ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2018, Vol. 55 ›› Issue (8): 1631-1640.doi: 10.7544/issn1000-1239.2018.20180233

所属专题: 2018数据挖掘前沿进展专题

• 人工智能 • 上一篇    下一篇

面向短文本分析的分布式表示模型

梁吉业,乔洁,曹付元,刘晓琳   

  1. (山西大学计算机与信息技术学院 太原 030006) (计算智能与中文信息处理教育部重点实验室(山西大学) 太原 030006) (ljy@sxu.edu.cn)
  • 出版日期: 2018-08-01
  • 基金资助: 
    国家自然科学基金项目(U1435212,61432011,61573229);山西省重点科技攻关项目(MQ2014-09) This work was supported by the National Natural Science Foundation of China (U1435212, 61432011, 61573229) and the Key Scientific and Technological Project of Shanxi Province (MQ2014-09).

A Distributed Representation Model for Short Text Analysis

Liang Jiye, Qiao Jie, Cao Fuyuan,Liu Xiaolin   

  1. (School of Computer and Information Technology, Shanxi University, Taiyuan 030006) (Key Laboratory of Computational Intelligence and Chinese Information Processing (Shanxi University), Ministry of Education, Taiyuan 030006)
  • Online: 2018-08-01

摘要: 短文本的分布式表示已经成为文本数据挖掘的一项重要任务.然而,直接应用分布式表示模型Paragraph Vector尚有不足,其根本原因是其在训练过程中并没有利用到语料库级别的信息,从而不能有效改善短文本中语境信息不足的情况.鉴于此,提出了一种面向短文本分析的分布式表示模型——词对主题句向量模型(biterm topic paragraph vector, BTPV),该模型通过将词对主题模型(biterm topic model, BTM)得出的主题信息融入Paragraph Vector中,不仅使得模型训练过程中利用到了全局语料库的信息,而且还利用BTM显性的主题表示完善了Paragraph Vector隐性的空间向量.实验采用爬取到的热门新闻评论作为数据集,并选用K-Means聚类算法对各模型的短文本表示效果进行比较.实验结果表明,基于BTPV模型的分布式表示较常见的分布式向量化模型word2vec和Paragraph Vector来说能取得更好的短文本聚类效果,从而显现出该模型面向短文本分析的优势.

关键词: 分布式表示, 短文本, 文本分析, 句向量, 词对主题模型

Abstract: The distributed representation of short texts has become an important task in text mining. However, the direct application of the traditional Paragraph Vector may not be suitable, and the fundamental reason is that it does not make use of the information of corpus in training process, so it can not effectively improve the situation of insufficient contextual information in short texts. In view of this, in this paper we propose a novel distributed representation model for short texts called BTPV (biterm topic paragraph vector). BTPV adds the topic information of BTM (biterm topic model) to the Paragraph Vector model. This method not only uses the global information of corpus, but also perfects the implicit vector of Paragraph Vector with the explicit topic information of BTM. At last, we crawl popular news comments from the Internet as experimental data sets, using K-Means clustering algorithm to compare the models’ representation performance. Experimental results have shown that the BTPV model can get better clustering results compared with the common distributed representation models such as word2vec and Paragraph Vector, which indicates the advantage of the proposed model for short text analysis.

Key words: distributed representation, short text, document analysis, paragraph vector, biterm topic model (BTM)

中图分类号: