Proceedings of the LFG 02 Conference National Technical - CSLI ...

More documents

Recommendations

Info

LFG02 – Kuhn: Corpus-based Learning in Stochastic OT-LFG effect of constraints of different importance, and which is for instance underlying the predecessor of OT, Harmony Grammar). The OT hypothesis of strict ranking is motivated for the very reason of making the system more restrictive, such that clearly testable typological predictions of the system follow from the assumption of a particular set of constraints. To illustrate this point, let us briefly compare the way predictions are grounded in a ranking scheme and how this compares to a weighting scheme. If we have a constraint violation profile as in tableau (5) (with the ranking of the constraints open) and we observe candidate A in the data, we know that CONSTR. 3 must outrank the other two constraints: CONSTR. 3 { CONSTR. 1, CONSTR. 2 }— else candidate A would be the winner. This kind of configuration is called a ranking argument. The fact that candidate B incurs three violations of CONSTR. 3 and not just one is irrelevant: the only way that B will lose against A is when CONSTR. 3 is ranked highest. (5) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 candidate A * * candidate B *** Now, if in addition to (5), we observe the A and A candidate of (6-b) and (6-c) data for the same language, we get an inconsistency: (6-b) and (6-c) are ranking arguments for CONSTR. 2 CONSTR. 3, and CONSTR. 1 CONSTR. 3, respectively. (6) Under the ranking hypothesis, (a) is incompatible with (b) and (c) (a) Candidate set: CONSTR. 3 CONSTR. 1 CONSTR. 2 ☞ candidate A * * candidate B *** (b) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 ☞ candidate A * candidate B * (c) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 ☞ candidate A * candidate B * So, a small set of clear data is already very informative about an OT account, based on the ranking hypothesis. If we do observe all the data in (6) in a single language, we know that the constraint set assumed was inadequate; maybe an additional constraint or an entirely different set of constraints is needed. Now, under a constraint weighting regime no such clear conclusion about the symbolic part of the theory—the constraints and the candidate representations—can be drawn. The (a) type of data may be compatible with (b), with (c), with both, or none. Examples (7)–(9) illustrate this with different negative weights assumed for 244
LFG02 – Kuhn: Corpus-based Learning in Stochastic OT-LFG the constraints (the winner is defined to be the candidate with the greatest weighted ¡£¢ sum over violation marks, e.g., (7-b,A ) wins over (7-b,B ¥¤ ) since ). In all cases the (a) data are correctly ¦¨§ predicted, 1©��¦�§ since 2©¨� CONSTR. CONSTR. 3© CONSTR. . Note however that absolute constraint weights would lead to ¤��¦¨§ different rankings in each of the cases, as is suggested by the order of notation (which of course has no technical effect, since we are looking at a weighting system). (7) (a) is compatible with (b), but not with (c) (a) Candidate set: CONSTR. 1 CONSTR. 3 CONSTR. 2 �� ☞ �� cand. A * * �� cand. B *** (8) (a) is compatible with (b) and (c) (a) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 �� ☞�� cand. A * * cand. B *** (b) Candidate set: CONSTR. 1 CONSTR. 3 CONSTR. 2 �� ☞�� cand. A * cand. B * �� (b) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 �� ☞ �� cand. A * �� cand. B * (9) (a) is compatible with neither (b) nor (c) (a) Candidate set: CONSTR. 3 CONSTR. 1 CONSTR. 2 �� ☞ �� cand. A * * �� cand. B *** not compatible (b) Candidate set: CONSTR. 3 CONSTR. 1 CONSTR. 2 �� ☞�� cand. A * cand. B * �� not compatible (c) Candidate set: CONSTR. 1 CONSTR. 3 CONSTR. 2 �� ☞�� cand. A * cand. B * �� (c) Candidate set: CONSTR. 1 CONSTR. 2 CONSTR. 3 �� ☞ �� cand. A * �� cand. B * not compatible (c) Candidate set: CONSTR. 3 CONSTR. 1 CONSTR. 2 �� ☞�� cand. A * cand. B * �� As the example illustrated, the constraint weighting scheme has an undesirable property if we are interested in finding a linguistically motivated set of constraints for predicting a typological spectrum of languages: the effect of picking a particular constraint set is underdetermined—a readjustment of the constraint weights may have the 245
Page 1 and 2:
Proceedings of the LFG 02 Conferenc
Page 3 and 4:
Table of Contents Asudeh, Ash 1-18
Page 5 and 6:
The Syntax of Preverbal Particles a
Page 7 and 8:
2 The Problem: Five Conflicting Cla
Page 9 and 10:
(10) CP C ′ XP CP XP The outermos
Page 11 and 12:
(18) CP C ′ C ∅ IP CP IP I I
Page 13 and 14:
1-3, unless problematic assumptions
Page 15 and 16:
(4c) *After [IP last year [IP she r
Page 17 and 18:
The adjunction rule (36) does not b
Page 19 and 20:
(8c) chun isteacht duit [CP[CP nuai
Page 21 and 22:
work on non-projecting categories a
Page 23 and 24:
Coordination and Parallelism in Glu
Page 25 and 26:
Traditional Linear Duplication a, a
Page 27 and 28:
(8) fred : fe see : fe ⊸ je ⊸ s
Page 29 and 30:
us add the following additional con
Page 31 and 32:
3.2 A General Semantic Schema for C
Page 33 and 34:
approach also achieves correct resu
Page 35 and 36:
In summary, to show that two deriva
Page 37 and 38:
Extraction possible from initial (p
Page 39 and 40:
5.3.2 Across-the-Board Extraction T
Page 41 and 42:
Asudeh, Ash (in progress). Resumpti
Page 43 and 44:
Montague, Richard (1974). Formal ph
Page 45 and 46:
Abstract VP-chaining in Oriya (an I
Page 47 and 48:
(3) haata dho-i bhaata khaa-i mun s
Page 49 and 50:
Figure 4. Schematic representation
Page 51 and 52:
the medicine is applied. If overtly
Page 53 and 54:
The control equation in (13) establ
Page 55 and 56:
intervening object and thus ‘kept
Page 57 and 58:
Lexicon L(1): khi-aa-ga-laa: V L(2)
Page 59 and 60:
(20) VP -> VP V' α � *α SIT-IND
Page 61 and 62:
CONSTRAINT SYMMETRY IN OPTIMALITY T
Page 63 and 64:
‘Juan hit that dog with a stick
Page 65 and 66:
‘With what did Juan hit the dog?
Page 67 and 68:
The unexpectedly bad candidate is (
Page 69 and 70:
Amharic (Semitic, Ethiopia) is a pr
Page 71 and 72:
31) Wïšša-w bä=Daniel bet atäg
Page 73 and 74:
A functional PP must contain a head
Page 75 and 76:
appear): A less well known example
Page 77 and 78:
appear on the non-recursive side of
Page 79 and 80:
Publications. Smith Stark, Thomas.
Page 81 and 82:
Abstract The development of large c
Page 83 and 84:
subj : spec : det : pred : the head
Page 85 and 86:
could annotate the CFG rules extrac
Page 87 and 88:
strings in the WSJ section of the P
Page 89 and 90:
Annotation coverage is measured in
Page 91 and 92:
Our grammars are extracted from sec
Page 93 and 94:
Penn-II training set is discarded f
Page 95 and 96:
grammatical functions are more akin
Page 97 and 98:
forms to semi-automatically develop
Page 99 and 100:
Predicate Argument Structure. In: P
Page 101 and 102:
Abstract PARTITIVE NOUN PHRASES IN
Page 103 and 104:
possessive partitives in (7) and (8
Page 105 and 106:
Extraction of constituents of dativ
Page 107 and 108:
iológiát [[az egyik-en]NP __ ]NP
Page 109 and 110:
Unlike post-verbal constituent orde
Page 111 and 112:
uncertainty. Adopting this idea, I
Page 113 and 114:
accounts for genitive possessor nou
Page 115 and 116:
(38) Hungarian TOPIC/SCOPE/FOCUS pa
Page 117 and 118:
The rule in (41) shows that the top
Page 119 and 120:
Payne, John and Erika Chisarik. (20
Page 121 and 122:
��
Page 123 and 124:
��
Page 125 and 126:
��
Page 127 and 128:
��
Page 129 and 130:
��
Page 131 and 132:
��
Page 133 and 134:
��
Page 135 and 136:
GETARUN PARSER Rodolfo Delmonte Abs
Page 137 and 138:
vital aid in the actual development
Page 139 and 140:
levels of representations required
Page 141 and 142:
sentence starts with a verb we let
Page 143 and 144:
adjunct_cp negat pro=CLI, {Case=acc
Page 145 and 146:
4. Parsing Strategies and P
Page 147 and 148:
qualified. The following list of pr
Page 149 and 150:
structure rules, after the head has
Page 151 and 152:
subordinators, and to arguments of
Page 153 and 154:
Hobbs J.R., D.E.Appelt, J.Bear, M.T
Page 155 and 156:
SECTION II: ITALIAN TEXTS o La stor
Page 157 and 158:
Mario criticava luigi perché ha ro
Page 159 and 160:
1. Introduction Since the beginning
Page 161 and 162:
. ⎡FOCUS ⎤ ⎢PRED ‘wonder (
Page 163 and 164:
(12) *ng- oingerang a mlarngii a be
Page 165 and 166:
(17) a. Hine ha- simla še kaniti.
Page 167 and 168:
have resumptive pronouns. Under the
Page 169 and 170:
way to license the same f-structure
Page 171 and 172:
(33) a. Robin met the man that/who
Page 173 and 174:
licensed in the normal way by funct
Page 175 and 176:
description of the situation. Borer
Page 177 and 178:
http://cslipublications.stanford.ed
Page 179 and 180:
A (Discourse) Functional Analysis o
Page 181 and 182:
A multi-factorial LFG analysis of a
Page 183 and 184:
(9) CP (↑ SUBJ)= ↓ ↑=↓ NP C
Page 185 and 186:
Word Order Properties SGF coordinat
Page 187 and 188:
4.1 Asymmetric Analyses Heycock and
Page 189 and 190:
(25) (a) S 〈 der Jäger ging in d
Page 191 and 192:
In sum, the Subject Functor Lineari
Page 193 and 194:
first and second conjunct are disti
Page 195 and 196:
In the Glue Semantics approach (see
Page 197 and 198: In the optimality-based model, mism
Page 199 and 200: complementiser (COMPL) may be exten
Page 201 and 202: BULGARIAN WORD ORDER AND THE ROLE O
Page 203 and 204: Bulgarian word order and the role o
Page 225 and 226: 1 Introduction This paper deals wit
Page 227 and 228: (4) To nifiko itan raf-to (*apo ton
Page 229 and 230: That is, in example (6) the partici
Page 231 and 232: (22) To spiti fenotan egataleli-men
Page 233 and 234: state denoted by the adjective appe
Page 235 and 236: in Modern Greek bears the intrinsic
Page 237 and 238: In addition, the lexical representa
Page 239 and 240: In this way, we capture the intuiti
Page 241 and 242: (57) To fagito ine magire-meno/*mag
Page 243 and 244: Corpus-based Learning in Stochastic
Page 245 and 246: LFG02 - Kuhn: Corpus-based Learning
Page 247: LFG02 - Kuhn: Corpus-based Learning
Page 263 and 264: 1. Introduction In this paper I wil
Page 265 and 266: This strategy has been outlined by
Page 267 and 268: The essence of the new approach is
Page 269 and 270: 3.2. Control relations in Hungarian
Page 271 and 272: complements of the postposition ut
Page 273 and 274: offered a more detailed discussion,
Page 275 and 276: lexical form which argument of this
Page 277 and 278: Berkeley. pp. 18. (On-line publicat
Page 279 and 280: 0. Abstract This paper attempts to
Page 281 and 282: 'understand', oppgi 'state', opplev
Page 283 and 284: This property of raising to object
Page 285 and 286: (32) ved/fra å svømme by/from to
Page 287 and 288: (45) *Til å sove ser han ut to (PR
Page 289 and 290: (57) Å gjøre det er blitt tenkt p
Page 291 and 292: Second argument: The preposition do
Page 293 and 294: In Lødrup (2002), I show that most
Page 295 and 296: Teleman, Ulf, Staffan Hellberg and
Page 297 and 298: Abstract Majority of Bantu language
Page 299 and 300:
No Object Agreement Bresnan and Mch
Page 301 and 302:
(7) a. N-ä-ï- `m-lyì-á k-èlyá
Page 303 and 304:
epresent 72% of the total of 312 se
Page 305 and 306:
We can characterize more concretely
Page 307 and 308:
(19) Markedness Subhierarchies a. *
Page 309 and 310:
4.1 Hierarchy of Morphological Feat
Page 311 and 312:
(30) Harmonic Alignment of (28) wit
Page 313 and 314:
topic-anaphoricity (as a bound pron
Page 315 and 316:
the present analysis highlights the
Page 317 and 318:
Grimshaw, Jane. 2001. Economy of st
Page 319 and 320:
CLITICS AND PHRASAL AFFIXATION IN C
Page 321 and 322:
(2) (a) ...wedi eu gweld nhw. ASP C
Page 323 and 324:
(7) IP (↑GF) = ↓ ↑ = ↓ NP I
Page 325 and 326:
(14) (a) Jovan je čitao knjigu. Jo
Page 327 and 328:
The Principle of Morphological Comp
Page 329 and 330:
3.2 Serbian Pronominal Clitics A se
Page 331 and 332:
Example (35) has the c-structure in
Page 333 and 334:
(43) (a) čovek-: N (↑PRED) = ‘
Page 335 and 336:
(52) (a) jučer: N (ADJ↑) (↑PRE
Page 337 and 338:
Case Marking and Subject Extraction
Page 339 and 340:
objects occur in VP or in I (“obj
Page 341 and 342:
Coordinated structures (14) Peter o
Page 343 and 344:
(20) Han hadde de trodd e ville kom
Page 345 and 346:
(28) a. * Peter forventer [I’ [I
Page 347 and 348:
One disadvantage of this approach i
Page 349 and 350:
6.2 Case Assignment and Empty Categ
Page 351 and 352:
Lets us consider the binding condit
Page 353 and 354:
claims that subject extraction is n
Page 355 and 356:
ham: N ( (SUBJ £ ) ¡ ((COMP SUBJ
Page 357 and 358:
[Vikner and Schwartz, 1996] Vikner,
Page 359 and 360:
. Mary-ka ttal-lul sakwa-lul mek-I-
Page 361 and 362:
In (4a), the object of the simple d
Page 363 and 364:
3. Morphological Causatives in Kore
Page 365 and 366:
. Mary-ka ttal-pakkey sakwa-ul an m
Page 367 and 368:
Firstly, objects in MC are asymmetr
Page 369 and 370:
The asymmetries of double accusativ
Page 371 and 372:
Abstract The paper describes a pres
Page 373 and 374:
construction - in good style. • I
Page 375 and 376:
Figure 4: Attributed adjectives (
Page 377 and 378:
Aspects of the syntax of psychologi
Page 379 and 380:
Despite the syntactic differences e
Page 381 and 382:
e. (exists independently of the eve
Page 383 and 384:
(10) Poco a poco, me enfadó que pe
Page 385 and 386:
to (3), we can thematically charact
Page 387 and 388:
we can derive the grammatical funct
Page 389 and 390:
‘This book interests him (John).
Page 391 and 392:
unaccusative verb, while interesar
Page 393 and 394:
Roegiest, E. (1999). Objet direct p
Page 395 and 396:
Abstract The aim of this paper is t
Page 397 and 398:
(6) Pierre lui a fait écrire une l
Page 399 and 400:
dative NPs that the examples (4) to
Page 401 and 402:
verb. In some languages, like Engli
Page 403 and 404:
(15) *J’ai convaincu de lui tél
Page 405 and 406:
VP � V + V + (NP) + (PP) (��
Page 407 and 408:
a number of signs (words or phrases
Page 409 and 410:
��
Page 411 and 412:
DAVIS, A. & P.-J. KOENIG (2000) Lin
Page 413 and 414:
1. Introduction Subsumption and Equ
Page 415 and 416:
an XCOMP attribute whose value is t
Page 417 and 418:
(8) a. S|VP Æ NP* (V’) ( S|VP )
Page 419 and 420:
inconsistency with erzählen. But i
Page 421 and 422:
We revise the fronted S|VP expansio
Page 423 and 424:
5. The interaction between PVPF and
Page 425 and 426:
(28) Ein Aussenseiter versuchte hie
Page 427 and 428:
. * Ein Aussenseiter gewonnen haben
Page 429 and 430:
The use of subsumption addresses fu
Page 431 and 432:
TIGER TRANSFER Utilizing LFG Parses
Page 433 and 434:
is an additional way of ensuring co
Page 435 and 436:
"tiger_4548: Hier herrscht Demokrat
Page 437 and 438:
y a multi-set of OT marks, thereby
Page 439 and 440:
sumably they cannot be fixed in pri
Page 441 and 442:
¢¡¤£¦¥¦§©¨�£¦§�¥
Page 443 and 444:
Macros and Templates The system all
Page 445 and 446:
Ambiguous Predicates A simple examp
Page 447 and 448:
HD−OUT TI−FORM herrscht, TI−M
Page 449 and 450:
(ignoring PP-attachment for the eva
Page 451:
Kameyama, Megumi, Ryo Ochitani, and
show all

Proceedings of the LFG 02 Conference National Technical - CSLI ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?