Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rational Retrieval Acts: Leveraging Pragmatic Reasoning to Improve Sparse Retrieval

Arthur Satouf, Gabriel Ben Zenou, Benjamin Piwowarski, Habiboulaye Amadou Boubacar, Pablo Piantanida

TL;DR

This work introduces Rational Retrieval Acts (RRA), an adaptation of the Rational Speech Acts framework to sparse information retrieval. By modeling literal and pragmatic speaker/listener dynamics over token-document interactions and exploiting sparsity, RRA produces pragmatic document representations that better contrast relevant documents, improving out-of-domain performance on BEIR without increasing inference cost. The approach uses a single RSA iteration with a tunable pragmatism parameter and a pre-transformation function to transform initial token-document weights into an RSA-ready lexicon. Empirical results show consistent gains for multiple sparse IR models, with notable boosts on datasets featuring longer queries, and a case study demonstrating improved token weighting in practice. The work highlights the potential of collection-aware representations and outlines offline RSA re-computation as a limitation, suggesting avenues for future collection-dependent or incremental updates.

Abstract

Current sparse neural information retrieval (IR) methods, and to a lesser extent more traditional models such as BM25, do not take into account the document collection and the complex interplay between different term weights when representing a single document. In this paper, we show how the Rational Speech Acts (RSA), a linguistics framework used to minimize the number of features to be communicated when identifying an object in a set, can be adapted to the IR case -- and in particular to the high number of potential features (here, tokens). RSA dynamically modulates token-document interactions by considering the influence of other documents in the dataset, better contrasting document representations. Experiments show that incorporating RSA consistently improves multiple sparse retrieval models and achieves state-of-the-art performance on out-of-domain datasets from the BEIR benchmark. https://github.com/arthur-75/Rational-Retrieval-Acts

Rational Retrieval Acts: Leveraging Pragmatic Reasoning to Improve Sparse Retrieval

TL;DR

This work introduces Rational Retrieval Acts (RRA), an adaptation of the Rational Speech Acts framework to sparse information retrieval. By modeling literal and pragmatic speaker/listener dynamics over token-document interactions and exploiting sparsity, RRA produces pragmatic document representations that better contrast relevant documents, improving out-of-domain performance on BEIR without increasing inference cost. The approach uses a single RSA iteration with a tunable pragmatism parameter and a pre-transformation function to transform initial token-document weights into an RSA-ready lexicon. Empirical results show consistent gains for multiple sparse IR models, with notable boosts on datasets featuring longer queries, and a case study demonstrating improved token weighting in practice. The work highlights the potential of collection-aware representations and outlines offline RSA re-computation as a limitation, suggesting avenues for future collection-dependent or incremental updates.

Abstract

Current sparse neural information retrieval (IR) methods, and to a lesser extent more traditional models such as BM25, do not take into account the document collection and the complex interplay between different term weights when representing a single document. In this paper, we show how the Rational Speech Acts (RSA), a linguistics framework used to minimize the number of features to be communicated when identifying an object in a set, can be adapted to the IR case -- and in particular to the high number of potential features (here, tokens). RSA dynamically modulates token-document interactions by considering the influence of other documents in the dataset, better contrasting document representations. Experiments show that incorporating RSA consistently improves multiple sparse retrieval models and achieves state-of-the-art performance on out-of-domain datasets from the BEIR benchmark. https://github.com/arthur-75/Rational-Retrieval-Acts
Paper Structure (12 sections, 10 equations, 2 figures, 2 tables)

This paper contains 12 sections, 10 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The Rational Retrieval Acts (RRA)
  3. Integrating the RSA
  4. Leveraging the sparsity in IR
  5. Scoring documents
  6. Experiments
  7. Setup
  8. Pre-transformation function
  9. Baselines and datasets
  10. Results
  11. Case Study: SPARTA on SciFact.
  12. Discussion

Figures (2)

  • Figure 1: Rational Retrieval Acts (RRA) architecture. Document relevance $L_1(d|t)$ is computed via RSA, assuming a pragmatic user (speaker) query. Final ranking uses the dot product between query and document representations (Eq.\ref{['eq:rsv']})
  • Figure 2: Impact of $\alpha$ on S-RRA performance (nDCG@10). Non-optimal $\alpha$ values can significantly degrade results.