Research On Text Sentiment Prediction Based On Active Learning And Transfer Learning

With the wide usage of e-commerce platform,not only can users enjoy the convenience,but also they can comment on the product through the forum.Through these comments,ordinary users can understand the product reputation,so as to make a reasonable production marketing choice.Producers can grasp the market trends quickly,so as to make the correct marketing decision.Therefore,product opinion mining and sentiment analysis is an effective means to solve this problem.Traditional supervised learning method is applied to the static single domain data,requires a large number of labeled data.The transfer learning method can use other areas or other period labeled data to learn classification model to solve the problem of insufficient training samples.In order to improve the prediction performance,because of the difference between the data in different areas or different periods,the active learning is used to optimize the classification model.The main research contents are as follows:(1)The problem analysis of text sentiment predictionBased on the corpus,this paper analyzed the problems existed in the sentiment analysis research from three aspects:traditional text representation limitations,review text language expression diversity and different review concerns of different period.And the corresponding solution method is proposed(2)Cross-domain sentiment prediction based on active learning and transfer learningGiven language expression diversity question of static cross-domain review text,a method of cross-domain sentiment prediction based on active learning and transfer learning is proposed.Firstly,this paper trained classification model on the labeled source domain,selected the texts with high confidence value as initial seed texts of the classification model.To speed up the optimization of the classification model,we added the high confidence text and low confidence text to the training data set in each iteration.Then the feature set was extracted to transform feature space through sentimental dictionary,evaluation collocation rules and assist feature words.Finally,the optimization classification model was used to classify on test data set.Compared with Active-Dynamic,the average accuracy of Active-Semi-Dynamic increases by 2.75 percentage points.The experimental results show that attaching high confidence samples can enrich the training samples and feature information,contribute the classification model train.Compared with Active-BOW,the average accuracy of Active-Semi-Dynamic increases by 2.79 percentage points.The experimental results show that sentiment words are extracted by using sentiment dictionary and syntactic dependency analysis,can describe the text sentiment more accurately,and improve the effect of cross domain text sentiment prediction.(3)Time series review sentiment prediction based on active learning and transfer learningGiven different concern question of dynamic time series review,a method of time series review sentiment prediction based on active learning and transfer learning is proposed.This paper used transfer learning idea to get initial marked samples of unlabeled data of present period through the labeled data of prior period.In the process of active learning,we used smote algorithm to balance training data set.Then,the optimization classification model was used to predict the sentiment orientation of present period car review.Compared with UN SMOTE,the SMOTE algorithm average accuracy increases by 4.32 percentage points.The experimental results show that adding new samples to minority class can balance training corpus,improve the car review sentiment prediction effect.At the same time,the method realizes the sentiment prediction of mixed reviews.