Master Thesis Topic: Active Learning for Entity Alignment
Knowledge graphs are a versatile method to store information. One way to represent them is as a collection of triples (subject, predicate, object) that encode relations (i.e. the predicate) between entities (i.e., subject, object). Popular examples comprise Wikidata, or YAGO. With growing knowledge graphs, inserting information while preserving correctness becomes a cumbersome task. One aspect encountered here is to merge knowlegde graphs from different sources. As the labelling of entities (and predicates) might not directly coincide on both sides, aligning entities is of major importance, i.e. finding correspondences between entities from one graph in the other one (cf. e.g. here). Manually doing so might ensure high quality but soon becomes infeasible when confronted with large graphs.
Hence, this thesis aims to explore the application of active learning methods to the task of entity alignment. To this end, several existing general heuristics for active learning are reviewed, as well as specialised methods for graphs. In extensive experiments, the effect of the number of labels given on the final accuracy is studied.