Coupled Learning of Textual Extraction Patterns

In this paper, we focus on methods of coupling the semi-supervised learning of information extractors that extract information (e.g., City(X) and AthletePlaysForTeam (X, Y)) from free text using textual extraction patterns (e.g., “mayor of X” and “Y star quarterback X”). We identify three general types of coupling among target functions that can be combined to form a dense network of coupled learning problems. We then present an approach in which the input to the learner is an ontology defining a set of target categories and relations to be learned, a handful of seed examples for each, and a set of constraints that couple the various categories and relations (e.g., Person and Sport are mutually exclusive). We show that given this input and millions of unlabeled documents, a semi-supervised learning procedure can achieve very significant accuracy improvements by coupling the training of textual pattern-based extractors for dozens of categories and relations. Based on results reported here, we hypothesize that even greater accuracy improvements will be possible by forming a larger and denser network of inter-constrained learning tasks. The main research contributions of the paper are: (1) this work is the first to couple the simultaneous semi supervised training of both category and relation textual pattern-based extractors and (2) this work proposes that learning many tasks and coupling them as much as possible leads to higher accuracy semi-supervised learning, and provides web-scale experimental evidence to support that point.