Sparse Representation for Sentiment Analysis
Zohre Karimi, Rohollah Ramezani · 2020
Sentiment analysis has been attracted many attentions in recent years. Various studies have been done with the main focus on feature representation and classification. Term Frequency-Inverse Document Frequency is the most common statistical feature which has been used in the literature that extracts high dimensional sparse features.In this paper, we propose to reduce data dimensionality by considering the assumption that high dimensional sentiment data lies on some low dimensional manifolds. This assumption is confirmed by numerous studies on high dimensional data and text mining. Spectral embedding is applied for dimensionality reduction. Since the performance of spectral embedding is heavily dependent on the graph of representing manifold(s), it is proposed to enforce sparse representation for generating suitable graph. The proposed method is compared with common dimensionality reduction methods in sentiment analysis on different classifiers and the results demonstrate that it significantly outperforms other methods.