Efficient Entity Disambiguation via Similarity Hashing

The task of Named Entity Disambiguation (NED), which maps mentions of ambiguous names in natural language onto a set of known entities, has been an important issue in many areas including machine translation and information extraction. Working with a huge amount of data (e.g. more than three million entities in Yago), some parts in an NED system which estimate the probability of a mention matching an entity, the similarity between a mention and an entity and the coherence among entity candidates for all mentions together might become bottlenecks. Thus, it is challenging for an interactive NED system to reach not only high accuracy but also efficiency. This thesis presents an efficient way of disambiguating named entities by similarity hashing. Our framework is integrated with AIDA which is an on-line tool for entity detection and disambiguation developed at Max-Planck Institute for Informatics. We apply various state-of-the-art approaches, for example Locality Sensitive Hashing (LSH) and Spectral Hashing, to some forms of similarity search problem such as near-duplicate search for mention-entity matching, and especially related pair detection for entity-entity mapping which is not the default application of using hashing techniques due to the usually low similarities between entities.