Challenging the Assumption of Structure-based embeddings in Few- and Zero-shot Knowledge Graph Completion

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6300–6309 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6300 Challenging the Assumption of Structure-based embeddings in Few- and Zero-shot Knowledge Graph Completion Filip Cornell1,3, Chenda Zhang2, ˇSar¯unas Girdzijauskas1, Jussi Karlgren3 1KTH Royal Institute of Technology, 2Carnegie Mellon University, 3Gavagai fcornell@kth.se, chendaz@andrew.cmu.edu, sarunasg@kth.se, jussi@lingvi.st Abstract In this paper, we report experiments on Few- and Zero-shot Knowledge Graph completion, where the objective is to add missing relational links between entities into an existing Knowledge Graph with few or no previous examples of the relation in question. While previous work has used pre-trained embeddings based on the structure of the graph as input for a neural network, nobody has, to the best of our knowledge, addressed the task by only using textual descriptive data associated with the entities and relations, much since current standard benchmark data sets lack such information. We therefore enrich the benchmark data sets for these tasks by collecting textual description data to provide a new resource for future research to bridge the gap between structural and textual Knowledge Graph completion. Our results show that we can improve the results for Knowledge Graph completion for both Few- and Zero-shot scenarios with up to a two-fold increase of all metrics in the Zero-shot setting. From a more general perspective, our experiments demonstrate the value of using textual resources to enrich more formal representations of human knowledge and in the utility of transfer learning from textual data and text collections to enrich and maintain knowledge resources. Keywords: Knowledge Graph completion, Meta-learning, Zero-shot learning, textual enrichment 1. Introduction Knowledge graphs are formalized representations of world knowledge, useful in many tasks such as Ques- tion Answering, Language Representation Learning and Recommender Systems (Ji et al., 2021). They con- sist of nodes representing entities of interest and di- rected edges representing different relations between the entities. A knowledge graph is frequently incom- plete: not every relation has been observed at the time the its construction (Xiong et al., 2018; Min et al., 2013). This is the basis for the Knowledge Graph com- pletion task: to infer missing links, given a graph. More formally, Knowledge Graphs can be described as a set of triplets (h, r, t) ∈ T with a head entity h, a tail entity t, and the relationship r defining their connec- tion. A Knowledge Graph can therefore be described as KG ⊆ E ×R×E, where E denotes the set of entities and R is the set of relationship types. Knowledge Graph completion is mostly addressed by scoring candidate triplets (h, r, t) for relation r by some assigned validity score. True or valid triplet candidates rank higher by the validity score and false ones lower. Previous work (Bordes et al., 2013; Yang et al., 2015; Dettmers et al., 2018; Sun et al., 2019) have moved a Knowledge Graph to a Euclidean vector space, assigning each entity and relation a d- dimensional vector. These …

Content truncated to 500 words