Aspect-based Sentiment Analysis Research Based On Graph Convolutional Neural Network

With the development of science and technology,social networks have become increasingly prosperous.How to mine the emotions expressed by users from massive text information has attracted widespread attention from scholars and has become a research hotspot in the field of Natural Language Processing（NLP）.Sentiment analysis is divided into three categories: document level,sentence level and aspect level.Document and sentence sentiment analysis take the whole document or specific sentences in the document as the analysis object respectively.It is often used for text content affective recognition in scenes such as product evaluation,film review and video subtitles.Aspect-based Sentiment Analysis（ABSA）aims to judge the emotional polarity of words or phrases in a sentence,so as to analyze the emotional color contained in the sentence or document in a more detailed manner.The word or phrase here is usually called "Aspect".At present,aspect-based sentiment analysis based on deep neural network has become the mainstream of research.Combining it with other coding structures,the deep network model has been favored by many scholars.However,the current research methods on affective analysis can not pay attention to the context and grammatical information of aspect words at the same time,and the semantic information of sentences is not introduced,so the model can not correctly judge the affective tendency of aspect.Traditional emotion analysis methods can extract data features with fixed geometric relationships.In reality,there are many data information containing graph structure.Graph convolution neural network can effectively capture the dependency of graph structure.Therefore,based on the graph convolution neural network,this paper fuses the different neighborhood information of aspects and the semantic information of text,and then constructs different aspect-based sentiment analysis models to capture the key features to determine the emotional polarity and improve the model performance.The main work of this paper includes:（1）Propose an aspect-based sentiment analysis model with Embedding Different Neighborhood Representations（EDNR）.The convolutional neural network is used in combination with the nearest neighbor strategy to obtain aspect neighborhood information,which reduces the influence of distant irrelevant information on sentiment polarity.Meanwhile,a graph convolutional neural network is used to extract node neighborhood information to obtain the grammatical dependencies of sentences.After fusing the two kinds of neighborhood representation information,the attention mechanism is used to pay special attention to the important information that determines the sentiment polarity of the aspect.An information evaluation coefficient is also proposed to evaluate the influence of contextual and grammatical information on sentiment polarity.Experiments are conducted on 5 public datasets,and the experimental results demonstrate the effectiveness of the EDNR model.（2）Propose an aspect-based sentiment analysis model Incorporating Prior Knowledge（IPK）.A knowledge graph between words is constructed,and a graph convolutional neural network is used to capture the dependency information of the knowledge graph.According to the input text,get the part-of-speech information of the word.By fusing part-of-speech features with knowledge information,the final prior knowledge feature representation can be obtained.The IPK model provides corresponding semantic and part-of-speech information for each word,reduces the adverse effects caused by the vague expression of emotional information in sentences,and effectively discriminates the emotional polarity of aspects in complex natural sentences.Comparing experiments and analysis with other models on the standard data set,the experimental results show that the model can obtain higher accuracy and F1 value.The model proposed in this paper can effectively extract data features with graph structure,obtain richer emotional information coding,improve the effect of aspect-based sentiment analysis of the network,and promote the application of ABSA tasks in social platforms,e-commerce platforms or public opinion platforms.