Robust open intent classification in many-shot and few-shot scenarios

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Open Intent Classification (OIC) is crucial for dialogue systems to handle both known and novel user intents.
• Current OIC methods using BERT fine-tuning face challenges with underfitting/overfitting in many-shot and few-shot scenarios due to fixed feature dimensions.
• Effective modeling of data distribution and precise decision boundary definition remain significant challenges in OIC.

Purpose of the Study:

• To propose a Robust Open Intent Classification (ROIC) method effective in both many-shot and few-shot settings.
• To enhance the modeling of data distribution and improve the precision of decision boundaries for OIC.
• To address the limitations of existing OIC approaches in handling diverse intent scenarios.

Main Methods:

• Implemented distance-aware contrastive learning by upgrading pairwise distances to a complete distance matrix to capture global relationships.
• Employed hard negative sampling mining to enhance inter-class separability and intra-class aggregation.
• Introduced feature space transformation to map representations into an appropriate space for robust decision boundary construction.

Main Results:

• The proposed ROIC method demonstrated significant effectiveness in both many-shot and few-shot scenarios.
• ROIC successfully improved the modeling of data distribution and the definition of decision boundaries.
• Experimental results on standard benchmarks validated the robustness and efficacy of the ROIC approach.

Conclusions:

• ROIC offers a robust solution for Open Intent Classification, outperforming existing methods in both many-shot and few-shot learning.
• The combination of distance-aware contrastive learning and feature space transformation effectively addresses OIC challenges.
• This research contributes to more accurate and reliable dialogue systems capable of handling a wider range of user intents.