Sentiment analysis on Twitter through topic-based lexicon expansion

Supervised learning approaches are domain-dependent and it is costly to obtain labeled training data from different domains. Lexiconbased approaches enjoy stable performance across domains, but often cannot capture domain-dependent features. It is also hard for lexicon-based classifiers to identify the polarities of abbreviations and misspellings, which are common in short informal social text but usually not found in general sentiment lexicons. We propose to overcome this limitation by expanding a general lexicon with domain-dependent opinion words as well as abbreviations and informal opinion expressions. The expanded terms are automatically selected based on their mutual information with emoticons. As there is an abundant amount of emoticon-bearing tweets on Twitter, our approach provides a way to do domain-dependent sentiment analysis without the cost of data annotation. We show that our technique leads to statistically significant improvements in classification accuracies across 56 topics with a state-of-the-art lexicon-based classifier. We also present the expanded terms, and show the most representative opinion expressions obtained from co-occurrence with emoticons.