Little research has addressed how systems can use the learning process of self-explanation to provide scaffolding or feedback. Here, we propose a model automatically generating sample self-explanations with knowledge components required to solve a math quiz. The proposed model contains three steps: vectorization, clustering, and extraction. In an experiment using 1434 self-explanation answers from 25 quizzes, we found 72% of the quizzes generated sample answers with all necessary knowledge components. The similarity between human-created and machine-generated sentences was 0.719, with a significant correlation of R = 0.48 for the best performing generation model by BERTScore. These results suggest that our model can generate sample answers with the necessary key knowledge components and be further improved by using the BERTScore.
- Automatic summarization