Research On Aspect Level Text Sentiment Analysis Based On Deep Learning

In recent years,the global Internet has developed rapidly,and a large number of texts containing users’ emotional tendencies have emerged.Timely and accurate analysis of the emotions shown in such texts can effectively help to grasp the people’s emotional attitudes towards commercial products and policies and regulations,and provide a strong basis for subsequent changes in product marketing strategies or policy regulation.Traditional text sentiment analysis analyzes text sentiment at a coarsegrained level,which can no longer meet the needs of current practical applications.Therefore,aspectbased text sentiment analysis,which can judge text sentiment at a fine-grained level,has become a hot spot in natural language processing research.This thesis focuses on the problem of insufficient text information mining in existing aspect-based sentiment analysis methods.The main work is as follows:(1)An aspect-based sentiment analysis model based on multi-channel attention fusion network is proposed.This model uses the BERT pre-training model to convert the initial text data into text word vectors with context information.The position information of the text sequence is captured by the bidirectional gated recurrent unit neural network,and the text features are extracted.Further build a multi-channel attention feature fusion layer,introduce a text convolutional neural network to extract text local information of different sizes in multiple channels,and introduce an attention mechanism in each channel,and use the context’s attention information for aspect words to improve the model accuracy.Experimental results show that this model has achieved good results on public datasets.(2)An aspect-based sentiment analysis model based on interactive attention and multi-channel residual network is proposed.This model introduces the aspect word enhancement mechanism to enhance the aspect meaning of the text,and constructs a feature extraction layer composed of a bidirectional gated recurrent unit neural network and an interactive attention mechanism for preliminary text feature extraction.Further use the text convolutional residual network and the multihead self-attention mechanism to construct a multi-channel residual layer to mine the feature vector for multi-fine-grained hidden text information.Experimental results show that this model has a certain effect improvement compared with several benchmark models.(3)Based on the proposed aspect-level text sentiment analysis model,this thesis designs and implements a text sentiment analysis system.Firstly,the needs of users are analyzed,and the overall architecture of the system is designed.Then,the system is developed on the relevant development platform,and the function display of the relevant modules is further given to verify the practicability of the system.