Implicit feature identification for opinion mining

Abstract

In opinion mining area, mining consumer reviews can give a finer-grained understanding of consumer needs, which can efficiently help companies and merchants to improve the quality of their products and services. However, identifying features on which consumers express their opinions and sentiments is not always a simple task. Some existing approaches that attempt to extract implicit features using opinion words as clues or co-occurrence techniques lead to unsatisfactory results, and that due to the ambiguity caused by common opinion words which are often expressed on various features. In this paper, we propose an approach based on Association Rule Mining (ARM) and classification techniques. The first step consists of creating from a corpus, a set of association rules regrouping explicit feature-opinion pairs. The second step consists to use this set to build a classification model able to predict for each given set of opinion words the appropriate target. Tested on many classifiers, experimental results show that our approach performs better when incorporating many opinion words rather than using single ones.