Automated Verb Sense Identification

This task seeks to develop an automated, domain-portable module for verb sense identification (VSI) in natural language processing (NLP). Since verbs are the primary carriers of event concepts, VSI, a primary sub-task of Word Sense Disambiguation (WSD),has significant impact on Information Extraction (IE). Currently, only keyword-based IE, mainly identification of named entities and simple, general events, is available. Concept-based extraction of events requires support from WSD.A scalable VSI system will extend InfoXtract, an IE system incorporating machine learning and grammar development toolkits. Language-specific, key-verb-based representation of extracted events will be translated into an `interlingua' based on actionconcepts.Previous systems for sense disambiguation are limited in terms of scalability due to their reliance on human-annotated corpora. The proposed approach is based on unsupervised machine learning from an InfoXtract-parsed corpus. Decision list models will betrained on an InfoXtract-parsed corpus from which a domain-dependent, WordNet-like thesaurus is constructed. Bootstrap training, using initial seeds, is performed on a new corpus without human intervention by leveraging the core IE engine.Deliverables include: (i) a verb thesaurus construction tool: given a representative training corpus, this tool will construct a domain-specific verb thesaurus, (ii) a VSI system customizable for a specific domain: both lexical knowledge as well ashigher-level selectional restriction rules will be employed in this system; and (iii) an information extraction application demonstrating the effective use of the VSI system in both the extraction and querying phase: the prototype will be demonstratedusing a news corpus such as the Foreign Broadcast Information Service (FBIS). The resulting component technology can be used in a variety of text processing and information applications, with the ability to automatically tune to any domain. This includesapplications such as Event Extraction, Concept Indexing, Question Answering, and Automatic Summarization. The technology will assist computerized information systems to make key information extracted from large volumes of free text accessible toinformation analysts and decision makers.