TiNID: A Transfer and Interpretable LLM-Enhanced Framework for New Intent Discovery

Shun Zhang, Chaoran Yan, Jian Yang, Wei Zhang, Changyu Ren, Tongliang Li, Jiaqi Bai, Zhoujun Li

Abstract: New Intent Discovery (NID) is an essential task in open-world learning, tasked with the identification and classification of both known and novel intents using a combination of limited labeled data and large-scale unlabeled data. Traditional methods only transfer knowledge implicitly through shared representation spaces ignoring explicit knowledge transfer patterns and can not assign novel intents with appropriate category labels without common knowledge. Inspired by the excellent generalization ability of large language models (LLMs), we introduce TiNID, a Transfer and interpretable LLM-enhanced framework comprised of a lightweight NID model (explicit knowledge transfer module) designed for explicit knowledge transfer to enhance the recognition of new intent categories and an LLM (reliable knowledge interpretation module) to generate descriptive category names. Specifically, the explicit knowledge transfer module is tailored to distill and propagate class relation knowledge from known to novel intents, facilitating the transfer of clear-cut knowledge about predictive distributions. The reliable knowledge interpretation module focuses on selecting characteristic samples from clusters related to new categories. It then employs the in-context learning capabilities of LLMs to assign fitting and precise names to these clusters. Our experimental analyses across multiple benchmarks reveal that TiNID significantly outperforms contemporary state-of-the-art models, achieving an average improvement of 2.26% on novel categories. Additionally, it demonstrates a remarkable ability to generate accurate category names for newly recognized clusters. Finally, we present a comprehensive evaluation of different LLMs in NID, outlining their strengths and weaknesses (The code is available at https://github.com/Zkdc/TiNID) .