Research On Multi-domain Text Sentiment Analysis Algorithm

In practical applications,sentiment analysis is usually related to domains.Due to the semantic differences between domains,the sentiment polarity expressed by the same word in different domains may be contradictory.Therefore,a well-trained model in a specific field may not work well in another field.In addition,there may be common information between domains,so it is redundant to train a separate sentiment classifier for each domain.Therefore,the work of this paper takes this as a starting point,and aims to study how to make full use of the limited training data in multiple domains to improve the classification performance in all domains.At the same time,this paper will focus on the cross-domain text sentiment analysis,and explore how to transfer sentiment knowledge from multiple domains with rich labeled training data to unlabeled domains.Therefore,in view of the above problems,the work of this paper is summarized as follows:Firstly,based on multi-task learning,this paper proposes a novel multi-domain text sentiment analysis model.This paper cleverly uses the coarse-grained and fine-grained label information of sentiment text,and conducts joint training through two closely related tasks.In the multi-domain text sentiment analysis benchmark dataset,compared with the single-task model and the simple private-public model,the multi-task learning model has achieved leading results.Secondly,this paper turns to the cross-domain text sentiment analysis scenario,aiming to transfer sentiment knowledge to the unlabeled domain.Therefore,this paper introduces adversarial training,and abandons the commonly used domain adaptation methods.Instead,we propose a method called label knowledge transfer learning to transfer label knowledge from source domain to target domain.At the same time,in order to improve the confidence of the label knowledge,we add an adapter to the model,and use internal and external voting mechanism to provide to provide labels with higher confidence for unlabeled target domain samples.In Amazon reviews and fdu-mtl datasets,compared with all the baseline models,it has a significant improvement.Finally,through comparative experiments,it is verified that our proposed method of label knowledge transfer is effective.Finally,this paper introduces the pre-training model BERT as the base model of multi-domain text sentiment analysis,and also focuses on the scene of cross-domain text sentiment analysis.Through a two-stage training model,the pre-training model BERT is fine-tuned in the first stage,and sentiment knowledge is transferred to the target domain in the second stage.This paper designs two combination methods of average value and target domain oriented to transfer the sentiment knowledge to the target domain by using the fine-tuned source domain model in the first stage.The final experimental results also show the excellent performance of the pre-training model,and through the comparative experiment,we find that only through the fine-tuning of the pre-training model,can we obtain significant performance improvement.