作者：刘清堂，肖俊吉，蒋新宇，蒋如意，马鑫倩

出版刊物：The 15th International Conference on Educational and Information Technology

出版时间：2026年

内容摘要：

Traditional classroom dialogue analysis relies on manual coding, which does not adapt to the characteristics of local n-gram features, strong context dependence and speaker ID in dialogue classification, which affects the discrimination ability and classification performance of the model. This study proposes a BiLSTM-based hybrid text classification model according to the characteristics of classroom dialogue. The model introduces CNN in the encoder to capture local n-gram features, and uses BiLSTM to get the global context. Sentence-level information is aggregated via multi-strategy pooling, and a MLP classification head is employed to enhance the expressive capacity of nonlinear. In addition, speaker is integrated into the input as a prefix to make use of the speaker information in the classroom dialogue, and verifies it under the speaker on/off experiment. The experimental results show that the Macro-F1 model improves by 2.62% on IRE (S off); On IRE (S on), Macro-F1 reached 90.29%. Ablation experiments verified the synergistic improvement effect of each module. On the three public datasets, the model MacroF1 improved by 1.53%, 4.58% and 3.86%. The research shows that the proposed method provides an effective scheme for the automatic analysis of classroom dialogue.

传统课堂对话分析依赖人工编码，难以适应对话分类中局部n-gram特征、强上下文依赖和说话人ID的特点，影响了模型的判别能力和分类性能。本研究根据课堂对话的特点，提出了一种基于BiLSTM的混合文本分类模型。该模型在编码器中引入CNN以捕获局部n-gram特征，并使用BiLSTM获取全局上下文。通过多策略池化聚合句子级信息，并采用MLP分类头增强非线性表达能力。此外，将说话人作为前缀整合到输入中，以利用课堂对话中的说话人信息，并在说话人开启/关闭实验中加以验证。实验结果表明，该模型的Macro-F1在IRE（说话人关闭）上提升了2.62%；在IRE（说话人开启）上，Macro-F1达到90.29%。消融实验验证了各模块的协同提升效果。在三个公开数据集上，该模型的Macro-F1分别提升了1.53%、4.58%和3.86%。研究表明，所提方法为课堂对话的自动分析提供了一种有效方案。