Conceptual structures such as frames--which identify the participant structure (i.e., the semantic predicate argument structure) of states, actions, and processes--have significant potential in knowledge-intensive tasks (for example, machine translation, information retrieval, information extraction, text summarization, natural language generation). For the most part, prior development of such structures has been either of limited scope and/or based on manual analysis, which is subject to lack of systematicity. What is needed are methods of deriving conceptual structures that are both methodologically regular and reasonably universal in scope. The goal of my dissertation research is (1) to devise a methodology for automatically identifying an inventory of semantic predicate argument structures and (2) to use the methodology to establish a basic thesaurus of such structures.
This work makes the foundational assumption that conceptual structure is manifest in language. More specifically, it draws on the dual observations (1) that the various verb (sense)s that may be used to communicate about a given conceptual structure (as buy, sell, pay, cost, spend, etc., could be used to communicate about a given commercial transaction) tend to accept more or less the same sets of lexical items among their syntactic arguments and (2) that the semantic classes of these lexical items tend to define the conceptual structure's semantic arguments (in the commercial transaction example, these would be a buyer, a seller, merchandise, and money).
The basic approach involves two steps: (1) identifying conceptual structures by finding sets of verb senses whose syntactic arguments tend to accept lexical items in common (by clustering verb senses based on nouns used in their definitions), and then (2) analyzing the semantic characteristics of the sets of nouns-corresponding-to-arguments in order to identify the (semantic) argument structure of the conceptual structures. Thus, the overall goals of the dissertation fall within the knowledge representation domain, while the means used to achieve those goals come out of computational linguistics.
For the colloquium series schedule, see the UMD Computational Linguistics Colloquium Series web page at http://umiacs.umd.edu/~resnik/cl_colloquium/. If you are interested in meeting with the speaker, please contact Philip Resnik (resnik@umiacs.umd.edu).