# Description-Enhanced Label Embedding Contrastive Learning for Text Classification

**Authors:** Kun Zhang, Le Wu, Guangyi Lv, Enhong Chen, Shulan Ruan, Jing Liu, Zhiqiang Zhang, Jun Zhou, Meng Wang  
**arXiv:** [2306.08817](https://arxiv.org/abs/2306.08817)  
**Date:** 2023-06-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.08817v1/

## Key points

- Paper focuses on text classification in natural language processing
- Existing methods overlook the importance of label information
- Authors propose a novel approach called R2-Net for text classification
- R2-Net incorporates Self-Supervised Learning (SSL) and introduces a self-supervised Relation of Relation (R2) classification task
- Triplet loss is used to enhance analysis of differences and connections among labels
- External knowledge from WordNet is incorporated for label semantic learning
- Paper is organized into different sections covering related work, formal definitions, technical details, experiments, discussions, and conclusions
- Experimental results demonstrate the effectiveness of the proposed method in utilizing label information for improved text classification performance.

## Summary

This paper focuses on the task of text classification in natural language processing. While deep neural networks, particularly Pre-trained Language Models (PLMs), have shown impressive performance in this area, most existing methods overlook the importance of label information. They either treat labels as meaningless one-hot vectors or use basic embedding methods to learn label representations during model training. This limits the semantic information and guidance that labels can provide. To address this issue, the authors propose a novel approach called Relation of Relation Learning Network (R2-Net) for text classification. The R2-Net incorporates Self-Supervised Learning (SSL) and introduces a self-supervised Relation of Relation (R2) classification task to better utilize label information from a one-hot perspective. The authors also employ triplet loss to enhance the analysis of differences and connections among labels. Additionally, they incorporate external knowledge from WordNet to obtain multi-aspect descriptions for label semantic learning. The paper is organized as follows: Section II provides an overview of related work in text classification and contrastive learning. Section III presents formal definitions of text classification and the proposed R2 classification task. Sections IV and V describe the technical details of R2-Net and its extension to a Description-Enhanced Label Embedding network (DELE). The experiments and detailed analysis are presented in Section VI, followed by discussions and conclusions in Sections VII and VIII. In summary, this paper proposes a novel approach for text classification that leverages SSL, R2 classification, triplet loss, and external knowledge from WordNet. The experimental results demonstrate the effectiveness of the proposed method in utilizing label information for improved text classification performance.