Research On Aspect-based Sentiment Analysis Method Based On Deep Learning

The current Internet has entered an era of intelligent development,with a large amount of text data containing personal emotional colors widely disseminated in social networks.Sentiment analysis is a basic and traditional data mining task for mining these emotional information.With the continuous exploration and development of deep learning technology in the field of natural language processing,the research on sentiment analysis is also increasingly fine-grained.Aspect-based sentiment analysis is a more challenging fine-grained sentiment classification task,which ranges from the single sentiment classification task to the multi-extraction task.Its research tasks are more refined and complex.For aspect-based sentiment analysis of single tasks,current research methods usually require the fusion of external information,which can easily generate a large amount of non emotional noise during the fusion process and lack effective utilization of long-distance word grammar information.For aspect-based sentiment analysis of multi-extraction task,a large number of current research methods focus on the extension of graph convolutional networks,and relatively lack the research and utilization of combining positional information,which easily overlooks the directionality of grammar dependencies and the dependencies between subtasks.Therefore,this article conducts research from multiple task perspectives of aspect-based sentiment analysis,and proposes corresponding solutions based on the characteristics and specific problems of different tasks.The main research content of this article is as follows:A lightweight feature-enhanced dual GRU model is proposed to solve the problem of additional noise caused by the direct concatenation of aspect term information.This method exploits the aspect-related GRU and position-related GRU with fewer parameters,to guide the learning way of position information and feature information of aspect terms,and constructs the modeling context sequence of context-related GRU.The proposed improved GRU can control the inflow of inessential information,selectively learn aspect information,and reduce the additional noise introduced in the process of directly introducing information.In order to enhance the dependency between context and aspect terms,a context-related GRU is constructed,and the attention mechanism is used to improve the dependency between aspect terms information and context grammar,so as to force the model to focus on the important sequence features between aspect terms and highly relevant context,to judge the emotional polarity of specific aspect terms in the aspect-based emotion analysis task.The experiment was carried out with conflict labels to verify the effectiveness of the our model in the singletask aspect-based sentiment analysis,and to verify the rationality of different GRU variants through ablation study.An aggregated graph convolutional networks based on extended graph dependency is proposed to solve the problem of poor node representation and long-range syntactic dependencies in convolutional graphs.This method introduces two different ways of aggregation functions to iteratively update the local neighborhood representation of each node based on the original graph convolution,and models aspect information and context information respectively.Considering the long-distance dependency problem in current grammatical relations,this method constructs a contextual sub-neighborhood matrix to broaden the "receptive field" of the target nodes,and then uses an attention mechanism to capture the sentiment dependency between different nodes’ feature information and predict the sentiment polarity.Experiments on a large number of Chinese and English datasets show that AGCN’s ability to aggregate features is better than the original graph convolution model and can effectively improve the effectiveness and accuracy of sentiment judgment for aspect terms.Aiming at the problem of location information missing in the co-extraction task,a location-aware graph convolution network for co-extraction tasks is proposed to extract aspect terms and predict their sentiment polarities in a mutually reinforcing way,and solve the problem of missing position information of aspect words and the internal dependency between joint extraction tasks.This method abandons the specific position index of the original aspect terms,but constructs the weight matrix graph with aspect terms position information through the number of node connections in the syntax dependency tree,and uses the convolution function to model the information features of the aspect terms to achieve the aspect terms extraction task.In addition,by considering the long-distance dependency and reconstructing the syntax dependency tree,a location-aware graph-based convolutional network is proposed to aggregate more context-related affective dependency features.The improved performance of the model on the aspect terms extraction task is conducive to the downstream judgment of the emotional polarity of the aspect word,and the correct sentiment polarity also helps the model improve the accuracy of the aspect terms extraction task.Based on four benchmark data sets to evaluate our model,the experimental results show that our method can achieve good performance in different aspect-based sentiment analysis tasks.A unified dependency-enhanced graph convolutional network is proposed for the interdependencies between subtasks in the triple extraction task,which aims to extract aspect terms,opinion terms,and the corresponding sentiment polarity in sentences,and can simultaneously deal with complex dependency problems between multiple objects,as well as corresponding type dependency and directional dependency problems.This method reintegrates the directional dependency in the syntax dependency tree,integrates the dependency type information to represent the features of rich context sequence,and then defines the specific dependency between contexts according to the directional dependency tree,and constructs the aggregator function to capture the syntactic features in the direction.The final underlying classification structure uses different forms of classifiers to deal with the dependency between subtasks and achieve multiple extraction tasks in aspect-based sentiment analysis.Through comparative experiments and sensitivity analysis on different extraction tasks,the feasibility and effectiveness of our model on multiple extraction tasks in aspectbased sentiment analysis are verified.