Social Emotion Prediction with Attention-based Hierarchical Neural Network

Text emotion mining has attracted many researchers’ attention in recent years. Most existing studies focus on identifying emotion of the text’s author. In this paper, we are concerned with the problem of predicting reader’s emotion response from the text content of news article, aka social emotion prediction. Social emotion prediction can not only be used for analyzing people’s reaction to news events, but also help news producers to forecast the underlying influence of their unpublished news articles. We propose a hierarchical LSTM network with attention mechanism to identify the relation between news article and evoked emotion of reader. Compared with existing word-level and topic-level methods which don’t consider the order of words, our method with hierarchical LSTM structure is more efficient to construct the news article representation. With attention mechanism, our method can also find out readers’ emotion reaction to specific word or phrase and generate social emotion lexicon. Through experiments on real-world datasets, we demonstrate that our approach outperforms word-level and topic-level baselines, and a state-of-the-art word2vec-based method.