- Keywords: open knowledge graph, link prediction, relational learning
- TL;DR: We study link prediction in open knowledge graphs and provide a dataset, an evaluation protocol and baseline models for this task.
- Abstract: We study the problem of link prediction in open knowledge graphs (OKGs) produced by open information extraction systems. In contrast to traditional knowledge graphs, OKGs do not use a predefined or closed vocabulary of entities and relations. The same entity can be mentioned in many different ways, and relations can be semantically close but not equal (e.g. ``visited by'' and ``frequently visited by''). The open link prediction problem is as follows: Given a mention of an entity and an open relation (e.g., ``A. Einstein'' and ``is a native of''), predict correct entity mentions (e.g., ``Imperial Germany'') using solely the OKG. Models that solve the open link prediction task successfully need to be able to reason about different entity mentions, as well as the relations between them, in a fully automatic, unsupervised way. We propose a benchmark consisting of an OKG dataset and an evaluation protocol for this task, and discuss and evaluate various baseline models. We found that the task is difficult to solve using current methods, but combinations of token-based link prediction models appear to be a promising direction for further research.
- Archival status: Archival
- Subject areas: Question Answering, Relational AI