Phrasal subconstructions
A constructionalist grammar design, exemplified with Norwegian and English
Thesis for the degree of Philosophiae Doctor Trondheim, June 2009
Norwegian University of Science and Technology Faculty of Arts
Department of Language and Communication Studies
Petter Haugereid
Norwegian University of Science and Technology Thesis for the degree of Philosophiae Doctor Faculty Arts
Department of Language and Communication Studies
© Petter Haugereid
ISBN 978-82-471-1629-6 (printed ver.) ISBN 978-82-471-1630-2 (electronic ver.) ISSN 1503-8181
Doctoral theses at NTNU, 2009:122
Printed by NTNU-trykk
1 Introdution 1
1.1 Theoretial assumptions . . . 2
1.2 Five subonstrutions. . . 4
1.3 A onstrution-onstrainingmehanism . . . 11
1.4 Syntati strutures . . . 15
1.5 Layout of the thesis . . . 19
I Argument Struture 21 2 HPSG, LFG, CG, and GB/Minimalism 23 2.1 Introdution . . . 23
2.1.1 Unergative and unausativeverbs . . . 24
2.1.2 Other alternations . . . 25
2.1.3 Voie . . . 27
2.2 HPSG . . . 28
2.3 LFG and the LexialMapping Theory . . . 31
2.4 Constrution Grammar . . . 37
2.5 GB/Minimalism . . . 39
2.5.1 Passive inGB/Minimalism . . . 39
2.5.2 Hale and Keyser's theory . . . 40
2.5.3 First Phase Syntax . . . 44
2.5.4 Minimalism -Borer's neo-onstrutionistapproah. . . 47
2.6 Comparison . . . 49
2.7 Some methodologial onsiderations . . . 50
2.7.1 Remarks to HPSG . . . 51
2.7.2 Remarks to LFG/LMT . . . 51
2.7.3 Remarks to ConstrutionGrammar . . . 54
2.7.4 Remarks to Haleand Keyser'stheory . . . 55
2.7.5 Remarks to First Phase Syntax . . . 56
2.7.6 Remarks to Borer's neo-onstrutionalist approah . . . 57
2.8 Summary . . . 57
3 Subonstrutions 59 3.1 Some syntati tests . . . 59
3.2 Five subonstrutions. . . 62
3.2.1 ARG1 . . . 62
3.2.2 ARG2 . . . 63
3.2.3 ARG3 . . . 65
3.2.4 ARG4 . . . 66
3.2.5 ARG5 . . . 67
3.3 Levin's EnglishVerb Classes and Alternations . . . 68
3.3.1 The Causative/InhoativeAlternation . . . 68
3.3.2 The InduedAtion Alternation . . . 68
3.3.3 The Substane/Soure Alternation . . . 69
3.3.4 Intransitive/Transitive Alternations . . . 70
3.3.5 Conative and Preposition Drop Alternations . . . 72
3.3.6 Dative and Benefative Alternations . . . 72
3.3.7 Loativeand similar alternations . . . 73
3.3.8 Delimiter Alternations . . . 74
3.3.9 Other Alternations . . . 76
3.4 BRRs and semantirepresentations . . . 78
3.4.1 MRS . . . 79
3.4.2 BRR/RMRS . . . 80
3.5 A hierarhy of subonstrutions . . . 82
3.6 Summary . . . 85
4 Valene 87 4.1 Valene inHPSG . . . 88
4.2.1 The GrammarMatrix . . . 91
4.2.2 Norsyg - some data . . . 93
4.3 The linkingtypes . . . 94
4.3.1 Four valene features . . . 94
4.3.2 A hierarhy of linkingtypes . . . 95
4.3.3 Paking of argument frames . . . 96
4.3.4 Introdutory remarks onthe ompositionof subonstrutions. . 98
4.4 Lexial types inNorsyg . . . 100
4.5 Expansion of the lexion . . . 105
4.5.1 Adaptation of Norsk Ordbank . . . 105
4.5.2 Unknown words . . . 107
4.5.3 Lexion aquisition . . . 108
4.6 Constrution-onstraining mehanism vs. lexialism . . . 112
4.7 Comparison with the RASP system . . . 117
4.8 Other Norwegian omputationalresoures . . . 120
4.8.1 TROLL . . . 120
4.8.2 NorKompLeks . . . 122
4.8.3 NorSoure . . . 124
4.8.4 NorGram . . . 125
4.9 Summary . . . 126
II The realization of argument struture in the syntax 129 5 Methodology 131 5.1 Preliminary analyses . . . 131
5.2 Some remarks onsyntati strutures . . . 137
5.2.1 Introdutory remarks ontree strutures. . . 137
5.2.2 Linguisti onsiderations . . . 141
5.2.3 Cognitiveonsiderations . . . 144
5.2.4 Computationalonsiderations . . . 145
5.3 Summary . . . 148
6 Basi syntati strutures 151
6.2 The ller rule . . . 155
6.3 The fore rules . . . 156
6.4 Some simple analyses . . . 157
6.4.1 Analysis of atransitive sentene . . . 159
6.4.2 Analysis of aresultativesentene . . . 163
6.5 The merge rule . . . 168
6.6 Subordinatelauses and relative lauses . . . 173
6.6.1 Subordinatelauses . . . 174
6.6.2 Relative lauses . . . 180
6.7 Innitivallauses and smalllauses . . . 183
6.7.1 Unexpressed subjets . . . 183
6.7.2 Analyses of innitivallauses and small lauses . . . 184
6.7.3 Raising and ontrol . . . 191
6.7.4 Remarks on raising . . . 194
6.8 The modierrules . . . 197
6.9 Long distane dependenies . . . 202
6.9.1 The trae approah . . . 202
6.9.2 Reetion of extrationpath . . . 202
6.9.3 The lexialapproah . . . 204
6.9.4 Some problems . . . 204
6.9.5 The approah taken inNorsyg . . . 206
6.10 Summary . . . 207
7 Passive and Presentation 209 7.1 Passive . . . 209
7.1.1 Data . . . 209
7.1.2 The passive types . . . 211
7.1.3 Analysis . . . 213
7.2 The presentational onstrution . . . 216
7.2.1 Some data . . . 216
7.2.2 The presentationalrules . . . 217
8 Coordination 219
8.1 Coordination of VPs . . . 219
8.1.1 Data . . . 219
8.1.2 Analysis . . . 220
8.2 Coordination of Vs . . . 223
8.2.1 Data . . . 223
8.2.2 Analysis . . . 225
8.3 Ellipsis . . . 227
8.4 Pseudo-oordination . . . 227
8.4.1 Sub-oordination . . . 228
8.4.2 The Empty ObjetConstrution . . . 230
8.4.3 Analysis . . . 231
8.5 Summary . . . 238
9 Comparison with GB 241 9.1 GB as presented inCarnie 2007 . . . 242
9.2 A GB analysis based on Norwegiandata . . . 246
9.3 Three positions for verbs . . . 249
9.3.1 The position orrespondingto C. . . 250
9.3.2 The position orrespondingto T. . . 252
9.3.3 The position orrespondingto V. . . 253
9.4 An aount of basi lause struture inEnglish. . . 254
9.4.1 Bloking main verbs fromappearing beforethe subjet . . . 255
9.4.2 Assuming anempty omplementizer. . . 256
9.5 Dierene between Norsyg and GB . . . 259
9.5.1 Dierene inparsing strategy . . . 260
9.5.2 Innitivallauses and `skewed' syntati-semantirelations . . . 262
9.6 Summary . . . 264
10 Sentene adverbials 265 10.1 Data . . . 266
10.1.1 Sentene adverbials in dierent lause types . . . 267
10.1.2 Sentene adverbials and the arguments . . . 268
10.3 The approah taken inNorsyg . . . 274
10.3.1 Analysis of sentene adverbials indierentlause types . . . 275
10.3.2 Analysis of sentene adverbials and the arguments . . . 276
10.4 Summary . . . 279
11 Conlusion 281 A Norsyg 287 A.1 Download . . . 287
A.2 Short desription . . . 288
A.2.1 Composing argumentstruture in the syntax . . . 288
A.2.2 Left-branhing tree strutures . . . 289
A.3 Data . . . 292
A.4 Coverage . . . 293
A.5 NorKompLeks test sentenes . . . 294
A.6 Tehnial details about ase and linking . . . 297
A.6.1 The linkingmehanism . . . 297
A.6.2 Case . . . 301
B Demo grammars for English and German 305 B.1 English demo grammar . . . 305
B.2 German demo grammar . . . 308
C Example sentenes of the thesis 315 C.1 Norwegian example sentenes . . . 315
C.2 English example sentenes . . . 322
D BRRs of example trees 331 D.1 BRRs of example trees in Chapter 4 . . . 331
D.2 BRRs of example trees in Chapter 5 . . . 332
D.3 BRRs of example trees in Chapter 6 . . . 335
D.4 BRRs of example trees in Chapter 7 . . . 342
D.5 BRRs of example trees in Chapter 9 . . . 344
D.6 BRRs of example trees in Chapter 10 . . . 347
D.7 BRRs of example trees in Appendix B . . . 351
Aknowledgements
Doing a PhD an be ompared to making a trip to a mountain top in the relatively
at landsapes of Lierne. You start out in high spirit, seeing the mountain top in the
distane. You walkfor a longwhile, through moss and belts offorest until you stop to
makea break. Youstill see the mountains ina distane. Aftersome more walkingyou
start approahing the foot of the mountain, and you don't see the top any more. By
thenyou are gettingalittletired, but youthinkthat the top annotbe very far away,
so you limb energetially, only to nd out that what you thought was the top turns
out to be a hill next to the real mountain. At this point you have to deide whether
toreturn orwhether you want toget tothe realtop. Soyourest fora short while, ll
your bottlewith water fromastream,and deidetolimbthe lastroky partuntilyou
nallyreah the highest top. It turns out, as you get to the top and an see far into
the distane, that the limbing ofthis roky part is not the biggesteort. It is getting
to and from the foot of the mountain that is the hallenge. And for a PhD student,
thisis the part whereyouget help andsupportthat you needfrom olleagues,friends,
and family.
My supervisor during my period as a PhD student has been Lars Hellan. With
patiene and insight, he has helped me to understand what it really is that I wanted
todo. In additiontobeing a mentor, he has inluded meinseveral of his projetsand
introduedme toa great network of peopleworking with grammarengineering.
Through the projets DeepThought, SanMatrix and espeially LOGON, whih
employed me as aPhD student, I got the hane to present and disuss my ideas to a
large group of people, inluding Lars Ahrenberg, Dorothee Beermann, Emily Bender,
Felix Bildhauer, John Carrol, Berthold Crysmann, Ann Copestake, Lua Dini, Helge
Dyvik, Andreas Eisele, Liv Ellingsen, Dan Flikinger, Dario Gonella, Hannes Hirzel,
Per Anker Jensen, Lars Johnsen, Valia Kordoni, Daniela Kurz, Gunn Inger Lyse, Jan
Tore Lønning, Giampaulo Mazzini, Paul Meurer, Stefan Müller, Torbjørn Nordgård,
Stephan Oepen, Vitoria Rosèn, Melanie Siegel, Anders Søgaard, Jesse Tseng, Hans
Uszkoreit, Erik Velldal, and Ben Waldron. Thanks to all of you! Dan Flikinger has
inpartiular been agreat inspiration duringmy periodas aPhD student. Whenever I
havepresented myworkwithhiminthe audiene,hehas,inhisowngentleway,helped
metosee howmyanalyses ould beimproved.
Jørn Almberg, Getahun Amare, Nana Amfo, Kaja Borthen, Jonathan Brindle, Heidi
Brøseth,RositsaDekova,TszvetanaDimitrova,Arne-KjellFoldvik,AnneSandøFrank,
ThorsteinFretheim,JanikeFurberg,SnefridHolm,OlaHuseth,JaquesKoreman,Ota
Ogie,ReinOveSikveland,SiriSimonsen,JosteinVen,andTesfayeWondwosen. During
my stayin Saarbrüken February-August2004 I enjoyed the ompany of Peter Dienes,
Frederik Fouvry, Shravan Vasishth, Yi Zhang and Zhiping Zheng. Thanks to Martha
LarsonforinvitingmetoBonn(andmakingmeaware ofEmptyObjetConstrutions),
thankstoBertholdCrysmann,whohelpedmetomotivatemyleft-branhingstrutures,
andspeialthankstoValiaKordoniandHansUszkoreitforaommodatingmeatCoLi.
I espeially want to thank Emily Bender, Kaja Borthen, Anders Søgaard, Stephan
Oepen, and Ben Waldron for reading drafts of the thesis and giving very valuable
feedbak. Andalsothanks toDorothee Beermann,AnnCopestake,Stefan Müller,who
intheir initialreport provided mewith many helpful omments onmy thesis.
Finally, I would like to thank my family for always believing in me, and a speial
thanks tomy wife, Lilian,for her support and faith.
Introdution
Current omputational grammars designed within the HPSG and LFG frameworks
suer from an inreasing amount of analyses of sentenes parsed, and inreasing
proessingtime, assentene lengthextends beyond thatof8-10 words. Suhgrammars
do not purport to reet the psyhologial reality of what happens in sentene
proessing, and so far, no theory adequately overs this area. I nevertheless feel it
asalegitimateonern thatthe ratherexplosiveproessingdemands witnessed insuh
grammarsbear nointuitivesimilaritytowhat happens when weatually usesentenes
of normal length (whih may well be 20-30 words). Part of the disrepany an be
attributedtopragmatis: muhof the proessing loadhingesonsubstantiveambiguity
ofthewordsused,andinatuallanguageuse,wenormallyhavenoproblemdetermining
the relevant meaning of any lexial item uniquely. The aount of this belongs to
theoriesof disourseandpragmatis,and shouldnotaet thedesignofomputational
grammars,whihdeal with modulesof word ombinatorisatsentene level. However,
even with this aspet sorted away, proessing demands remainhaving to dowith non-
loality of information, manifesting itself in multiple lexial entries even when no real
ambiguity is in question, and umbersome strategies and massive hypothesis-building
inparsing.
Inthisthesis,Itry,withdeparturepointinformalismsasalludedtoabove,todene
designsoflexionbuildingand syntatianalysiswhihwillreduetheproessingloads
ofaparsing mehanism signiantly. I buildagrammarof Norwegiantoillustrateand
verify my proposals.
This grammar model may seem unorthodox in many ways, but in presenting it, I
provide evidene and motivation that would be relevant in any standard analysis. No
appeal to psyhologial reality is made throughout, exept one partiular paragraph
whereIrelatetotheissue. Thus, themodelpresentedistobeevaluatedasanystandard
analysis and implementation should be; only, the reader may bear in mind that what
motivates the various sub-proposals being olleted into this partiular whole, is the
intuitionmentioned.
1.1 Theoretial assumptions
One of the dierenes between Constrution Grammar (CG) on the one hand and
lexialist frameworks like HPSG and LFG on the other, is that in the analysis of
verbal onstrutions, the former posit onstrutionalframes as `primitive' entities into
whih the individual verbs will aommodate their semantis, whereas in the latter
frameworks,theorrespondingtypeofentityisoftenreferredtoas`argumentstruture',
and isassumed to be propagatedintothe grammarthrough the speiations(`lexial
frames', or `subat restritions') of the individual verbs. In the analyti pratie in
suh grammars, these lexial frames are distinguished as `lexial types' or `maros'
and dened at an abstrat level, and only in turn assoiated with the individual
verbs; hene it might be questioned whether the dierene originally mentioned is of
mainlyrhetorialsignianeratherthanrepresenting adiereneininsightsaboutthe
interplaybetweengrammarandthelexion. Inthepresentthesis,Iwilltrytoshowthat
thediereneanindeedbemodelledinsuhaway astoprovideinterestinglydierent
designs of grammar. I will do this using the overall arhiteture of HPSG grammars,
but inside of this arhiteture, develop a mehanism by means of whih the over-all
grammatial onguration in whih a verb ours, rather than its predened lexial
frame, is what induesits argument struture. I will show that this design provides a
more eient parsing grammar than one using the `lexialist' design, and argue that
alsoononeptual and empirialgrounds, this designis advantageous.
In this enterprise, the grammarengineering aspet is the most important one, and
is the area where I hope to be ontributing something new by this thesis. However,
the modelIdevelopan befully appreiatedonlyonthe bakground of my theoretial
views of grammar.
My theoretial view of grammar makes a sharp distintion between `form' and
validatedaspets of what is alled`syntax'. Grammar,in my view, is onstituted only
by these omponents, exluding semantis; as far as syntax is onerned, I thereby
standverymuh onthe side of`autonomous syntax',maintainedby Chomsky allsine
Chomsky(1957). Toavoidonfusionwith moreinlusiveoneptionsof`syntax'found
inthe literature, I willrefer tomy notionas one of `strit syntax' when neessary.
Myviewof the Lexionasonneted toagrammaris thatitshouldhighlightthose
properties or parameters whih are highlighted in the grammar, and only subsidiarily
exposeotherpropertiesoflexialitems(thusquiteunlikeanenylopedia,forinstane).
It follows that by my view of Grammar, in the lexion, only those properties whih
reet parameters of morphology and strit syntax should be represented. Valene
properties of verbs are in my view mostly a reetion of their meaning, and therefore
nota properaspet of grammar: `argumentstruture' isthus not partof stritsyntax,
and valene requirements should not be part of verb entries in the Lexion.
However, I reognize that for most parsing grammars, a omponent of `valene' or
`argument struture' may be desirable: a parsinggrammar is, in many respets, more
a `performane' than a `ompetene' onstrut, and thereby ombining omponents
whih on a strit view should be kept apart. To the extent that `argument struture'
ought to be represented in the verb lexion of a parsing grammar and reeted in
the parsing mehanisms, I want to do that in suh a way that in a lexial entry,
this type of information is easily detahable, almost to be regarded as an `add-on'
property. This `add-on' nature of argument struture speiation is what models my
onstrutional view of grammar, in that what `adds' the speiation in question is
information provided by the environment of the verb, i.e., the onstrution in whih
the verb ours.
The parameters of speiation onstituting argument struture are of the same
typeasthoseunderlyingthe `GrammatialRelations'ofLFG,andrelationalprimitives
ofRelationalGrammar-see Setion1.2below-and donot involvesemantiproperties
suhas`roles'ofpartiipantsand thelike. Sinetheriterialbasis fortheGrammatial
Relations are onstrutional environments, my formal term for grammatial relations
is subonstrutions. Subonstrutions are realized by morpho-syntati signs suh as
syntati rules, inetions and funtion words.
My avoidane of semanti assumptions in syntax also has as a onsequene that I
omit the more standardlyassumed levels of onstituent struture representation (suh
sineI believe thatthe strutures proposedtoa largeextentreet assumptionsabout
`logial form'. Thus, my assignment of `Grammatial Relations' to a string will be
based mostly onlinear order, and not supposing any `onstituent struture bakbone'
previouslyassigned, asin standard LFG and HPSGgrammars.
The grammar implementation I am providing is alled Norsyg, 1
whih has been
developed sine 2002. The grammar is a typed feature struture grammar, and it is
implemented with the LKB system, whih is a standard software for implementing
typed feature struture grammars, typially HPSG grammars. I employ the over-all
arhiteture of the `HPSG Grammar Matrix', however only up to the point where
the `onstrutional' design is dened. At this point, what populates the mehanisms
representing semantisinastandardHPSG/MatrixgrammarsuhasMRS(seeSetion
3.4),isadisplayof GrammatialRelations,and thus, notionally,moreonapar withan
LFGf-struture rather than with anHPSG semanti struture. The feature geometry
employed is similar towhat is used in the HPSGliterature, but a new mehanism for
assigning and onstraining the expeted argument frames of verbs, involving a type
hierarhy of onstrution and subonstrutiontypes, will be presented.
1.2 Five subonstrutions
A onstrution serves as a skeleton that open lass lexial items t into. On the view
outlined above, the `argument struture' of an open lass lexial item is projeted
from the onstrution it ours in. This grammatial onguration is a onstellation
of funtional signs like inetions, funtion words (i.e., `strit syntax') and (more
abstratly)rules.
2
Inordertogettherelationbetweenaonstrutionandtheindividual
funtional signs that together express the onstrution, I assume that a onstrution
an be deomposed intosubonstrutions.
3
Asantiipatedabove,asubonstrutionisloselytiedtothenotionof`Grammatial
Relation'. AGrammatialRelationisalwaysrealizedthrough asyntati onstellation
1
SeeAppendixA.
2
InthisthesisI makeadistintionbetweenwhatI referto asfuntionalsigns, namelyinetions,
losedlasslexialitems,andsyntatirulesontheonehand,andopen lass lexial items,whihare
uninetedadjetives,nounsandverbs.
3
Sine I assume that subonstrutions are expressed by what I refer to as funtional signs (see
footnote 2), I sometimesrefer to subonstrutions asphrasal subonstrutions, in order to separate
themfrom what isreferredto aslexial onstrutions(seeSag etal. (2003,Chapter 16),and Müller
for instane, `subjet-of' is realizedthrough aonstellation depitable as
(1)
S
NP VP
inalanguagelikeNorwegian,andsimilarlyforotherfuntions. Aonstellationlikethat
in(1),whihmaybealledaloalsubtree,willhere bereferredtoasasubonstrution,
anda GRwillbe seen asorresponding tothe set of subonstrutionswhihrealize it.
SuhaviewonGRsrelativetorealizingonstellationsissimilartothewayinwhihLFG
orrelates GRs with C-struture onstellations, through, in the PS-rules, annotating
these onstellations for the GRs they indue. For instane, for the onstellation (1),
the PS-rule in an LFG grammar would provide the following annotation stating that
the onstellationrealizes the `subjet-of' GR: (2)
(2)
S
NP VP
↑
SUBJ =↓
The ounterpart of this notation in the present work is outlined in this setion, in
Chapter 3, and in Setion 6.1. A omparison between our representation of GRs and
the `f-struture' in LFG is given in Setion2.7.2.
I assume that there are ve kinds of subonstrutions, and that a onstrution
anbeaonstellationofzero tovesubonstrutions. The subonstrutionsare alled
arg1-sign,arg2-sign,arg3-sign,arg4-signandarg5-sign. Thesevesubonstrutionsare
signs with a syntati expression and a semantiontent. As mentioned, the syntati
expression is either a funtion word, aninetion, ora rule. The subonstrutions are
not expressed as open lass lexialitems like verbs, nouns,or adjetives.
Thesemantiontent ofthesubonstrutionsareParsons-style underlyingevents.
Parsons (1990), argues that a transitive sentene like (3a) an be given the semanti
representations in (3) or (3d) rather than the traditional semanti representation in
(3b). In (3) the binary relationStabbedhas been given an underlying event analysis
with three underlying events. The prediate is the rst underlying event (Stabbing),
the rst argument is the seond underlying event (Subj), and the seond argument is
the third underlying event (Obj). (3d) is a representation with themati roles instead
(3) a. Brutus stabbed Caesar.
b. (
∃
e)[Stabbed(B,C)℄. (
∃
e)[Stabbing(e) & Subj(e,B) &Obj(e,C)℄d. (
∃
e)[Stabbing(e) & Agent(e,B) & Theme(e,C)℄WhileParsons uses funtional terms suh as Subj and Obj, or thematirole names
suh as Agent, Theme, Goal, Benefative, Instrumentand Experiener (Parsons, 1990,
7172) for the underlying events, I willuse the relation names arg1-rel, arg2-rel, arg3-
rel, arg4-rel and arg5-rel for the underlyingevents. These represent underlying events
that are not meant to orrespond diretly to themati roles (Fillmore, 1968), but
in ombination with the meaning of the main verb and the arguments, they an be
interpreted asthematiroles.
Theunderlyingeventsprovidedbythesyntatielements(arg1-relarg5-rel)isasfar
asI willgo into semantideomposition. In order to get thematirole interpretation,
or more elaborate semanti deomposition as in Jakendo (1990), I assume that the
underlying events will have to be interpreted in onjuntion with the meaning of the
verb and the meaningof the arguments. This isoutside the sope of this thesis.
I do not have as an ambition to let my analysis yield meanings or semanti
representations of sentenes. Aording to Frege's priniple of ompositionality, the
meaningof a sentene is determined by the meaning of the onstituents as wellas the
struture of the onstituents. In this thesis, I will only look into the struture of the
onstituents. The meaningof the onstituents willnot be taken into onsideration. So
one ofthe twofators, whihaording to the prinipleof ompositionality are needed
togiveasemantirepresentation ofasentene,ismissing. Thisdoesnotmeanthatthe
outputisompletelydetahed frommeaning, onlythat itrepresents apartialmeaning,
namely the meaning provided by the struture. (The onstituents will be represented
as well, but only as unanalyzable prediates.) Given that representations produed
by the grammar are assumed to give meaning only if interpreted in onjuntion with
the meaning of the onstituents, I have hosen to refer to them as a Basi Relation
Representations (BRRs).
4
The BRR of (3a) is given in Figure 1.1. It represents the
4
ThistermwassuggestedtomebyLarsHellan. Ihavealsoonsideredothertermssuhasstrutural
semantirepresentationsorolletionsofGrammatialRelations. However,thesetermsarepotentially
onfusing or misleading. The term strutural semanti representation may be seenas nonsensial if
onedoesnotadhere to theviewthatmeaningis ompositional. Thetermolletion of Grammatial
stabbing event deomposed into three underlying events _stab_v_rel, arg1_rel, and
arg2_rel. Theunderlyingeventsarelinkedbymeansofahandle(h1)(seeSetion3.4on
semanti representations for more details). The indiesof the two partiipants Brutus
and Caesar are bound by the underlying events arg1_rel and arg2_rel, respetively.
The bindingof the indies implies that the representation is an indexed BRR.
LTOP
h 1
h
INDEX
e 2
e
RELS
*
_stab_v_rel
LBL
h 1
ARG0
e2
,
arg1_rel
LBL
h 1
ARG0
x3
,
arg2_rel
LBL
h 1
ARG0
x4
,
brutus
LBL h
ARG0
x 3
,
aesar
LBL h
ARG0
x 4
+
Figure 1.1: Indexed BRR of Brutus stabbed Caesar
The subonstrutions an to some extent be illustrated by argument struture
features used in LFG (see Bresnan (2001, 302321), and disussion in Setion 2.3).
In LFG, argument struture is assumed to be lexially speied, and the semanti
argument roles arry features, [
±
o℄ and [±
r℄, whih onstrain the way the argumentroles are mapped onto argument funtions in f-strutures. The feature [r℄ maps the
argumentrole ontoanunrestritedsyntatifuntion,thatis,eithersubjetorobjet.
5
Obliques and restrited objets are [+r℄. The feature [o℄ maps arguments onto non-
objetive syntatifuntions(subjetsandobliques). Thefeature[+o℄ mapsarguments
ontoobjets and restrited objets.
The subonstrutions an more diretly be illustrated by means of Grammatial
Relations in Relational Grammar (Blake, 1990). In Relational Grammar, strata
represent the grammatial relations of a verb by means of ars labelled 1 (subjet),
2 (diret objet), and 3 (indiret objet). In addition there are oblique relations
(inluding benefative, loative, and instrumental). The Initial Stratum shows the
deep grammatial relations of a verb, and the Final Stratum shows the surfae
arepresentationofsurfaegrammatialrelationslikeF-strutureinLFG.TheGrammatialRelation
`Subjet'doesforexamplenotappearin therepresentations.
5
These funtions arereferredto asunrestritedsinetheyaordingto thetheorydonotneedto
haveasemantirole. Raisedandexpletiveargumentsarepresentedasexamplesofsyntatifuntions
with no semantirole. It should benoted that in this thesis, raised arguments are assumed to be
grammatial relations. The Initial Stratum may be idential to the Final Stratum
of ars. This is the ase in ative, transitive lauses, where the initial 1 is the nal 1
and the initial 2 is the nal 2. There may also be revaluations of ars. In that ase,
the Final Stratum is dierent from the Initial Stratum. This is the ase in a passive
transitivelausewheretheinitial1isdemoted tohmeurandtheinitial2ispromoted
to1. There may be more than one revaluation. The subonstrutions assumed in this
thesis orrespond toars in the InitialStratum inRelationalGrammar.
6
Thearg1-signisasubonstrutionthatorrespondstotherealizationofanexternal
argument, or deep struture subjet, in GB. It orresponds to the realization of an
(agent) argument with the [o℄ feature in LFG. It orresponds to the realization of
an argument whih has a 1-ar in the initial stratum in Relational Grammar. When
this subonstrution is used, it implies that the event of the main verb has something
that an be interpreted as a auser or initiator (an arg1-rel underlying event). The
information that the event has an arg1-rel is assumed to ome from the syntax, and
notfrom the main verb. In an ative main lause, the arg1-signis expressed as a rule
that links the subjet to the head projetion (see (4a)), and in a passive lause, this
subonstrution is expressed as the passive auxiliary or the passive morphology (see
(4b)). In aninnitivalativelause, thearg1-signisexpressed asthe innitivalmarker
(see (4)).
(4) a. John smashed the ball.
b. The ball wassmashed.
. (John tried) to smash the ball.
The arg2-sign is asubonstrution that orresponds tothe realizationof the diret
objet internal argument in GB. In LFG it orresponds to the realization of an
(patient/theme) argument with the [r℄ feature, or an(patient/theme) argument with
the [+o℄ feature if there is another (beneiary) argument with the [r℄ feature. It
orrespondstothe realization ofanargumentwhihhas an2-ar intheinitialstratum
in Relational Grammar. The arg2-sign expresses that the event of the main verb
has something that an be interpreted as a theme or patient argument (an arg2-rel
underlying event). Again, the information that the event has an arg2-rel underlying
event, omes from the syntax, and not from the main verb. The arg2-sign is usually
6
It should alsohere be notedthat in theapproah presentedin this thesis, raisedargumentsare
realizedas a rule that attahes the diret objet to the head projetion (see (5a)). In
unausative and passive lauses, the rule attahes the subjet to the head projetion
(see (5b) and (5)). In an innitival unausative orpassive lause the arg2-signmay
berealized asthe innitivalmarker (see (5d)).
(5) a. John smashed the ball.
b. The boat arrived.
. The ball wassmashed.
d. (The ar needed) to bewashed.
Thearg3-signisasubonstrutionthatorresponds totherealizationofanindiret
objet internal argument in GB. In LFG it orresponds to the realization of a
(beneiary) argument with the [-r℄ feature. It orresponds to the realization of an
argument with a 3-ar in the initial stratum in Relational Grammar. The arg3-sign
expresses that theeventhappens inthe (dis)favorof somebody(an arg3-relunderlying
event). Also here, the information that the event has an arg3-rel underlying event is
ontributed bythe syntax,and not by the mainverb. The arg3-signisusually realized
asa rule that attahes the indiret objet to the head projetion (see (6a)), but if the
lause is passive, it may be the subjet that the rule attahes to the head projetion
(see(6b)). Thearg3-signmayalsoberealizedasthe innitivalmarkerinaditransitive
passivelause (see (6)).
(6) a. John gave Mary abook.
b. Mary was given the book.
. (Mary wanted) to be given abook.
The arg4-signisasubonstrutionthat attahes adelimitertothe headprojetion.
Itorresponds tothe realizationofagoal/loativeobliqueinGB, LFG,andRelational
Grammar. A delimiter is a goal phrase as in (7a) or a resultative as in (7b). The
arg4-signexpresses that there is something that an be interpreted as an end point or
endstatefortheargumentrealizedbythearg2-sign(ifrealized)(anarg4-relunderlying
event). It isimportant tonotie that the informationabout there being an arg4-rel is
assumed to ome from the syntax, and not from the main verb (or from the delimiter
itself). The arg4-sign is realized by a rule that attahes the delimiter to the head
(7) a. John smashed the ballout of the room.
b. John hammered the metal at.
The arg5-sign is a subonstrution that attahes PP arguments that are not
delimiters, to the head projetion. It orresponds to the realization of for example
aninstrumentobliqueinLFGandRelationalGrammar. Anarg5-signan expressthat
the event has an instrument as in(8).
(8) John puntured the balloonwith a needle.
Table 1.1summarizes what argument realizations the subonstrutions orrespond
toin GB,LFG, 7
and RelationalGrammar(RG).
8
Subonstr. GB LFG RG
arg1-sign externalargument agent[o℄ initial1-ar
arg2-sign internaldir obj patient/theme [r℄ initial2-ar
arg3-sign internalindir obj beneiary [r℄ initial3-ar
arg4-sign oblique oblique oblique
arg5-sign oblique oblique oblique
Table 1.1: Subonstrutions orresponding to argument realizations in GB, LFG, and
RelationalGrammar
The ve subonstrutions an be ombined to form a wide range of onstrutions.
An intransitive sentene like (9a), has only an arg1-sign. This means that it has an
arg1-onstrution. A transitive sentene, like(9b), has two subonstrutions, anarg1-
sign and anarg2-sign. This means that ithas anarg12-onstrution. An unausative
sentene like(9) onlyhas anarg2-sign,whihmeans thatit has anarg2-onstrution.
Aditransitivesentenelike(9d) has three subonstrutions, anarg1-sign,anarg2-sign
and an arg3-sign. This means that it has an arg123-onstrution. A transitive lause
with a PP omplement like (9e) has an arg1-sign, an arg2-sign and an arg4-sign (the
PPto Mary isa delimiter). This means that ithas anarg124-onstrution.
7
Ifthereisabeneiaryargumentwiththe[r℄feature,theargumentrealizationorrespondingto
thearg2-signisapatient/themewiththe[+o℄feature.
8
The dierene between an oblique realized asan arg4-sign, and an oblique realized asan arg5-
sign an be understood by means of the distintion made between subsequent and anteedent roles
in Croft (1991, 184191). Croft refers to the roles benefative, malefative, reipient, and result as
subsequentand therolesinstrumental,manner,means, omitative, passiveagent, ergative,and ause
asanteedent. The subsequent roles are assumed to follow the objet in the ausal hain and the
anteedentroles are assumed to preede them. In this thesis, the arg4-signis assumed to realize a
(9) a.
John smiles. (arg1-onstrution)
b.
John smashed the ball. (arg12-onstrution)
.
The boat arrived. (arg2-onstrution)
d.
John gave Mary abook. (arg123-onstrution)
e.
John gave a book toMary. (arg124-onstrution)
Sine the onstrutions are reations of the syntax, a lexialentry an be allowed
to enter all possible onstrutions simply by not onstraining it. A verb like drip is
nottightlyonnetedtoapartiular onstrution. Therangeofonstrutionsthatthis
verb an enter, an easily be aounted for. (I willpresent onstrutionsthat dripan
enter inChapter 3.)
1.3 A onstrution-onstraining mehanism
The grammarI am presenting has a mehanism whih makes it possible to onstrain
verbs insuh a way that they onlyenter onstrutions that one would expet them to
appear in. Averblikeeatisnormallyallowedintoanarg1-onstrution(see(10a))and
anarg12-onstrution(see(10b)). Giventhatthesearetheonstrutionsonewantsthe
verb toappear in,the verb an be provided with the lexial onstraint arg1-12, whih
meansthatitiseitherallowedintothearg1-onstrutionorthearg12-onstrution,but
noother onstrution.
(10) a.
John eats. (arg1-onstrution)
b.
John eats anapple. (arg12-onstrution)
The onstrution-onstraining mehanism involves 8 top types, one positive type
and one negative type for eah of the rst four subonstrutions.
9
(The positive types
are named arg1+ (arg1 plus), arg2+ (arg2 plus), arg3+ (arg3 plus), and arg4+ (arg4
plus), and the negative types are named arg1 (arg1 minus), arg2 (arg2 minus),
arg3 (arg3 minus), and arg4 (arg4 minus).) The types indiate whether or not a
subonstrutionispresentinalause. Bydefault,alauseisassignedthefournegative
types. Foreahsubonstrutionthatappliesinthelause,thenegativetypeisswithed
9
The arg5-sign is not a part of the mehanism. The PPs realized by the arg5-sign are in the
toa positive type. Soif an arg1 sign applies, the type arg1 is swithed toarg1+. An
intransitive lause like (10a) has one subonstrution, the arg1-sign, and it therefore
has the types arg1+, arg2, arg3, and arg4. The uniation of these types gives the
onstrution type arg1, as shown in Figure 1.2.
10
A transitive lause like (10b) has
twosubonstrutions, thearg1-signand thearg2-sign,andhas thetypesarg1+, arg2+,
arg3 and arg4.
link
arg1 arg2 arg3 arg4 arg1+ arg2+ arg3+ arg4+
arg1
Figure1.2: Supertypes of the onstrution type arg1 inthe link hierarhy
In order to limit the number of possible onstrutions a verb an enter, a set of
intermediate types is introdued. The hierarhy in Figure 1.3 illustrates one suh
type, namely arg1-12. It inherits from arg1+, arg3, and arg4, and it has two
subtypes, the onstrution types arg1 (whih inherits from arg1-12 and arg2) and
arg12 (whih inherits from arg1-12 and arg2+). The intermediate types represent
lexialinformation assoiated with verbs, and they are unied with the four (positive
ornegative) subonstrutiontypes of the lause. This fores a verb speied with the
arg1-12 type toour in lauses with the arg1-onstrutionor the arg12-onstrution.
Theonstrution-onstrainingmehanismisnot apartofstritsyntax. Itsfuntion
is to prevent odd sentenes rather than ungrammatial sentenes.
11
However, suh a
10
AmoreextendedversionofthehierarhyisgiveninFigure4.9,p. 96. Thefullhierarhyisgiven
inthele`nor.tdl'innorsyg,under `valenetypes'andinludes128types.
11
I believethatthere should beadistintionmadebetweentheungrammatialityofexampleslike
(xi)ontheonehand,andtheoddityofexampleslike(xii)ontheother. Whiletheexamplesin(xi)are
unaeptablebeauseofsyntatierrors(in(xia)thereisapasttensedverbinaninnitivallause,and
in(xib) thedetermineradoesnotagreewiththenoun men), theexamplesin(xii) areunaeptable
beausethemainverbsenteronstrutionsthattheyarenotompatiblewith. Thesyntatistrutures
inthelatterexamples,Iargue,aregrammatial.
(xi) a. *Johntries toslept.
b. *Amensmiles.
link
arg1 arg2 arg3 arg4 arg1+ arg2+ arg3+ arg4+
arg1-12
arg1 arg12
Figure1.3: A partial link hierarhy
mehanism is neessary in order tokeep the searh spae of a parser at a manageable
level. When implementing a grammar, one has to attend to the grammar both as
a linguisti theory and as a parser. This raises onerns that not always unite. For
example,inprinipleIwouldliketoallowallverbs(ormaybeevenallopenlasslexial
items)toenter allonstrutions,but inarealimplementation,thiswillmakethe parser
tooslow. The onstrution-onstrainingmehanism isdesigned for these onerns; see
Chapter 4. (Some of the more tehnial aspets of the implemented grammar are also
disussed inAppendix A.6).
To give an idea of the type of system I amproposing inthese respets, imaginean
LFG-like grammar diering from any atual LFG grammar in not obeying priniples
(xii) a. #Johnsleptthear.
b. #Johnadmires.
Theommon judgmentof exampleslike(xiiia) and (xiiib) is that (xiiia) is grammatial, whereas
(xiiib)isungrammatial.
(xiii) a. Johnlledthemouth withhoolate.
b. #Johnsmiledthemouthwithhoolate.
Iarguethat(xiiib)isnotreallyungrammatial,ratherthat itisveryodd. (Iwilllaterinthethesis
starveryoddsenteneslike(xiia),(xiib)and(xiiib),eventhoughIlaimtheyarenotungrammatial.)
Itispossibletogetsomemeaningoutof(xiiib)byoerion. ForexamplethatJohnausedhismouth
to be lledwith hoolate by smiling. Orthat John used hoolate to turn his mouth into asmile.
The term ungrammatial I reserve for sentenes like (xia) and (xib). These sentenes ould never
be grammatial,irrespetiveof the meaningassigned to theopenlass lexial items. I will however
usethetermungrammatialaboutsentenesthatare veryodd lateron,simply beausethat isthe
of Completeness and Coherene (see Bresnan (2001, 63)). In the lexial entries for
verbs in suh a grammar - being f-struture skeletons - there would be no GR-lists
provided with the attribute`PRED' (justlines like`PRED `kik). Insuhagrammar,
due to phrase struture rules like (2), syntati ombination would still populate the
f-struture with whatever GRs were enountered, and the resulting f-struture would
provide areordofthe GRssyntatially enoded inthe onstrutionparsed;however,
withoutany mehanismheking whether suh anassembly of GRsis aepted by the
verb inquestion. ThisisinspirithowIwould likeagrammartofuntion. However, for
onernsmentioned,wemaywanttoinludeonstraintsineahlexialentryonerning
admissibleGRs. In the imaginary LFG grammarin question, one would then add the
relevant speiation inside the PRED value, e.g., `PRED `kik(Subj, Obj). In my
system,Isimilarlyhaveone versionof lexialentries wherenothingissaid aboutwhih
arg-types a verb may ombine with, and one line in whih, for `kik', for instane, I
an insert the speiation `arg12' (f. above). So far, though, this might seem just
a pointless exerise of notational inventiveness. What are ruial ontributions by my
system are the following,however:
In the rst plae, in ases where a given verb has many environments, LFG and
standardHPSGwillpositasmanyentriesforthatverbasithasframes. Mydeployment
of a type system as skethed, on the ontrary, will allow me to have only one entry,
whih still aommodates all the frames. This will be shown in Setions 4.3 and 4.4.
Seondly, this same type-design willallow meto use the atual parsingof a orpus as
a way of inrementally dening the sum of frames inwhih a verb an enter, but as a
resolutionproess workingrelativetothe one singleentryrequired. This willbeshown
inSetion4.5. Althoughthelatterpointhasnotyetbeenarriedoutonalargeorpus,
the mehanism is lear, and I see these two points as valuable tehnial ontributions
toparsing designand grammar engineeringin general.
Thewayinwhihtheunitarytypedenitionsmentionedabovedependonresolution
by the syntati environment, may raise the question whether this mehanism would
apply also for a grammar where disriminants of multi-frame verbs involve semanti
strutureinadditiontoGFs. Ofrelevaneareasesofnon-isomorphybetween semanti
and syntati struture. Having stated that I willnot be onerned with semantis, it
follows that I will not try to represent the `skewed' syntax-semantis relationship of
sentenes like I believe himto be sik orHe seems sik. By standard assumptions,
latter `seem (he be sik)', thus `believe' here being logially a two-plae prediate
and `seem' a one-plae prediate. As far as arg-roles in my system are onerned,
`believe' willhave three arg-rolesand `seem' two in these examples, sine the analysis
addresses syntati struture exlusively. By these resolutions, I obviously will not
get any semanti struture beyond what stands in a one-to-one relation to the GF
strutures. Atleastatitspresent stageofdevelopment,this anbeseen asalimitation
ofmysystem, andI disusswhatitmaytake forittoopewiththesetypesof`skewed'
onstrutionsin Setion6.7.4.
1.4 Exo-skeletal grammar and left-branhing synta-
ti strutures
I propose that the approah I am taking an bealled an exo-skeletal approahin the
sense of Borer (2005a, 15). This term is borrowed from zoology, where it is used to
desribeanimalsthathavetheirskeletonontheoutside. Theoppositeofexo-skeletal is
endo-skeletal, whih is used todesribeanimals with the skeleton inside the body, like
humans. In anexo-skeletal grammar, the funtional signs (funtion words, inetions
and rules) are given more emphasis, while the role of the open lexion (lexial entries
of nouns, verbs and adjetives) is played down. In an endo-skeletal grammar,it is the
lexemes that dene what is outside, and the argument struture is xed in the verb
lexeme.
Inanexo-skeletalgrammar,thegrammaraninprinipleonlygenerategrammatial
sentenes even if the open lass lexial items do not have any subategorization
onstraints. This is an advantage that an exo-skeletal grammar has to a stritly
endo-skeletal grammar, whih ruially relies on the subategorization onstraints of
open lass lexial items. If the subategorization onstraints of the open lass lexial
items in an endo-skeletal grammar were left out, there would be nothing preventing
ungrammatialsentenes like(14) frombeing generated.
(14) *John eats anapple Mary that he smiles.
Theideasabout exo-skeletalgrammarthatI presentinthis thesis,areimplemented
inthe Norsyg grammar. The mainobjetive of this grammaris this: I havewanted to
ordertoaountforverbswithmorethanoneonstrution. Thus, nomatterhowmany
argument frames a verb an our in (and provided they are not distint in terms of
`stritsyntax'),the lexionwillprovidejustoneentryfortheverb, andthemultipliity
of frames willbe indued from the dierent onstrutionalenvironments solely.
To put this another way, Norsyg isdierent from lexialist grammarsin that open
lass lexial items are unonstrained by default. Restritions an be made if there is
a need for it. The ommon proedure in lexialist grammars is to be very restritive
by default, that is, only to allow one onstrution on a lexial entry, and then reate
mehanisms that produe other possible onstrutions, mainly by means of multiple
lexialentries orlexialrules.
Syntati strutures are assumed in general to be left-branhing (see Figure 1.4),
rather than mixed left- and right-branhing (enter-embedded) (see Figure 1.5), as
assumedinHPSGandLFG,orright-branhing(see Figure1.6),asassumedinversions
ofGB/MinimalismusingLarsonianshells(Larson,1988;Culiover,1997). Withaleft-
branhingstruture, therst onstituentwillappearatthebottomofthetree (likethe
node ain Figure1.4), andthe lastonstituentwillbethe lastdaughter of the top rule
(like the node d).
A
B
C
a b
d
Figure 1.4: Left-
branhing tree
A
a B
C
b
d
Figure 1.5: Mixed
left- and right-
branhing tree
A
a B
b C
d
Figure 1.6: Right-
branhing tree
Left-branhing syntati strutures make it possible to give an aount of long
distane dependenies where the ller appears at the bottom of the tree, and the
extration site -ommandsthe ller. Thatis, the positionthat the lleris assumed to
be extrated from,is situated higher up the tree, as a sister of one of the anestors of
theller. Theinformationthatthere isalongdistane dependeny,passes throughthe
nodes intervening between the ller and the extration site. Ifthere is a long distane
dependeny between the node a and d in the tree in Figure 1.4, this information will
mixed left- and right-branhing tree struture as shown in Figure1.5, a long distane
dependeny between a and d will not be loal to the nodes b and , sine it does not
pass through their mother (C). In a right-branhing tree struture as shown in Figure
1.6, the informationthat there isa longdistane dependeny between thenodes aand
d, is again loal to the nodes b and , sine it passes through their mothers B and C.
In some languages (Sag (2005) mentions amongother languages Chamorro and Irish),
long distane dependenies are registered by verbs or omplementizers. This indiates
that suh onstituentshave loalaess to long distane dependenies.
The left-branhingstrutures allowfor inrementalparsing,with abottom-up, left-
to-rightparsingstrategy. ThenodesofthetreeinFigure1.4arethenenumeratedinthe
order shown in(15a). Also right-branhing strutures (often used inGB/Minimalism)
allow for inremental parsing, if they are parsed with a left-orner parsing strategy.
The nodes of the tree in Figure1.6 are then enumerated in the order shown in (15b).
Mixed left- and right-branhing tree strutures (used in LFG and HPSG) do not lend
themselvestoinrementalparsinginthesamewaysinethesekindsofstruturesrequire
storage proportional to the height of the tree. (I will return to parsing strategies in
Setion5.2.)
(15) a. a, b, C, , B, d, A
b. a, A, b, B, , C, d
Given the left-branhing strutures assumed in this thesis, 12
the traditionalnotion
of a syntati onstituent, is not appliable. What traditionally is oneived of as a
syntationstituent(awordoraphrasewhihanbereplaedbyapronoun,whihan
befronted, or whih may bepossible tooordinate) is ratherreeted as aonstituent
inthe BasiRelation Representation. Syntati strutures inthis thesis are to alarge
part determined by the exo-skeletal nature of the grammar. A main verb may for
example be regarded more as a modier than as the syntati head of a lause. A
omplementizer may form a onstituent together with the matrix lause, rather than
forming a onstituent together with the rest of the subordinate lause. The syntati
strutures reet how words and phrases ombine and form new onstituents, but as
mentioned, these onstituentsare not neessarily onstituents in the traditionalsense.
12
There are someaseswhere the left-branhing tree struturesare notemployedin the analyses,
likeinPPsandsomeasesofoordination,buttheseaseshavenotbeenthefousofmystudy. Iwill
alsomakeuseofastakinordertoaountforembeddedlauses. Thisimpliesthatparsingwillnot
Rather, they an be seen as the history of ombinations of words and phrases of a
sentene.
The grammar formalism I am presenting in this thesis borrows ideas from several
grammatialtheories,inludingHPSG,ConstrutionGrammar,LFG,andGB.Thefat
that the grammar is a typed feature struture grammar and designed for bottom-up
hart parsing (Kay, 1986), is due to the fat that it is implemented with the LKB
system (Copestake, 2002). Sine the formalism was developed from the Grammar
Matrix(Bender et al.,2002),the terminologyused torepresent grammatialobjets is
toa large degreetaken fromHPSG.
The ideaof one lexialentry perstem (and nolexialrules)and that onstrutions
havemeaningindependentofthewordsthatappearinthemisinspiredbyConstrution
Grammar,but while onstrutions in my grammarformalism an bedeomposed into
subonstrutions, onstrutions in Constrution Grammar are seen as entities that
annotbe analyzed further (see Setion2.4).
As already mentioned, the grammatial relations assumed to hold between a
prediate and its arguments an be ompared to the grammatial relations used in
LFG,but there is noone to one orrespondene.
Apart from apparent similarities to HPSG, Constrution Grammar, and LFG,
the grammar formalism is maybe best oneived of as a monostratal variant of GB
(Chomsky, 1986) where surfae grammatial relations, deep grammatial relations,
and movements are represented at one level. Movement to the speier position of
C(aountingforwh-movement/longdistanedependeniesinGB)isaounted forby
meansof the perolation of a feature slashas inHPSG (butas I willshowin Setion
6.9,theapproahinthisgrammarformalismdiersinseveralrespetstotheapproahes
inHPSG).Movementto anargumentpositionas assumed inaseswhere anargument
reeivesthematirolefromoneverbandase fromanotherverb(aounting forraising
onstrutionsandsmalllausesinGB)isnotpossible. Instead,thegrammarformalism
allowsforanargumenttoberealizedtwieintheseases. Thisorrespondstoassuming
anargumentsimilartoPROinGB.(SeedisussioninSetions6.7.4and9.5.2.) Passive
is aounted for by assuming that what orresponds to the external argument in GB
is realized by the passive auxiliary or the passive morpheme (see Setion 7.1). The
formalism does not imply anything orresponding to head movement in GB (V to T
and/or T to C movement), but ertain positions orrespond to C, T, and V, and the
Setion9.3).
Theleft-branhingtree struturesresultintree struturesompletelydierentfrom
the right-branhing strutures known from GB (and from syntati strutures in any
other theory, exept perhaps from CCG), and onstituents in the traditional sense
are not formed. Still, given the dierent parsing strategies assoiated with the two
approahes(abottom-up,left-to-rightparsingstrategyintheapproahpresentedinthis
thesis vs. a left orner parsing strategy argued to be appropriate for GB (see Setion
5.2)), preterminals are enumerated in the same order. This will be demonstrated in
Chapter 9.
1.5 Layout of the thesis
The rst part of the thesis inludes Chapters 24 and deals with argument struture.
In Chapter 2, I introdue some entral notions in the disussion around argument
struture, suh as unausativity and unergativity, valene alternations and voie. I
disuss how HPSG, LFG, Constrution Grammar, and three versions of Minimalism
deal with argument struture. I look at how muh argument struture information
the theories assume is present in the lexion, and how muh they assume an be
redued to syntax, and I situate the theories on a sale lexialist <> non-lexialist
(or endo-skeletal <> exo-skeletal). In Chapter 3, I go through a number of the
valene alternations and onstrutions presented in Levin(1993), and showhow these
alternations an be aounted for syntatially with the ve subonstrutions that I
am assuming. I will present the Basi Relation Representations (BRRs) that are
employedinthegrammar. InSetion3.5,Isuggestfourbasisigntypeswhihrepresent
the realization of the rst four subonstrutions. I show how dierent syntati
instantiationsofthesubonstrutionsinheritfromthebasisigns. InChapter4,Ishow
howvalenean berepresented inagrammarformalismwhereargumentstruture an
be inferred from funtional signs. I introdue four valene features, one for eah of
therst foursubonstrutions. Thesewillarry positiveand negativevalues,reeting
whether the argumentis realizedornot. Further, I introdue ahierarhy whih allows
metogiveaompat representationofpossibleonstrutionsfor alexeme. I givesome
examplesoflexialentrytypes, andpresentmethodsforexpandingthelexion. Finally
I ompare my approah to a lexialist version of the grammar, the Robust Aurate
The seond part of the thesis inludes Chapters 510. In this part I show how an
exo-skeletal grammar may be strutured. I present analyses of a range of linguisti
phenomena. Chapter 5 gives a preliminary introdution to the syntati strutures I
amassuming. I will present some methodologial onsiderationsonerning linguisti,
oneptual, and omputational aspets of the approah. In Chapter 6, I present the
basi syntati interior of a grammarforNorwegian. I suggest six main kindsof rules.
First, the valene rules, whih realizethe rst four kinds of subonstrutions. Seond,
the llerrule, whihllsinthe extrated onstituent. Third,the merge rule, whihfor
example ombines a projetion headed by a omplementizer or an auxiliary with the
main verb. Forth, the subordination rules, where embedded lauses are entered. Fifth,
thelause boundaryrules,whihmarkthe boundaryof thelauses. Sixth, themodier
rules, whih let a modier modify a head projetion. The hapter gives analyses of
mainlauses, subordinatelauses,relativelauses and innitivallauses. There isalso
a setion on long distane dependenies. In Chapter 7, I present analyses of passive
andpresentation. InChapter 8,I presentfourkindsof oordinationinNorwegian,and
arguethatitisanadvantagetouseanexo-skeletalgrammarinforexampletheanalysis
ofoordinationofVs. InChapter9,Iomparethe analysispresentedinChapter6with
GB, and use the omparison to illustrate how syntati strutures of basi lauses in
Englishanbeaountedfor. InChapter10,Ipresentananalysisofsenteneadverbials
inNorwegian inlight of the analysis presented inChapter 6.
Appendix A has information about the Norsyg grammar, where the analysis
presented inthis thesisis implemented. Appendix B has informationabout anEnglish
and a German demo grammar, whih I have developed in order to illustrate how the
analysisan beextended toother languages. All the Norwegian and English examples
in this thesis are gathered in the les `ex.items' and `eng-ex.items', distributed with
Norsyg,andtheresultsofbathparsesofthesesenteneswiththeNorsyggrammarand
the English demo grammar are given in Appendix C. Basi Relation Representations
(BRRs)of all analyses onduted with the Norsyg grammarand with the English and
Germandemo grammarsare given inAppendix D.
Argument Struture
Argument struture in HPSG, LFG,
Constrution Grammar, and
Minimalism
2.1 Introdution
InthishapterIwilllookathowHPSG(Head-DrivenPhraseStrutureGrammar),LFG
(Lexial Funtional Grammar), Constrution Grammar, and three approahes within
Minimalismtreat argument struture and valene alternations. The three Minimalist
approahes are Hale and Keyser's Prolegomenon to a Theory of Argument Struture,
Ramhand's First Phase Syntax and Borer's neo-onstrutionist approah. I have
hosen three Minimalistapproahes that span from a lexialist approah to argument
struture to a strit non-lexialistapproah to argument struture. I willpresent how
the theories aount for the most basi argument frames of intransitive verbs (both
unergativeand unausative), transitive verbs, and ditransitiveverbs. I willalso show
howthey dovalene alternationslikepassive, theausative/inhoative alternationand
resultative onstrutions.
1
I aim at situating the frameworks on a sale lexialist
non-lexialistby lassifying them with regardto three main riteria:
1
StudiesbyBoguraevandBrisoe(1989)andManning(2003)showthatitisdiulttogivegood
riteriafor when valenealternations anapply. Corpusevidene presentedin Bangaloreand Joshi
(1999)showsthat lexial items on average are assoiated with asmany as 47 supertags,whih are
bundlesofphrasestrutureinformationanddependenyinformation.
1. Variable behavior verbs 2
Whether the alternation between unausativity and
unergativity of the same verb is treated as part of the lexion or as part of the
syntax.
2. Valene alternations Whether alternations suh as the dierene in arity, 3
the ausative/inhoative alternation, the dative alternation, the spray/load
alternation and the resultative onstrution are aounted for lexially or
syntatially. 4
3. VoieWhether ative,passiveand middlevoieis treatedlexially,orasapart
of the syntax.
Generallyspeaking,frameworkslikeHPSGandLFGwillbeshowntolassifymostly
as lexialist with regard to all three riteria. The Minimalist frameworks I will be
onsidering dierwith regard to the three riteria. Before I disuss the frameworks in
detail,I willpresentsome linguistinotionsthat I willuse inthis setion. Muhof the
materialI present is taken from orinspired by Levin (1993). I willonsider argument
frames that our in Norwegian and English.
2.1.1 Unergative and unausative verbs
Thedierenebetweenunergativeandunausativeverbshasbeenanissueinlinguistis
fora long time (see Jespersen (1924,164-167),Fillmore(1968), Perlmutter (1978)and
Levinand Hovav (1995)).
Unergative (or real intransitive)verbs are verbs like smile, laugh and sing. These
verbs may passivizeinNorwegian. They an not transitivizeinthe sensethat aauser
isaddedtothe event. This isillustrated by (16)where (16a)is grammatialand (16b)
isungrammatial.
(16) a. The man smiled.
b. * Mary smiledthe man. (On the interpretation that Mary ausedthe man to
smile)
2
I havetakenthisnotionfromBorer(2005b,30-46).
3
By dierene in arity I mean whether a verb an shiftbetweenintransitive and transitive, and
transitiveandditransitive.
4
Variable behavioris not treated aspart of valene alternations sine variable behaviorin some
theoriesannotbeaountedforbymeansofoneroot/lexialitem,whileinotheralternationsitan.
Unausativeverbsonthe otherhandare intransitiveverbs likearrive,dieandfall.
These verbs annot passivize. An intuition behind this group of verbs is that their
argumentorresponds totheobjetofatransitivelause. Ifweinludethe intransitive
versionsofverbslikebreak,widen,andraktotheunausativeverbs,weseethatthese
verbsmaytransitivizebyaddingaauser,asillustratedin(17)wheretheauserMaryis
addedin(17b). Theobjetofthe ausativizedversionorrespondtothe subjetinthe
intransitive version. This phenomenon is often referred to as the ausative/inhoative
alternation.
(17) a. The glass broke.
b. Mary broke the glass.
It ispossiblefor an unergativeverb to have an objet addedwhile maintainingthe
semanti role of the subjet asillustrated in(18a). An objet like a big smile in(18a)
isusually referredtoasa ognate objet. Unausativeverbs onthe otherhand annot
havesuhobjets, as(18b)illustrates. Inorderfor(18b)tobegrammatial,thesubjet
annotbe the argument that isbeing broken, as itis in(17a).
(18) a. Mary smiled abig smile.
b. * The glass broke arak. (On the interpretationthat the glass isbreaking)
Someverbs areambiguousbetween anunausative andanunergativereading,like
drip in (19). Either the subjet is the soure of the dripping, as in (19a) (unergative
reading), or the subjet is what is dripping, the theme, as in (19b) (unausative
reading). These verbs, as said above, are alled variablebehaviorverbs.
(19) a. The roofdrips.
b. Water drips (fromthe roof).
Data suh as those presented in examples (16)-(18) have made linguists propose
that the syntati subjet of an unausative verb as in (17a) is really an underlying
objet or internal argument of the verb, sine this argument funtions as objet if a
auseris added asin (17b)(see for exampleFillmore(1968); Perlmutter (1978)).
2.1.2 Other alternations
Transitive verbs and unausative verbs an have the resultative onstrution, as
(typially a PP or an adjetive) prediates over the underlying objet. In (20b),
the prediative element prediates over the objet of an ative transitive verb, and
in (21b), it prediates over the subjet of an unausative verb. An unergative verb
(whihdoesnothaveanunderlyingobjet)annotexpresstheresultativeonstrution,
asillustrated in (22).
(20) a. John hammered the metal.
b. John hammered the metal at.
(21) a. The riverfroze.
b. The riverfroze solid.
(22) a. The man smiles.
b. * The man smiles happy. (On the interpretation where the man beomes
happy)
Someovertlytransitiveverbs likeeat,readand paintmay have anunderstood objet
thatmayormaynotbeexpressed, asillustratedwiththe pairin(23). This isone form
of alternationin arity.
(23) a. John ate the apple.
b. John ate.
The dative alternation is an alternation between a ditransitive verb, as in (24a),
and a transitive verb with a PP omplement, as in (24b). The indiret objet of
the ditransitive verb (Mary) orresponds to the prepositional objet of the transitive
verb. The indiretobjetof the ditransitiveverb must be somethingthat an takethe
diret objet into its possession. This interpretation is not neessarily present for the
prepositional objet of the transitiveverb (see Pinker (1989, 48)).
(24) a. John gave Mary anapple.
b. John gave anapple toMary.
Thespray/loadalternationisanalternationbetween twotransitiveverbs withaPP
omplement. In one variantthe objetis the argumentwhose loationis hanged, and
the PP is the new loation (see (25a)). In the other variant the objet is the loation
(25) a. John loaded hay onto the wagon.
b. John loaded the wagon with hay.
2.1.3 Voie
Englishhas ative, passive,and middlevoie, as illustrated in(26).
5
(26) a. The buther uts the meat.
b. The meat was ut (by the buther).
. The meat uts easily.
The transitiveverb ut an bethe main verb of lauses with allthree voies. (26a)
is an ative sentene. So far in this setion all sentenes have been ative. (26b) is a
passivesentene. Passive isusually eitherperiphrasti,asin English(passiveauxiliary
+ past partiiple) or morphologial (marked with an ax on the main verb). When
a lause is passive, as in (26b), the subjet of the orresponding ative lause (in this
ase the buther) is expressed in an optional PP
by
. Some other element is realized asthe subjet. In English, this will be the objet that in ative is losest to the verb
(i.e. the meat in (26b)). Even though the agent may not be expressed, there is still a
notionof some auser of the situationexpressed. Inthis sense, passive sentenes dier
fromsenteneswithunausativeverbs (see(17a))wherethereisnonotionofaauser.
((17a)doesnotonveythat the breakingeventis ausedby anyone oranything,it just
hap pended.)
(26)isasentenewithmiddlevoie. Asentenewithmiddlevoiehasnopartiular
markingin English exept that itusually ontains an adverb likeeasilyin(26).
6
The
subjet of the orresponding ative lause (the buther) is not expressed. Stillthere is
a notion of ausation, whih is not present in the unausative lause. Compare for
5
I here hange perspetive and present voie as a property of lauses, rather than a property
inherenttoverbs. I ouldalsohavetakenthelauseperspetiveforthealternationsIhavepresented
in theprevious setions, but sinemost of theliterature seemsto treat these alternationsaslexial
alternations,ratherthanassyntatialternations,Ihaveusedthelexial perspetive.
6
Norwegiandoesnot havemiddle voie. Instead of middle, the sequene let + reexive+ main
verbisusedasin (xxvii). LanguageslikeSpanishandRussian markmiddlewithareexivesux.
(xxvii)Studiet
study-def lar
lets seg
itself lett
easily
kombinere
ombine med
with en
a jobb.
job
`Thestudyombineseasilywithajob.'
example The up broke with The up broke easily. In the latter example there is a
notion of something external to the up that made it break, while this notion is not
available inThe up broke.
Havingskethed the intuitionsbehindverb alternationsandvoie,I nowproeedto
adisussion of dierenttheoretial frameworksand how they relate tothe phenomena
I have presented.
2.2 HPSG
InHPSG, the argumentframeof a verb isto alarge extent determinedwhen the verb
enters the syntax. A lexial item is a sign onsisting of phonologial, syntati and
semantiinformation,as illustrated inFigure 2.1.
7
phon
D
admire
E
at
head verb
val
subj
D
NP
1
E
omps
D
NP
2
E
ont
|
restr*
pred _admire_rel
arg1
1
arg2
2
+
Figure2.1: Lexialentry for the verb admire
Thephonologialinformationisusuallyrepresented asalistofstrings(seethevalue
of phonin Figure 2.1) The syntati informationis represented asa feature struture
as value of the feature at. The semanti information is represented as the value of
ont.
8
The verb admire is transitive, and this is reeted on the valene lists subj
and omps.
9
The subjlist ontains an NP (the subjet) and the omps listontains
7
There are dierent naming onventions for features in HPSG. I will be using the ontology of
featuresthatisusedinPollardandSag(1994),Chapter9. ThesefeaturesarealsousedintheEnglish
ResoureGrammar(ERG)Flikinger(2000).
8
Inparts oftheliteraturethefeaturessynandsemareusedinsteadofatandont.
9
In my presentation of HPSG I use the valene features subj and omps as in Borsley (1996).
Inparts ofthe HPSG literature,there is onlyonevalene list,subat(e.g. Pollardand Sag (1994)
(Chapter1-8),Müller(2002)),whileinotherpartsoftheliteraturethefeaturearg-st(orarg-s)has
