Open Bachelor Thesis
Neural entity typing [1,2] aims at associating every entity mention in a context with multiple types. For example, an entity mention “Obama” in a sentence could be associated with multiple types “persons”, “politician”, “lawyer” and “author”. Usually, these types could be organized in a hierarchical structure, e.g., directed acyclic graphs (DAG).
Previous works embed types into a high-dimensional Euclidean [1] or hyperbolic space [2]. However, the prediction scores of these methods might be inconsistent with the inherent type hierarchy, e.g., an input mention must be labeled as "person" if it is already labeled as "politician" and "lawyer". In this work, we propose to associate each type with a hyperbolic region and model type-type relations as regional inclusions. The classification score is then modeled by the confidence of an instance being inside of the regions. The method would work on settings in which the type hierarchies are either explicitly given or unknown but can be derived from labeled data (e.g., by co-occurrence). We will evaluate the method in both settings.