Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

QAGCN: Answering Multi-Relation Questions via Single-Step Implicit Reasoning over Knowledge Graphs

Ruijie Wang, Luca Rossetto, Michael Cochez, Abraham Bernstein

TL;DR

QAGCN tackles multi-relation QA over knowledge graphs by enabling end-to-end single-step implicit reasoning through a novel Question-Aware GCN that jointly embeds questions and KG entities. The model constructs a question-focused subgraph, encodes both the question and subgraph with BERT-backed embeddings and attention-guided GCN layers, and searches for answers via distance in embedding space plus a relation-path based reranking mechanism. Empirical results on PQ, PQL, and MetaQA show competitive or superior performance relative to state-of-the-art explicit reasoning methods, with notably faster training than NSM and generally efficient inference times. The work demonstrates that careful attention to question-driven message propagation and path-level semantics can yield strong QA performance with a simpler, more scalable architecture, making it practical for real-world KG QA deployments.

Abstract

Multi-relation question answering (QA) is a challenging task, where given questions usually require long reasoning chains in KGs that consist of multiple relations. Recently, methods with explicit multi-step reasoning over KGs have been prominently used in this task and have demonstrated promising performance. Examples include methods that perform stepwise label propagation through KG triples and methods that navigate over KG triples based on reinforcement learning. A main weakness of these methods is that their reasoning mechanisms are usually complex and difficult to implement or train. In this paper, we argue that multi-relation QA can be achieved via end-to-end single-step implicit reasoning, which is simpler, more efficient, and easier to adopt. We propose QAGCN -- a Question-Aware Graph Convolutional Network (GCN)-based method that includes a novel GCN architecture with controlled question-dependent message propagation for the implicit reasoning. Extensive experiments have been conducted, where QAGCN achieved competitive and even superior performance compared to state-of-the-art explicit-reasoning methods. Our code and pre-trained models are available in the repository: https://github.com/ruijie-wang-uzh/QAGCN

QAGCN: Answering Multi-Relation Questions via Single-Step Implicit Reasoning over Knowledge Graphs

TL;DR

QAGCN tackles multi-relation QA over knowledge graphs by enabling end-to-end single-step implicit reasoning through a novel Question-Aware GCN that jointly embeds questions and KG entities. The model constructs a question-focused subgraph, encodes both the question and subgraph with BERT-backed embeddings and attention-guided GCN layers, and searches for answers via distance in embedding space plus a relation-path based reranking mechanism. Empirical results on PQ, PQL, and MetaQA show competitive or superior performance relative to state-of-the-art explicit reasoning methods, with notably faster training than NSM and generally efficient inference times. The work demonstrates that careful attention to question-driven message propagation and path-level semantics can yield strong QA performance with a simpler, more scalable architecture, making it practical for real-world KG QA deployments.

Abstract

Multi-relation question answering (QA) is a challenging task, where given questions usually require long reasoning chains in KGs that consist of multiple relations. Recently, methods with explicit multi-step reasoning over KGs have been prominently used in this task and have demonstrated promising performance. Examples include methods that perform stepwise label propagation through KG triples and methods that navigate over KG triples based on reinforcement learning. A main weakness of these methods is that their reasoning mechanisms are usually complex and difficult to implement or train. In this paper, we argue that multi-relation QA can be achieved via end-to-end single-step implicit reasoning, which is simpler, more efficient, and easier to adopt. We propose QAGCN -- a Question-Aware Graph Convolutional Network (GCN)-based method that includes a novel GCN architecture with controlled question-dependent message propagation for the implicit reasoning. Extensive experiments have been conducted, where QAGCN achieved competitive and even superior performance compared to state-of-the-art explicit-reasoning methods. Our code and pre-trained models are available in the repository: https://github.com/ruijie-wang-uzh/QAGCN
Paper Structure (16 sections, 3 equations, 3 figures, 5 tables)

This paper contains 16 sections, 3 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. QAGCN -- A Question-Aware GCN-Based Question Answering Model
  4. Task Definition
  5. The Question-Aware GCN
  6. Subgraph Extraction
  7. Question Encoding
  8. Subgraph Encoding
  9. Answer Search
  10. Experiments
  11. Effectiveness Evaluation
  12. Comparison with SOTA Reasoning-based Method --- NSM
  13. Efficiency Evaluation
  14. Ablation Study
  15. Case Study with Embedding Visualization
  16. ...and 1 more sections

Figures (3)

  • Figure 1: An overview of the QAGCN model.
  • Figure 2: Training curves of QAGCN and NSM on PQL-2hop and PQL-3hop.
  • Figure 3: Visualization of the embedding space after each layer of the Q&G-Encoders. Please note that the axes do not have a same scale. The purple point indicates the question embedding after each layer. The blue points indicate the top-3 candidates finally retrieved. The orange points indicate all other entities in the subgraph.