$R^2$-CoD: Understanding Text-Graph Complementarity in Relational Reasoning via Knowledge Co-Distillation

2025-08-02

AI Summary

This work investigates the complementary mechanisms between textual and graph-structured representations in relational reasoning tasks and their implications for hybrid modeling. We propose Knowledge Co-Distillation (CoD), a unified framework that jointly optimizes a text encoder and a graph neural network, dynamically tracking the evolution of their latent spaces across five relational reasoning tasks. Through representation analysis, we characterize stage-wise alignment and divergence patterns, systematically identifying the conditions and intrinsic drivers underlying modality complementarity. Experiments demonstrate that CoD not only improves multi-task performance but also yields interpretable insights into cross-modal synergy: text enhances semantic generalization, while graph structure enforces logical constraints—complementarity peaks at specific training stages. To our knowledge, this is the first interpretability framework for multimodal relational reasoning grounded in latent-space dynamics.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Machine Learning: Multimodal Learning
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Relational reasoning lies at the core of many NLP tasks, drawing on complementary signals from text and graphs. While prior research has investigated how to leverage this dual complementarity, a detailed and systematic understanding of text-graph interplay and its effect on hybrid models remains underexplored. We take an analysis-driven approach to investigate text-graph representation complementarity via a unified architecture that supports knowledge co-distillation (CoD). We explore five tasks involving relational reasoning that differ in how text and graph structures encode the information needed to solve that task. By tracking how these dual representations evolve during training, we uncover interpretable patterns of alignment and divergence, and provide insights into when and why their integration is beneficial.

Problem

• Understanding text-graph complementarity in relational reasoning
• Investigating interplay of text-graph representations in hybrid models
• Analyzing alignment and divergence of dual representations during training

Innovation

• Knowledge co-distillation for text-graph complementarity
• Unified architecture supporting relational reasoning analysis
• Tracking dual representation evolution during training