Verb Sense Disambiguation Using Integrated Feature Experts

"We propose an innovative combination of machine learning techniques coupled with an extensible framework of integrated feature type classifiers (called "feature experts") capable of performing high-accuracy verb sense disambiguation. Our DisambiguatingIntegrated Verb Sense Experts (DIVERSE) system combines a committee of feature experts with a common representational ontology bootstrapped for a new domain using cooperative learning techniques and an unannotated domain corpus. The result will be anadaptive, user-trainable system that leverages multiple predictive features (to be selected from a comprehensive survey) to generate sense tags and associated confidence levels with respect to a predetermined catalog of verb senses. Active learning andbootstrapping techniques minimize system reliance on a non-linguist user during ontology tuning. An iterative development cycle ensures that the system is rapidly trained to sufficient levels of accuracy within a new task domain. Our extensible frameworkprovides a solid foundation for a systematic evaluation of feature experts operating in isolation and in coordination with other experts. Phase I research and development of a proof-of-concept limited prototype will demonstrate the feasibility and utilityof DIVERSE's verb sense identification capability and will lay the groundwork for its Phase II implementation and eventual commercialization. DIVERSE is applicable across language processi