|Stochastic Systems Group|
Learning annotated hierarchies from relational data
Collins and Quillian (1969) introduced Semantic Networks, a knowledge representation that captured the intuition that (i) objects in many real-world domains can be organized into hierarchies, where each internal node picks out a category of objects, and (ii) features of objects should be specified at the most general category to admit generalization. However, this representation suffers from several drawbacks; the strict logical foundation of Semantic Networks makes them brittle in the presence of noise, there exists no principled proposal to learn them from data, and they can only represent features of individual object, not relationships between objects. We introduce a Bayesian model and learning algorithm for _annotated hierarchies_, a hierarchical extension of the stochastic block model used for statistical relational learning. Like Semantic Networks, annotated hierarchies organize objects into nested categories and specify which categories are relevant for modeling each feature. Unlike Semantic Networks, they also model relations between objects, handle noisy and missing data, and provide predictive semantics. We describe the model and show that our learning algorithm discovers interpretable structure in several real-world data sets.
Joint work with Charles Kemp, Vikash Mansinghka and Joshua Tenenbaum.
Problems with this site should be emailed to firstname.lastname@example.org