Fouvry F (fouvry@llsun10.essex.ac.uk)
Fri, 9 May 1997 11:52:32 +0100
[ Despite the header, this message is only posted to the HPSG-list. ] Dear HPSG'ers, dear ALE users, For my master's dissertation Computational Linguistics, I worked on an implementation for a Dutch HPSG Grammar in ALE. I encountered some problems which were hard/impossible(?) to solve with the existing situation and I wondered whether anyone has had the same difficulties, and if there are any answers. The problems had to do with the word order difficulties in Dutch (which was one of the reasons why I excluded sentence and verb modification from my grammar). In order to be able better to cope with the word order, I split up comps in lcomps and rcomps, for the Mittelfeld and the Nachfeld respectively (in the style of Bouma--Van Noord, "A lexicalist approach of the Dutch verbal complex" where there is a DIR feature). Some complements (participles and prepositional objects) can be placed more freely than nominal complements; the former are allowed to go both in rcomps and lcomps, the latter only in lcomps (in a subordinate clause). - In a main (V2) clause, they all come after the conjugated verb, except of course for the phrase that precedes the verb. - In the case of argument inheritance, only lcomps can be inherited; the rcomps of a complement need to be satisfied before it can be selected by a head. This strategy has the effect that verbal complements always follow their head. Constructions like in German "... dass er das buch lesen wird" are possible in Dutch, but they are somewhat quaint, so I excluded them. It would be possible though to incorporate them (allow selection with non-empty rcomps), since it only applies to a closed class of verbs. For the same word order reasons, schemata 1 and 2 were replaced by LP-specific variants of schema 3: - verb-initial clauses use the `normal' schema 3; - V2-clauses use schema 3 with extraction of one complement (i.e. the functionality of schema 1 goes to schema 6 (head filler) and that of schema 2 is taken over by schema 3). To make it more general, the valence and the nonlocal feature principle cooperate to pass on the extracted value without using lexical rules (in ALE the lexical rules are applied at compile time). The phrase structure rules now work as follows: if there is a verb, then it looks what complements it got. It is allowed that one of them is not present at the moment of application of the rule, and this one will `come out' of the valence principle and `go into' the nonlocal feature principle where the value is unioned with the inherited slash set. This treatment can make the insertion of an (often disambiguating) "er" in subject position if the subject is extracted, a lot easier (this is about sentences like (a) "Er staat een man in de straat." "There is a man in the street." (b) "... dat (er) een man in de straat staat." "... that (there) a man in the street is.") This is based on some of Gerrit Rentier's papers. WH-questions and relative clauses are handled in the same way (i.e. with this extraction). - verb final constructions are obtained with an unspecified list of complements. For the implementation, I used the solution that is presented in the ALE manual, viz. to instantiate the cats> list with a definite clause giving the list a finite number of elements (<5). This made the grammar less efficient (and especially so with empty categories), but the results were acceptable. (And it allowed me to treat cross serial dependencies.) [Perhaps it would have been a better solution to use a rule specific head-selecting definite clause ?] Up to there, everything was fine. The problem: I ran into a spurious ambiguity: in the case of certain verbs that allow two constructions in subordinate clauses which collapse in the main clause (due to linearisation effects): (1) ... dat hij een huis probeert te bouwen. <... that he a house tries to build.> (2) ... dat hij probeert een huis te bouwen. <... that he tries a house to build.> (3) Hij probeert een huis te bouwen. <He tries a house to build.> Those verbs (here "proberen" - to try) are in the ANS (the standard Dutch grammar) classified as NVG (non-obligatory group forming), which means that they are not obligatorily followed by the infinitive. In (1) it is followed by the infinitive, but not in (2). Both word orders collapse in a V2 clause: first the verb `goes' in front, and then the subject or the complements. Or in a diagram: (4) Hij[NP] probeert[V] een_huis[NP] te[MRK] bouwen[V] --------------VP-------------- ("proberen" has 1 argument, like in (2) - lcomps of "bouwen" satisfied) (5) Hij[NP] probeert[V] een_huis[NP] te[MRK] bouwen[V] -----NP----- -------V--------- (argument inheritance: "proberen" has 2 arguments, like in (1)) This effect seems unavoidable to me, unless I start tinkering with the rule sequences, which would make the grammar messy, to say the least. Anything I missed? Any solutions? Any comments? [ A more detailed description is on the Web: http://clwww.essex.ac.uk/~fouvry/thesis-aoti-english.html http://clwww.essex.ac.uk/~fouvry/thesis-aoti-nederlands.html ] -- Frederik Fouvry PhD student Department of Language and Linguistics University of Essex E-mail fouvry@essex.ac.uk Wivenhoe Park URL http://clwww.essex.ac.uk/~fouvry/ Colchester ESSEX CO4 3SQ Tel. +44 1206 87 20 91 United Kingdom Fax +44 1206 87 20 85
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:06 PST