Research On Social Emotion Classification Based On Topic-enhanced Neural Networks

In the mobile Internet era,the generation and dissemination of online news has become very rapid.How to predict public opinion in a timely and accurate manner has become a problem that cannot be ignored in today's society.The technique of social emotion classification has received extensive attention from academic circles in recent years.Social emotion classification aims to predict the distribution of numerous readers' emotions,such as happiness,sadness,disgust,etc.,after they read the news.This thesis studies the social emotion classification methods based on topic-enhanced neural networks,which specifically includes the following three aspects:(1)Propose a method of social emotion classification based on the topic-semantics fusion feature: The core idea is to propose a fusion gate to simultaneously use document topic feature and document semantic feature in social emotion classification.Specifically,for semantic learning,we propose a dependency embedded recursive neural network(DERNN)to use the dependency syntactic knowledge of sentences;for topic learning,this method uses the LDA topic model to extract the document topic distribution,and uses the neural network to obtain the topic feature vector.Finally,the fusion gate is used to fuse the text semantic vector and the topic vector.Experiments show that this method can effectively integrate the topic feature and the semantic feature,greatly improving the performance of social emotion classification,and also prove the superiority of the proposed DERNN for semantic encoding.(2)Propose a method of social emotion classification based on the end-to-end topic-enhanced self-attention network: This method has solved the problem of separation of topic learning and emotion classification in the social emotion classification task for the first time.It designs a unified neural network framework to jointly train the topic model and the emotion classification in a multi-task learning manner.Specifically,we propose a neural topic model for learning the topic feature,and also propose a topic-enhanced self-attention mechanism that fuses the semantic feature and the topic feature.Experiments indicate that this end-to-end neural network can not only significantly improve the performance of social emotion classification,but also has an obvious positive effect on topic modeling.At the same time,the method has a good performance in robustness and training efficiency.(3)Propose a method of social emotion classification based on the semanticsdriven topic encoder: this method integrates the fine-grained word-level semantic feature in topic encoding.We propose a semantic-driven topic encoder to compose the embedding vector of each topic,and obtain the document topic vector based on the topic embedding and the document topic distribution.We also propose a forward self-attention network that considers word order to learn the text semantic feature.Experiments show that our proposed semantic-driven topic encoder surpasses the traditional probability-based topic encoder,and can effectively improve the performance of social emotion classification.At the same time,the method has a good performance in robustness and training efficiency.This thesis has improved the performance of social emotion classification by a large margin from the above three aspects,and provides important reference significance for the follow-up work in the field of social emotion classification.In the future,we can conduct further research from the following perspectives: the prior relations between emotions,event modeling,large-scale language model,and long text encoding.