Shuly Wintner (shuly@csa.cs.technion.ac.il)
Tue, 6 Aug 1996 09:56:57 +0300
Hello, I wish to thank all those who responded to my previous note, both publicly and in private messages. In light of the responses I can now say that there are at least two different issues in question here: (1) what is the "underlying mathematical formalism" for HPSG? (2) Given the formalism, which is probably Turing-equivalent in power (so that "one can do one's taxes with it"), what subset of the grammars are indeed HPSG grammars? Why are these questions important? First, I believe we must have a good understanding of the theory we work with in order for us to use it effectively. Second, as people might have different views regarding these questions, there certainly is a place for some standardization, or at least a common vocabulary, in order to get better cooperation among different researchers. I understand that the second question is hard to answer, and indeed maybe it doesn't have an answer. But the first question is crucial, especially if you are planning any implementation for HPSG (such as a grammar compiler, a parser, a generator etc.) I'm afraid that P&S94 doesn't elaborate enough on this point, and I'll give an example below. If we agree that the principles are stated as "implicative constraints within a certain feature logic" (in other words, the formalism is "based on typed feature logic with recursive types and relational constraints"), then each principle has the form: if X is compatible with description_1 then X must be compatible with description_2 We must clarify what "compatible" means and what the description language is, but I want to point out a specific problem. Consider the subcategorization principle of P&S94: if X is a headed phrase, then DTRS|HEAD-DTR|SYNSEM|LOC|CAT|SUBCAT is the concatenation of SYNSEM|LOC|CAT|SUBCAT with the list of SYNSEM values of DTRS|COMP-DTRS Can the consequence of the implication be considered a description? In particular, "the list of SYNSEM values" isn't well defined - the number of elements is unknown. Of course, one can write relational constraints for 'concatenate', if one has access to something like Prolog (or the definite clause component of ALE). Do we want to assume that such a component is inherently part of the underlying formalism? I think we'd better do without it. Indeed, some pre-defined relations (say, 'append', 'set-union' etc.) might be required, but not the full mechanism of some logic programming language. I will be happy to get any comments, either by private e-mail or on the list. Shuly Shuly Wintner shuly@cs.technion.ac.il Computer Science shuly@techunix.bitnet Technion, Israel Institute of Technology tel: +972-4-8294315 Haifa 32000, Israel fax: +972-4-8294353 http://www.cs.technion.ac.il/~shuly
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:33:28 PST