Research On Aspect-based Sentiment Analysis Method For User Online Reviews

With the advent of the Internet era,various social and e-commerce network platforms have entered people’s lives.While enjoying the immediacy and convenience of the Internet,people also generate a large amount of comment data on the Internet.These users’ online comment data contain a large number of opinions,which contain rich scientific research value and social business value.Opinion mining and sentiment analysis of these data has become one of the research hotspots in the field of natural language processing.In the process of opinion mining and sentiment analysis of these users’ online comment data,because the text of the user’s online comment data is short,highly targeted,and contains multiple specific aspects of emotional polarity,deeper understanding and modeling are required.Therefore,aspect-level sentiment analysis methods have attracted more and more attention of researchers.In the research process of aspect-level sentiment analysis methods,compared with traditional algorithms based on sentiment dictionaries and traditional machine learning,deep learning does not rely on artificially constructed features,and automatically learns and extracts features through neural networks,which is very suitable for complex language and text data.It has become the mainstream method for aspect-level sentiment analysis.Based on this,this paper uses the deep learning method to build an aspect-level sentiment analysis model,analyzes the single-domain and cross-domain user online comment data involving transfer learning,and designs two aspect-level sentiment analysis models.The main work and results of the paper are as follows:(1)In the single-domain aspect-level sentiment analysis,the graph convolutional network model based on dependency tree has inaccurate results of text dependency analysis during modeling,and the informal expression of users’ online comment data results in insufficient grammatical information.A bi-graph convolution aspect-level sentiment analysis model with enhanced syntactic-semantic information.The model conducts research from both syntactic and semantic perspectives,first using Bi LSTM as a sentence encoder to extract hidden representations of aspect words and context.Then,syntactically,on the basis of the previous output of the dependency parser,a dependency probability matrix weighted by the grammatical distance is constructed,and the useful information of the grammatical structure is obtained by combining the position distance feature,and the feature extraction of the graph convolutional network based on the dependency tree is improved.The ability to solve the inaccuracy of text-dependent parsing results;Semantically,an attention adjacency matrix is obtained by introducing an attention mechanism and then combined with a graph convolutional network to complete feature extraction,which solves the problem of insufficient grammatical information in online comment data.Then the feature information of the two modules is fused to improve the predictive ability of the semantic and grammatical feature-based network model for the sentiment polarity of aspect words.Finally,experiments are conducted on three public aspect-level sentiment analysis datasets and the latest research methods,and the comparative results demonstrate the effectiveness of the model.(2)In cross-domain aspect-level sentiment analysis,the models trained in specific domains have domain dependencies and distribution differences between different domains,resulting in poor model generalization ability.A cross-domain aspect-level sentiment analysis model based on attention mechanism and domain adversarial network is proposed.First,the sentence pairs composed of aspect words and sentences are used as the input of the pretrained model BERT algorithm.Then in the feature extraction layer,the dependency syntactic knowledge and the interactive attention mechanism are used to improve the feature extraction ability of the source domain and the target domain,which is convenient for the acquisition of shared features in the subsequent domains.Finally,the sentiment classification module is used to classify the source domain data.The domain classifier module realizes the feature confusion of the source domain and target domain data,improves the domain shared feature extraction by adversarial training,and resolves domain dependencies and distribution differences.Improve model generalization ability.Finally,the validity of the model is verified through experimental comparison.