Research On Multi-aspect Opinion Mining From Review Texts

Opinion mining on review texts refers to mining the subjective information related to the authors’opinions or standpoints in the review texts and analyzing the sentiment orientation of the information. With review texts being widely used in people’s life, mining the overall opinions only could not satisfy people’s requirement of the information about fine-grained aspects in review texts. Therefore, multi-aspect opinion mining from review texts has become an important research issue.The goal of multi-aspect opinion mining from review texts is to discover the multiple aspects of the entity discussed in the reviews of the given dataset and obtain the latent rating and weight for each aspect in each review based on the given overall ratings and the contents of the review texts.There are two stages in the multi-aspect opinion mining task. The first stage is the aspect extraction of the multiple aspects, and the second stage is the analysis of aspect ratings and weights for each review.In this paper, we implement the Boot-strapping based method for aspect extraction. We propose and implement the Local LDA based method for aspect extraction, which is unsupervised and can obtain the multiple aspects of the entity discussed in the review texts by modeling and analyzing the sentences in the review texts.We apply the LRR model for modeling the generation of overall ratings for the review texts, thus obtaining the latent aspect ratings and weights.The experimental results demonstrate that the proposed "Local LDA+LRR" method can effectively do the task of aspect ratings and weights analysis on review texts, and it performs significantly better than the "Boot-strapping+LRR" method.