Combining Data Augmentation With Graph Neural Network For Aspect Based Sentiment Analysis

In recent years,with the development of Big Data rapidly.Many online comments with user sentiment have appeared on media platforms and various online shopping websites.These reviews contain a large amount of emotional information,which can not only help businesses discover users’ emotional tendencies towards goods or services,but also serve as a basis for improving the performance and quality of goods.Text sentiment analysis is a research hot-point in Natural Language Processing（NLP）and has a wide application prospects.Text sentiment analysis can be divided into three categories according to the granularity attribute: Text level,Sentence level and Aspect Level.Compared with text-level and sentence-level sentiment analysis tasks,Aspect Based Sentiment Analysis（ABSA）has more granular information,which is generally reflected in the user’s emotional evaluation of a certain aspect of the goods.This evaluation information can provide more detailed direction and ideas for merchants to improve commodity performance and service delivery.Therefore,ABSA has higher research and business value.ABSA is to judge the sentimental polarity of a given facet in the input text.Currently,the common method is to code the input sentences using a neural network,extract the relationship between the facet words and the contextual emotional words,and make comments based on the related sentimental words to make emotional judgments on specific facets.However,most existing models only pay attention to the contextual position relationship between words in a sentence,which results in poor matching performance for long-distance non-adjacent words,ignoring the grammatical structure of the sentence and failing to reflect the grammatical structure of the text.In addition,the methods to solve the ABSA problems often use the neural network,which will have its own dependence on the quality and quantity of training marker data.In the absence of training data or labeling is not ideal,the neural network will often perform poorly,even not robust and stable.Based on the above background and problems,this paper presents an ABSA combining data augment and Graph Neural Network（GNN）.The main contents of this paper are as follows:（1）This paper presents a new model for ABSA,called RGAT-BAT which combines adversarial training with relational graph attention network（RGAT）.RGAT have a multilayer architecture,each layer uses neighbor feature codes and updates the representation of nodes in the graph.RGAT can be used to effectively utilize sentence grammar information and word dependency in the structure of semantic grammar tree.Based on using grammar information,integrating information that depends on the tag itself can more accurately capture the relationship between words and effectively solve the matching problem of long distance non-adjacent words.Adversarial training is a kind of data augment technology.It can dynamically create new training data from previously marked data in model training and can be trained by disturbing the neural network to improve the resistance to malicious samples and enhance the robustness and stability of the neural network.At the same time,adversarial training is also a regularization method,which can solve the problem of fitting and improve the generalization ability of the model.（2）This paper presents a model（RGAT-Mixup）that combines Mixup with RGAT.Using Mixup technology in conjunction with diagrams to focus on the network for ABSA.Mixup technology is an effective tool for expanding smaller datasets in data enhancement methods,which can be used to match the amount of tagged data required for in-depth learning models and improve model performance in limited data volumes.Using Mixup in conjunction with diagrams can help reduce the cost of model development,solve the problem of insufficient tag data,and ensure model performance.The model presented in this paper was experimentally validated in the Sem Eval2014-Task4 Restaurant,Sem Eval 2014-Task4 Laptop,and Twitter reviews datasets.The experimental results show that this method can effectively improve the accuracy of emotional judgment and model stability.