Introducing RezoJDM16k: a French KnowledgeGraph DataSet for Link Prediction

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5163–5169 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5163 Introducing RezoJDM16k: a French Knowledge Graph DataSet for Link Prediction Mehdi Mirzapour1, Waleed Ragheb2,5, Mohammad Javad Saeedizade3, K´evin Cousot4, H´el`ene Jacquenet1, Lawrence Carbon1, Mathieu Lafourcade2 ContentSide1; LIRMM, Univ Montpellier2; IUST3; Emvista4; FCAI, Cairo Univ5 R&D Dept, Lyon, France1; CNRS, Montpellier, France2; Tehran, Iran3 R&D Dept, Montpellier, France4; Cairo, Egypt5 {first.last}@contentside.com1; {first.last}@lirmm.com2 m saeedizade@comp.iust.ac.ir3; kevin.cousot@emvista.com4 Abstract Knowledge graphs applications, in industry and academia, motivate substantial research directions towards large-scale infor- mation extraction from various types of resources. Nowadays, most of the available knowledge graphs are either in English or multilingual. In this paper, we introduce RezoJDM16k, a French knowledge graph dataset based on RezoJDM (Lafourcade, 2007). With 16k nodes, 832k triplets and 53 relation types, RezoJDM16k can be employed in many NLP downstream tasks for the French language such as machine translation, question-answering and recommendation systems. In addition, we provide strong knowledge graph embedding baselines that are used in link prediction task for future benchmarking. Compared to the state-of-the-art English knowledge graph datasets used in link prediction, RezoJDM16k shows a similar promising predictive behavior. Keywords: language resource, knowledge graph dataset, link prediction, knowledge graph embedding, knowledge graph completion, lexical-semantic network 1. Introduction Knowledge Graphs (KGs) are structured representa- tions of semantic information mainly used for different tasks in artificial intelligence such as information extraction, search engines, question answering, and recommendation systems. KGs are often represented as multi-relational graphs with nodes and different types of edges. In a KG, each link is a triplet of the form (head, relation, tail). A triplet is a semantic representation of external world fact in which head and tail are entities. They are connected by a rela- tion which acts as the semantic predicate between the entities. For instance the triplet (hunt r agent −−−−−→ lion) indicates that lion is the agent of verb hunt, and the (coffee r carac −−−−−→ hot) indicates that coffee has the characteristic of being hot. There is no lim- itation to have a large number of facts collected in KGs. Although the basic idea behind KGs seems very promising, there is a crucial problem about KGs that makes them challenging to utilize: they are always incomplete (Wang et al., 2021). We can observe a lot of missing information (links or relations) between the entities in KGs. Moreover, real-world data are often dynamic and evolving, which makes it hard to build complete KGs (Cai et al., 2018; Arora, 2020). This is the reason behind the necessity of predicting missing information in KGs to make them as complete as possible. This task is called Link Prediction or Graph Completion which semantically refer to a unique notion. The most successful approach to address Link Prediction problem is based on Knowledge Graph Embedding (KGE) methods which transform KGs into a low-dimensional vector space. This transforma- tion, in principle, should preserve the structure of the KG and …

Content truncated to 500 words