Research On Aspect-Based Sentiment Analysis Based On Attentional Mechanism

With the development of deep learning,researchers have made use of neural networks to conduct text level sentiment analysis,and obtained very good results.But then the research fell into a bottleneck,the previous technology could not handle the text containing complex emotions,especially when the text contains several contradictory emotions.Therefore,a more fine-grained sentiment analysis is the key to solving the current problem.Therefore,the research direction shifts from the text level sentiment analysis to aspects level sentiment analysis.The main research content is aspect-level sentiment analysis.We put forward three models.In the work of word-level interactive attention network,considering the aspect words are usually not just a word,if average the word vector simply,the network will lose detail information of aspect words.The attention from each aspect word to other words of the sentence is not same,and the attention from each context word to aspects word is not same too.Therefore,a word-level interactive attention network is proposed to solve these problems.The experimental results show that our model is much better than the conventional attention-based network.The research of aspect level sentiment analysis mostly divides the input sequence into aspect sequence and text sequence,and models the relationship between them.However,considering the distance between text and aspect words is an important factor,a new idea is proposed: the local context of aspect words should contain more important information than the global context.Therefore,it's very important to determine which aspect a context word modifies.To solve this problem,we propose a netword based on local self-attention mechanism,which aims to help the model capture local context information,so as to better conduct sentiment analysis for a given aspect.The results on three commonly used datasets show that our model has a great improvement over other algorithm.About the method based on graph convolutional attention,we propose a method to update the adjacency matrix by means of graph attentional mode,so as to carry out graph convolution calculation with attention.Make full use of word dependencies.The affective features are obtained from the grammatical context of aspect words through dependency diagrams.The results on the three data sets show that the results of our model are much better than those of previous models based on graph convolution or graph attention.