Prefix Data Augmentation for Contrastive Learning of Unsupervised Sentence Embedding

This paper presents prefix data augmentation (Prd) as an innovative method for enhancing sentence embedding learning through unsupervised contrastive learning. The framework, dubbed PrdSimCSE, uses Prd to create both positive and negative sample pairs. By appending positive and negative prefixes to a sentence, the basis for contrastive learning is formed, outperforming the baseline unsupervised SimCSE. PrdSimCSE is positioned within a probabilistic framework that expands the semantic similarity event space and generates superior negative samples, contributing to more accurate semantic similarity estimations. The model's efficacy is validated on standard semantic similarity tasks, showing a notable improvement over that of existing unsupervised models, specifically a 1.08% enhancement in performance on BERTbase. Through detailed experiments, the effectiveness of positive and negative prefixes in data augmentation and their impact on the learning model are explored, and the broader implications of prefix data augmentation are discussed for unsupervised sentence embedding learning.