Learning to Borrow– Relation Representation for Without-Mention Entity-Pairs for Knowledge Graph Completion

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2887 - 2898 July 10-15, 2022 ©2022 Association for Computational Linguistics Learning to Borrow – Relation Representation for Without-Mention Entity-Pairs for Knowledge Graph Completion Huda Hakami1, Mona Hakami2, Angrosh Mandya3 and Danushka Bollegala3,4 1Department of Computer Science, Taif University, Saudi Arabia 2King Saud University, Saudi Arabia 3University of Liverpool, UK 4Amazon hahakami@tu.edu.sa, mohakami@ksu.edu.sa {angrosh,danushka}@liverpool.ac.uk Abstract Prior work on integrating text corpora with knowledge graphs (KGs) to improve Knowl- edge Graph Embedding (KGE) have obtained good performance for entities that co-occur in sentences in text corpora. Such sentences (tex- tual mentions of entity-pairs) are represented as Lexicalised Dependency Paths (LDPs) be- tween two entities. However, it is not possible to represent relations between entities that do not co-occur in a single sentence using LDPs. In this paper, we propose and evaluate sev- eral methods to address this problem, where we borrow LDPs from the entity pairs that co- occur in sentences in the corpus (i.e. with men- tion entity pairs) to represent entity pairs that do not co-occur in any sentence in the cor- pus (i.e. without mention entity pairs). We propose a supervised borrowing method, Su- perBorrow, that learns to score the suitability of an LDP to represent a without-mention en- tity pair using pre-trained entity embeddings and contextualised LDP representations. Ex- perimental results show that SuperBorrow im- proves the link prediction performance of mul- tiple widely-used prior KGE methods such as TransE, DistMult, ComplEx and RotatE. 1 Introduction Knowledge Graphs (KGs) are a structured form of information that underline the relationships between real-world entities (Ehrlinger and W¨oß, 2016; Kroetsch and Weikum, 2016; Paulheim, 2017). A KG is represented using a set of rela- tional tuples of the form (h, r, t), where r repre- sents the relation between the head entity h and the tail entity t. For example, the relational tu- ple (Joe Biden, president-of, US) indicates that the president-of relation holds between Joe Biden and US. There exists a large number of publicly avail- able and widely used KGs, such as Freebase (Bol- lacker et al., 2008), DBpedia (Auer et al., 2007), and YAGO ontology (Suchanek et al., 2007). KGs have been effectively applied in various NLP tasks such as, relation extraction (Riedel et al., 2013; We- ston et al., 2013), question answering (Das et al., 2017; Sydorova et al., 2019), and dialogue sys- tems (Xu et al., 2020). However, most KGs suffer from data sparseness as many relations between entities are not explicitly represented (Min et al., 2013). To overcome the sparsity problem, Knowledge Graph Embedding (KGE) models learn representa- tions (a.k.a. embeddings) for entities and relations in a given KG in a vector space, which can then be used to infer missing links between entities (Bordes et al., 2013; Nickel et al., 2015; Wang et al., 2017). Such models are trained to predict relations that are likely to exist between entities (known as link prediction …

Content truncated to 500 words