The study addresses the challenge of few-shot intent classification and slot filling due to the lack of finely labeled data. To overcome this, a new method is proposed which decouples the transfer of general semantic representation and domain-specific knowledge. The method enhances the performance of these tasks by capturing intent-slot relations and slot-slot relations. The results from experiments on Snips and FewJoint datasets show a significant improvement in the joint accuracy metric, demonstrating the effectiveness of the proposed method.
Publication date: 22 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.13495
