Latent syntactic structure-based sentiment analysis

Abstract

People share their opinions about things like products, movies and services using social media channels. The analysis of these textual contents for sentiments is a gold mine for marketing experts, thus automatic sentiment analysis is a popular area of applied artificial intelligence. We propose a latent syntactic structure-based approach for sentiment analysis which requires only sentence-level polarity labels for training. Our experiments on three domains (movie, IT products, restaurant) show that a sentiment analyzer that exploits syntactic parses and has access only to sentence-level polarity annotation for in-domain sentences can outperform state-of-the-art models that were trained on out-domain parse trees with sentiment annotation for each node of the trees. In practice, millions of sentence-level polarity annotations are usually available for a particular domain thus our approach is applicable for training a sentiment analyzer for a new domain while it can exploit the syntactic structure of sentences as well.