[Paper Review] Covidia: COVID-19 Interdisciplinary Academic Knowledge Graph

Cheng Deng, Jiaxin Ding+4 more|arXiv (Cornell University)|Apr 14, 2023

Biomedical Text Mining and OntologiesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

Covidia is a large-scale, interdisciplinary academic knowledge graph for COVID-19 research, integrating 1.5 million papers across biology, medicine, computer science, sociology, and other fields. It uses contrastive learning for multi-label paper classification, BERT-based entity and relation extraction, and a novel ontology scheme to enable cross-domain knowledge integration, benchmarking community discovery and link prediction with high-fidelity retrieval across disciplines.

ABSTRACT

The pandemic of COVID-19 has inspired extensive works across different research fields. Existing literature and knowledge platforms on COVID-19 only focus on collecting papers on biology and medicine, neglecting the interdisciplinary efforts, which hurdles knowledge sharing and research collaborations between fields to address the problem. Studying interdisciplinary researches requires effective paper category classification and efficient cross-domain knowledge extraction and integration. In this work, we propose Covidia, COVID-19 interdisciplinary academic knowledge graph to bridge the gap between knowledge of COVID-19 on different domains. We design frameworks based on contrastive learning for disciplinary classification, and propose a new academic knowledge graph scheme for entity extraction, relation classification and ontology management in accordance with interdisciplinary researches. Based on Covidia, we also establish knowledge discovery benchmarks for finding COVID-19 research communities and predicting potential links.

Motivation & Objective

• To address the fragmentation of interdisciplinary COVID-19 research by integrating knowledge across biology, medicine, computer science, and social sciences.
• To overcome the limitations of existing literature platforms that focus only on biomedical publications, which exclude 48% of all COVID-19 papers.
• To develop a unified academic knowledge graph that supports multi-domain entity extraction, relation classification, and disciplinary categorization.
• To enable effective information retrieval and knowledge discovery across disciplines through structured, semantically enriched representations.
• To establish benchmarks for community detection and link prediction in interdisciplinary research networks.

Proposed method

• Proposes a contrastive learning-based multi-label paper classification model to assign papers to multiple disciplines, improving cross-domain categorization accuracy.
• Introduces a BERT-based entity extraction model enhanced with Wikipedia alignment to capture discipline-specific meanings and improve open-domain adaptability.
• Employs a segment embedding-based relation extraction model to identify relations between extracted entities in interdisciplinary contexts.
• Designs a novel academic knowledge graph schema with dual layers: ontology (concepts and relations) and instance (papers, entities, and bibliometric data) linked via shared references.
• Uses geohash to represent spatio-temporal information and OWL:sameAs axioms to link internal entities to external knowledge bases.
• Builds four social science networks—coauthor, citation, author-paper, and paper-author networks—using bibliographic data for network science analysis.

Experimental results

Research questions

• RQ1 How can interdisciplinary COVID-19 research papers be effectively classified across multiple domains, especially when they span beyond biology and medicine?
• RQ2 What techniques enable accurate and context-aware entity and relation extraction in heterogeneous academic domains?
• RQ3 How can a unified knowledge graph integrate entities and relations from diverse disciplines while preserving disciplinary semantics?
• RQ4 What are the structural and statistical properties of interdisciplinary research networks in the context of COVID-19?
• RQ5 To what extent can the knowledge graph support knowledge discovery tasks such as community detection and link prediction?

Key findings

• Covidia integrates 1.5 million COVID-19 research papers, with 48% originating from non-biomedical venues, highlighting the need for interdisciplinary knowledge platforms.
• The contrastive learning-based classification model improves multi-label categorization of papers across diverse disciplines, enabling accurate disciplinary tagging.
• The BERT-based entity extraction model, enhanced with Wikipedia alignment, achieves high precision in identifying domain-specific entities across open domains.
• The knowledge graph supports complex queries such as one-hop (e.g., papers mentioning a specific knowledge point), two-hop (e.g., illustrations in a domain), and three-hop (e.g., places studied by an affiliation), which are unanswerable by standard search engines.
• Network analysis reveals power-law degree distributions in coauthor, citation, and author-paper networks, with α values ranging from 1.377 to 1.511, indicating scale-free network characteristics.
• The system enables benchmarking for community detection and link prediction, demonstrating utility in identifying research communities and uncovering hidden interdisciplinary connections.