Research On Aspect-Level Sentiment Analysis Based On Deep Learning

With the rapid development of online shopping,a large number of review information related to commodities with research value has been generated.This information has important guiding significance for the development of the platform,the reputation of the merchant and the shopping experience of consumers.The amount of these comments is huge,and it is difficult to analyze and deal with it by manpower alone.Sentiment analysis is a technique that uses natural language processing related techniques to study these commodity reviews and extract valuable emotional information from the reviews.However,ordinary sentiment analysis cannot identify the sentiment of different aspect-terms in the same comment,while aspect-level sentiment analysis can identify different comment sentiment in the comments and the different sentiment of these review aspect-terms,which is a more fine-grained sentiment analysis task.For the aspect-level sentiment analysis,this paper mainly does the following work.(1)A evaluation aspect-term extraction model based on GCLSTM-CRF is constructed for the evaluation aspect-term extraction task.The model uses CNN to extract local window information,LSTM extracts the context-dependent information of the comment,and then automatically combines the two kinds of information with the gate structure.Finally,the learned feature representation is sent to the CRF layer to globally search for the optimal tag sequence.After experimental comparison,the evaluation aspect-term extraction model based on GCLSTM-CRF makes full use of local information and context information,and has achieved good results in evaluation aspect-term extraction.(2)The M-IAN model based on attention mechanism and a DIMN model based on multiple attention interaction are constructed for the aspect-level sentiment polarity discriminant task.The M-IAN model uses LSTM to model the comment context and the comment aspect-term,respectively,and uses the attention mechanism to extract the semantic information between the aspect-term and the context.The DIMN multi-layered attention interaction model uses the LSTM hidden layer output of the aspect-term and the context as the memory,adding two linear layer structures,and continuously transmitting and capturing multiple attention information with the memory structure,solving the problem that the complex sentence features are difficult to capture.After experimental comparison,the aspect-level sentiment polarity discriminant model DIMN fully utilizes the multiple attention information to solve the problem that the complex sentence features are difficult to capture,and the recognition accuracy is greatly improved compared with the comparison model.