中文社交化短文本情感分析与话题挖掘研究

Sentiment analysis and topic mining for large scale text have always been the core research points in business intelligence field. With the rapid growth of social media, users generate more and more reviews on a product, a news event or a life experience. The huge-volume, high-dimensional, high-sparse characteristics of these social media and the lack of Chinese corpora post severe challenges to the traditional sentiment analysis and topic mining algorithms. In this project, we try to handle the issue by cross-lingual sentiment analysis and non-negative matrix factorization (NMF) with soft orthogonal constraint techniques. Specifically, 1) To avoid high sparseness of large scale social short texts, we combine word co-occurrence matrix and orthogonalization process to propose an unsupervised feature selection algorithm; 2) We propose a semi-supervised learning algorithm with space transfer for cross-lingual sentiment analysis of Chinese short text, which employs existing abundant English sentiment corpora and extends structural correspondence learning; 3) To conquer the weakness of topic coherence and indicators of traditional topic model methods, we introduce soft orthogonal constraints into NMF-based topic mining model of social short text; 4) We will finally develop a prototypical system, which will be used for product-users’ micro blog reviews to verify its effectiveness. The project will provide great values in terms of both theories and practices to business intelligence under social media environment.

面向大规模文本的情感分析与话题挖掘一直以来都是商务智能领域的核心研究问题。随着社交媒体的迅速发展，用户对某一产品、新闻事件或生活体验的评价越来越丰富。这些社交化短文本的海量、高维、高稀疏特点以及中文语料的匮乏对传统的情感分析和话题挖掘算法提出了严峻的挑战。本项目尝试采用跨语言的情感分析和软正交约束的非负矩阵分解技术来解决上述问题。具体包括：1）针对社交短文本高稀疏的特点，将词语共现矩阵和正交化相结合，研究大规模短文本的非监督特征选择问题；2）利用现有的、丰富的英文情感语料和结构对应学习算法，研究并提出空间迁移的跨语言短文本半监督情感分类算法；3）针对传统方法挖掘的话题语义内聚性不强及指示词不突出问题，引入软正交约束，研究基于非负矩阵分解的短文本话题挖掘算法；4）研发系统原型，并针对产品-用户微博评论展开应用研究。本项目将为社会化媒体环境下的商务智能理论与实践提供有益参考。

大规模文本的情感分析与话题挖掘是产品评论领域的研究热点，本项目主要利用迁移学习和非负矩阵分解技术来解决现有的问题。具体成果包括：1）研究并提出一种基于随机投影和GS的特征选择算法，并在大规模短文本数据集上进行了实验的验证；2）利用现有的、丰富的英文情感语料和结构对应学习算法，研究并提出一种软关联的迁移学习算法，解决跨领域或者跨语言的产品评论情感分类问题；3）提出一种互补的方面级观点挖掘算法，为产品评论意见挖掘提供支持；4）提出一种基于稀疏表示的软正交约束的非负矩阵分解的层次主题算法，实现文本数据的层级话题分析。本项目的成果为社会化媒体环境下的商务智能理论与实践提供有益参考。