Context-aware sentiment analysis of social media

Abstract

The lexicon-based approach to opinion mining is typically preferred where training data is difficult to obtain or cross domain robustness of algorithms is of essence. However, this approach suffers from the semantic gap between the polarity with which a sentiment-bearing term appears in the text (i.e. contextual polarity) and its prior polarity captured by the lexicon. This is further exacerbated when mining is applied to social media. Here, we propose an approach to address this semantic gap. Firstly, by accounting for the influence of surrounding terms to a sentiment bearing term (local context). Secondly, by accounting for content and context disagreement between the lexicon and the domain in which it is applied (global context). This is achieved by generating a domain-focused lexicon using distant-supervision and integrating its scores with a generic lexicon (SentiWordNet). Evaluation results from sentiment classification over social media content extracted from three different platforms show benefits of accounting for local and global contexts, both individually and in combination. We also present some promising results from our investigation into the cross-platform transferability of our approach.