Improving entity disambiguation with a vector space semantic tagger
Abstract
This note describes a few suggestions for improving entity disambiguation using a vector-based semantic tagger, trained using the Skipgram model. The suggestions include assuming non uniform distributions for the probability distribution of the entity, different ways of building vectors for the document, and using a neural network architecture. We exemplify the suggestions on a running example: the disambiguation of entity Boston, which may be referring to the famous city in Massachusetts US, or a town in Lincolnshire UK, or as we will see in the examples below, a few other places in the US and UK.