Research On Fine-grained Sentiment Analysis Of Weibo Comments Based On Topic Model

Because Weibo can collect noticeable events in time,and the public can also express their opinions through comments,it is widely respected by the public.Weibo has gradually occupied an important position in the majority of social media platforms.Weibo comments are generally characterized by high-dimensionality and sparse semantics,and generally contain strong emotions.By studying Weibo comments,the public's thoughts and attitudes can be effectively observed.The fine-grained sentiment analysis of Weibo comments related to a certain event can clearly understand the public's various emotional tendencies towards the event.This thesis is dedicated to the topic model-based fine-grained sentiment analysis of Weibo comments,which can be divided into aspects extraction and aspect sentiment analysis tasks.In the aspect extraction task,the topic model's analysis method based on syntactic rules is used to label the comments.This thesis uses syntactic rules to split and expand the context of comments involving many aspects and too short comments.At the same time,the number of comments is relatively small,and the comments are characterized by strong context dependence and sparse features.In order to solve the problem that comments are highly context-dependent and common comments contain many topics,this thesis proposes a GBs-LDA topic model based on generalized Polya Urn Model(GPU),BERT model and sen LDA.Firstly,the model uses BERT language model to represent words semantically,and generates semantically related word sets according to the cosine distance of each word vector.Then,the comments are processed by the sentence upgrading module to get the comment sentences with stronger themes.Finally,the GPU model is used to complete the sentence features extension according to the semantically related word set,and then the topic extraction,that is,aspect extraction,is carried out through the improved LDA model.In the aspect sentiment analysis task,the sentiment polarity classification is performed on the comments extracted from the aspect.This thesis proposes a pretrained BERT-sen model fused with sentiment dictionary.Here,firstly input the comment clauses with aspect labels and the Weibo comment dataset into the BERT-sen model for emotion polarity training,and then predict the emotion polarity of the comment clauses,and then count the number of emotion polarity of comments in various aspects.So as to achieve fine-grained sentiment analysis of comments.Experiments show that the fine-grained sentiment analysis model designed in this thesis exhibits good semantic analysis capabilities and sentiment analysis capabilities.Compared with the traditional model in which a comment belongs to a topic,the topic model proposed in this thesis can obtain more comprehensive and accurate clustering results.The topic mode plays a foundational role for the overall fine-grained sentiment analysis model to obtain better results.At the same time,the BERT-sen model based on the fusion sentiment dictionary expands the characteristics of comments through the sentiment dictionary,which can more accurately reach the sentiment polarity.