# Adaptive Re-Ranking with a Corpus Graph

**Authors:** Sean MacAvaney, Nicola Tonellotto, Craig Macdonald  
**arXiv:** [2208.08942](https://arxiv.org/abs/2208.08942)  
**Date:** 2022-08-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.08942v1/

## Key points

- The paper introduces Graph-based Adaptive Re-ranking (GAR) as a novel approach to improving re-ranking pipelines in search systems.
- GAR has shown significant improvements in precision- and recall-oriented measures compared to traditional re-ranking methods.
- The method is based on the clustering hypothesis and involves continuously adding similar documents to the candidate pool during re-ranking.
- GAR is compatible with existing techniques like dense retrieval, robust in terms of hyperparameters, and adds minimal computational and storage costs.
- Experiments on the MS MARCO passage ranking dataset showed promising results, with GAR enhancing nDCG of a BM25 candidate pool by up to 8% when combined with a monoT5 ranker.

## Summary

The paper "Adaptive Re-Ranking with a Corpus Graph" by Sean MacAvaney, Nicola Tonellotto, and Craig Macdonald introduces a novel approach to improving the performance of re-ranking pipelines in search systems. The proposed method is known as Graph-based Adaptive Re-ranking (GAR) and has shown significant improvements in precision- and recall-oriented measures compared to traditional re-ranking methods. <br><br>
Re-ranking pipelines typically involve assigning new ranking scores to documents or passages from an initial pool of candidates. These pipelines are limited by the recall of the initial candidate pool, as documents not identified initially cannot be re-ranked. To overcome this limitation, the authors propose a method based on the clustering hypothesis. Their approach involves continuously adding documents to the candidate pool that are most similar to the highest-scoring documents at each stage of re-ranking. This feedback process adapts the pool to include potentially high-ranking documents that were not present in the initial set and boosts the scores of deeper-lying documents that may have been overlooked due to budget constraints.<br><br>
GAR is also compatible with various existing techniques such as dense retrieval and is robust in terms of hyperparameters. It adds minimal computational and storage costs while showing promising results in experiments on the MS MARCO passage ranking dataset. When combined with a monoT5 ranker, GAR was able to enhance the nDCG of a BM25 candidate pool by up to 8%. Overall, this innovative approach presents a promising solution for enhancing search system performance through adaptive re-ranking strategies based on corpus graphs.