Contextual Document Embeddings

TL;DR

This paper introduces methods for creating contextualized document embeddings that incorporate neighboring document information, leading to improved retrieval performance especially out-of-domain, and achieves state-of-the-art results on the MTEB benchmark.

Contribution

The paper proposes two novel approaches for contextualized document embeddings, explicitly integrating neighbor context into the embedding process, and demonstrates their effectiveness over traditional biencoders.

Findings

Both methods outperform biencoders in various settings.

Significant improvements observed especially out-of-domain.

Achieved state-of-the-art results on MTEB benchmark.

Abstract

Dense document embeddings are central to neural retrieval. The dominant paradigm is to train and construct embeddings by running encoders directly on individual documents. In this work, we argue that these embeddings, while effective, are implicitly out-of-context for targeted use cases of retrieval, and that a contextualized document embedding should take into account both the document and neighboring documents in context - analogous to contextualized word embeddings. We propose two complementary methods for contextualized document embeddings: first, an alternative contrastive learning objective that explicitly incorporates the document neighbors into the intra-batch contextual loss; second, a new contextual architecture that explicitly encodes neighbor document information into the encoded representation. Results show that both methods achieve better performance than biencoders in several settings, with differences especially pronounced out-of-domain. We achieve state-of-the-art results on the MTEB benchmark with no hard negative mining, score distillation, dataset-specific instructions, intra-GPU example-sharing, or extremely large batch sizes. Our method can be applied to improve performance on any contrastive learning dataset and any biencoder.