Research On Aspect-based Sentiment Analysis For Weibo Text

With the vigorous development of the Internet,various online social media platforms have gradually come into people’s attention.As a popular social software nowadays,Sina Weibo is very popular due to its large number of users and wide range of information coverage.Internet users can share their views on hot topics in real time through Weibo,and massive text data is constantly generated.In today’s increasingly complex network language,compared to traditional coarse grained emotional analysis,Aspect-Based Sentiment Analysis(ABSA)can capture the emotional tendencies of netizens from different perspectives under specific topics,mining more comprehensive and richer information from the text.This not only provides enterprises with marketing and management decisions,but also assists national and government agencies in timely monitoring the current trend of public opinion.Therefore,it is of practical significance to conduct aspect level emotional analysis on Weibo texts.Single deep learning models,such as Recurrent Neural Network(RNN)and Convolutional Neural Network(CNN),have limitations in ABSA tasks and cannot fully extract text features.This paper takes Weibo text as the main research object,and makes relevant improvements based on RNN and CNN models to improve model performance.The main work contents are as follows:(1)Obtain all comment texts under specific microblog entries through crawler technology,and conduct data cleaning based on the characteristics of the microblog text.Cut clauses on the comment text,annotate the dataset with aspects and emotional polarity through programs and manual operations,and create a microblog text dataset that includes all aspects.(2)To address the problem that a single RNN cannot extract local text features and that a single CNN cannot learn long-term dependencies,this paper proposes an AGGCNN model based on dual channel feature extraction.The model utilizes a tower structure using Text Convolutional Neural Network(Text CNN)and Bidirectional Gate Recurrent Unit(Bi-GRU)modules to extract features from both global and local relationships in parallel,reducing feature loss to a certain extent.The model introduces a gating mechanism and an attention mechanism to further filter contextual information unrelated to aspect words,while embedding aspect word information in the feature extraction stage,fully realizing the interaction between specific aspects and context.Compared to the benchmark model,the classification performance of the AG-GCNN model proposed in this article has improved on the public dataset Sem Eval2015.Comparative experiments have proven the effectiveness of the model.In addition,the AG-GCNN model also exhibits good classification results on the microblog dataset produced,and the experimental results verify the practicality and generalization of the model.(3)This paper uses the method based on word frequency and combined with manual screening to preliminarily extract aspect words in the comment text,and then clusters the extracted aspect words to obtain various aspect categories for specific microblog topics.Meanwhile,the comments on various aspects of the topic are statistically summarized and analyzed to obtain netizens’ views and emotional tendencies towards the topic.The paper also provides suggestions on the problems reflected in this social topic.