基于用户评论的自动摘要的研究和分析

发布时间：2018-09-04 14:03

【摘要】：电子商务在迅猛发展,网络信息日益增多。随着越来越多的人在网上购物,如何提高用户体验加强商家与用户、用户与用户之间的信息交流成了一个重要的问题。用户购物后留下的评论是用户之间和用户与商家之间信息反馈的重要平台,因此本文提出了对用户评论进行研究。对用户评论的挖掘和传统的文本挖掘有所不同,因为用户评论一般远远短于普通文本长度,聚焦的信息点更细。这其中涉及了很多自然语言处理,机器学习和数据挖掘的技术。随着机器学习的发展,特别是深度学习的兴起,很多问题都得到了进一步的深层次研究。本文结合自然语言处理基础知识,关联挖掘算法,层次聚类模型,神经网络和决策树算法等对评论自动摘要做了新的研究。本文针对中文的特性,对提取英文中评论特征的Apriori算法做了改进,取得了不错的效果,证明了该方式的可行性。提出了使用词激活力模型对评论特征聚类,比传统聚类模型有更强的适用性。针对评论句的情感分析,基于word2vec使用了递归自编码的神经网络,比传统朴素贝叶斯分类器,F1值提升了大约8%。最后提出使用决策树建立特征的层次模型,更好的组织了摘要的展示。
[Abstract]:E-commerce is developing rapidly, network information is increasing day by day. With more and more people shopping on the Internet, how to improve the user experience and enhance the exchange of information between merchants and users has become an important issue. The comments left by users after shopping is an important platform for information feedback between users and merchants, so this paper proposes a research on user comments. The mining of user comments is different from that of traditional text mining because user comments are usually much shorter than ordinary text and focus on more detailed information points. This involves many natural language processing, machine learning and data mining techniques. With the development of machine learning, especially the rise of deep learning, many problems have been further studied. Based on the basic knowledge of natural language processing, association mining algorithm, hierarchical clustering model, neural network and decision tree algorithm, this paper makes a new research on automatic summary of comments. According to the characteristics of Chinese, this paper improves the Apriori algorithm for extracting comment features in English, and obtains good results, which proves the feasibility of this method. In this paper, the word activation force model is applied to comment feature clustering, which is more applicable than the traditional clustering model. For the emotional analysis of comment sentences, the recurrent self-coding neural network is used based on word2vec, which is about 8 times higher than the traditional naive Bayesian classifier. Finally, a hierarchical model based on decision tree is proposed to better organize the presentation of the summary.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【参考文献】