cmp-lg/ Nov 1994

Transcription

1 Extractio i Dutc wit Lexical Rules Gerrit Retier Istitute for Laguage Tecology ad Artificial Itelligece Tilburg Uiversity, PO Box 90, 000 LE Tilburg, Te Neterlads retier@kub.l cmp-lg/900 Nov 99 Abstract Dutc argumet extractio is aalyzed i te versio of Head-drive Prase Structure Grammar wic is sketced i [Pollard&Sag(99)], capter 9. I tis versio of HPSG te existece of traces is deied ad istead extractio iformatio is itroduced by lexical rules operatig o te features of lexical eads. We preset suc lexical rules to accout for Dutc verb secod ad te igly idiosycratic properties of Dutc prepositio stradig, tus illustratig te descriptive power of suc lexical rules. We also geeralize te aalysis to accout for te beaviour of Dutc P 0 w.r.t. euter proous ad defie P 0 i te ierarcical lexico so tat P 0 is disallowed to locally gover ay proou wic deotes a euter referet. Dutc Verb Secod ad Prepositio Stradig Dutc displays a itriguig idiosycracy wit respect to prepositio stradig, as ca be read from te cotrast betwee (b) ad (c); () a. Aa To welke stictig scekt Beatrix et uis? wic foudatio doates Beatrix te ouse? To wic foudatio does Beatrix doate te ouse? b. * Wat scekt Beatrix et uis aa? Wat[R ] doates Beatrix te ouse to? Wat does Beatrix doate te ouse to? c. Waar scekt Beatrix et uis aa? Wat[R+] doates Beatrix te ouse to? Tis researc report is a sligtly corrected versio of [Retier(99b)]. Te relevat correctios cocer sematic tecicalities i Sect. ad a erroeous defiitio i a aalysis i Sect.. Autor was, over te relevat periods of time, sposored by EC projects ESPRIT P (PLUS) ad P (DANDELION), a travel grat from te Neterlads Orgaizatio for Scietific Researc (NWO) ad a travel grat from SHELL NEDERLAND. Taks for may useful suggestios ad for detailed commets o earlier versios of tis paper to Bart Geurts, Josée Heemskerk, Paola Moacesi, Jo Nerboe, Carl Pollard, Iva Sag, Cris Sijtsma, Wietske Sijtsma, Craig Tiersc, Margriet Verlide, attedats of CLIN III, Tilburg, te Neterlads, October 9 for commets o [Retier(99)], ad attedats of te First Iteratioal Worksop o HPSG i Columbus, Oio, USA, August 9 (were a earlier versio of tis paper was read) ad te audiece of KONVENS9 i Viea, Austria, September 9.

2 Frotig of a PP o-subject argumet is always possible wit te PP as a wole (pied pipig, (a)), modulo te usual islad costraits. Tis is just a istace of verb secod, wic we sall discuss i Sect.. Te iterestig idiosycracy owever is tat prepositio stradig is impossible wit a certai variat of te WH-proou, but grammatical wit aoter ((b) vs. (c)). Startig wit [va Riemsdijk(98)] tis cotrast ad similar cotrasts wit euter demostrative, relative ad clitic proous i Dutc ave bee explaied by te sytactically relevat absece vs. presece of te poeme /R/. I sectios ad we will give a lexicalist accout of te above ad related cotrasts. Te accout will be lexical i te sese tat it will itroduce o additioal mecaisms, priciples or prase structure rules ito HPSG but istead carefully defies te local ad olocal selectio properties of prepositios wile itroducig mior costraits o tree idepedetly motivated lexical rules. I Sect., owever, we tetatively suggest a additioal immediate domiace scema to accout for filler-like costituets i te Mittelfeld (cf. [Retier(99)]). Subjects, Complemets ad Dutc Clause Structure Istead of te feature SUBCAT, wic is put forward to list a ead s locally selected argumets i [Pollard&Sag(98)], we will adopt te divisio of argumets as subjects ad osubjects wic is motivated for Eglis i [Borsley(98)]. I Cap. 9 of [Pollard&Sag(99)] tis approac to local selectio of argumets is developed furter ad leads to te postulatio of te Valece Priciple. Te Valece Priciple refers to te valece features SUBJ ad COMPS troug F i te followig defiitio: () Valece Priciple [Pollard&Sag(99)], Cap. 9, pp.8 I a eaded prase, for eac valece feature F, te F value of te ead-daugter is te cocateatio of te prase s F value wit te list of SYNSEM values of te F-daugters value. Of course we wat a teory of valece to be uiversal. If we assume it for Eglis for te reasos give by [Borsley(98)] ad Cap. 9 of [Pollard&Sag(99)], te we sould be able to succesfully implemet it i our aalysis of Dutc as well. Furtermore, wit Dutc tis divisio allows for a iterestig aalysis of Cross Serial Depedecies, as discussed i [Retier(99a)]. Terefore we assume lexical etries for Dutc fiite verbs like scekt ( doates ) to look like (), were we also iclude te NONLOCAL features tat pass o extractio iformatio; te use of tese features will be extesively illustrated i te remaider of te paper. Te effect of te Valece Priciple o a eaded prasal sig tat is eaded by, e.g., te lexical sig i () is tat tis sig ca oly be regarded as complete or saturated if te lexical sig is combied wit te appropriate argumets; Specifically, te fact tat we list te least oblique argumet of a govered verb i a separate feature SUBJ, allows for a particularly ituitive aalysis of te fact tat, i Dutc cross serial costructios, te goverig verb assigs case to tat argumet of te govered verb, but ot to ay of te oter argumets of te govered verb. For discussio, cf. [Retier(99a)].

3 S NP[NOM] NP[ACC] PP[AAN] SUBJ i COMPS, i Beatrix et uis Scekt P[COMPS i] NP[ACC] aa aar () PHON scekt i SYNSEM LEX + LOCAL NONLOCAL HEAD verb VFORM FIN SUBJ NP[CASE NOM] i i COMPS NP[CASE ACC], PP[PFORM AAN] i INHERITED TO-BIND SLASH SLASH i fg i fg I te case of () te appropriate argumets are a omiative subject, a accusative object ad a PP-complemet wic is eaded by aa ( to ). If we assume a flat clause structure aalysis ([Pollard(fortc.)], [Nerboe(99)]) for Dutc, ad assume lexical sigs like (), te te followig immediate domiace statemets suffice to describe te fragmet we are cocered wit; () a. (Scema I) a [SUBJ i,comps i] prase wit daugters of sort ead-subjcomps-struc i wic te ead-daugter is a lexical sig b. (Scema II) a [COMPS i] prase wit daugters of sort ead-comps-struc i wic te ead-daugter is a lexical sig c. (Scema III) a prase of sort filler-ead-struc wit a filler-daugter tat as a LOCAL value toke-idetical to bot te INHERjSLASH ad te TO-BINDjSLASH value of te ead-daugter, were te ead-daugter is a fiite setece Togeter immediate domiace scemata I ad II, te demads made by te Valece Priciple ad te selectioal requiremets made by te lexical etry for scekt give rise to prase structure aalyses of Dutc yes/o-iterrogatives ad PP s as i te figure above. Here ead-subj-comps-struc idicates tat te daugters of te prase iclude a ead, a subject ad complemets, ot ecessarily i tat order; ead-comps-struc idicates te same, but witout te subject daugter. Cf. Cap. 9 of [Pollard&Sag(99)] for detailed discussio. I tis figure ad trougout te paper, recurrig i s idicate structure sarig, tat is toke-idetity of iformatio, as is commo usage i HPSG.

4 Verb Secod as Argumet Extractio witout Traces I Cap. 9 of [Pollard&Sag(99)] a teory of extractio wic does ot employ ay otio of traces or empty categories is suggested. Tis teory is furter motivated by [Sag&Fodor(99)], wo argue covicigly tat o teory exteral evidece for traces exists. Also, tey poit out several advatages of suc a lexically based teory of extractio, for istace w.r.t. isolated idiosycracies tat ivite a aalysis as lexical exceptios. S NP[NOM] SUBJ i COMPS i INHERjSLASH TO-BjSLASH o o Beatrix NP[ACC] PP[AAN] SUBJ i COMPS, i INHERjSLASH scekt o et uis P[COMPS i] NP[ACC] aa aar We argue tat suc a teory, togeter wit te assumptios from Sect., allows for a elegat traceless versio of te aalysis of verb secod i Germaic wic is discussed i [Pollard(fortc.)]. Followig Pollard, we assume tat i HPSG, te verb secod peomeo sould be modeled as extractio from a ead-iitial (flat) clause. I te trace-less model of extractio, we ca te describe a setece wit te subject i te Vorfeld by assumig tat te Subject Extractio Lexical Rule for Dutc sould read as follows; () Subject Extractio Lexical Rule Dutc " LOCjSUBJ i NONLOCjINHERjSLASH f g # ) " LOCjSUBJ i NONLOCjINHERjSLASH f g A declarative Dutc setece wit a umarked teme will te ave a structure as i te figure above, were te top is licesed by ID-Scema III (cf. (c)) ad te middle is licesed by ID-Scema II (cf. (b)). Te SELR, listed i (), as applied to te etry i () to produce te omopoous etry for scekt wic is te ead i te tree above. Ay value of te feature INHER (for INHERITED ) is subject to te Nolocal Feature Priciple: #

5 () Nolocal Feature Priciple [Pollard&Sag(99)], Cap., pp. For eac NONLOCAL feature, te INHER value o te moter is te uio of te INHER values o te daugters mius te TO-BIND value o te ead-daugter Te effect of te NFP is tat te olocal selectio iformatio percolates up i te structure util it ca be associated wit a appropriate filler. Tis appropriateess is forced by te demad of toke-idetity metioed i te immediate domiace scema wic combies te filler wit te clause wic it is extracted from, tat is, Scema III. Tis approac also allows for a traceless aalysis of o-subject argumets i te Vorfeld; we will discuss tis i its relatio to Dutc prepositio stradig. A Traceless Accout of Dutc Prepositio Stradig As discussed i te itroductio, Dutc displays a iterestig idiosycracy wit respect to extractio from prepositioal prases, illustrated i () wit te cotrast betwee (b) ad (c). Suc cotrasts are usually attributed to te presece vs. te absece of te poeme /R/ i a sytactically relevat way. If we idicate tis property as [R+] ad [R ] o te appropriate lexical items, we migt accout for tis idiosycracy by adoptig a modified versio of te Complemet Extractio Lexical Rule (cf. Cap. 9, [Pollard&Sag(99)]) for Dutc prepositioal eads; () Prepositio Complemet Extractio Lexical Rule First Versio LOCjHEAD prep LOCjHEAD prep LOCjCOMPS [R ] i ) LOCjCOMPS i NONLOCjINHERjSLASH f g NONLOCjINHERjSLASH f [R+] g Tis lexical rule is restricted by te occurece of [R+] i suc a way tat te itroduced extractio iformatio o a PP will always cocer a extracted elemet wic is [R+]. Troug () we ca accout for te cotrast betwee (c) ad (b) i a straigtforward maer; we merely ave to state i te lexico tat waar is [R+] ad tat wat is [R ]. Te aalysis of (c) is give i te figure ext page (were P 0 is derived troug te PCELR). Note furtermore tat wit () we propose a substatial extesio to te traceless teory of extractio troug lexical rules. Tis is so sice ere we place idiosycratic restrictios ot o te ead wic liceses te ubouded depedecy, but o te elemet wic is extracted itself. To allow for (8a), ad extractio of o-subject argumets of verbs i geeral, Dutc must ave a separate Complemet Extractio Lexical Rule for verbs. As we ca see from te cotrast wit (8b), verbs ca oly be olocally related to proous wic are [R ]: For a extesive discussio of data, cf. a.o. [va Riemsdijk(98)]. We assume tat i te lexico, all NP argumets wic are o COMPS are [R ] accoutig for Beatrix waardeert dat/*daar. Furter restrictios o NP argumets come ito play wit P 0, cf. Sect.. Tat is, we impose oter restrictios ta te sytactic ad sematic selectio restrictios wic are stadardly imposed by te lexical ead wic selects te argumet. N.B.; Te uderscores i (8) ad trougout tis paper are ot to be uderstood as traces or ay oter kid of empty categories. Te uderscores idicate, for expository reasos, positios were a argumet could ave bee locally realized, but is t locally realized because it is istead olocally realized elsewere.

6 S NP[R+] SUBJ i COMPS i INHERjSLASH TO-BjSLASH o o Waar NP[NOM] NP[ACC] P COMPS i INHERjSLASH o SUBJ i COMPS, i Beatrix et uis aa INHERjSLASH fg scekt (8) a. Dat zal Peggy waarscijlijk Tat[R ] will Peggy probably Peggy will probably appreciate tat b. * Daar zal Peggy waarscijlijk Tat[R+] will Peggy probably c. Daar zal Peggy waarscijlijk [op Tat[R+] will Peggy probably o Peggy will probably cout o tat waardere appreciate waardere appreciate ] rekee cout Topicalizatio of a [R+] demostrative proou ca obviously oly be licesed by P 0, cf. (8c). Terefore, we propose tat i Dutc, ay lexical rule wic itroduces a ubouded depedecy o a verb will costrai te itroduced elemet to be [R ]; tis is reflected i te Dutc rule for verb complemet extractio i (9). (9) Verb Complemet Extractio Lexical Rule LOCjHEAD verb LOCjCOMPS : : :, [R ],: : : i NONLOCjINHERjSLASH f g ) LOCjHEAD verb LOCjCOMPS : : : : : : i NONLOCjINHERjSLASH f [R ] g Te same costrait sould ad ca be built ito te Dutc SELR ().

7 A Geeralizatio of te Aalysis O our accout so far, all examples i (0) sould be grammatical; (0) a. Hij eeft op et slecte weer gereked He as o te bad weater couted b. Hij eeft op em/ aar gereked He as o im[r ]/ er[r ] couted c. * Hij eeft op et/ daar/ dat gereked He as o it[r ]/ tat[r+]/ tat[r ] couted However, te facts i (0) seem to geeralize to te observatio tat except for full NPs, oly proous wit male or female geder ca be locally govered by P 0. I te ugrammatical costructios give i (0c) te proous deote referets of euter geder. I [Pollard&Sag(99)], it is assumed tat o-predicative P 0 is sematically vacuous. Cosequetly, te CONTENT of a projectio of P 0 is structure sared wit te CONTENTvalue of its NP-argumet. 8 From (0), it seems te tat a Dutc prepositio wit a CONTENT-value of te type of a proou wic as NEUTER as te value of GENDER caot gover tat proou locally. Istead, suc a P 0 sould always itroduce a extractio ad tus select its argumet olocally istead of locally. 9 Here we propose to capture tis geeralizatio by, firstly, imposig a egative costrait o te sematics of o-predicative P 0 i te ierarcical lexico for Dutc: 0 () PHON SYNSEM X i LOCAL NONLOCAL CATEGORY HEAD prep PFORM X COMPS NP[ACC][R ]: i CONTENT NOT INDEX ppro INHERITED TO-BIND SLASH fg SLASH fg i i i GENDER NEUTER i Witout workig out te details, we assume tat we ave a ierarcical lexico wic elimiates redudacies troug ieritace, as suggested i Cap. 8 of [Pollard&Sag(98)]. We orgaize te lexico so tat all lexical etries for P 0 are as i (), so tat tey ca licese ay complemet locally as a value of COMPS if it is a accusative NP wic is [R ] ad ot a euter proou. Suc etries are appropriate for ay Dutc o-predicative P 0 wic locally govers a complemet wic is euter but ot proomial (cf. (0a)) or proomial but ot euter 8 Te feature CONTENT of ay omial object rages over om-obj or a subtype of it; oe suc subtype is ppro. Cf. [Pollard&Sag(99)], Cap., pp.-, for discussio of furter details. 9 Suc olocal selectio of P 0 is always for a euter proou; cf. te examples i (c), (8c), (b) ad, o our tetative aalysis preseted i Sect., also i (). 0 I (), () ad (), te colo, :, abbreviates te pat SYNSEMjLOCjCONTENT.

8 (cf. 0b)). Te egative costrait o te value of te CONTENT feature correctly excludes ugrammatical costructios ivolvig euter proou objects, like (0c). Furtermore, etries for P 0 like () correspod to a PP wic must locally select its argumet; o P 0 wit sematics like () will be allowed to appear i a slased form. Tis is guarateed because, i te lexical rule approac to extractio, categories oly appear slased if a lexical rule produces tem as suc. Sice ay PP licesed as a istace of () is specified as empty for te feature SLASH, o PP licesed by a istace of () will be olocally related to its complemet. Te questio as to ow we sould licese grammatical cases of prepositio stradig is of course still aswered troug te Prepositio Complemet Extractio Lexical Rule. But we sould cage it i suc a maer tat it will erase te egative costrait o te value of CONT, tus givig rise to te desired results; () Prepositio Complemet Extractio Lexical Rule Fial HEAD prep COMPS [R ]: CONT NOTppro[INDEX[GEND NEUT]] INHERjSLASH f g i ) HEAD prep COMPS i CONT ppro[index[gend NEUT]] INHERjSLASH f [R+]: Applicatio of tis PCELR to etries like i () gives rise to etries like (): () PHON X i SYNSEM LOCAL NONLOCAL CATEGORY CONTENT INHERITED TO-BIND HEAD prep PFORM X COMPS i i INDEX GENDER NEUTER ppro SLASH SLASH fg i i NP[ACC][R+]: o Te etries wic are like () te are assumed to take part i all grammatical cases of prepositio stradig (cf., e.g., te tree for (c); also, (8c), (b) ad ()). Extesios to te Aalysis Similar lexical rules ca explai obviously related cotrasts betwee two variats of te Dutc relative proous (cf. ()); te oly differece will be tat tis PCELR sould make referece to te olocal feature REL (Cap., [Pollard&Sag(99)]), ad ot to SLASH. () a. * Het slecte weer wat [Peggy [op Te bad weater wat[r ] Peggy o Te spell of rai tat Peggy as couted o b. Het slecte weer waar [Peggy [op Te bad weater wat[r+] Peggy o 8 ] eeft as ] eeft as gereked] couted gereked] couted g

9 Furtermore, te Nolocal Feature Priciple, we examied closely, allows fillers to be sisters to te argumets from wic tey are extracted. Te relevat fact is, tat bot te filler ad te argumet(s) from wic it is extracted sould be allowed as sisters of te ead-daugter, by some additioal immediate domiace scema; () (Scema IIIb) (Additioal) a prase of sort filler-(subj)-comps-ead-struc wit a filler-daugter tat as a LOC value toke-idetical to te TO-BINDjSLASH value of te ead-daugter ad te INHERjSLASH value of some COMPS-daugter, were te ead-daugter is a fiite lexical verb ad te filler is [R+] Tis allows for structures were fillers ca be at te same level i te tree as te eads to wic tey are olocally related. If we allow for suc prasal structures, te te versio of te PCELR i () will also accout for grammatical costructios i Dutc were te demostrative proou daar or te clitic proou prepositioal object er appear i te Mittelfeld, ot te Vorfeld; () a. Peggy oeft daar / * dat iet [op Peggy as tat[r+]/ tat[r ] ot o Peggy sould t cout o tat b. Beatrix scekt er / * et gee uis [aa Beatrix doates it[r+]/ it[r ] o ouse to Beatrix does t doate a ouse to it ] te rekee to cout Tis aalysis, toug stipulative, gives a atural accout of te relatio betwee te prepositio ad its proou object. Firstly, it explais te discotiuity betwee te P 0 ad its object i (), cf. ext page figure. Secodly, it is cosistet wit te geeralizatio tat P 0 caot locally gover ay euter proou, cf. (0). We claim tis is a appealig advatage of te above aalysis of Dutc prepositio stradig sice it allows for a completely uified accout of te distributio of R-proous i Dutc. Refereces [Borsley(98)] R.D. Borsley Subjects ad Complemets i HPSG, CSLI Report 0, Staford Uiversity, USA [Nerboe(99)] J. Nerboe Partial Verb Prases ad Spurious Ambiguities, i: J. Nerboe, K. Netter ad C. Pollard, (eds.), Germa i HPSG, CSLI Lecture Notes, Cicago Uiversity Press, USA More closely te i [Retier(99b)]; i [Retier(99b)], we icorrectly assumed tat te NFP sould be revised i order to accout for data suc as te data i (). It is, of course, a goal of te teory to reduce ad ot icrease te umber of immediate domiace scemata. Peraps it is tecically feasible to costrue IIIb i () as a prasal subtype wic multiply ierits from te prasal types I, II ad III i (). Furter discussio of oter empirical argumets for suc a uificatio of te accout: [Retier(99)]. ] 9

10 S NP[NOM] SUBJ i COMPS i INHjSLASH TOBjSLASH o o Beatrix SUBJ i COMPS INHjSLASH TOBjSLASH, i o o NP[R+] NP[ACC] P COMPS i INHjSLASH o scekt er gee uis aa [Pollard(fortc.)] C. Pollard O Head No-Movemet, i: But, H. & va Horck, A. (eds.), Proceedigs of te Symposium o Discotiuous Costituecy, Mouto-de Gruyter, Germay [Pollard&Sag(98)] C. Pollard, I. A. Sag Iformatio-Based Sytax ad Sematics; Fudametals CSLI Lecture Notes, Staford Uiversity, USA [Pollard&Sag(99)] C. Pollard, I. A. Sag Head-drive Prase Structure Grammar, Uiversity of Cicago Press ad CSLI Publicatios, USA [Retier(99)] G. Retier Dutc Object Clitics, Prepositio Stradig ad Across-te- Board Extractio, i: Sijtsma, W. & Zweekorst, O. (eds.), Papers from Computatioal Liguistics i te Neterlads (CLIN) III, 99, Tilburg, te Neterlads [Retier(99a)] G. Retier A Lexicalist Approac to Dutc Cross Serial Depedecies, i: Proceedigs of te 0t Regioal Meetig of te Cicago Liguistic Society, CLS, Cicago, USA. Also: ITK Researc Report 0, Tilburg Uiversity, te Neterlads [Retier(99b)] G. Retier A Lexicalist Approac to Dutc Prepositio Stradig, i: H. Trost, (ed.), Proceedigs of KONVENS9, Lecture Notes i Computer Sciece, Spriger-Verlag, Heidelberg, Germay [va Riemsdijk(98)] H. va Riemsdijk A Case Study i Sytactic Markedess Foris, Dordrect, te Neterlads [Sag&Fodor(99)] I.A. Sag, J.D. Fodor, Extractio witout Traces, i: Proceedigs of t West Coast Coferece o Formal Liguistics, USA 0