Margin-based angular losses for lightweight text classification: Lessons from face recognition

doi:10.1016/j.asoc.2025.114431

Margin-based angular losses for lightweight text classification: Lessons from face recognition

2025-12-13

0

PRE

AI

A

Arash Yousefi Jordehi

S

Seyed Abolghasem Mirroshandel *

DOI：10.1016/j.asoc.2025.114431

Original

Original request for help

Save

Abstract

En 中文

• Novel text classification framework inspired by margin-based face recognition losses (e.g., CosFace, ArcFace, SphereFace). • Our method models class labels as dynamic prototypes in angular space, promoting intra-class compactness and inter-class separability. • A dual-objective contrastive mechanism enhances semantic clustering via attractive and repulsive angular forces. • The framework operates in a single stage, enabling inference during training, and avoids complex contrastive pre-training procedures. • Extensive experiments demonstrate superior performance and efficiency across multiple benchmark datasets.

Margin-based angular losses for lightweight text classification: Lessons from face recognition

Abstract

Journal

Organization

Cited Papers