Construction and Canonicalization of Economic Knowledge Graphs with LLMs

Hanieh Khorashadizadeh, Nandana Mihindukulasooriya, Nilufar Ranji, Morteza Kamaladdini Ezzabady, Frédéric Ieng, Jinghua Groppe, Farah Benamara, Sven Groppe

Abstract: Ontology-based knowledge graphs, such as YAGO and Wikidata, rely on pre-defined schemas to organize and connect information. While effective, these systems are inherently domain-specific, requiring tailored ontologies that are costly, time-consuming, and demand expert knowledge to develop. To address these limitations, Open Information Extraction (OpenIE) offers a complementary approach by extracting structured information directly from unstructured text without needing a predefined schema. However, OpenIE results in a vast number of relations, often leading to redundancy and inconsistencies. To overcome this, we propose a novel approach that leverages Large Language Models (LLMs) for constructing a knowledge graph and for canonicalizing relations within it. Our method includes generating question-answer pairs from text, extracting triples from these pairs, and applying a two-step canonicalization process to ensure consistency and reduce redundancy. This paper presents our approach in detail, exploring related work, the construction of the knowledge graph, the canonicalization process, and the evaluation of our methods.