Domain-matched Pre-training Tasks for Dense Retrieval

TL;DR

This paper demonstrates that domain-matched pre-training tasks significantly improve dense retrieval performance by using large-scale synthetic questions and Reddit data, overcoming previous limitations in pre-training for retrieval tasks.

Contribution

It introduces a novel pre-training setup tailored for dense retrieval, utilizing large domain-specific datasets to achieve substantial performance gains.

Findings

Pre-training on synthetic questions improves retrieval accuracy.

Using Reddit post-comment pairs enhances dialogue retrieval performance.

The approach outperforms supervised baselines on multiple benchmarks.

Abstract

Pre-training on larger datasets with ever increasing model size is now a proven recipe for increased performance across almost all NLP tasks. A notable exception is information retrieval, where additional pre-training has so far failed to produce convincing results. We show that, with the right pre-training setup, this barrier can be overcome. We demonstrate this by pre-training large bi-encoder models on 1) a recently released set of 65 million synthetically generated questions, and 2) 200 million post-comment pairs from a preexisting dataset of Reddit conversations made available by pushshift.io. We evaluate on a set of information retrieval and dialogue retrieval benchmarks, showing substantial improvements over supervised baselines.