Embedding And Clustering Your Data Can Improve Contrastive Pretraining

TL;DR

This paper proposes a novel data stratification method using clustering to improve contrastive pretraining of text embeddings, demonstrating significant performance gains on retrieval tasks.

Contribution

It introduces a clustering-based data organization approach for contrastive pretraining, extending beyond source-based stratification to semantic clusters within sources.

Findings

Increased NDCG@10 on MSMARCO dataset

Clustering improves contrastive pretraining effectiveness

Connects clustering with TAS-B and ANCE methodologies

Abstract

Recent studies of large-scale contrastive pretraining in the text embedding domain show that using single-source minibatches, rather than mixed-source minibatches, can substantially improve overall model accuracy. In this work, we explore extending training data stratification beyond source granularity by leveraging a pretrained text embedding model and the classic k-means clustering algorithm to further split training data apart by the semantic clusters within each source. Experimentally, we observe a notable increase in NDCG@10 when pretraining a BERT-based text embedding model on query-passage pairs from the MSMARCO passage retrieval dataset. Additionally, we conceptually connect our clustering approach to both the Topic Aware Sampling (TAS) aspect of the TAS-B methodology and the nearest-neighbor-based hard-negative mining aspect of the ANCE methodology and discuss how this unified view motivates future lines of research on the organization of contrastive pretraining data.