Aspect-based Sentiment Analysis Research Based On Graph Convolution

With the increasing requirements of users for text sentiment analysis,sentiment analysis is changing from a coarse-grained level to a fine-grained level.Aspect-based sentiment analysis aims to identify the sentiment category of each specific aspect/target in a given text.As one of the important sub-tasks of sentiment analysis,it can dig out more specific and deeper emotions of different aspects for users.Therefore,it has become an important basis for decision-making in many fields,has great application value,and has become a hot research direction in natural language processing.Important challenges of this task include how to better model the semantic relationship between aspects and contexts,and to dig deeper into the grammatical dependence of the text.Aiming at the above two issues of aspect-based sentiment analysis,this paper studies graph convolutional networks that can obtain text grammatical information,and proposes a graph convolution model combined with multi-head attention mechanism and a graph convolution model combined with gated convolution,the specific tasks are:(1)In order to make full use of the grammatical dependence information of the text to improve the model's emotional polarity prediction performance,this paper proposes an aspect-based sentiment analysis model based on the multi-head attention mechanism and graph convolution.This model combines the advantages of graph convolutional network to effectively extract text grammatical information and Bi-LSTM network to extract long-distance timing dependent information;introduces the position information of aspect words to improve the performance of information retrieval;multi-head attention mechanism is introduced to learn the influence between aspect words and context,and effectively capture the interactive semantic information of aspect and context.(2)In order to enrich the local relationship information to further improve the aspect sentiment polarity prediction performance of the graph convolution model,this paper proposes an aspect-level sentiment analysis model based on gated dilated convolution and graph convolution.This model uses gated dilated convolutional network to enrich local relation feature extraction information,and at the same time incorporates aspect information when encoding context;aspect-specific graph convolutional network can better capture grammatical information and long-distance word dependence,and further extract aspect features;introduce position-aware transformation to reduce noise and deviation in the process of graph convolution.(3)Experiments on the Twitter comment dataset and the restaurant and laptop comment corpus from the Sem Eval-2014 ABSA dataset have proved the effectiveness of our proposed models.The performance of the model based on the multi-head attention mechanism and graph convolution is better than most classic models.The model based on gated dilated convolution and graph convolution effectively improves the performance of the aspect-specific graph convolution model,and the effect after using the Bert pre-training model for word embedding is better than most of the latest effective models.