基于ERNIE的MOOC评论情感分析的研究

Research on Sentiment Analysis of MOOC Reviews Based on ERNIE

摘要: MOOC平台提供了丰富的优质课程，促进了在线学习。 然而，课程完成率较低，需根据学生评论 进行改进。 由于学生对教师的尊重，常通过反语、 讽刺等隐性情感表达不满，传统情感分析方法 难以识别这些情感。 因此，开展针对学生评论的情感分析研究至关重要。 本研究提出了一种新的 情感分析方法，旨在提升复杂情感和隐含情感的识别能力。 首先基于ERNIE模型进行微调，生成 动态词向量，并将其与FastText和NMF生成的静态词向量融合，形成综合特征。 同时，结合、 部情感知识库SenticNet来增强特征表达。 通过语法和语义分析的双图卷积网络进一步优化特征 表示，并引入双仿射变换促进信息交互，最后通过softmax层进行情感极性分类。 实验结果表明， 该方法在自建数据集上的准确率达到89.5%，F1值为88.8%，验证了动态与静态特征融合及情感 知识引入的有效性。

Abstract: MOOC platforms provide a wealth of high-quality courses, promoting online learning. However, the completion rate of courses is relatively low, necessitating improvements based on student reviews. Due to students’ respect for teachers, they often express dissatisfaction through implicit expressions such as irony and sarcasm, which are dif- ficult to identify using traditional sentiment analysis methods. Therefore, conducting sentiment analysis research on student reviews is crucial. This study proposes a new sentiment analysis method aimed at enhancing the recognition ability of complex and implicit emotions. Firstly, fine-tuning is performed based on the ERNIE model to generate dynamic word vectors, which are then fused with static word vectors gen- erated by FastText and NMF to form comprehensive features. At the same time, external sentiment knowledge base SenticNet is incorporated to enhance feature rep- resentation. The feature representation is further optimized through a dual graph convolutional network that combines grammatical and semantic analysis, and a du- al affine transformation is introduced to facilitate information interaction. Finally, sentiment polarity classification is performed through a softmax layer. Experimental results show that the accuracy of this method on a self-built dataset reaches 89.5%, with an F1 score of 88.8%, validating the effectiveness of the fusion of dynamic and static features and the introduction of sentiment knowledge.