From - Mon Feb 9 11:02:02 2004 X-Mozilla-Status: 0001 X-Mozilla-Status2: 00000000 Path: acsu.buffalo.edu!rapaport From: rapaport@cse.buffalo.edu (William J. Rapaport) Newsgroups: sunyab.cse.740 Subject: QUESTIONS ABOUT CVA Date: Thu, 5 Feb 2004 15:16:40 -0500 (EST) Organization: Computer Science and Engineering Lines: 78 Sender: Ncs@buffalo.edu Distribution: sunyab Message-ID: NNTP-Posting-Host: wasat.cse.buffalo.edu X-Trace: prometheus.acsu.buffalo.edu 1076012200 1747 128.205.32.15 (5 Feb 2004 20:16:40 GMT) X-Complaints-To: abuse@buffalo.edu NNTP-Posting-Date: Thu, 5 Feb 2004 20:16:40 +0000 (UTC) X-Newsreader: trn 4.0-test76 (Apr 2, 2001) Xref: acsu.buffalo.edu sunyab.cse.740:66 ------------------------------------------------------------------------ Subject: QUESTIONS ABOUT CVA ------------------------------------------------------------------------ A student writes: > I have some questions about CVA project: > > 1. For SNePS natural language processing, there is a lexicon file. All > the word you input has to be defined in the lexicon file, or SNePS > will not respond. So, if the new word is not defined in the lexicon, > how can SNePS respond? Good question. First, although there is a natural-language front end to SNePS, we won't be using it in this part of the project. Here's why: In order to (semantically) parse a sentence into a SNePS representation, we need to know what the representation should be. The definition algorithms only work on the SNePS representation, ignoring (for now) any grammatical information that might come from a syntactic parsing. Consequently, as a research strategy, we've concentrated just on the knowledge representation and reasoning part of the project, and not on the linguistic part. (One exception has been a project by Vik Rao on modifying a computational grammar of a fragment of English (in particular, an ATN grammar that interfaces with SNePS) to take the "brachet" demo in English as input, to generate the SNePS representations as output, then to run the definition algorithm on it, and to generate the response in English.) The problem of not having the unknown word in the lexicon can be handled in two ways. Vik handled it (as I recall) merely by putting "brachet" in as a noun. But Hunt & Koplas 1988 is a tech report that shows how to modify the grammar and lexicon to automatically insert new words; this paper is online and accessiple from the CVA website. > 2. When SNePS figures out the meaning of a unknown word, will it write the > information to the lexicon file? No. For one thing, the lexicon only contains grammatical information, not semantic. For another, we (currently) are following Johnson-Laird's advice *not* to assert (i.e., "memorize" or "learn") the definition. (Actually, we tried this once, with a "word-definition" case frame. When we modified the definition algorithm to use it, that's all it did: it reported the stored definition.) > 3. For natural language processing, SNePS only recognizes the words > defined in the lexicon file. It doesn't know the meaning of the word. > Lexicon file doesn't provide the meaings of words. Depends on what you mean by "the meaning". The lexicon does not contain a dictionary-like definition (or a Fregean sense or referent, either). But, on a "conceptual role semantics" theory of meaning, all the meaning is in the network. Our algorithms are designed to "recover" (or "construct") a dictionary-like definition from the entire network. For more on "CRS" theories, see Rapaport 2002 on the CVA website. > In CVA project, we give > the new word meaning. How about other words? Shall we need the meaning of > other words? You might need some of this as "prior knowledge" to be included in the network along with the representation of the textual context. One way we've been handling this is via "meaning postulates", by which we mean either necessary or sufficient conditions. (A definition typically provides both necessary and sufficient conditions for a term; but in practice, only one of these is really needed. As for *which* part, that will depend on the text you're working on and the word whose (partial) meaning you need. > If yes, Does CVA project means add more vocabulary to a > dictionary? No, first because there's no dictionary; second, because the only thing that is going to be added is information that will help in providing a definition of the unknown word. Take a look at some of the other reports that have come out of the project to get a clearer idea. I recommend in particular the report on the word "harbinger", Goldfain 2003.