Research On Video Bullet Chatting Orientation Based On Multidimensional Sentiment Dictionary And Deep Learning

As the comment function of bullet chatting is put into use and recognized by users on video playing platforms,the amount of bullet chatting in videos increases.However,it is quite difficult to extract and analyze the emotional information of bullet chatting data in specific fields by using the existing emotion dictionary.This thesis takes the video bullet chatting of Bilibili as the research object.the purpose of this study is to the problem that the conventional emotion dictionary does not apply to the corpus in the field of video bullet chatting and has a relatively single dimension in emotion analysis,the method is based primarily on a multi-dimensional sentiment dictionary associated with deep learning be put forward study and analyze the bullet chatting sentiment orientation.The results of the research studies on the emotional orientation of the bullet chatting at Bilibili are obvious,indicating the superiority of the combination of the two methods.The research work of this thesis mainly keeps to the following aspects:(1)Selection of computing algorithm for the similarity between words.Firstly,the pros and cons of two commonly used similarity calculation algorithms in the Chinese context are analyzed.Then,the semantic similarity calculation algorithm between words in Cilin and CNKI is improved to improve the accuracy of similarity calculation between words,laying a technical foundation for the subsequent combination of multiple emotion dictionaries with different dimensions.(2)Bullet chatting orientation classification based on multidimensional sentiment dictionary and semantic rules.Firstly,the fusion similarity algorithm is used to merge several emotion dictionaries according to the seven dimensions of emotion.Then,the multi-dimensional emotion dictionary is expanded by collecting annual bullet chatting,emoticons,and emoticons of Bilibili,and manually screening the emoticons of movies of different styles.However,it is limited to relying only on the dictionary to analyze the emotional tendency of bullet chatting.Therefore,this thesis analyzes bullet chat based on the semantic rules of bullet chat and provides different processing methods for common sentence patterns and determiners.Finally,based on the self-constructed multi-dimensional emotion dictionary combined with semantic rules,an improved method to calculate the emotion value of bullet chatting is used to complete the sentiment tendency analysis of bullet chatting.(3)Projectile bullet chatting orientation Classification based on deep learning.Firstly,deep learning Bert pre-training model combined with Softmax classifier is selected for bullet chatting without obvious emotional tendencies and complex emotion.Then,the model training process is fine-tuned to complete the classification research on the emotional orientation of bullet screen content.Finally,the robustness and generalization of Bert model are improved by introducing adversarial training FGM.To sum up,this thesis has completed the tendentious research on the video bullet chatting of Bilibili,which is helpful to extract the emotional information carried by videos and provide a video retrieval method to meet the personalized needs of viewers and increase the user viscosity of video playing platforms.