基于Word2Vec的中文问句检索技术研究及系统实现

发布时间：2018-06-25 13:26

本文选题：交互式智能客服 + 词向量　；参考：《哈尔滨工业大学》2016年硕士论文

【摘要】：随着电子商务等互联网服务的快速发展,为了提高线上服务质量,企业、单位对客服的需求日益增长。而在人口老龄化的趋势下人工客服成本不断增加,交互式智能客服技术的发展为企业、单位提供了新的解决思路。智能客服的目标是提供如人工客服般的交互体验,在与用户的交互过程中,自动地完成对用户问句的识别、检索与回答。本文利用中文词向量等工具,研究基于企业实际的客服语料情境下的用户问句分类、问句检索等问题。本文研究内容主要有如下三方面:(1)为了提供更好的用户体验,本文在检索用户问句前,对用户问句的分类方法进行研究。首先本文收集了大量中文语料实现中文词向量的训练,其次提出了一个面向智能客服场景的两层分类体系。最后研究了在不同参数词向量特征基础上增量组合词汇特征与句法特征的分类效果。(2)在问句分类的基础上,本文将词向量信息引入到问句检索中。利用词向量中隐含的语义信息间接计算两个句子的相似度,对Lucene检索出的候选结果进行重排序,从而实现检索效果的优化。针对一些语义不全的问句,利用基于依存句法分析方法以及Text Rank方法的关键词抽取方法识别出问句中的关键词,并利用自定义同义词词典、中文词向量以及局部相关性反馈信息对其进行查询扩展。(3)实现了一个基于本文提出的问句分类方法及问句检索方法的中文问句检索系统,实现对用户问句的分类与检索。同时针对一些语料缺失的场景,系统中集成了爬虫处理模块,可以结构化地抽取互联网中存在的问答对信息。
[Abstract]:With the rapid development of electronic commerce and other Internet services, in order to improve the quality of online service, the demand of enterprises and units for customer service is increasing day by day. Under the trend of aging population, the cost of artificial customer service is increasing. The development of interactive intelligent customer service technology provides a new solution for enterprises and units. The goal of intelligent customer service is to provide interactive experience such as artificial customer service. In the process of interaction with users, the identification, retrieval and answer of user questions are accomplished automatically. In this paper, we use Chinese word vector and other tools to study the classification and retrieval of user questions in the context of customer service corpus. The main contents of this paper are as follows: (1) in order to provide a better user experience, this paper studies the classification method of user question before retrieving user question. Firstly, a large number of Chinese corpus is collected to train Chinese word vectors, and then a two-layer classification system for intelligent customer service scene is proposed. Finally, the classification effect of incremental combination of lexical and syntactic features based on different parameter word vector features is studied. (2) on the basis of question sentence classification, this paper introduces word vector information into question retrieval. The similarity between the two sentences is calculated indirectly by using the semantic information implied in the word vector, and the candidate results retrieved by Lucene are reordered to optimize the retrieval effect. For some questions with incomplete semantics, keyword extraction method based on dependency syntactic analysis and text Rank method is used to identify the keywords in question sentences, and a custom synonym dictionary is used. The Chinese word vector and local relevance feedback information are extended. (3) A Chinese question retrieval system based on the question classification method and question retrieval method proposed in this paper is implemented to achieve the classification and retrieval of user questions. At the same time, the crawler processing module is integrated to extract the information of question and answer pairs in the Internet.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】