隐主题模型下产品评论观点的凝聚与量化

Latent Topic-Model Based Review-Opinion Clustering and Review Mining

随着电子商务应用的不断深入,用户评论大量涌现,质量堪忧。本研究的目标,是从纷繁杂乱的海量商品用户评论中有效凝练出有价值内容,最大程度地发挥其商业应用价值。论文从信息凝练与整合的视角对用户评论汇总问题展开探索。面向中文领域,重点研究了基于用户观点的产品主题特征汇聚以及用户观点量化两个核心问题,提出并构建了基于特征序列描述的观点聚类模型Opinion_LDA,实现了基于主题模型的用户观点的自动聚类,同时利用依存句法分析及词法修饰关系对用户评价观点进行了量化。算法的效果及实现策略通过了系统实验的评测和检验。从应用的角度,完成了基于用户观点的商品"在线口碑"的信息凝聚以及产品性能的全方位汇总。

With the rapid growth of the lnternet, a wealth of user-generated product reviews has been spread in the web. However, product reviews posted at commercial websites mostly vary in quality. In this paper, we attempt to solve the problem of extracting and integrating useful content from the massive product reviews in order to make the online reviews useful for commercial applications. For this purpose, great attention has been given to the review summarization of Chinese language. The two critical problems for the summarization task are the user-opinion based product-aspect clustering and the user-opinion measurement. Opinion-LDA, a Latent Aspect Model based on se- quence of feature words, has been built to aggregate users＇ opinion automatically. Moreover, polarity lexicon and modified relationship between terms have been employed for measuring user opinion. The proposed model and method have been tested in a real tablet compute-specific dataset of reviews. The result indicates the effectiveness and feasibility of the method for the task of review summary