SEMANTIC EMOTION-TOPIC MODEL BASED SOCIAL EMOTION MINING

With the booming of social media users, more and more short texts with emotion labels appear, which contain users' rich emotions and opinions about social events or enterprise products. Social emotion mining on social media corpus can help government or enterprise make their decisions. Emotion mining models involve statistical-based and graph-based approaches. Among them, the former approaches are more popular, e.g. Latent Dirichlet Allocation (LDA)-based Emotion Topic Model. However, they are suffering from low retrieval performance, such as the bad accuracy and the poor interpretability, due to them only considering the bag-of-words or the emotion labels in social media corpus. In this paper, we propose a LDA-based Semantic Emotion-Topic Model (SETM) combining emotion labels and inter-word relations to enhance the retrieval performance of social emotion mining result. The performance influence of four factors on SETM are considered, i.e., association relations, computing time, topic number and semantic interpretability. Experimental results show that the accuracy of our proposed model is 0.750, compared with 0.606, 0.663 and 0.680 of Emotion Topic Model (ETM), Multi-label Supervised Topic Model (MSTM) and Sentiment Latent Topic Model (SLTM) respectively. Besides, the computing time of our model is reduced by 87.81% through limiting word frequency, and its accuracy is 0.703, compared with 0.501, 0.648 and 0.642 of the above baseline methods. Thus, the proposed model has broad prospects in social emotion mining area.