|Date||Nov 15, 2013|
|Title||Modeling Large-Scale Knowledge Bases and Connecting Them to Text|
Huge amounts of complex data can be represented as multi-relational data, that is,graphs whose nodes stand for concepts and edges for relations among them. Inparticular, a subset of such data, termed knowledge bases (KBs) becameessential tools for storing, manipulating and accessing information in variousdomains ranging from search (e.g. Google Knowledge Graph) or bioinformatics(e.g. GeneOntology) to recommender systems (e.g. IMDB). However, KB datatypically cumulate many difficulties (large numbers of relation types -- somebeing significantly more represented than others, noisy and incomplete data, orlarge scale dimensions with up to millions of entities and billions of edgesfor real-world KBs), that make them hard to be fruitfully inserted intoexisting frameworks. This talk will present two research directions. First, wewill present new approaches for learning representations of large-scale KB data using energy-basedmethods, which allow for visualizing and completing them. Then, we willintroduce how such representations can be efficiently used to connect KB totext, and hence to improve relation extraction systems. This is a joint workwith Google, Université de Montréal, INRIA and Xerox.
Antoine Bordes is a CNRS researcher in the Heudiasyc laboratory of the University ofTechnology of Compiegne in France. In 2010, he was a postdoctoral fellow inYoshua Bengio's lab of Université de Montréal. He received his PhD in machinelearning from Pierre & Marie Curie University in Paris in early 2010. From2004 to 2009, he collaborated regularly with the Machine Learning department ofNEC Labs of America in Princeton. He received two awards for best PhD from theFrench Association for Artificial Intelligence and from the French ArmamentAgency. Antoine's current research concerns large-scale machine learningapplied to natural language processing and information extraction, and isfunded by the French National Research Agency.