Sentiment analysis of microblogs with rich emoticons

Abstract

Sentiment analysis for social media can help to explore deeper insight into the attitudes, opinions, and emotions behind the posts. Existing work usually analyze the emoticons and texts of the posts separately, and ignore the impact of emoticons on the emotional polarity of texts. As a result, the polarity of the posts could be marked inaccurately in the scenarios where the polarity of the texts relies on the contextual information of the emoticons. To address this issue, we propose a model, WnhBert-Bi-LSTM, for microblog sentiment analysis. The model trains phrase and emoticon embedding on a large-scale corpus composed of 280,000 Chinese microblogs, and uses the self-attention mechanism to evaluate the impact of emoticons on the overall emotional polarity. By converting emoticons into tractable features, the emoticons can be analyzed jointly with the texts to explore their feature interaction. Evaluations on 8,965 sina microblog posts show that the accuracy of our model is 3.19% higher than the baseline models. In addition, we constructed and open-sourced a new emoticon label corpus with more widely used words and more comprehensive emoticon data than the existing corpus.