Research On Product Features Extraction And Sentiment Classification

With the development of information technology and the Web 2.0, the amount of reviews of products is increasing rapidly. People need to acquire the sentiment orientation of a product from the massive reviews, with the help of sentiment analysis, to form their own opinions about a product. But traditional sentiment analysis, either on the document or the sentence level, is gradually not suitable for this task when people want to know each special property of a product in more detailed. So, people pay more attention to the feature-level sentiment analysis.This thesis conducts some deep research on feature-level sentiment analysis from both theoretical and practical aspects, which includes:(1) Product features extraction on the Lasso-based Feature Selection Method.We proposed a method to automatically extract product features from review texts by the Lasso-based feature selection. Firstly, to the target review texts and comparing review texts, we build a L1-norm regularization linear classifier (Lasso) which can also select several weighted features. This kind of features can be tagged as the special product features in the target reviews. Then, sort the extracted features by the frequency they appeared in the review texts, and remove some features by a default frequent threshold. Finally, we acquire the product features by some combination and PMI pruning. The experiment results show that our method is very competitive, and more suitable for the Chinese review texts.(2) Opinion words extraction on the dependency relationship and DCDT (Derogatory and Commendatory Direction Tendency).We firstly apply the dependency parsing and DCDT rules to acquire the opinion words relevant to each extracted feature. Then judge the sentiment polarity of each opinion word by the sentiment dictionary. Finally, we gather the polarity of each product feature, and calculate the global sentiment of each of them. The experiment results show that our method can acquire the opinions of each part and property of product efficiently, and is very useful for customers to make decision.