Abstract:
In the Natural Language Processing domain, a thriving research topic is
learning from large-scale knowledge graphs such as Wikidata, DBpedia, WordNet,
or Freebase. These graphs contain large scales of triple-based facts such as (Obama,
born in, Hawaii). Several models have been proposed for representation learning of
the graphs’ entities and relations, which makes it possible to assess the plausibility
of facts not yet present in the graph (Knowledge Graph Completion).
In this thesis, we explore methods that can be used to extend these knowledge
graph completion models to determine the plausability of facts which contain
entities that were not present in the knowledge graph, a task commonly referred
to as Open-World Knowledge Graph Completion. More specifically, we present a
novel extension which performs this task by utilizing the textual descriptiosn of the
unknown entities. A transformation is learned to map the embeddings of an entity’s
name and description to the graph-based embedding space. We demonstrate
competitive results on several datasets including FB20k, DBPedia50k and our new
dataset FB15k-237-OWE. Our approach exploits the full knowledge graph structure
even when textual descriptions are scarce, does not require a joint training on graph
and text, and can be applied to any embedding-based Knowledge Graph Completion
model.
