Tapping into sociological lexicons for sentiment polarity classification

Abstract

Sentiment Analysis, or the extraction of emotional content from text, has been a prominent research topic for a decade. Numerous annotated lexicons have been created for identification and classification of emotions (or affect) in text. This extraction of emotional content from text makes possible emotion-aware Information Retrieval, which is especially important with the growing popularity of user-generated content like blogs, tweets, and wikis. This paper introduces a new source of high quality manual annotations that can be used for sentiment extraction. A subfield of sociology symbolic interactionism, more precisely Affect Control Theory (ACT), measures the emotional meanings we associate with various concepts. Research in this field produces multi-dimensional manual annotations of words much like those used in Sentiment Analysis. We compare these annotations with SentiWordNet and WordNet-Affect, lexicons produced for Sentiment Analysis, in the task of text polarity classification and show that classifier trained on the ACT lexicon outperforms the other two.