Research On Aspect-level Text Sentiment Analysis Methods Based On Graph Convolutional Neural Networks

In recent years,with the rapid development and popularity of the Internet and social media,a large amount of data containing rich sentiment information has emerged.Analyzing these data to mine and extract valuable information has become the focus of attention in various industries today.As an important subfield in the field of natural language,aspect-level sentiment analysis is different from traditional sentiment analysis that only focuses on the overall text.It performs fine-grained analysis of the emotional information expressed by each aspect object in the text,and has greater application value,and its main application scenarios are in various scenarios such as personalized recommendation,customer service,and market research.Graph convolutional neural networks have shown significant advantages in aspect-level sentiment analysis tasks and are widely used in aspect-level sentiment analysis.In this thesis,we delve into the aspect-level sentiment analysis task,refine and propose two aspect-level text sentiment analysis models based on graph convolutional neural network methods,and then combine these two approaches to design and implement a system for aspect-level sentiment analysis functions.The main research components include:(1)Approaches based on syntactic dependencies tend to ignore the types of dependencies in sentences,and the importance of different dependencies is different,so ignoring dependency weight information can affect the judgement of sentiment polarity.In this thesis,we propose DWAGCN,an aspect-level sentiment analysis model based on a graph convolutional neural network model that automatically learns dependency weights.The model encodes dependencies,automatically learns dependency weights through dependency types,and employs an attention mechanism to enhance sentiment features around aspect terms,thereby improving the accuracy of sentiment polarity prediction.Experiments were conducted on four relevant benchmark datasets,and the experimental results show that the DWAGCN model performs better than traditional syntactic dependency tree approaches in aspect-level sentiment analysis tasks and achieves superior performance.These experimental results demonstrate the reliability and validity of the model.(2)To address the problem that existing interactive graph convolutional networks often ignore semantic relevance,the SInter GCN model is proposed,which is an improved aspect-level sentiment analysis model for interactive graph convolutional networks.The model combines syntactic and semantic structures to capture the relationships between key aspects and context.Dependency graphs are first constructed for each sentence,and then the structural graphs are combined with the associated semantic graphs.Dependencies between aspect words and other aspects were then modelled using aspect-oriented GCN to interactively extract the sentiment features of aspect words.Experiments were conducted on four relevant benchmark datasets,and the results showed that this model outperformed existing interactive methods on all four benchmark datasets.This validates the validity and feasibility of the model.(3)Using the two graph convolutional neural network models proposed in this thesis,an aspectual word sentiment analysis system is designed and implemented.With functions such as text uploading,pre-processing,dependency graph construction and aspectual word sentiment polarity prediction,the system lays the foundation for further large-scale text processing and provides a convenient tool for aspect-level sentiment analysis.