Core Concepts
Robust and Adaptive Prototypical Learning Framework for New Intent Discovery.
Abstract
The article introduces a Robust and Adaptive Prototypical Learning (RAP) framework for New Intent Discovery (NID). It addresses the limitations of existing methods by focusing on within-cluster compactness and between-cluster separation. The RAP framework consists of Robust Prototypical Attracting Learning (RPAL) and Adaptive Prototypical Dispersing Learning (APDL) to optimize intent representations. Experimental results show significant improvements over current state-of-the-art methods, even outperforming large language models.
Abstract:
New Intent Discovery aims to recognize known and infer new intent categories.
Existing methods lack cluster-friendly representations.
RAP proposes RPAL and APDL to enhance within-cluster compactness and between-cluster separation.
Experimental results demonstrate substantial improvements over current methods.
Introduction:
Conventional intent detection in dialogue systems focuses on pre-defined intents.
New Intent Discovery is essential for handling new intents outside existing categories.
Early works adopt unsupervised clustering, while recent studies explore semi-supervised settings.
Approach:
Problem Definition: NID follows an open-world setting to recognize all intents with limited labeled data.
Intent Representation Learning: Pre-trained BERT model used for feature extraction and fine-tuning on labeled data.
Categorical Prototypes Generation: Class prototypes computed as representative embeddings within each class.
Robust Prototypical Attracting: RPAL minimizes instance-to-prototype distances for within-cluster compactness.
Adaptive Prototypical Dispersing: APDL maximizes prototype-to-prototype distances for between-cluster dispersion.
Dynamic Prototypes Update: Exponential moving average algorithm used to update class prototypes continuously.
Multitask Learning: Joint optimization of RPAL, APDL, and cross-entropy loss for NID task.
Data Extraction:
"Experimental results evaluated on three challenging benchmarks...average +5.5% improvement."
"Extensive experiments on three benchmark datasets show that our model establishes state-of-the-art performance..."
Stats
Experimental results evaluated on three challenging benchmarks...average +5.5% improvement.
Quotes
"The proposed RAP significantly outperforms the previous unsupervised and semi-supervised baselines..."
"Our method consistently outperforms ChatGPT3.5 across all datasets..."