IMPRS Publications :: Thesis :: DeMelo, Gerard

systems are increasingly dependent on knowledge about words and the entities they represent. This thesis investigates novel methods for automatically building large repositories of knowledge that capture semantic relationships between words, names, and entities, in many different languages. Three major contributions are made, each involving graph algorithms and statistical techniques that combine evidence from multiple sources of information.
The lexical integration method involves learning models that disambiguate word meanings based on contextual information in a graph, thereby providing a means to connect words to the entities that they denote. The entity integration method combines semantic items from different sources into a single unified registry of entities by reconciling equivalence and distinctness information and solving a combinatorial optimization problem. Finally, the taxonomic integration method adds a comprehensive and coherent taxonomic hierarchy on top of this registry,
capturing how different entities relate to each other.
Together, these methods can be used to produce a large-scale multilingual
knowledge base semantically describing over 5 million entities
and over 16 million natural language words and names in more than 200
different languages.