Aspect Extraction in Product Reviews via an Improved Unsupervised Method

摘要

This paper studied on aspects extraction from product reviews by unsupervised topic model, which is an important subtask of opinion mining. The topic distribution of topic model, such as LDA, leans to the high-frequency words since the words in the document comply with the characteristics of power law distribution, which leads to that most of the words that can represent topics are overwhelmed by a small number of high-frequency words, and consequently, the topic expressive ability is reduced. To solve these problems, an unsupervised method is proposed by us in this paper, and on the basis of Sentence-LDA topic model, a new Unsupervised Weighted LDA model (UW-LDA) based on the weighted topic model is obtained through weighting on feature words by a improved Gaussian function. Finally, the experiments on two aspects of the review corpus of several products from different domains show that our proposed model has made a satisfactory result.