Research On The Method Of Mining Web Comments Based On Transfer Learning

Online comments are an important source of text data,and the process of mining viewpoints on them is the process of extracting valuable information.However,online comments have rich content and diverse formats,making it difficult to annotate data,making it even more difficult to establish separate datasets for the fields,majors,and categories they cover.In order to reduce the labor cost of building data sets separately and improve the comprehensive performance of the model in the task of opinion mining,this paper conducts research on the method of opinion mining for online comments based on transfer learning.In response to the current research status of online comment viewpoint mining,this article proposes the problem of few samples and domain adaptation:1.The small sample problem refers to the problem where there is less labeled data in the target domain.Currently,there are only a few fields such as online shopping that have certain labeled data,and most fields of online comment viewpoint mining have not established a complete dataset.2.The problem of domain adaptation refers to the problem that the model has poor ability of opinion mining in the composite domain.The internal information entropy of the composite domain is high,including multiple sub domains with low correlation.It is difficult for the model to perform well in all sub domain opinion mining at the same time.In order to solve the above problems,the main work of this article is as follows:1.To solve the problem of few samples,this paper designs a model MWCA BiGRU based on multi granularity word vectors and parallel attention mechanism,which has achieved better results in viewpoint mining experiments in fixed fields compared to advanced models in recent years;In the data attenuation experiment,the model maintained a certain problem-solving ability while reducing the amount of data,demonstrating the effectiveness of multi granularity word vectors and parallel attention mechanisms in small sample problems.In order to better solve the problem of small sample size through cross domain methods,this article adds cross domain data and unlabeled target domain data to the input in the case of small sample size,based on MWCA-BiGRU.This better captures the relationship between comment text and target evaluation objects,and achieves good results in the small sample experiment.At the same time,this chapter demonstrates the impact of model input data matching on cross domain knowledge transfer.2.In order to solve the problem of domain adaptation,the paper designs and implements a prompt tuning model based on pre trained models.This model proposes a flexible prompt method that can learn prompt templates,enabling the model to generate corresponding templates based on target domain data,paying more attention to the characteristics of the target domain,and demonstrating strong viewpoint mining ability.The semi supervised learning method based on confrontation training indicates that the optimization model has a strong Domain Adaptation ability on the data set of composite domain.Based on transfer learning technology,this paper proposes innovative algorithms and solutions to the problem of few samples and domain adaptation of network comment opinion mining,and has achieved better results in both open data sets and self built data sets.