Precise Zero-Shot Dense Retrieval without Relevance Labels

By Luyu Gao et al

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
1 Introduction
2 Related Works
3 Methodology
4 Experiments
4.1 Setup
4.2 Web Search
4.3 Low Resource Retrieval

Summary

This paper introduces a method called HyDE for precise zero-shot dense retrieval without relevance labels. The authors propose using Hypothetical Document Embeddings (HyDE) to tackle the challenge of zero-shot learning and encoding relevance. HyDE involves generating hypothetical documents using an instruction-following language model and encoding them with a contrastive encoder. The experiments conducted show that HyDE outperforms existing unsupervised dense retrievers and performs comparably to fine-tuned models across various tasks and languages.