Similarity Dependency Dirichlet Process for Aspect-Based Sentiment Analysis

Aspect-base Sentiment Analysis is a core component in Review Recommendation System. With the booming of customers’ reviews online, an efficient sentiment analysis algorithm will substantially enhance a review recommendation system’s performance, providing users with more helpful and informative reviews. Recently, two kinds of LDA derived models, namely Word Model and Phrase Model, take the dominant positions in this field. However, the requirement of exact aspect number underlies the usability and flexibility of these LDA extended models. Although, Dirichlet Process(DP), which can help to automatically generate the number of aspects, has been on trial, its random word assignment mechanism makes the result unsatisfying. This paper proposes a model named Similarity Dependency Dirichlet Process(SDDP) to cope with the above problems. SDDP inherits the merits of DP to automatically determine the number of aspects, but it exploits the semantic similarities between words to infer aspects and sentiments, alleviating the random assignment problem in DP. Furthermore, based on SDDP, this paper builds a word model W-SDDP and a phrase model P-SDDP respectively to detect aspects and sentiments from two different perspectives. Finally we experiment both our two models on 6 datasets, and compare with other 6 currently most popular models. The result shows that both W-SDDP and P-SDDP outperform the other 6 models, indicating SDDP is a promising model for sentiment analysis.