For the problem of insufficient feature interaction between intent classification and slot filling in spoken language understanding tasks, this paper proposes a method that uses ChatGPT to generate more diverse samples, combined with a contrastive learning approach, to improve the model architecture and strengthen the interaction between intent and slot features. Specifically, prompts are designed for ChatGPT to generate diverse synthetic data with the same slots but different intents, and with the same intents but different slots. A contrastive learning module is further designed, in which positive and negative intent-slot sample pairs are constructed via the ChatGPT-based mixed data augmentation method. The feature space distribution is optimized using a weighted InfoNCE loss, enhancing the aggregation of similar features and the separation of dissimilar ones. Meanwhile, a multi-task joint training framework is employed to simultaneously optimize the cross-entropy loss for intent classification and the contrastive loss, enabling deeper semantic interaction between intents and slots, thereby improving the overall model performance. Experimental results on the ATIS and SNIPS datasets demonstrate that the proposed method significantly outperforms traditional baseline models in both intent detection accuracy and slot filling F1 score. In addition, ablation studies confirm the effectiveness of the contrastive learning and mixed data augmentation components. Overall, this work introduces a contrastive learning mechanism to effectively address the insufficient label-feature interaction in spoken language understanding tasks, offering a novel approach for optimizing multi-task dialogue systems.
Yang et al. (Fri,) studied this question.