Entity typing using distributional semantics and DBpedia

Recognising entities in a text and linking them to an external resource is a vital step in creating a structured resource (e.g. a knowledge base) from text. This allows semantic querying over a dataset, for example selecting all politicians or football players. However, traditional named entity recognition systems only distinguish a limited number of entity types (such as Person, Organisation and Location) and entity linking has the limitation that often not all entities found in a text can be linked to a knowledge base. This creates a gap in coverage between what is in the text and what can be annotated with fine grained types. This paper presents an approach to detect entity types using DBpedia type information and distributional semantics. The distributional semantics paradigm assumes that similar words occur in similar contexts. We exploit this by comparing entities with an unknown type to entities for which the type is known and assign the type of the most similar set of entities to the entity with the unknown type. We demonstrate our approach on seven different named entity linking datasets. To the best of our knowledge, our approach is the first to combine word embeddings with external type information for this task. Our results show that this task is challenging but not impossible and performance improves when narrowing the search space by adding more context to the entities in the form of topic information.