Research Of Multiple Semantic Object Sentiment Analysis For Social Media

In the Web 2.0 era,the internet gives more initiative to netizens,and has promoted the rapid development of social media.There are a large number of social media users active on various social media platforms every day,such as Microblog,WeChat,Forum and some shopping websites with social function,and generate a large number of user-involved reviews on people,organizations,events,products and services,etc.These reviews often contain sentiment orientation expressed by users.And such large amounts of reviews that contain subjective emotions of users have became a valuable resource for decision support.From ordinary users to enterprises to governments,they can meet their needs by mining those subjective reviews.Contrapose to growing demands of texts analysis and review resources that grow increasingly complicated from social media,this thesis focuses on finegrained sentiment analysis of social media reviews,and research on multiple semantic objects sentiment analysis for social media.Different opinion targets are extracted from large amounts of social media reviews firstly.Then automated sentiment analysis are conducted to solve the problem of sentiment orientation recognition of multiple specific semantic objects and clarify the semantic sentiment directionality in social media reviews.Current research has applied deep learning methods on fine-grained sentiment analysis tasks,and has achieved many important theories and applications.However,due to the complexity of social media reviews,deep learning methods still have difficulties in identifying different semantic objects and sentiment terms from social media reviews and fully mining the sentiment feature information of social media reviews to determining sentiment polarity of different semantic objects.Therefore,this thesis improve the algorithm or architecture of model based on deep neural networks to improve the performance of model in fine-grained sentiment analysis task.The main research points of our thesis are as follows:(1)Aimed at extracting different semantic objects and sentiment terms accurately and efficiently from social media reviews,we study how to extract specific opinion targets and sentiment terms from social media reviews,and propose an opinion targets and sentiment terms extraction model based on self-attention.The model takes global dependencies of review sequence into consideration,and captures global dependencies and learns internal structure of review sequence effectively by self-attention.Finally,opinion targets and sentiment terms are automatically extracted by performing tag prediction on review sequence.(2)Aimed at determining semantic directionality of sentiment in social media reviews and determining sentiment polarity of different semantic objects in social media reviews,we propose a joint attention LSTM network(JAT-LSTM)for aspect-level sentiment analysis.The model takes contextual information,opinion targets information and sentiment terms information into consideration,and combines attention mechanism to construct a joint attention LSTM network for aspect-level sentiment analysis task to improve aspect-level sentiment classification accuracy.