Multi-feature Fusion Based Model For Aspect Based Sentiment Analysis

Along with the rapid development of mobile internet,internet users are not only the user of content,but also the creator.China has a large number of groups of internet users,which is growing at a fast speed.Internet users frequently express their comments and opinions on various applications.Faced with such large-scale emotionally text,it is worth to deeply explore the potential value of the content to serve the society.Aspect-based sentiment analysis is a fine-grained sentiment analysis task,which aims to analyze and predict the sentiment polarity(Positive,Neutral,Negative,etc.)of the aspects.The current approach uses neural networks to encode the sentences and uses attention mechanisms to operate the relationship weights among context for classification.The prediction of aspect-based sentiment analysis is sensitive to both the position and context,and aspect present different sentiment polarity in different sentence structures.With the development of neural networks of structure,syntactic structures generated by external parsers have been shown to provide valid structural information,But the utilization of structural features of sentences by existing sentiment analysis methods has yet to be improved.Moreover,although graph neural networks have been used in recent studies to solve tasks which related to dependent syntax with excellent results,the dependency label information on the edges of the dependent syntactic tree,which also contains important relational feature representations and is useful for disambiguation,is still neglected.To solve the above problems,this thesis proposes an edge-enhanced based graph convolutional networks for aspect-based sentiment analysis.The model defines the syntactic dependency tree of sentence as a graph into the graph convolutional network,which provides the model with powerful and reliable syntactic structure features,and the syntactic structure can effectively help the model to obtain the semantic relevance among contexts.Then the dependency label information on the edges of the syntactic dependency tree is combined by an edge enhanced module,which strengthens the representation of the relationship features among words and makes the update feature of nodes no longer limited to neighborhood nodes.Experiments conducted on five standard datasets show that the model proposed in this thesis can effectively exploit the syntactic structure and relational features in sentences,the model can train an aspect representation that more accurately reflects the relationship among words,in addition,it can achieve higher accuracy rates using this aspect representation.Besides,this thesis also design and develop a comment-oriented aspectbased sentiment analysis system using BE-GCN,which can effectively predict the sentiment polarity of comment in specified aspects and display the sentiment analysis results with an easy-to-use interface.