Opinion Analysis through Constraint Optimisation

Abstract

Opinion lexicon plays a vital role in sentiment classification. A previous study shows that a compositional model can be eective in sentiment classification. But such a model has been only applied using hand-crafted composition rules. The need for hand-crafted rules arise when dealing with conflicting polarity values within the same phrase. In this thesis, we show that an alternative is to employ a weighted polarity lexicon. There are several key advantages of a weighted polarity lexicon. Firstly, compositionality rules simply become linear sums without requiring conflict resolution rules. Secondly, a weighted polarity lexicon can be automatically learnt from review data using constraint optimisation. Thirdly, instead of providing just a binary positive or negative output, our model can be used to provide a graded overall sentiment. Our experiments show that our model provides state-of-the-art opinion classication.