New Intent Discovery for Educational Dialogue Texts via Semantic-Aware Data Augmentation

Understanding educational intents underlying teacher-student dialogues is of great importance for discovering effective instructional patterns and optimizing teaching strategies. Traditionally, analyzing these intents relies on static, predefined coding schemes. However, these schemes struggle to adapt to diverse educational scenarios; furthermore, manually refining them to capture new educational intents entails prohibitive expert costs. Despite its importance, the use of data-driven techniques to automatically discover new educational intents from large-scale unlabeled corpora remains under-explored. Therefore, a New Educational Intent Discovery (NEID) method is needed. In this paper, we propose a novel transfer learning framework that extends a two-stage New Intent Discovery (NID) method to NEID with enhanced semantic-aware discriminative representation. First, a multi-task pre-training with context and speaker information injections is adopted for representation learning. Then, we design a reverse strategy under a knowledge distillation paradigm for guiding the student model to autonomously identify noisy words. Finally, the student model performs Semantic-Aware Data Augmentation (SADA) to apply adaptive noise perturbation to noisy words for contrastive clustering, which transfers the prior knowledge from labeled to unlabeled data. Experimental results on two educational dialogue datasets, MathDial and TalkMoves, demonstrate the effectiveness of the proposed method. Notably, under the experimental settings most aligned with educational scenarios, our method improves NMI performance by 19.83% on TalkMoves compared with the original NID framework. Our code is available at: https://github.com/ZhuXY247/NEID_SADA.git.