FewJoint: few-shot learning for joint dialogue understanding

Area of Science:

• Machine Learning
• Natural Language Processing
• Artificial Intelligence

Background:

• Few-shot learning (FSL) is crucial for advancing machine learning.
• Dialogue understanding, encompassing intent detection and slot filling, benefits from joint learning.
• Joint learning in few-shot dialogue understanding faces challenges due to data sparsity and limited research.

Purpose of the Study:

• Introduce FewJoint, the first benchmark for few-shot joint dialogue understanding.
• Provide a new corpus and code platform to facilitate FSL research in this area.
• Develop a novel method to address challenges in few-shot joint learning for dialogue understanding.

Main Methods:

• Proposed FewJoint benchmark with a corpus of 59 industrial API dialogue domains.
• Developed a novel trust gating mechanism to guide slot filling with intent information.
• Implemented a Reptile-based meta-learning strategy for improved generalization in few-shot domains.

Main Results:

• The proposed method significantly improves performance on two datasets.
• Achieved new state-of-the-art results in few-shot joint dialogue understanding.
• Demonstrated the effectiveness of the trust gating mechanism in ensuring high-quality information sharing.

Conclusions:

• FewJoint benchmark and proposed method advance few-shot joint dialogue understanding research.
• The novel approach effectively tackles data sparsity and noisy sharing in few-shot settings.
• The method shows strong generalization capabilities for unseen few-shot domains.