A Memory Efficient Baseline for Open Domain Question Answering

TL;DR

This paper proposes methods to significantly reduce the memory footprint of dense retrieval systems in open-domain question answering, enabling competitive performance with less than 6GB of memory.

Contribution

It introduces three strategies—dimension reduction, vector quantization, passage filtering—to decrease index size without sacrificing accuracy.

Findings

Achieved competitive QA performance with under 6GB memory

Demonstrated effectiveness of memory reduction techniques on TriviaQA and NaturalQuestions

Provided practical solutions for memory-efficient dense retrieval systems

Abstract

Recently, retrieval systems based on dense representations have led to important improvements in open-domain question answering, and related tasks. While very effective, this approach is also memory intensive, as the dense vectors for the whole knowledge source need to be kept in memory. In this paper, we study how the memory footprint of dense retriever-reader systems can be reduced. We consider three strategies to reduce the index size: dimension reduction, vector quantization and passage filtering. We evaluate our approach on two question answering benchmarks: TriviaQA and NaturalQuestions, showing that it is possible to get competitive systems using less than 6Gb of memory.