An Unsupervised Sentiment Classification Method Based on Multi-Level Fuzzy Computing and Multi-Criteria Fusion

Abstract

With the rapid growth of user-generated content, unsupervised methods that do not require label training data have gradually become a research focus in the field of sentiment classification and natural language processing. But the performance of unsupervised methods is unsatisfactory. This is because the ambiguity of sentiment polarity and the fuzziness of sentiment intensity are usually ignored in existing unsupervised methods. To address these problems, we propose an unsupervised sentiment classification method based on multi-level fuzzy computing and multi-criteria fusion which involves three innovations. Firstly, we come up with a multi-level computing model to compute the sentiment intensity of reviews for partly reducing the ambiguity of sentiment polarity. Secondly, to further decrease the ambiguity of sentiment polarity, a multi-criteria fusion strategy based on sentiment category credibility and domain category representativeness is proposed. Thirdly, a fuzzy classifier is constructed to solve the fuzziness of sentiment intensity. Furthermore, a self-supervised method using pseudo-labeled training data is proposed to learn the optimum parameters of the fuzzy classifier. Experimental results in three different domain balanced datasets and one unbalanced dataset proved that our method improves 12.35% more accuracy than the competitive baselines in sentiment classification.

Keywords

• Unsupervised sentiment classification
• pseudo-labeled data
• fuzzy sets
• self-learning