Researchers boost short text clustering with semantic-aware optimal transport

A major challenge in short text clustering is generating reliable pseudo-labels, a task often tackled via Optimal Transport (OT). However, existing OT methods focus solely on aligning global cluster structures and ignore semantic similarities between individual samples. This can assign different pseudo-labels to semantically similar texts, leading to poor clustering performance. A new paper from Zhihao Yao and colleagues proposes a framework called “Harmonizing Global Structure and Local Consistency” that directly addresses this gap.

The approach first designs an instance-level attention mechanism to model pairwise semantic relationships between samples. These semantic affinities are then embedded into the OT formulation, making the transport process aware of local neighborhood contexts. The resulting pseudo-labels naturally balance sample-to-sample consistency with sample-to-cluster global structure. Experiments across multiple short text clustering datasets show the method consistently outperforms prior state-of-the-art OT-based and non-OT clustering techniques. The code has been open-sourced, enabling rapid adoption and further research.