利用依存句法树分析选择句法结构特征,联合表层特征构造深度神经网络,学习问题、候选信息句和候选信息三者的内在关联信息。实验结果表明,该方法可自主学习到句子的句法语义特征,有较好的信息抽取性能。最后本文设计并实现了信息抽取的应用实例问答系统,将文中提出的深度神经网络的方法运用到问答系统中,通过实践验证,没有制定较复杂句法语义特征,问答系统具有较好的答案抽取性能。
[Abstract]:The development of the Internet and the explosive growth of network information bring more comprehensive and timely information to people, but also make users quick. Finding exactly what information is needed becomes more difficult. Information extraction can be retrieved from vast amounts of data and returned to users with more accurate and concise information. Text classification can reduce the selection space of information extraction, and can make different strategies for different information types, which is an indispensable prerequisite step for information extraction. The syntactic and semantic understanding of natural language is the key of text classification and information extraction. It is feasible to understand natural language. By using the idea of deep learning, we can learn the syntactic and semantic features of the text actively, and then learn the depth features of the extracted information, and reduce the difficulty of making artificial features. In this paper, based on the advantages of CNN model and traditional syntactic tree, a depth neural network model is constructed for text classification and information extraction. The main work of this paper is to improve the traditional convolution neural network model in text classification, to propose a multi-granularity convolution kernel method, and to combine the LSTM model with long and short memory artificial neural network model to make use of the advantages of the two models. A new learning model, L-MFCNN, is proposed to study word order semantics and deep feature mining. The experimental results show that this method still has good performance without making complicated artificial feature rules. In this paper, the lexical vector is used to represent the question sentence and candidate information sentence, and the long and short memory neural network (LSTM) is used to study the semantic correlation features of the question sentence and candidate information sentence, and then the dependency syntax tree is used to analyze and select the syntactic structure features. The results of experiments show that this method can learn syntactic and semantic features of sentences independently, and combine surface features to construct the internal correlation information of depth neural network, learning problem, candidate information sentence and candidate information, and the experimental results show that the proposed method can learn syntactic and semantic features of sentences independently. Finally, this paper designs and implements the application case answering system of information extraction, and applies the method of depth neural network in this paper to the question and answer system, which is verified by practice. Without complex syntactic and semantic features, the question answering system has better performance of answer extraction.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】