基于预训练语言模型的用户评论情感分析

打开文本图片集

中图分类号：TP391 文献标识码：A

文章编号：2096-4706（2025）18-0059-06

Abstract： With the development of the Internet economy，text sentiment analysis has significant value in fields such as businessinteligence.Inviewofthelimitationsof tradiionalmethods （suchasTF-IDFandWord2Vec）inprocessingcomplex semantics，this paper studies the advantagesof pre-trained models （such as the ERNIE model and the BERTmodel） in Chinese sentiment analysis.Through comparative experiments onuser takeawayreviewdatasets，itis found that theERNIEmodel significantlyimprovestheclassificatioaccuracyandrobustnesscomparedwithtraditionalmethods intheMLPclassifierof the samestructure.Inorder tofurther breakthroughthelimitationsofasinglemodel，thispaperproposesadynamicgatingdualmodel fusion methodofusing the domainadaptabilityof BERTand theuniversal semantic abilityofERNIE，and theoutput oftheBERTpoling layerisgeneratedbythesigmoidfunctiontogeneratedynamic weights，thenthecontributionof the two models isrealized inreal-timeadaptive adjustment.Experiments show thatthe fusion model surpases the single model in precision，recallandFale，eielylaingdmaspicityndgeeralatioability.Iisfdtisf this method in Chinese sentiment analysis tasks.

Keywords： sentiment analysis;BERT model; ERNIE model; dynamic gating

0 引言

随着互联网和大数据技术的飞速发展，情感分析任务在商业、社会、学术等多个领域的重要性日益凸显[1]，广泛应用于舆情监测、产品评价、智能客服、金融分析等领域。（剩余10977字）