Entity Ambiguity Resolution with Topic Models

We address the problem of ambiguity resolution in associating facts to the entities with multiple senses. A semi-supervised graphical model assigns the best categories for entities given the context. Problem Statement Analyzing the results of current algorithms for extracting facts about entities [1] shows that despite the mutual exclusion constraints that are enforced, there is still a need for resolving ambiguity and determining which set of facts belong to what sense of the entity. For example, for the entity “Tigers” we have the categories “Mammal” and “Sports Team”. While there is a mutex rule in the seed ontology for these two categories, the entity contains extracted fact for “teamPlaysAgainstTeam” which only applies to “Sports Team” sense of this word. In this project, we will add the support for multiple senses of the entity and also define the learning method that can recognize when ambiguity exists and also the ability to associate facts with the correct sense of the entity. This work can also potentially allow us to infer the mutex rules from the data.