Intent Clustering with Shared Pseudo-Labels

TL;DR

This work addresses unsupervised short-text intent clustering without relying on labeled data or known cluster counts $K$. It introduces a training-free, label-free method that constructs a shared pseudo-label set via open-source LLMs and refines text representations through iterative multi-label classification, guided by the central assumption that label-augmented embeddings of texts from the same cluster become more similar. Empirically, the approach matches or surpasses state-of-the-art baselines on four benchmarks (Bank77, CLINC150, Mtop, Massive) and remains robust across multiple open-source LLMs (e.g., LLaMA, Qwen, Gemma), demonstrating practical utility in low-resource settings. The method also yields human-readable pseudo-labels, aiding interpretability and post-processing, with ablations confirming the critical role of the iterative labeling step. Future work could adapt the framework to longer texts and explore extensions for broader information-access tasks.

Abstract

In this paper, we propose an intuitive, training-free and label-free method for intent clustering that makes minimal assumptions using lightweight and open-source LLMs. Many current approaches rely on commercial LLMs, which are costly, and offer limited transparency. Additionally, their methods often explicitly depend on knowing the number of clusters in advance, which is often not the case in realistic settings. To address these challenges, instead of asking the LLM to match similar text directly, we first ask it to generate pseudo-labels for each text, and then perform multi-label classification in this pseudo-label set for each text. This approach is based on the hypothesis that texts belonging to the same cluster will share more labels, and will therefore be closer when encoded into embeddings. These pseudo-labels are more human-readable than direct similarity matches. Our evaluation on four benchmark sets shows that our approach achieves results comparable to and better than recent baselines, while remaining simple and computationally efficient. Our findings indicate that our method can be applied in low-resource scenarios and is stable across multiple models and datasets.

Figures (1)

• Figure 1: The method pipeline. Construction stage: Pseudo-labels for each text are derived using an LLM and its closest texts. Iterative stage: Multi-label classification is performed for each text until its pseudo-labels converge.