Detmar Meurers (dm@sfs.nphil.uni-tuebingen.de)
Thu, 7 May 1998 18:55:40 +0200 (MET DST)
Dear Victoria, >I would be very grateful to obtain any information on good platform >for *real* HPSG grammar and lexicon implementation. what information one can supply here depends a lot on what you mean by "real HPSG grammar". I take it that with "real" you mean "true to the linguistic theory"? Under that perspective, one supposedly wants to take an HPSG theory as formulated by the linguist, e.g. the one included in the appendix of Pollard and Sag (1994), and run it in a computational system. The most important thing in such an enterprise is to establish a direct, formal link between the semantics of a grammar in the computational system you use and the denotational semantics of an HPSG theory as for example provided by King (1989,1994). But this is easier said than done, (at least) for two reasons: 1. Theoretical: HPSG is an evolving formalism, and not all the mechanisms used by linguists when writing down "an HPSG theory" are properly understood (or understandable). For other mechanisms, only one of several possible formalizations have been fully worked out, like the description level formalization of lexical rules I proposed in 95 as opposed to a meta-level possibility. This specific formalization can then be used as the basis of a computational treatment, like the one Guido Minnen and I proposed for lexical rules in Computational Linguistics 23:4 (which to my knowledge is the only computational approach dealing with lexical rules as written down in HPSG linguistics). But in general, there is no agreement on how which mechanism should be interpreted. This means that it is advisable for the linguist to explicitly say what formalizations (s)he assumes. Then one can look for a computational setup realizing those specifications. 2. Computational: In the HPSG architecture, the phrase structure relationship between mother and daughters has been replaced by general constraints allowing any kind of composition of the phonology of a constituent, not only concatenation of the phonology of a known number of daughters. Indeed, a number of authors have argued that theses increased linearization possibilities are linguistically motivated (Reape, Kathol, Richter/Sailer, ...). But this liberty comes at the cost of abandoning basically all of the efficient and termination ensuring algorithms, which rely on the possibility to index by phonology. As a result, a computational setup like the ConTroll system we develop in T"ubingen, which attempts to realize the idea of implementing HPSG as close as possible to the linguistic theory, will have to deal with a significant number of problems which don't arise or have already been solved for ordinary phrase structure based grammars. In the current ConTroll system, the user therefore has to make a number of theoretically unmotivated specifications to ensure tractability - but we're working on that ;-) If you're interested in the issues involved in developing a system such that it closely corresponds to the semantics of HPSG, take a look at the following pages: http://www.sfs.nphil.uni-tuebingen.de/~dm/papers.html http://www.sfs.nphil.uni-tuebingen.de/controll/ http://www.sfs.nphil.uni-tuebingen.de/~tg/ The first one also includes a report I wrote in 1994 entitled "On Implementing an HPSG theory - Aspects of the logical architecture, the formalization and the implementation of head-driven phrase structure grammars", which discusses some of the foundational, formal and computational issues involved in recoding an HPSG theory as a phrase structure grammar or logic program. Finally, a detailed survey of some systems for the implementation of HPSG-like grammars was written by Leonard Bolc, Krzysztof Czuba, Anna Kupsc, Malgorzata Marciniak, Agnieszka Mykowiecka and Adam Przepiorkowski. It can be retrieved from: http://www.sfs.nphil.uni-tuebingen.de/~adamp/Papers/1996-survey/index.html One should keep in mind, though, that most of the computational systems available today were never actually intended for the implementation of true-to-theory HPSG grammars, but as general-purpose computational tools to implement general feature-based grammars. So even though one should be aware of the many places where all currently available systems differ from HPSG theory, these differences should not be misunderstood as general shortcomings of such systems. Sonnigen Gru3, Detmar ---------------------- Detmar Meurers Universit"at T"ubingen Seminar f"ur Sprachwissenschaft Abt. Computerlinguistik Kleine Wilhelmstr. 113 D-72074 Tuebingen Tel. +49-(0)7071-2978489 http://www.sfs.nphil.uni-tuebingen.de/~dm/ ----------------------
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:38:19 PST