Research On Fine-grained Sentiment Classification Model Based On Deep Learning

With the development of user-oriented internet product models,fine-grained sentiment analysis tasks have gained broader application value in fields such as social media and e-commerce.This thesis mainly focuses on aspect-based sentiment classification tasks in fine-grained sentiment analysis,which judges the corresponding sentiment polarity of different entities and objects in sentence-level text.Traditional methods based on recurrent neural networks have the problems of vanishing and exploding gradients.Models based on attention mechanisms lack the ability to explain syntactic constraints and long-range word dependency relationships.Methods based on graph networks often face parsing errors and noise problems introduced by syntax structures that are irrelevant to sentiment judgment.This thesis will focus on the shortcomings of these methods and explore a series of improvement and optimization strategies,with the following main research contents:1.Based on the motivation of integrating multiple syntactic and semantic information to enhance the energy absorption of fine-grained sentiment analysis models,this thesis proposes a graph network extension model,SSE-GAT(Semantic Syntactic Enhanced GAT).The model aims to use word vectors,dependency relationships,part-of-speech tags,and position features in the text as the basis for sentiment judgment.Comparative experiments with current mainstream models show that SSE-GAT can better capture sequence-based contextual semantic information and graph-based syntactic structure information.2.Based on the motivation of optimizing syntactic dependency trees to integrate multiple parser output results and reduce erroneous parsing noise,this thesis explores a syntax dependency tree optimization method for aspect words.Through reconstruction and pruning operations,a Reshaped Pruned Dependency Tree(RP tree)structure is obtained,and a merging mechanism called Multiple-Priority-Based Merge(Mpb-Merge)is proposed based on the summary of experiments with Stanza,Biaffine,and LAL dependency parsers to fuse conflicting dependency trees.3.Regarding the special structure of the aspect-oriented RP dependency trees,this thesis further proposes a Double-Distance Enhanced GAT(DDE-GAT)model,which introduces relative position information related to aspects in the text sequence and dependency tree to enhance the performance of the graph network model on the RP dependency tree structure.Two different fusion mechanisms based on a voter and the Mpb-Merge algorithm are provided,and a series of ablation experiments demonstrate that the dependency tree optimization mechanism proposed in this thesis can effectively improve the performance of the model in aspect-level sentiment analysis tasks.4.Based on the motivation of providing a visual process for dependency tree-related fine-grained sentiment analysis tasks,this thesis modularizes the workflow of SSE-GAT and DDE-GAT models into five stages: data input,pre-processing,dependency tree construction,model prediction,and result visualization.An interactive prototype system focusing on front-end user experience is also designed.