摘要
针对中文教学评论的情感分析,提出了基于预训练语言模型MacBERT的方面级中文教学评论无监督情感分析框架。首先,针对每个教学方面和情感极性,通过预训练语言模型构建语义一致的类别词表;然后,利用构建的词表和词性标签,基于重合率矩阵对训练语料库中的部分评论句进行标注;最后,利用已标注的评论句构建神经网络,通过MacBERT提取测试数据中方面类别和情感类别的联合隐藏特征,并利用抗噪声损失函数完成方面抽取和情感分类。实验结果表明,该模型在方面抽取和情感分类任务上的宏观F1值分别为78.10%和89.20%,说明该模型能够从学生反馈中准确完成方面类别抽取并确定每个评论句的情感极性。
An aspect-level unsupervised sentiment analysis framework for Chinese teaching comments based on the pre-trained language model MacBERT is proposed for sentiment analysis.Firstly,for each teaching aspect and emotional polarity,a semantically consistent category vocabulary table is constructed through pre-trained language models.Then,part of the comment sentences in the training corpus is annotated using the constructed vocabulary and part-of-speech tags and based on an overlap rate matrix.Finally,a neural network is constructed using annotated comment sentences to extract joint hidden features of aspect and emotion from the test data through MacBERT,and accurate aspect extraction and emotion classification are achieved using a noise-resistant loss function.The experimental results show that the macro F1 values of the model in aspect extraction and sentiment classification tasks are 78.10%and 89.20%respectively,indicating that the model can accurately extract aspect categories from student teaching feedback and determine the emotional polarity of each comment sentence.
作者
顾明
王晓勇
胡胜利
GU Ming;WANG Xiaoyong;HU Shengli(School of Economics and Management,Huainan Union University,Huainan Anhui 232038,China;School of Information Engineering,Huainan Union University,Huainan Anhui 232038,China;School of Computer Science and Engineering,Anhui University of Science Technology,Huainan Anhui 232001,China)
出处
《重庆科技大学学报（自然科学版）》
CAS
2024年第5期72-78,共7页
Journal of Chongqing University of Science and Technology(Natural Sciences Edition)
基金
安徽省重点科研项目“基于数据挖掘技术下高职院校状态数据采集平台中的数据二次应用技术研究”(KJ2021A1306)
淮南联合大学校级科研项目“基于改进深度学习模型的在线学习评论情感分析”(JYB2208)。
