Sentiment Classification Research Based On Semi-supervised Learning

With the rapid development of the Internet and e-commerce,a large number of sentimental texts have emerged on the Internet.Identifying the sentiments of these texts correctly is conducive to the realization of multiple intelligent text processing tasks,such as public opinion monitoring and product recommendation.Therefore,sentiment analysis has attracted the attention of researchers in the field of natural language processing gradually.The sentiment classification studied in this paper is a basic task of sentiment analysis,which aims to automatically analyze the text and judge the sentiment category it belongs to.However,the study of sentiment classification requires a large number of labeled samples,and obtaining such labeled samples is costly and time-consuming.In order to solve the problems above,this paper starts with the sentiment classification method based on semi-supervised learning and the study includes following three aspects:First,this paper proposes a semi-supervised sentiment classification method based on auxiliary task learning.The method automatically labels a large number of unlabeled samples,and then constructs two tasks,which are sentiment classification on manual annotation samples(main task)and sentiment classification on automatic annotation samples(auxiliary task).These two tasks share the auxiliary LSTM layer to assist the main task to complete sentiment classification.In addition,this paper also considers the loss information between the models,and constructs the loss function of the two tasks for joint learning.Experimental results show that this method has obvious advantages over other traditional semi-supervised sentiment classification methods.Second,this paper proposes a semi-supervised sentiment classification method based on bilingual adversarial learning.This method does not need to obtain the labels of unlabeled samples,but the sample information of unlabeled samples through the adversarial learning of the classifier and the discriminator.The function of the classifier is to make labeled samples and unlabeled samples in the same distribution,while the discriminator is used to distinguish whether the input sample is labeled or unlabeled.In addition,this paper also uses bilingual information,which is to improve the experimental performance of semi-supervised sentiment classification by joint learning of Chinese and English adversarial neural networks.Experimental results show that this method has demonstrated the significant improvement compared to other baselines.Finally,this paper proposes a sentiment classification method on question-answer text,which is a semi-supervised learning method based on network embedding.The method adds the sentiment information of labeled samples to the network embedding,and constructs a heterogeneous network learning word vector which is composed of labeled samples of word-word network,question text-word network,answer text-word network,sentiment label-word network and unlabeled samples of word-word network,question text-word network,answer text-word network.Then,the obtained network embedding is applied to the Hierarchical Matching Network Model.Experimental results show that this method has achieved better results in question-answer sentiment classification.