Feature Selection Using Multi-objective Optimization for Aspect Based Sentiment Analysis

In this paper, we propose a system for aspect-based sentiment analysis (ABSA) by incorporating the concepts of multi-objective optimization (MOO), distributional thesaurus (DT) and unsupervised lexical induction. The task can be thought of as a sequence of processes such as aspect term extraction, opinion target expression identification and sentiment classification. We use MOO for selecting the most relevant features, and demonstrate that classification with the resulting feature set can improve classification accuracy on many datasets. As base learning algorithms we make use of Support Vector Machines (SVM) for sentiment classification and Conditional Random Fields (CRF) for aspect term and opinion target expression extraction tasks. Distributional thesaurus and unsupervised DT prove to be effective with enhanced performance. Experiments on benchmark setups of SemEval-2014 and SemEval-2016 shared tasks show that we achieve the state of the art on aspect-based sentiment analysis for several languages.