IOPC: Aligning Semantic and Cluster Centers for Few-shot Short Text Clustering

ACL ARR 2025 February Submission3869 Authors

Abstract: In short text clustering, determining appropriate cluster centers is crucial. However, due to the limitations of short text representation quality, traditional methods often struggle to align cluster center with the core concept (_semantic center_) of each category, degrading clustering performance. To address this issue, we propose __IOPC__, a novel few-shot learning framework that achieves the alignment through two key modules. First, to capture effective semantics, we introduce an Interaction-enhanced Optimal Transport (__IEOT__) that leverages semantic interactions between samples to generate confident pseudo-labels. Based on these high-quality pseudo-labels, pseudo-semantic centers (_prototypes_) can be obtained. Furthermore, we propose Prototype-based Contrastive Learning (__PBCL__) to optimize text representations towards their corresponding prototypes. As training progresses, the continuous updating of pseudo-labels and prototypes gradually reduces the gap between cluster centers and semantic centers, improving clustering performance and stability. Extensive experiments on eight benchmark datasets show that IOPC outperforms state-of-the-art methods, achieving up to 7.34\% improvement in accuracy on challenging Biomedical datasets and excelling in clustering stability and efficiency. The code is available at: https://anonymous.4open.science/r/IOPC-origin-2522.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: few-shot learning,representation learning

Languages Studied: English

Submission Number: 3869