Discriminative representation learning via attention-enhanced contrastive learning for short text clustering

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Mining

Background:

• Contrastive learning is popular for short text clustering.
• A key challenge is the false negative separation problem, where similar samples are incorrectly distinguished.

Purpose of the Study:

• To propose a novel method, Attention-Enhanced Contrastive Learning (AECL), for discriminative representation learning in short text clustering.
• To address the false negative separation issue inherent in contrastive learning.

Main Methods:

• AECL employs two modules: a contrastive learning module and a pseudo-label assisting module.
• Both modules use a sample-level attention mechanism for similarity extraction and feature aggregation.
• This creates consistent representations, forming effective positive samples and reliable supervision.

Main Results:

• AECL successfully addresses the false negative separation problem.
• The method generates discriminative representations for improved short text clustering.
• Experimental results show AECL outperforms state-of-the-art methods.

Conclusions:

• AECL offers an effective solution for short text clustering by enhancing contrastive learning.
• The proposed attention mechanism and dual-module approach improve representation quality and clustering accuracy.