[Paper Review] Open Relation Modeling: Learning to Define Relations between Entities

Jie Huang, Kevin Chen–Chuan Chang+2 more|arXiv (Cornell University)|Aug 20, 2021

Topic Modeling4 citations

TL;DR

This paper introduces Open Relation Modeling, a novel task that generates coherent, interpretable sentences describing general relations between entities—especially those not directly linked in knowledge graphs. By fine-tuning pre-trained language models (PLMs) to generate definition-like descriptions conditioned on entity pairs and incorporating multi-hop reasoning paths from knowledge graphs with a confidence-based selection mechanism, the model produces concise, informative, and human-understandable relation descriptions, significantly improving interpretability over existing relation extraction methods.

ABSTRACT

Relations between entities can be represented by different instances, e.g., a sentence containing both entities or a fact in a Knowledge Graph (KG). However, these instances may not well capture the general relations between entities, may be difficult to understand by humans, even may not be found due to the incompleteness of the knowledge source. In this paper, we introduce the Open Relation Modeling problem - given two entities, generate a coherent sentence describing the relation between them. To solve this problem, we propose to teach machines to generate definition-like relation descriptions by letting them learn from defining entities. Specifically, we fine-tune Pre-trained Language Models (PLMs) to produce definitions conditioned on extracted entity pairs. To help PLMs reason between entities and provide additional relational knowledge to PLMs for open relation modeling, we incorporate reasoning paths in KGs and include a reasoning path selection mechanism. Experimental results show that our model can generate concise but informative relation descriptions that capture the representative characteristics of entities.

Motivation & Objective

• To address the lack of interpretable, general relation descriptions between entities, especially when they are not co-occurring in text or connected in knowledge graphs.
• To develop a method that generates human-readable, informative relation descriptions without requiring pre-specified relation types.
• To improve interpretability and openness in relation modeling by teaching PLMs to learn from entity definitions and reasoning paths.
• To enable modeling of multi-hop and implicit relations beyond direct facts in existing knowledge graphs.

Proposed method

• Fine-tune pre-trained language models (PLMs) to generate definition-like descriptions conditioned on extracted entity pairs.
• Collect entity definitions from text and extract entity pairs from these definitions to create a supervised signal for training.
• Incorporate multi-hop reasoning paths from knowledge graphs to provide relational knowledge and improve PLM reasoning between entities.
• Design a confidence-based reasoning path selection mechanism to filter and prioritize informative and interpretable paths for PLM input.
• Train the model end-to-end to generate concise, contextually accurate relation descriptions that reflect the representative characteristics of the entity pair.
• Leverage pre-trained model knowledge, entity definitions, and external reasoning paths to enhance the informativeness and coherence of generated descriptions.

Experimental results

Research questions

• RQ1 Can pre-trained language models be effectively fine-tuned to generate coherent, definition-like relation descriptions between entities without pre-specified relation types?
• RQ2 How can knowledge graph reasoning paths be effectively integrated into PLMs to improve the quality and informativeness of generated relation descriptions?
• RQ3 Can a confidence-based path selection mechanism improve the interpretability and relevance of reasoning paths used in open relation modeling?
• RQ4 To what extent can the model generate meaningful descriptions for entity pairs that are not directly connected in knowledge graphs or co-occurring in corpora?
• RQ5 How does the proposed method compare to existing relation extraction or text-to-text generation approaches in terms of interpretability and coverage of implicit or multi-hop relations?

Key findings

• The model generates concise and informative relation descriptions that effectively capture the representative characteristics of entity pairs, significantly improving interpretability over traditional relation extraction methods.
• Incorporating reasoning paths from knowledge graphs enhances the model’s ability to generate accurate and contextually relevant descriptions, especially for multi-hop or implicit relations.
• The confidence-based reasoning path selection mechanism successfully filters and prioritizes informative and interpretable paths, improving the quality of generated descriptions.
• The model performs well on entity pairs not directly connected in knowledge graphs, demonstrating its openness to previously unseen or incomplete relations.
• Fine-tuning PLMs on entity definitions enables them to learn general relation patterns, resulting in more coherent and semantically rich descriptions compared to zero-shot or prompt-based generation.
• Qualitative results show that generated descriptions are human-understandable and capture key aspects of the entities’ relationship, such as 'data mining is a process of extracting and discovering patterns in large data sets...' for 'data mining' and 'database'.