比较任务修正少样本细粒度情感分类方法  

Comparison task correction approach to few-shot learning for fine-grained sentiment classification

作　　者：金旭 崔艳荣[1] 陈杰[1] 陈佳力 JIN Xu;CUI Yan-rong;CHEN Jie;CHEN Jia-li(School of Computer Science,Yangtze University,Jingzhou 434023,China)

机构地区：[1]长江大学计算机科学学院,湖北荆州434023

出　　处：《计算机工程与设计》2024年第11期3420-3426,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62077018)。

摘　　要：针对主流少样本学习方法在细粒度情感分类任务中难以区分情感程度相似样本的问题,提出一种基于比较任务的少样本细粒度文本情感分类的修正方法。对于比较任务的训练阶段,构建训练集样本的正负样本并设计一套比较任务模板。在比较任务的预测阶段设计一种将测试集样本与训练集样本情感程度比较结果进行投票分类的方法。利用比较任务的结果对单句分类结果进行修正,取得更稳定的结果。在细粒度情感分类数据集SST5及Amazon Product数据集上进行实验,结果表明,修正方法相较于主流方法获得更优的性能与稳定性。To address the problem of difficulty in distinguishing samples with similar sentiment degrees in fine-grained sentiment classification tasks using mainstream few-shot learning methods,a modified few-shot fine-grained text sentiment classification method based on comparison tasks was proposed.During the training phase of the comparison task,positive and negative samples for training set samples were constructed,and a set of comparison task templates was designed.During the prediction phase of the comparison task,a voting classification method was designed to compare the sentiment degrees of test set samples with those of training set samples.The results of the single sentence classification were corrected using the comparison task results to obtain more stable results.Experiments were conducted on the fine-grained sentiment classification dataset(SST5)and Amazon Product dataset.The results show that the modified method achieves better performance and stability compared to the mainstream methods.

关 键 词：细粒度 情感分类 少样本 比较任务 修正方法 自然语言处理 文本分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]