Planning Ahead in Generative Retrieval: Guiding Autoregressive Generation through Simultaneous Decoding

TL;DR

This paper presents PAG, a novel approach for guiding autoregressive document identifier generation in generative retrieval models through simultaneous decoding, significantly improving retrieval performance and speed.

Contribution

PAG introduces a set-based and sequential identifier construction method, enhancing generative retrieval with optimized decoding strategies and relevance-based representations.

Findings

15.6% MRR improvement on MS MARCO

22x faster query latency

Outperforms state-of-the-art generative retrieval models

Abstract

This paper introduces PAG-a novel optimization and decoding approach that guides autoregressive generation of document identifiers in generative retrieval models through simultaneous decoding. To this aim, PAG constructs a set-based and sequential identifier for each document. Motivated by the bag-of-words assumption in information retrieval, the set-based identifier is built on lexical tokens. The sequential identifier, on the other hand, is obtained via quantizing relevance-based representations of documents. Extensive experiments on MSMARCO and TREC Deep Learning Track data reveal that PAG outperforms the state-of-the-art generative retrieval model by a large margin (e.g., 15.6% MRR improvements on MS MARCO), while achieving 22x speed up in terms of query latency.