Instruction-Guided Distribution Maximization for General Few-Shot Intent Recognition.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Intent recognition is crucial for understanding user needs from natural language.
• Few-shot intent recognition faces challenges due to data scarcity.
• Current methods treat single- and multi-label intent recognition separately, limiting generalization.

Purpose of the Study:

• To unify single- and multi-label intent recognition into a single natural language generation task.
• To improve the generalization and robustness of few-shot intent recognition models.
• To address limitations of discriminative models requiring fine-tuning for new intents.

Main Methods:

• Re-framed intent recognition as a unified natural language generation task.
• Employed instruction pre-training for direct generalization to few-shot data.
• Introduced Instruction-Guided Distribution Maximization (IGDM) for enhanced robustness and generalization.
• Utilized a two-level optimization strategy: inner sample distribution maximization and outer model error minimization.

Main Results:

• The proposed Instruction-Guided Distribution Maximization (IGDM) method significantly enhances model robustness and generalization.
• IGDM achieves superior performance across 20 single-label and 7 multi-label intent recognition benchmarks.
• The unified generation approach allows direct generalization to few-shot target data without fine-tuning.

Conclusions:

• The unified generation approach for intent recognition overcomes limitations of separate classification tasks.
• Instruction-Guided Distribution Maximization (IGDM) provides a robust and generalizable solution for few-shot intent recognition.
• The proposed method demonstrates state-of-the-art performance across diverse intent recognition benchmarks.