Product reviews have an important feature that they contain aspects for a target product and sentiments for each aspect. To analyze product reviews more correctly and in more detail, aspects and sentiments have to be considered together. Previous stud...
Product reviews have an important feature that they contain aspects for a target product and sentiments for each aspect. To analyze product reviews more correctly and in more detail, aspects and sentiments have to be considered together. Previous studies utilized the concept of probabilistic topic model to model the feature and they treated aspects and sentiments as latent variables like topics in topic model. Various models for analyzing product review have been proposed as results of the studies and most of them rely on bag of words assumption. The models assume that a product review is a set of individual words and they can extract only unigrams. However, unigrams are not enough for people to figure out what reviewers intended to mention and how the reviewers felt. N-grams should be extracted to discover aspects and sentiments they came from. Further, it is required to detect exactly how many individual words combined for each n-gram to precisely estimate how much each n-gram is related to each aspect and sentiment. Classification model with syntactic features probably can help to detect the length of each n-gram. This dissertation proposes novel topic model that utilize Maximum Entropy model to estimate the length of each n-gram with features of lexicons and part of speech tags. Proposed model assume that a product review consists of sentences and each sentence is a set of n-grams. In proposed model, an aspect and a sentiment of each sentence is estimated by how much each n-grams in the sentence is related to each aspect and sentiment, and a sentiment of a product review is determined by the sentiments of sentences in the product review. When sentiment classification is performed, proposed model that based on n-grams shows better performance than unigram based model.