Research On Weibo Emotion Classification Technology Based On Topic And User Representation

With the rapid development of network technology and smart mobile devices,more and more netizens are posting their opinions and perceptions on social media platforms such as Weibo and Twitter every day.These subjective texts contain rich emotional knowledge.It has important scientific research and application value to identify the emotion in these texts.Although the emotion classification technology based on pre-training models such as BERT has achieved good results,there are still some shortcomings for social media text.For examp le,the pre-training model does not perform special processing on sentiment words,and has insufficient ability to capture sentiment information.And most of the existing methods ignore topic and user information,which is important auxiliary information for social media texts.This article proposes three improvements in response to these shortcomings.First of all,the pre-training model is improved.For sentiment-related tasks such as emotion classification,the pre-training model containing sentiment is obviously more advantageous than the general pre-training language model.Therefore,we propose a method to inject sentiment knowledge into the model in the pre training process using unlabeled data.Compared with other work that uses sentiment-related labeled information for pre-training,our method is more suitable for generalization to all fields for application.Secondly,for social media texts such as Weibo,in addition to the information contained in the Weibo text itself,the topic information in the Weibo and the information of the user who posted the Weibo are also very helpful for emotion classification.This paper proposes a method of fusing coarse and fine-grained topic representations for topic information.First,use news corpus to train topic classifiers for topic recognition on Weibo,and use topic category information as coarse-grained topic information;Then the automatic annotation tool is used to annotate the semantic role of the text and extract event representation as fine-grained topic information.And finally we use the attention mechanism to fuse coarse-grained topic information and fine-grained topic information in the process of text representation.For user information,we start from the user's basic data and historical microblogs.We use the LDA topic model to automatically mine the user's content preferences,and cluster the user groups based on the content preferences;at the same time,we combine the well-known psychological research results "Big Five Personality" to dig out the user's personality traits;then we use the user's personality characteristics and content preference characteristics to update the text representation,and the user's basic information can also provide a certain prior knowledge of emotional distribution.The final experimental results show that the three improvement mechanisms we proposed can effectively improve the performance indicators of emotion classification tasks,and the three methods can be used in combination to further improve the effect of emotion classification tasks.