Research On Review Helpfulness Based On Semantic And Sentiment Information

Online consumer reviews play an important role in decision-making, which provide users with a wealth of opinion and shopping guide, affecting user's shopping decisions. While with the development of e-commerce and more and more users are involved, the number of online reviews increase sharply, which lead user can't easily find out reliable and useful reviews from massive amount of reviews. In recent years, the research on review mainly involves three aspects: opinion spam, review helpfulness and opinion summarization. It's easy to recognize opinion spam using machine learning methods, opinion summarization and review helpfulness, which are different ways to organize reviews, have been studied for many years.This paper mainly completes the following works according to the problems existing in the research on opinion summarization and quality of reviews:(1)For problems on syntactic analysis in irregular texts and feature extraction using topic model, we propose a improved LDA model, which called SA-LDA, to combine syntactic analysis and topic model for aspect extraction.(2)For the clustering of feature words, the distance between feature words is measured by the method of combining the semantic similarity and opinion similarity.(3)Using the feature set and opinion set obtained by syntactic analysis to recognize opinion sentences, which act as input of the topic model, and then constructing must-link and cannot-link for topic learning, to ensure the precision of the model.(4)Explore the relationship between review helpfulness and opinion summarization, we propose a unsupervised model, which called OSUD, to predict the quality of reviews based on the idea that reviews which hold the same opinions share the same helpfulness value.(5)Explore the effects of user replies below reviews on opinion support and review helpfulness. User reply is more valuable than vote, for it provide us user opinions about a specific aspect of product, and it's also more credible.The experimental data is crawled from the Zhongguancun website, and we label the data by hand. The experimental results show that the feature extraction method not only ensures better recall score, but also improves precision score, and it's also useful for implicit feature extraction. The model for predicting review helpfulness based on opinion support shows an ideal result, it also provides users a meaningful explanation for experimental results.