Using Hyperbolic Geometry for FG-NET over Distantly Supervised data

Fine-Grained Named Entity Typing (\FGNET{}) classifies an entity mention into a fine range of entity types. A large number of entity types make it difficult to manually label the training data, thus distant supervision is used to automatically acquire the training data. Distant supervision incurs a lot of training noise which hinders the performance improvement of the FG-NET systems. In this paper, we propose to use hyperbolic geometry for FG-NET with the hope that it can help overcoming the noise incurred by distant supervision.