Research On Attention Convolutional Network For Fine-grained Sentiment Analysis

In recent years,with the rapid development of various types of e-commerce platforms,ecommerce content has been generated.The comments of users on social platforms and ecommerce platforms contain a great deal of text information,which describes the sentiment polarities of users towards certain events or goods from various aspects.The analysis of these sentiment information can help businesses to improve their products and services,help the government departments to understand the public opinion,and let the public understand the social evaluation of the event.Document-level sentiment analysis methods can obtain the sentiment polarity of the whole text,but it is difficult to analyze different aspects of sentiment.In the real world,the fine-grained sentiment can provide more informative semantics than document-level sentiment.Aspect-based sentiment analysis,as a fine-grained task of sentiment analysis,aims to obtain the sentiment polarity of all aspects in the sentence,which has high application value.This paper focuses on fine-grained sentiment analysis task based on attention and convolution mechanism.The main work of this paper is as follows:(1)In aspect-based sentiment analysis task,the standard convolution neural network is only used to obtain local information of context,which leads to insufficient information understanding.Therefore,we propose an aspect level sentiment analysis method based on Multi-angle Convolutional Transformation Neural Network(MCTNN).The method utilizes Multi-angel Convolutional Transformation(MACT)module,which maps context features to different channels of convolutional neural network,by applying attention mechanism to each channel to learn local information of context from multiple semantic spaces and angles.Finally,the method obtains the importance of each channel,and learns richer and more informative feature of text for fine-grained sentiment analysis.Experimental results show that the accuracy of aspect level sentiment analysis method based on multi angle convolution network is better than most mainstream aspect level sentiment analysis methods.(2)Aspect-based sentiment analysis aims to predict the emotional polarity of a given aspect word in a sentence.It often uses attention mechanism to extract features related to aspect words in a sentence,but attention mechanism will introduce noise,which affects the accuracy of sentiment analysis.To address this issue,we propose an aspect-based sentiment analysis method based on Semantic Perception and Refinement Network(SPRN).Firstly,Multi-head Self-attention(MHSA)mechanism is used to extract context semantic features,and contextbased aspect embedding method is combined to obtain aspect word semantic features.Furthermore,the gate mechanism is used to construct a dual refinement gate.The interaction between context and aspect word semantic features is realized through two processes,including preliminary refinement and reinforced refinement.The redundant features unrelated to aspect words are removed,and the sentiment semantic features related to aspect words in sentences are obtained.The experimental results show that the proposed semantic perception refinement network is effective for semantic perception and feature refinement,and achieves high accuracy in various benchmark datasets.(3)For the practical needs of aspect level sentiment analysis method,this paper also designs an aspect level sentiment analysis method based on semantic perception network.Firstly,the multi-head self-attention mechanism is used to encode sentences with long dependence,and then the local information is perceived by bit one-dimensional convolution to improve the efficiency of aspect level emotional feature acquisition.Based on the above methods,this paper further designs the Chinese public opinion analysis system and deploys it to the web page.In the specific implementation of the system,the crawler technology is used to collect Chinese public opinion data,and the back translation method is used to enhance the sample number of relatively small data categories,so that the sample number of each category is in balance,which is conducive to network training.Finally,combined with the flash backend framework,the trained Chinese public opinion analysis system is deployed to the web page to provide a visual interface for users.