Context-sensitive contrastive feature-based opinion summarisation of online reviews

Abstract

Contrastive opinion summarisation (COS) systems produce summary by selecting and aligning contrastive sentences from a set of positive and negative opinionated sentences. Most of the existing COS methods do not consider the implicit opinion present in a sentence while producing summary. Implicit opinion can be identified based on context terms present in a sentence. Therefore, a new COS approach called context-sensitive contrastive opinion summarisation is proposed. Initially linguistic rules are framed based on dependency relation to extract context-feature-opinion phrases. To automatically cluster the extracted context-feature-opinion phrases into contrastive arguments, a clustering algorithm is proposed. Context sensitive weight is calculated for each phrase based on their probability of occurrence in the concepts of ConceptNet. Clustering algorithm integrates context sensitivity with contrastive similarity for producing better arguments summary. Experimental conducted on car and product review datasets demonstrate that the context-sensitive clusters achieved good coverage and precision when compared to state-of-art approaches.