[Paper Review] Knowledge Base Relation Detection via Multi-View Matching

Yang Yu, Kazi Saidul Hasan+3 more|arXiv (Cornell University)|Mar 1, 2018

Topic Modeling27 references3 citations

TL;DR

This paper proposes a multi-view matching model for knowledge base relation detection that improves performance by jointly matching questions against relation names and their associated object entity types. By leveraging attention-based matching across multiple views and incorporating KB-structured metadata, the model achieves state-of-the-art results on both SimpleQuestions and WebQSP datasets, outperforming prior methods by 3.42 and 0.45 points respectively.

ABSTRACT

Relation detection is a core component for Knowledge Base Question Answering (KBQA). In this paper, we propose a KB relation detection model via multi-view matching which utilizes more useful information extracted from question and KB. The matching inside each view is through multiple perspectives to compare two input texts thoroughly. All these components are designed in an end-to-end trainable neural network model. Experiments on SimpleQuestions and WebQSP yield state-of-the-art results.

Motivation & Objective

• To address the challenge of low accuracy in KB relation detection due to large relation vocabularies and zero-shot generalization needs.
• To improve relation detection by incorporating structural KB metadata, such as object entity types, into the matching process.
• To design an end-to-end neural network that performs multi-perspective matching across multiple views of question and relation representations.
• To evaluate the model on real-world KBQA benchmarks, including WebQSP and SimpleQuestions, under open-vocabulary and multi-hop reasoning settings.
• To demonstrate that object entity types significantly enhance matching performance, especially in complex or open-domain scenarios.

Proposed method

• The model uses a multi-view matching framework where each view combines different aspects of question and relation representations, including relation names and object entity types.
• It applies attention mechanisms to perform multiple-perspective matching between question and relation embeddings, enabling fine-grained comparison across different textual and structural features.
• Entity mentions in questions are replaced with special tokens (e.g., [ENT]) to help the model focus on contextual clues rather than specific entities.
• The model integrates both word-level and character-level embeddings, though experiments show character embeddings provide minimal benefit in this KB-specific setting.
• All components are trained end-to-end using a supervised loss on positive and negative relation instances derived from KB triples.
• Object entity types are automatically extracted from Freebase triples and used as an additional view in the matching process.

Experimental results

Research questions

• RQ1 Can incorporating object entity types from the KB improve relation detection performance beyond matching on relation names alone?
• RQ2 How does multi-perspective matching across multiple views of question and relation representations affect detection accuracy?
• RQ3 Does replacing entity mentions in questions improve model generalization and focus on relevant linguistic cues?
• RQ4 To what extent do character-level embeddings contribute to relation detection in low-vocabulary KB scenarios?
• RQ5 Can the model generalize to unseen relations and complex, multi-hop questions in open-domain KBQA?

Key findings

• The proposed model achieves state-of-the-art performance on the WebQSP dataset with a relation detection accuracy of 85.12%, outperforming the previous best system by 3.42 points.
• On the SimpleQuestions dataset, the model achieves 96.78% accuracy, a 0.45-point improvement over the prior state-of-the-art.
• Using object entity types as an additional view consistently improves performance, with the best results obtained when they are matched separately from relation names.
• Replacing entity mentions in questions with [ENT] tokens improves model focus on contextual features, leading to better generalization.
• Character-level embeddings provide no significant benefit, likely due to the limited and less variable vocabulary in KB relations compared to general text.
• The model demonstrates strong robustness on complex, multi-relation questions in WebQSP, where the average number of candidate relations per question is high and reasoning paths are longer.