Cross-collection Multi-aspect Sentiment Analysis

This paper proposes the use of cross-collection topic models to achieve aspect-based sentiment analysis of multiple entities simultaneously. A topic refinement algorithm that enhances semantic interpretability of topics to match that of visually identifiable aspects is presented. It is shown that, with this refinement, topics elicited from cross-collection topic models align excellently with entity aspects. Finally, the utility of opinion words returned from cross-collection topic models in investigated in the task of sentiment analysis. It is concluded that the use of such words as features for sentiment analysis yields more accurate sentiment scores than supervised counterparts.