[论文解读] Dynamically Fused Graph Network for Multi-hop Reasoning

Yunxuan Xiao, Yanru Qu+5 more|arXiv (Cornell University)|May 16, 2019

Topic Modeling参考文献 21被引用 43

一句话总结

DFGN 引入了一个基于查询引导的动态实体图推理模型，用于多跳文本基础的问答，具备文档到图与图到文的融合以及用于解释推理链的弱监督掩码。在 HotpotQA 上的评估显示竞争力的结果。

ABSTRACT

Text-based question answering (TBQA) has been studied extensively in recent years. Most existing approaches focus on finding the answer to a question within a single paragraph. However, many difficult questions require multiple supporting evidence from scattered text among two or more documents. In this paper, we propose Dynamically Fused Graph Network(DFGN), a novel method to answer those questions requiring multiple scattered evidence and reasoning over them. Inspired by human's step-by-step reasoning behavior, DFGN includes a dynamic fusion layer that starts from the entities mentioned in the given query, explores along the entity graph dynamically built from the text, and gradually finds relevant supporting entities from the given documents. We evaluate DFGN on HotpotQA, a public TBQA dataset requiring multi-hop reasoning. DFGN achieves competitive results on the public board. Furthermore, our analysis shows DFGN produces interpretable reasoning chains.

研究动机与目标

• Motivate multi-hop QA where evidence is scattered across multiple documents.
• Propose a dynamic entity-graph reasoning framework guided by the query.
• Enable information fusion between documents and the entity graph to improve answer localization.
• Provide interpretable reasoning chains through learned entity masks and weak supervision.
• Evaluate on HotpotQA to demonstrate competitive performance and reasoning transparency.

提出的方法

• Construct a dynamic, query-guided entity graph from paragraph context and query.
• Encode query and context with BERT and Bi-Attention to get Q0 and C0.
• Use a fusion block to (a) map tokens to entities (Doc2Graph), (b) propagate over a dynamic graph with a soft start-entity mask, and (c) propagate back from graph to tokens (Graph2Doc).
• Apply dynamic graph attention to propagate information over a subgraph selected by a soft mask.
• Update the query representation across hops and use a cascade of LSTMs for predicting supporting facts, start/end positions, and answer type.
• Weakly supervise the mask predictions via heuristic start masks and BFS-based masks to guide reasoning chains.

实验结果

研究问题

• RQ1 Can a dynamically constructed, query-focused entity graph improve multi-hop reasoning over scattered evidence?
• RQ2 Does end-to-end training of a dynamic fusion and masking mechanism yield interpretable reasoning chains?
• RQ3 How does doc-to-graph and graph-to-doc fusion affect robustness to noisy paragraphs and open-domain QA?
• RQ4 What level of performance does DFGN achieve on HotpotQA compared with state-of-the-art models?

主要发现

Model	Answer_EM	Answer_F1	SupFact_EM	SupFact_F1	Joint_EM	Joint_F1
Baseline Model	45.60	59.02	20.32	64.49	10.83	40.16
GR ∗	52.92	66.71	52.37	84.11	31.77	58.47
DFGN(Ours)	55.17	68.49	49.85	81.06	31.87	58.23
QFE ∗	53.86	68.06	57.75	84.49	34.63	59.61
DFGN(Ours) †	56.31	69.69	51.50	81.62	33.62	59.82

• DFGN achieves competitive results on HotpotQA’s distractor setting (e.g., higher EM/F1 than baselines and competitive with unpublished models).
• Using a 2-layer fusion block yields significant gains; single-layer blocks degrade performance, highlighting multi-hop reasoning importance.
• Entity graphs built with a stronger NER (BERT NER) improve joint F1 by about 1.5% in their reported setting.
• Ablations show components like BFS-based supervision, exact entity masking, and query update contribute modest gains individually (~1–2% relative).
• ESP scores and analysis indicate the model can predict reasonable reasoning chains and exposes interpretable reasoning paths, though some cases miss supporting entities due to NER limits.