Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hierarchical Graph Network for Multi-hop Question Answering

Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuohang Wang, Jingjing Liu

TL;DR

The paper introduces Hierarchical Graph Network (HGN) for multi-hop question answering by constructing a hierarchical, heterogeneous graph with four node types (question, paragraph, sentence, entity) and performing GAT-based reasoning to jointly predict paragraphs, supporting facts, entities, and answer spans. By linking clues across granularity levels and guiding sub-tasks through graph propagation, HGN achieves state-of-the-art performance on HotpotQA in both Distractor and Fullwiki settings. Key contributions include the two-step paragraph selection, seven edge-types in a unified graph, and a gated attention mechanism that fuses graph-aware representations for final answer extraction. The work demonstrates strong empirical gains and offers analysis on ablations, LM effects, and generalizability to other multi-hop QA scenarios.

Abstract

In this paper, we present Hierarchical Graph Network (HGN) for multi-hop question answering. To aggregate clues from scattered texts across multiple paragraphs, a hierarchical graph is created by constructing nodes on different levels of granularity (questions, paragraphs, sentences, entities), the representations of which are initialized with pre-trained contextual encoders. Given this hierarchical graph, the initial node representations are updated through graph propagation, and multi-hop reasoning is performed via traversing through the graph edges for each subsequent sub-task (e.g., paragraph selection, supporting facts extraction, answer prediction). By weaving heterogeneous nodes into an integral unified graph, this hierarchical differentiation of node granularity enables HGN to support different question answering sub-tasks simultaneously. Experiments on the HotpotQA benchmark demonstrate that the proposed model achieves new state of the art, outperforming existing multi-hop QA approaches.

Hierarchical Graph Network for Multi-hop Question Answering

TL;DR

The paper introduces Hierarchical Graph Network (HGN) for multi-hop question answering by constructing a hierarchical, heterogeneous graph with four node types (question, paragraph, sentence, entity) and performing GAT-based reasoning to jointly predict paragraphs, supporting facts, entities, and answer spans. By linking clues across granularity levels and guiding sub-tasks through graph propagation, HGN achieves state-of-the-art performance on HotpotQA in both Distractor and Fullwiki settings. Key contributions include the two-step paragraph selection, seven edge-types in a unified graph, and a gated attention mechanism that fuses graph-aware representations for final answer extraction. The work demonstrates strong empirical gains and offers analysis on ablations, LM effects, and generalizability to other multi-hop QA scenarios.

Abstract

In this paper, we present Hierarchical Graph Network (HGN) for multi-hop question answering. To aggregate clues from scattered texts across multiple paragraphs, a hierarchical graph is created by constructing nodes on different levels of granularity (questions, paragraphs, sentences, entities), the representations of which are initialized with pre-trained contextual encoders. Given this hierarchical graph, the initial node representations are updated through graph propagation, and multi-hop reasoning is performed via traversing through the graph edges for each subsequent sub-task (e.g., paragraph selection, supporting facts extraction, answer prediction). By weaving heterogeneous nodes into an integral unified graph, this hierarchical differentiation of node granularity enables HGN to support different question answering sub-tasks simultaneously. Experiments on the HotpotQA benchmark demonstrate that the proposed model achieves new state of the art, outperforming existing multi-hop QA approaches.

Paper Structure

This paper contains 28 sections, 9 equations, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-Hop QA
  4. Graph Neural Network
  5. Hierarchical Coarse-to-Fine Modeling
  6. Hierarchical Graph Network
  7. Graph Construction
  8. Paragraph Selection
  9. Nodes and Edges
  10. Context Encoding
  11. Graph Reasoning
  12. Gated Attention
  13. Multi-task Prediction
  14. Experiments
  15. Dataset
  16. ...and 13 more sections

Figures (3)

  • Figure 1: An example of multi-hop question answering from HotpotQA. The model needs to identify relevant paragraphs, determine supporting facts, and then predict the answer correctly.
  • Figure 2: Model architecture of Hierarchical Graph Network. The constructed graph corresponds to the example in Figure \ref{['fig:example_question']}. Green, blue, orange, and brown colors represent paragraph (P), sentence (S), entity (E), and question (Q) nodes, respectively. Some entities and hyperlinks are omitted for simplicity.
  • Figure 3: Examples of supporting facts prediction in the HotpotQA Distractor setting.