Tuesday, April 16, 2013
9:30 AM
Computer Studies Bldg. Room 209
Ph.D. Thesis Proposal
Jansen Orfan
University of Rochester
JORDN Ontology of Reasonably Deep kNowledge JORDN
A major hurdle for deep language understanding is the lack of intermediate general knowledge to support inference. One way to overcome this issue is to extract such knowledge from large sources of raw text, like the Web. However, because of the nature of its source material, this approach is prone error and unlikely to learn about obscure concepts. Another approach is to use one of the few hand built lexical knowledge bases, such as WordNet, VerbNet, and FrameNet. The sources offer high quality knowledge with a better concept coverage but they have two drawbacks. First, they are, at least, partially incompatible with each other, which prevents using all of the knowledge in the three sources at once and forces researchers to ignore a significant portion of work in the field simply because it relies on a different lexical resource. Secondly, a bulk of the sources' deeper knowledge - like concept definitions, semantic relationships, fine-grained selectional restrictions, etc - is not machine readable.
In this proposal I present the JORDN Ontology of Reasonably Deep kNowledge (JORDN), a linguistic ontology meant to provide general knowledge for deep language understanding tasks and to offer greater coverage and deeper knowledge than any single source described above. JORDN will begin with a core ontology and be populated in a few ways. First it will unifying the machine readable knowledge in WordNet, VerbNet, and FrameNet. It will also apply knowledge extraction techniques, similar to those currently used by the Gloss system, to their non-machine readable knowledge. Finally, JORDN will use inferences made from what it has already learned to extract even more knowledge about the semantic relations that exist between concepts. Completing JORDN will provide the general knowledge quality and coverage that deep language understanding requires.