[Paper Review] Exploiting Hybrid Semantics of Relation Paths for Multi-hop Question Answering Over Knowledge Graphs

Zile Qiao, Wei Ye+4 more|arXiv (Cornell University)|Sep 2, 2022

Topic Modeling4 citations

TL;DR

This paper proposes TERP, a novel multi-hop knowledge graph question answering method that exploits hybrid semantics from off-the-shelf relation paths by fusing their explicit textual descriptions and implicit KG structural features via a question-aware rotate-and-scale entity link prediction framework. The method achieves state-of-the-art performance, especially in complex multi-hop scenarios, by systematically coordinating question semantics with relation path representations to improve answer entity selection.

ABSTRACT

Answering natural language questions on knowledge graphs (KGQA) remains a great challenge in terms of understanding complex questions via multi-hop reasoning. Previous efforts usually exploit large-scale entity-related text corpora or knowledge graph (KG) embeddings as auxiliary information to facilitate answer selection. However, the rich semantics implied in off-the-shelf relation paths between entities is far from well explored. This paper proposes improving multi-hop KGQA by exploiting relation paths' hybrid semantics. Specifically, we integrate explicit textual information and implicit KG structural features of relation paths based on a novel rotate-and-scale entity link prediction framework. Extensive experiments on three existing KGQA datasets demonstrate the superiority of our method, especially in multi-hop scenarios. Further investigation confirms our method's systematical coordination between questions and relation paths to identify answer entities.

Motivation & Objective

• Address the challenge of multi-hop knowledge graph question answering (KGQA) where reasoning over multiple relations is required but supervision is limited.
• Overcome the underutilization of relation paths—rich yet overlooked sources of auxiliary information in existing KGQA methods.
• Improve answer entity selection by systematically coordinating question semantics with complementary relation path semantics.
• Develop a robust framework that fuses both explicit textual and implicit structural features of relation paths in a question-aware manner.
• Demonstrate the superiority of hybrid relation path semantics over relying solely on textual or embedding-based features in multi-hop reasoning tasks.

Proposed method

• Integrate explicit textual descriptions of relation paths (e.g., relation names and descriptions) as auxiliary features to enrich representation.
• Leverage RotatE embeddings to capture implicit structural semantics of relations by modeling them as rotations in complex vector space.
• Synthesize hybrid relation path representations through element-wise multiplication of individual relation embeddings and concatenation with textual features.
• Design a novel rotate-and-scale entity link prediction framework that jointly models entity rotation and modulus scaling to better align question and relation path semantics.
• Project both question representations and hybrid relation path features into the rotate-and-scale framework for end-to-end answer prediction.
• Apply a learnable fusion mechanism with hyperparameter λ to balance textual and structural contributions, optimizing performance across varying question-path similarity levels.

Experimental results

Research questions

• RQ1 Can off-the-shelf relation paths, with their combined textual and structural semantics, significantly improve multi-hop KGQA performance?
• RQ2 How effectively can a question-aware fusion of textual and KG embedding-based features of relation paths enhance answer entity selection?
• RQ3 To what extent do relation paths provide complementary information for hard questions with low semantic overlap to candidate paths?
• RQ4 Does a rotate-and-scale framework outperform standard rotation-only link prediction in aligning questions with relevant relation paths?
• RQ5 How does the performance of the method vary with the similarity between question and relation path representations?

Key findings

• The proposed TERP model achieves state-of-the-art performance on three benchmark KGQA datasets: WebQSP, ComplexWebQuestions, and MetaQA, with significant improvements in multi-hop scenarios.
• Hits@1 scores on 2-hop questions in WebQSP reach 92.4 when λ = 0.6, demonstrating optimal balance between textual and structural features.
• The model shows the most substantial performance gain—+7.3 Hits@1—on the 'Very Low' similarity group in WebQSP, confirming that relation paths provide critical support for hard, low-similarity questions.
• Performance degrades when relation path information is ignored on high-similarity questions, indicating that noisy paths can hinder models if not properly filtered.
• The rotate-and-scale mechanism significantly improves robustness and accuracy over standard rotation-only models, particularly in multi-hop reasoning.
• The ablation study confirms that both textual and structural features of relation paths are essential, and their fusion via question-aware modeling leads to systematic coordination between questions and paths.