Source-linked AI summary
Dynamically Fused Graph Network for Multi-hop Reasoning
Yunxuan Xiao, Yanru Qu, Lin Qiu, Hao Zhou, Lei Li, Weinan Zhang, Yong Yu
TL;DR
Multi-hop text-based QA requires reasoning over scattered evidence, while prior approaches often focus on single-paragraph answers or restricted graph entities. DFGN dynamically selects and propagates over entity subgraphs while fusing graph and token representations, achieving competitive or leading HotpotQA results and interpretable reasoning chains.
Problem
Multi-hop text-based QA requires selecting and inferring among scattered evidence from multiple documents, including cases where answers are not entities in an extracted graph.
Method
DFGN dynamically constructs query-guided entity graphs, masks irrelevant subgraphs at each reasoning step, and fuses document and graph representations.
Results
DFGN achieves leading results on HotpotQA and competitive performance against state-of-the-art unpublished models.
Takeaways & Limitations
DFGN produces reliable, explainable reasoning chains through predicted entity-graph masks.
Takeaways & Limitations
31.3% of development cases lack a complete reasoning path because limited NER accuracy and incomplete graph construction leave supporting entities unreachable.
Abstract
from arXiv · showhide
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.
1 Introduction
Multi-hop text-based QA requires selecting and connecting scattered evidence across documents, beyond single-paragraph retrieval. DFGN addresses this with dynamic entity-graph reasoning, document–graph fusion, and interpretable reasoning chains.
- Existing QA methods often retrieve answers from a single paragraph, which rarely tests deep reasoning capabilities.
- Multi-hop text-based QA requires selecting at least two scattered pieces of evidence and inferring among them without a predefined knowledge base.
- DFGN dynamically builds query-guided entity graphs and masks irrelevant entities during multiple reasoning rounds.
- DFGN fuses document-to-graph and graph-to-document information iteratively, obtaining final answers from document tokens rather than only graph entities.
- Predicted graph masks induce interpretable reasoning chains, with weak supervision and a new metric for evaluating chains and entity graphs.
- The authors report competitive HotpotQA performance against state-of-the-art unpublished models.
2 Related work
Related work distinguishes text-based and multi-hop QA and reviews graph and sequential reasoning models. The paper focuses on reasoning over raw text without knowledge-base schema restrictions.
- Text-based Question Answering: Text-based QA uses raw text as supporting information and tests end-to-end extraction of relevant facts and reasoning.
- Multi-hop Reasoning: Multi-hop QA requires identifying multiple related facts and reasoning about them, unlike single-hop QA, which requires one extracted fact.
- Multi-hop Reasoning: WikiHop and ComplexWebQuestions are constrained by existing knowledge-base schemas, whereas HotpotQA permits answers formed from free text.
- Multi-hop Reasoning for QA: Prior graph models aggregate scattered entity information using graph recurrent or relational graph convolutional networks, but their effective reasoning over constructed graphs remains underexplored.
- Multi-hop Reasoning for QA: Sequential multi-hop models update query or state representations step by step and can link selected entities into interpretable reasoning chains.
3 Dynamically Fused Graph Network
DFGN answers multi-hop text-based questions by dynamically building and exploring an entity graph while repeatedly fusing graph and document information. Its components select relevant paragraphs, propagate reasoning across entities, update queries, and predict answers from document tokens.
- 3.1 Paragraph Selection: DFGN selects relevant paragraphs with a BERT-based sentence classifier that assigns each paragraph a relevance score between 0 and 1.Paragraphs containing at least one supporting sentence receive positive training labels.
- 3.2 Entity Graph: The model constructs an entity graph from named entities, linking entities co-occurring in sentences or sharing the same mention text.The graph is built from the provided context without assuming a global knowledge base.
- 3.3 Encoding Query and Context: DFGN encodes the concatenated query and context with BERT, then applies bi-attention to enhance their cross-interactions.Concatenating query and context performs better than passing them separately, while bi-attention improves over BERT encoding alone.
- 3.4 Reasoning with the Fusion Block: Each fusion block transfers token information to entities, propagates it through a dynamically masked graph, and returns entity information to document tokens.The graph-to-document flow enables localization of answer text spans in the context.
- 3.4 Reasoning with the Fusion Block: A soft mask restricts information propagation to query-relevant start entities, producing a dynamic sub-part of the entity graph for each reasoning step.Attention between query and entity embeddings predicts the mask, while masked entities are encouraged or penalized before propagation.
- 3.4 Reasoning with the Fusion Block: The model updates query embeddings from current-step entities and uses the final fused context representation for supporting-sentence, answer-span, and answer-type predictions.Four stacked LSTMs produce the output dimensions, and their cross-entropy losses are jointly optimized with weighted terms and weak mask supervision.
4 Experiments
DFGN is evaluated on HotpotQA’s distractor setting through benchmark comparison, ablations, graph-quality metrics, and case studies of reasoning chains and failure modes.
- Experimental Setup: DFGN is evaluated on HotpotQA in the distractor setting, excluding full-wiki experiments because information retrieval is considered the bottleneck.
- Implementation Details: 97% recall and 69% precision are achieved when selecting supporting facts during paragraph selection.
- Main Results: DFGN attains the leaderboard’s second-best result, with a 1.5% joint F1 gain from an entity graph built using a BERT NER model.
- Main Results: Each DFGN component contributes 1%–2% relative QA gains, while a 1-layer fusion block causes an obvious performance loss.
- Main Results: The model is relatively insensitive to noisy paragraphs, unlike the baseline, which gains more than 5% when given only gold paragraphs or supporting facts.
- Graph Construction and Reasoning Chains: 31.3% of development cases cannot complete reasoning because at least one supporting sentence is unreachable through the constructed entity graph.
- Graph Construction and Reasoning Chains: ESP evaluates whether predicted top-k entity paths hit supporting sentences, using exact match and recall over the selected reasoning chains.
- Case Study: Case studies show that a bridge entity can link reasoning chains, whereas missed starting entities or absent numerical computation produce failures.
5 Conclusion
DFGN addresses multi-hop reasoning by dynamically predicting sub-graphs at each step while fusing entity-level reasoning with token-level contexts. On HotpotQA, it achieves leading results and produces reliable, explainable reasoning chains.
- DFGN uses a dynamic fusion reasoning block based on graph neural networks to address multi-hop reasoning.
- DFGN dynamically predicts sub-graphs at each reasoning step instead of relying on a fixed graph.
- DFGN fuses entity-level reasoning with token-level contexts.
- DFGN achieves leading results on HotpotQA.
- DFGN produces reliable and explainable reasoning chains.