Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Neural Re-Ranking via Corpus Graph

Andrea Giuseppe Di Francesco, Christian Giannetti, Nicola Tonellotto, Fabrizio Silvestri

TL;DR

The paper tackles re-ranking by incorporating document distribution through a corpus graph, enabling query-time modeling of cross-document interactions. It introduces Graph Neural Re-Ranking (GNRR), which builds a query-induced corpus subgraph and uses a GNN to encode document interactions alongside an MLP for individual-query signals, producing refined relevance scores. Empirical results on MS MARCO and DL19/DL20/DLHard show consistent improvements over uni-variate baselines, with up to $5.8\%$ relative $AP$ gains, and ablation confirms the GNN component's contribution, though statistical significance is not established. This approach demonstrates the practical value of contextual document modeling in neural re-ranking and suggests avenues for future work with heterogeneous graphs and broader tasks.

Abstract

Re-ranking systems aim to reorder an initial list of documents to satisfy better the information needs associated with a user-provided query. Modern re-rankers predominantly rely on neural network models, which have proven highly effective in representing samples from various modalities. However, these models typically evaluate query-document pairs in isolation, neglecting the underlying document distribution that could enhance the quality of the re-ranked list. To address this limitation, we propose Graph Neural Re-Ranking (GNRR), a pipeline based on Graph Neural Networks (GNNs), that enables each query to consider documents distribution during inference. Our approach models document relationships through corpus subgraphs and encodes their representations using GNNs. Through extensive experiments, we demonstrate that GNNs effectively capture cross-document interactions, improving performance on popular ranking metrics. In TREC-DL19, we observe a relative improvement of 5.8% in Average Precision compared to our baseline. These findings suggest that integrating the GNN segment offers significant advantages, especially in scenarios where understanding the broader context of documents is crucial.

Graph Neural Re-Ranking via Corpus Graph

TL;DR

The paper tackles re-ranking by incorporating document distribution through a corpus graph, enabling query-time modeling of cross-document interactions. It introduces Graph Neural Re-Ranking (GNRR), which builds a query-induced corpus subgraph and uses a GNN to encode document interactions alongside an MLP for individual-query signals, producing refined relevance scores. Empirical results on MS MARCO and DL19/DL20/DLHard show consistent improvements over uni-variate baselines, with up to relative gains, and ablation confirms the GNN component's contribution, though statistical significance is not established. This approach demonstrates the practical value of contextual document modeling in neural re-ranking and suggests avenues for future work with heterogeneous graphs and broader tasks.

Abstract

Re-ranking systems aim to reorder an initial list of documents to satisfy better the information needs associated with a user-provided query. Modern re-rankers predominantly rely on neural network models, which have proven highly effective in representing samples from various modalities. However, these models typically evaluate query-document pairs in isolation, neglecting the underlying document distribution that could enhance the quality of the re-ranked list. To address this limitation, we propose Graph Neural Re-Ranking (GNRR), a pipeline based on Graph Neural Networks (GNNs), that enables each query to consider documents distribution during inference. Our approach models document relationships through corpus subgraphs and encodes their representations using GNNs. Through extensive experiments, we demonstrate that GNNs effectively capture cross-document interactions, improving performance on popular ranking metrics. In TREC-DL19, we observe a relative improvement of 5.8% in Average Precision compared to our baseline. These findings suggest that integrating the GNN segment offers significant advantages, especially in scenarios where understanding the broader context of documents is crucial.
Paper Structure (10 sections, 7 equations, 4 figures, 1 table)

This paper contains 10 sections, 7 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries
  4. Proposed Approach
  5. Experimental Set-up
  6. Datasets
  7. Baselines
  8. Further Details on Experiments and Hyperparameters
  9. Results
  10. Conclusions

Figures (4)

  • Figure 1: Overview of our Graph Neural Re-Ranking pipeline with $K = 3$. We distinguish three phases: Data Retrieval, Subgraph Construction, and Features and Scores Computation. These phases encompass both retrieval and re-ranking, with the final output shown on the right. Initially, $d_3$ is more relevant than $d_4$, but after the re-ranking phase, $d_4$ records a higher score and should be ranked second.
  • Figure 2: Ablation study on DL19.
  • Figure 3: Ablation study on DL20.
  • Figure 4: Ablation on DLHard.

Theorems & Definitions (1)

  • Definition 1