摘要：

思维过程的精准量化和思维品质的高效诊断是思维型教学智能化开展的难题。现有的思维分析方法普遍存在静态局限性, 割裂了事理逻辑和动态情境对思维的影响。人机对话作为思维外显和评估的重要载体, 为计算思维自动化评估提供了潜在可能。为提高人机对话环境下计算思维水平预测的准确性和可解释性, 构建基于BERT-异质图注意力网络(HAN)的计算思维自动化评估模型。采集人机对话过程中所获取的时序性文本作为学习者计算思维的外部表征, 通过BERT-HAN模型从人机对话文本数据中提取句子级语义特征表示, 将这些特征作为异质图的节点特征输入到HAN中。模型耦合了基于余弦相似度的句子语义特征和基于关系词列表的元路径嵌入, 进一步提取语句之间的语义关系。在此过程中, 通过注意力机制生成学习节点间的关系权重, 形成具有丰富语义信息的事理图谱。事理图谱的构建不仅考虑语句之间的直接关系, 还可以基于多头注意力机制灵活捕捉并处理异质图中不同关系类型的特征。最终, 根据这些特征, 利用Softmax分类器进行计算思维水平的识别和预测, 以实现自动化评估。实验结果表明, 该模型的预测准确率为0.869, 召回率为1, AUC值为0.998, 相较于BERT、TextCNN、LSTM-HAN等模型具有更好的性能。

关键词: 人机对话, 事理图谱, 计算思维, 文本分析, 异质图注意力网络

Abstract:

The precise quantification of cognitive processes and efficient diagnosis of cognitive qualities are outstanding challenges in the intelligent development of thinking-based pedagogy. Conventional methods for analyzing cognition typically exhibit static limitations, interfering with the influence of material logic and dynamic contexts on thinking. Human-machine dialogues, as vital mediums for external assessments of cognition, provide potential possibilities for automated assessment in computational thinking. To improve the accuracy and interpretability of predicting computational thinking levels within a human-machine dialogue environment, an automated computational thinking assessment model based on the Bidirectional Encoder Representations from Transformers-Heterogeneous graph Attention Network (BERT-HAN) is constructed. The temporal text acquired during human-machine dialogues is collected as an external representation of learners' computational thinking. Sentence-level semantic feature representations are extracted from the human-machine dialogue text data using the BERT-HAN model, which serves as node features in a heterogeneous graph fed into the HAN. The model integrates sentence semantic features, derived through cosine similarity, with meta-path embeddings generated from lists of relational words to further extract the semantic relationships between utterances. During this process, the relationship weights among learning nodes are computed through an attention mechanism, thereby creating an event graph enriched with semantic information. This construction of the event graph not only considers the direct relationship between utterances but also flexibly captures the features of different relationship types within the heterogeneous graph based on the multi-head attention mechanism. Leveraging these features, the level of computational thinking is identified and predicted using the Softmax classifier for automated assessment. Experimental results show that this model achieves a prediction accuracy of 0.869, a recall rate of 1, and an Area Under the receiver operating characteristic Curve (AUC) value of 0.998, surpassing the performance of BERT, Text Convolutional Neural Network (TextCNN), Long Short-Term Memory (LSTM)-HAN, and other models.

Key words: human-machine dialogue, event graph, computational thinking, text analysis, Heterogeneous graph Attention Network (HAN)