ITR: Framenet++: An On-Line Lexical Semantic Resource and its Application to Speech and Language Understanding

International Computer Science Institute, Berkeley CA

Investigators

Abstract

This is the first year funding of a three-year continuing award. Robust domain-independent language understanding is essential for multilingual information extraction, summarization, question answering, and automatic translation. With pervasive computing environments soon to come, language understanding will become even more indispensable for interacting with artifacts of widely different functionalities. The field of natural language understanding has made significant progress in the last fifteen years. A large part of this gain is due to the sophisticated combination of statistical algorithms with template-based algorithms tailored to specific domains like air-traffic information, travel scheduling, and business news. But any real solution to the problem of domain-independent understanding will require moving beyond template-based monolingual systems to more flexible, general purpose HCI systems via three key innovations: (1) a domain-independent semantic language as the back end for these understanding systems, replacing the current domain-restricted templates and slots; (2) rich semantic lexical databases which are broad enough to cover the necessary words for language engineering tasks, and deep enough in usable semantic information to support true domain-independent understanding; and (3) sophisticated techniques for performing this mapping. This project will develop these three components: a very large lexical database FrameNet++, a semantic language designed for domain-independent understanding tasks, and the tools for applying it to and evaluating it on key NLU applications. The semantic language and lexical database are based on formalizing the semantic frames and the semantic and syntactic combinatory properties - the valences - of a significant portion of the English lexicon. FrameNet++ will offer significantly richer semantic information than is available in current databases like COMLEX and WordNet, by characterizing the conceptual frames within which words are defined and identifying the semantic roles which the arguments of these words can take. These roles and frames are key to building domain-independent language understanding applications. The project will focus from the start on specific NLU applications: word sense disambiguation, information extraction, multilingual information extraction, and an eventual extension to text data mining. For each application, the PI and his team will apply the FrameNet++ system to improve the domain independence of the semantic components, using statistical algorithms for semantic annotation that we have already begun to implement. These applications will in turn provide a rich and realistic evaluation framework to guide FrameNet++ development, and will encourage potential users to apply it to a wide variety of tasks. The FrameNet++ database will be capable of serving many purposes. Provided with statistical information about frequencies of words, word/sense mappings, and combinatorial patterns linked to word senses, it will be usable in various automatic language understanding processes, including word sense disambiguation and information extraction. Since the formal semantic annotations are keyed to conceptual structures which are independent of any individual language, they are available for the creation of parallel lexicon databases of other languages. The semantic structures in the databases will facilitate matches from one language to another, in machine translation and machine-assisted translation, while the syntactic structures allow the production of appropriate grammatical sentences in the target language.