The Research On The Algorithm Of Social Emotion Ranking For Online Public Sentiment

Emotion modeling for online documents has always been a hotspot concerned by researchers. This thesis models the online users’ social emotions that are the emotional responses of the masses of network users on an online document(such as network news). In traditional literatures, most researches focus on classification of emotions from writer’s perspective. However, from the perspective of the masses of network users, there will be various responses on the same online news or network event and it is difficult to measure by single emotional category, it’s better to conduct emotions ranking. Therefore, this thesis develops social emotion ranking algorithms from the perspectives of the masses of network users in the application background of online public sentiment. The main task and contributions of this thesis include the following issues:We have proposed a novel social emotion ranking algorithm based on listwise loss minimization. In this method, the loss function is created on the list of emotional label, and minimizing the differences between the label ranking outputted by model and the list of preference given as the ground truth to gain the ranking model. Given an instance, the model can predict the social emotions aroused by the instance. Because of the variety of social emotions, the result of prediction will not only belong to certain one emotional category, but a ranking of several emotional labels.We have proposed a cost-sensitive label ranking algorithm to meet the fact that the loss is usually different owing to different wrong label ranking. The key point of the method is the construction of the cost-sensitive loss function. The whole task of label ranking is divided into a sequence of sub-task and in which sub cost-sensitive loss function is constructed, then they are liner-combined to obtain cost-sensitive loss function in the whole label ranking. This method has been applied in social emotions and achieved the expected result.We have proposed a label ranking algorithm by directly optimizing performance measure Normalized Discounted Cumulative Gain(NDCG) to meet the fact people only pay attention to ranking of the top-k labels instead of the whole label ranking in some practical applications related with label ranking. This method can construct the loss function on the random top-k labels to measure the ranking losses. Empirical experiments show that this approach achieves the better ranking result on the top-k labels.