Text and Code Embeddings by Contrastive Pre-Training

TL;DR

This paper demonstrates that contrastive pre-training on large-scale unsupervised data produces high-quality text and code embeddings, outperforming previous models in classification, semantic search, and code search tasks.

Contribution

The authors introduce a contrastive pre-training approach that yields state-of-the-art unsupervised text and code embeddings across multiple benchmarks.

Findings

Achieves 4% and 1.8% improvements in classification accuracy over previous models.

Attains 23.4%, 14.7%, and 10.6% improvements in semantic search benchmarks.

Obtains 20.8% improvement in code search performance.

Abstract

Text embeddings are useful features in many applications such as semantic search and computing text similarity. Previous work typically trains models customized for different use cases, varying in dataset choice, training objective and model architecture. In this work, we show that contrastive pre-training on unsupervised data at scale leads to high quality vector representations of text and code. The same unsupervised text embeddings that achieve new state-of-the-art results in linear-probe classification also display impressive semantic search capabilities and sometimes even perform competitively with fine-tuned models. On linear-probe classification accuracy averaging over 7 tasks, our best unsupervised model achieves a relative improvement of 4% and 1.8% over previous best unsupervised and supervised text embedding models respectively. The same text embeddings when evaluated on large-scale semantic search attains a relative improvement of 23.4%, 14.7%, and 10.6% over previous best unsupervised methods on MSMARCO, Natural Questions and TriviaQA benchmarks, respectively. Similarly to text embeddings, we train code embedding models on (text, code) pairs, obtaining a 20.8% relative improvement over prior best work on code search.