Identifying emotion topic — An unsupervised hybrid approach with Rhetorical Structure and Heuristic Classifier

Abstract

This paper describes an unsupervised hybrid approach to identify emotion topic(s) from English blog sentences. The baseline system is based on object related dependency relations from parsed constituents. However, the inclusion of the topic related thematic roles present in the verb based syntactic argument structure improves the performance of the baseline system. The argument structures are extracted using VerbNet. The unsupervised hybrid approach consists of two phases; firstly, the information of Rhetorical Structure (RS) is extracted to identify the target span corresponding to the emotional expression from each sentence. Secondly, as an individual target span contains one or more topics corresponding to an emotional expression, a Heuristic Classifier (HC) is designed to identify each of the topic spans associated in the target span. The classifier uses the information of Emotion Holder (EH), Named Entities (NE) and four types of Similarity features to identify the phrase level components of the topic spans. The system achieves average recall, precision and F-score of 60.37%, 57.49% and 58.88% respectively with respect to all emotion classes on 500 annotated sentences containing single or multiple emotion topics.