Research On Aspect-level Sentiment Analysis Method For User Review

The aspect-based sentiment analysis task involves identifying aspect words in user comments and determining the corresponding aspect’s sentiment polarity.Currently,there is a significant increase in the number of online user reviews,and sentiment analysis on specific aspects of review text can extract valuable emotional information for research and commercial purposes.As review texts become more complex,and customer demand for sentiment information from user reviews increases,aspect-based sentiment analysis has evolved into multiple fine-grained subtasks,deep learning-based sentiment analysis methods are widely used to solve these complex sentiment analysis subtasks and play an essential role in areas such as product marketing,online social networking,and government administration.The application of deep learning techniques has improved the accuracy of aspect-level sentiment analysis and enabled it to meet the needs of sentiment analysis in various complex situation.Current research in the area of aspect-level sentiment analysis still has inadequacies: the existing sentiment analysis models have low robustness and generalization ability,which may lead to erroneous analysis of the models when there are perturbations in the input;the models are not aided by lexical information to fully extract important local feature information in and around the aspect words;in addition,the semantic information and dependency relationships among words are not fully utilized,and the intrinsic connections among aspect words,sentiment polarity and opinion words are ignored.To address these issues,this thesis focuses on two aspects:(1)Aspect-level sentiment classification based on dynamic adversarial training and part-of-speech fusion.Aspect-level sentiment classification is a research focus in the field of aspect-based sentiment analysis,which aims to determine the sentiment polarity of a given aspect in a user’s comment sentence.In view of the current issues of low robustness of models and insufficient feature information extraction capabilities in aspect-level sentiment classification research,this thesis presents an aspect-level sentiment classification model combined with dynamic adversarial training and part-ofspeech fusion,which uses BERT-generated word vectors instead of the DPCNN model’s word vectors,and BERT can capture semantic information in sentences and longer distance dependencies.By using dynamic adversarial training techniques to generate dynamic adversarial examples for the training phase of the model,the robustness of the model has been improved,and the deep CNN model-DPCNN can fully extract context feature information without significantly increasing the computational load.Additionally,the part-of-speech fusion layer integrates important part-of-speech information with context information,extracts the features of the essential words and hidden semantic information near aspect words.This thesis conducted experimental comparisons between the proposed model and other baseline models on four publicly available datasets,and the experimental results validated the effectiveness of the DATPF model in this thesis.(2)Aspect sentiment triplet extraction based on dependency graph convolution and multi-layer attention.Opinion words are essential in explaining the judgment of sentiment polarity and have a deep connection with aspect words and sentiment polarity in aspect-based sentiment analysis.Thus,this thesis focuses on the aspect sentiment triple extraction task.The proposed aspect sentiment triplet extraction model is combined with dependency graph convolution and multi-layer attention,which addresses the neglect of the connection between triplet elements and the semantic information and dependency relationship between words in previous studies.First,introduce part-of-speech information and analyze the dependency relationships and syntactic structure of the sentence.Then,by using a multi-layer attention mechanism,extract the position information of aspect words and opinion words,and model the context to improve the model’s contextual awareness and obtain a feature representation of the target word containing contextual information.Furthermore,using graph convolutional networks,the text data is constructed into a text graph based on dependency relationships.Dependency graph convolution can fully learn the semantic information and dependency relationships in the sentence,and extract deep-level feature information.Finally,analyzing the word-pair relationship tags in the GTS grid and obtain the sentiment triplets by decoding the tag decoder.This thesis conducted experimental comparisons between the proposed model and other baseline models on four publicly available datasets,and the experimental results demonstrated the superiority of the AGGTS model in this thesis.