Observational Initialization of Type-Supervised Taggers

Recent work has sparked new interest in type-supervised part-of-speech tagging, a data setting in which no labeled sentences are available, but the set of allowed tags is known for each word type. This paper describes observational initialization, a novel technique for initializing EM when training a type-supervised HMM tagger. Our initializer allocates probability mass to unambiguous transitions in an unlabeled corpus, generating token-level observations from type-level supervision. Experimentally, observational initialization gives state-of-the-art type-supervised tagging accuracy, providing an error reduction of 56% over uniform initialization on the Penn English Treebank.