New Intent Discovery with Attracting and Dispersing Prototype

• DOI: 10.48550/arxiv.2403.16913
• Publication Date: 2024-03-25
• Authors: Shun Zhang; Jian Yang; Jiaqi Bai; Chaoran Yan; Tongliang Li; Yan Zhao; Zhoujun Li
• Keywords: FOS: Computer and information sciences; Computer Science - Computation and Language; Computation and Language (cs.CL)

New Intent Discovery (NID) aims to recognize known and infer new intent categories with the help of limited labeled large-scale unlabeled data. The task is addressed as a feature-clustering problem recent studies augment instance representation. However, existing methods fail capture cluster-friendly representations, since they show less capability effectively control coordinate within-cluster between-cluster distances. Tailored NID problem, we propose Robust Adaptive Prototypical learning (RAP) framework for globally distinct decision boundaries both categories. Specifically, robust prototypical attracting (RPAL) method designed compel instances gravitate toward their corresponding prototype, achieving greater compactness. To attain larger separation, another adaptive dispersing (APDL) devised maximize distance from prototype-to-prototype perspective. Experimental results evaluated on three challenging benchmarks (CLINC, BANKING, StackOverflow) our better representation demonstrate that RAP brings in substantial improvements over current state-of-the-art (even large language model) by margin (average +5.5% improvement).