Research On Chinese Sentiment Analysis Based On Deep Learning

According to the report of China Internet Network Information Center in April2020,the number of people online in my country has exceeded 900 million,and the Internet penetration rate has reached more than 60%.On the one hand,it shows that people’s lives have become more informatized and intelligent,But on the other hand,it also reminds that our society is facing an unprecedented challenge.How to deal with the explosion of information in the Internet age is an urgent problem that needs to be solved.Therefore,how to use Natural Language Processing(NLP)technology to analyze massive amounts of web comment text data has become a research hotspot for researchers.With the rapid development of hardware technology,deep neural network technology,also known as deep learning,has become possible in recent years and has achieved impressive results in the field of natural language processing.Domestic and foreign researchers have also applied this technology to Chinese text sentiment analysis tasks.However,previous studies have the following shortcomings.First,most of the existing deep learning methods used in Chinese sentiment analysis tasks lack prior knowledge of Chinese sentiment.Second,most of the models trained in previous studies can only complete a single text task.It ignores that natural language tasks are related,and the related information between tasks can be used.Third,in the past,the loss function weight σ of each task in the text-based multi-task joint loss function is mostly static and requires manual adjustment.This article mainly focuses on the shortcomings of the above-mentioned deep learning in Chinese text sentiment analysis tasks,and has carried out work.The main research and innovation contents are as follows:(1)This paper proposes a new text feature representation method B-ECM,which combines two feature vectors,one is the semantic feature vector encoded by the BERT pre-training model,the other is the text sentiment feature vector obtained by using the Chinese sentiment dictionary and sentiment rules.This paper conducts a comparison experiment with the benchmark method BERT coded text representation on three standard Chinese data sets.The experiment shows that B-ECM improves the accuracy of the benchmark model on these three Chinese data sets by an average of 0.6%.On the basis of this experiment,this paper conducted a comparison experiment of the benchmark method of word vector + ECM,and the accuracy of the emotion classification model that only used the word vector representation method was improved by 0.68% on average.The above two parts of the experimental results show the effectiveness of the B-ECM representation method and the ECM module alone.(2)This paper uses the idea of multi-task learning to solve the task of Chinese sentiment classification.Based on this idea,a multi-task model,MT-GSU,is proposed.Through experiments on a standard Chinese data set,the accuracy of the benchmark model is improved by 0.82%.The above experimental results show the feasibility of multi-task thinking to solve Chinese sentiment classification tasks and the effectiveness of the MT-GSU multi-task text sentiment classification model.On this basis,this paper proposes to apply the uncertainty and homovariance loss function combination strategy that has achieved good results in computer vision tasks to the multi-task model MTGSU,which improves the accuracy by 1.21% compared with the benchmark model.(3)This paper verifies the effectiveness of the combination of the three main innovations proposed in this paper through ablation experiments,and obtains the BECM as the feature representation method,and the MT-GSU with the uncertainty homovariance loss function combination strategy as the classifier The sentiment classification model has an accuracy improvement of 1.27% compared with the benchmark model.This article also conducted comparative experiments on standard data sets with the more excellent text sentiment classification models in recent years,and the results also showed that compared with these models,there are different degrees of improvement.Based on the work of the past and this article,it can be seen that deep learning is feasible and effective when applied to Chinese text sentiment classification.Training a pre-training model that is more suitable for Chinese text,exploring the mutual promotion of various natural language processing tasks and establishing a more standardized Chinese corpus are more important research directions in the future.