Proceedings - Österreichische Gesellschaft für Artificial Intelligence

More documents

Recommendations

Info

Compositionality in (high-dimensional) space Marco Baroni Università di Trento Abstract Formal semanticists have developed sophisticated compositional theories of sentential meaning, paying a lot of attention to those grammatical words (determiners, logical connectives, etc.) that constitute the functional scaffolding of sentences. Corpus-based computational linguists, on the other hand, have developed powerful distributional methods to induce representations of the meaning of contentrich words (nouns, verbs, etc.), typically discarding the functional scaffolding as "stop words". Since we do not communicate by logical formulas, nor, Tarzan-style, by flat concatenation of content words, a satisfactory model of the semantics of natural language should strike a balance between the two approaches. In this talk, I will present some recent proposals that try to get the best of both worlds by adapting the classic view of compositionality as function application developed by formal semanticists to distributional models of meaning. I will present preliminary evidence of the effectiveness of these methods in scaling up to the phrasal and sentential domains, and discuss to what extent the representations of phrases and sentences we get out of compositional distributional semantics are related to what formal semanticists are trying to capture. 20 <strong>Proceedings</strong> of KONVENS 2012 (Invited talks), Vienna, September 20, 2012
Data-driven Knowledge Extraction for the Food Domain Michael Wiegand, Benjamin Roth, Dietrich Klakow Spoken Language Systems Saarland University D-66123 Saarbrücken {michael.wiegand|benjamin.roth|dietrich.klakow}@lsv.uni-saarland.de Abstract In this paper, we examine methods to automatically extract domain-specific knowledge from the food domain from unlabeled natural language text. We employ different extraction methods ranging from surface patterns to co-occurrence measures applied on different parts of a document. We show that the effectiveness of a particular method depends very much on the relation type considered and that there is no single method that works equally well for every relation type. We also examine a combination of extraction methods and also consider relationships between different relation types. The extraction methods are applied both on a domain-specific corpus and the domain-independent factual knowledge base Wikipedia. Moreover, we examine an open-domain lexical ontology for suitability. 1 Introduction There has been only little research on natural language processing in the food domain even though there is a high commercial potential in automatically extracting knowledge involving food items. For example, such knowledge could be beneficial for virtual customer advice in a supermarket. The advisor might suggest products available in the shop that would potentially complement the items a customer has already in their shopping cart. Additionally, food items required for preparing a specific dish or typically consumed at a social occasion could be recommended. The advisor could also suggest an appropriate substitute for a product a customer would like to purchase if that product is out of stock. In this paper, we present methods to automatically extract knowledge from the food domain. We apply different relation extraction methods, such as simple manually designed surface patterns or statistical co-occurrence measures, on both a domain-specific corpus and the opendomain factual knowledge base Wikipedia. These large corpora are exclusively used as unlabeled data. In addition to the corpora, we also assess an open-domain lexical ontology. Moreover, we combine these methods and harness the relationship between different relation types. Since these methods only require a low level of linguistic processing, they have the advantage that they can provide responses in real time. We show that these individual methods have varying strength depending on which particular food relation type is considered. Our system has to solve the following task: It is given a partially instantiated relation, such as Ingredient-of(FOOD-ITEM=?, pancake). The system has to produce a ranked list of possible values that are valid arguments of the unspecified argument position. In the current example, this would correspond to listing ingredients that are necessary in order to prepare pancakes, such as eggs, flour, sugar and milk. The entities that are to be retrieved are always food items. Moreover, we only consider binary relations. The relation types we examine (such as Ingredient-of) are domain specific. 21 <strong>Proceedings</strong> of KONVENS 2012 (Main track: oral presentations), Vienna, September 19, 2012
Page 1 and 2: 11 th Conference on Natural Languag
Page 3 and 4: Table of Contents Main conference P
Page 5 and 6: Comparing variety corpora with vis-
Page 7 and 8: LexSem 2012: Workshop on Recent Dev
Page 9 and 10: KONVENS 2012 The 11 th Conference o
Page 11 and 12: Organizers: Jeremy Jancsary (ÖGAI/
Page 13 and 14: Ines Rehbein (Potsdam University) M
Page 15 and 16: Integrating viewpoints into newspap
Page 17 and 18: Phonetically aided syntactic parsin
Page 19: The Statistical Approach to Natural
Page 23 and 24: 4 The Different Relations Types In
Page 25 and 26: Relation Type #Patterns Examples Su
Page 27 and 28: Suits-to Served-with Substituted-by
Page 29 and 30: References Timothy Chklovski and Pa
Page 31 and 32: 1.2 Problem Definitions In a MRA, t
Page 33 and 34: oot person organisation product ...
Page 35 and 36: Group Organis. Persons Products all
Page 37 and 38: Method α α+β α+γ all SentiWS 0
Page 39 and 40: A Supervised POS Tagger for Written
Page 41 and 42: Our tagset mixes and matches tags a
Page 43 and 44: ecause the enclitic pronoun is an i
Page 45 and 46: stochastic rules as opposed to larg
Page 47 and 48: Kevin Duh, and Katrin Kirchhoff. 20
Page 49 and 50: forms, and that existentials can be
Page 51 and 52: Pseudo Verb Punctuation PNC
Page 53 and 54: NLP Workflow for On-line Definition
Page 55 and 56: nological phrases on the basis of p
Page 57 and 58: Figure 2: A screenshot of the entir
Page 59 and 60: When analyzing the evaluation sets
Page 61 and 62: Projecting Semantic Roles via Tai M
Page 63 and 64: quent prediction, return it (iii) i
Page 65 and 66: for the B/T/FT settings. This is pr
Page 67 and 68: ut it is implied by Table 3 and Fig
Page 69 and 70: References Wanxiang Che, Zhenghua L
Page 71 and 72:
Our ultimate goal is to develop com
Page 73 and 74:
ADORNING CAUSE_FLUIDIC_MOTION COTHE
Page 75 and 76:
English nouns, verbs, adjectives, a
Page 77 and 78:
considerably worse than FrameNet. T
Page 79 and 80:
Extraction. In Proceedings of the 1
Page 81 and 82:
lexicon from a document-derived voc
Page 83 and 84:
Initialize hypernym set H to contai
Page 85 and 86:
ulary distance. Let d(a, b) denote
Page 87 and 88:
the reduced lexicon (243) was neare
Page 89 and 90:
Adding nominal spice to SALSA - fra
Page 91 and 92:
SALSA verbs nouns 0 FE 0 2 1 FE 68
Page 93 and 94:
gen, which in one of its meanings i
Page 95 and 96:
In the next section we describe how
Page 97 and 98:
Acknowledgments This work was funde
Page 99 and 100:
most commonly used (Turney and Pant
Page 101 and 102:
anteed because the forward feature
Page 103 and 104:
5 000 feat ICA50 SVD50 SENNA Semcat
Page 105 and 106:
words bicycle and tricycle go with
Page 107 and 108:
of the Twelfth European Conference
Page 109 and 110:
incorporate the context information
Page 111 and 112:
those candidate phrases that are an
Page 113 and 114:
5.2 Evaluation Metrics for WPD The
Page 115 and 116:
Runs F-measures α=0.5 α=0.2 Pur.
Page 117 and 118:
Evaluations (SemEval-2007), June (2
Page 119 and 120:
2 Related Work The majority of pape
Page 121 and 122:
and Roth, 2009; Lin and Wu, 2009).
Page 123 and 124:
Features Dev Test CoNLL-2003 w 0 +
Page 125 and 126:
Feature Dev Test Test+ — 93.78 91
Page 127 and 128:
the Joint Conference of the 47th An
Page 129 and 130:
tion. It was first included as a sp
Page 131 and 132:
indicates the presence of the ellip
Page 133 and 134:
[Armada britainiarrak] [Ipar Irland
Page 135 and 136:
Información, (JOTRI 2003), pages 1
Page 137 and 138:
Phonetically aided Syntactic Parsin
Page 139 and 140:
mar network or PCFG. Grammar networ
Page 141 and 142:
phonetic confusion that has the con
Page 143 and 144:
in confusion in the network. This i
Page 145 and 146:
Linguistic analyses of the LAST MIN
Page 147 and 148:
Table 1: Comparison between corpora
Page 149 and 150:
Table 2: Most frequent POS patterns
Page 151 and 152:
Table 3: A problematic dialog situa
Page 153 and 154:
6 Summary We have presented the cur
Page 155 and 156:
S-restricted monotone alignments St
Page 157 and 158:
s k n i i: f phh ohehnh ih xh s k n
Page 159 and 160:
Dataset µ G µ P P [G>15] P [P >15
Page 161 and 162:
Perplexity H(L | P ) 2PKE-Uni 7.002
Page 163 and 164:
(1, 1) (2, 1) (3, 1) (4, 1) (1, 2)
Page 165 and 166:
Use of linguistic features for impr
Page 167 and 168:
whereas English has a verb-initial
Page 169 and 170:
Accuracy of Accuracy of Accuracy of
Page 171 and 172:
English Persian Train: Sentences 30
Page 173 and 174:
to compensate for sparsity of paral
Page 175 and 176:
Name # Documents # Tokens # Lemmas
Page 177 and 178:
Figure 1: Precision-Recall graphs c
Page 179 and 180:
A Multi-level Annotation Model for
Page 181 and 182:
and sentence level with the aim of
Page 183 and 184:
(Schmid, 1995; Schiller et al., 199
Page 185 and 186:
EI EII Tag AKR MARK Typography ITER
Page 187 and 188:
References Saima Aman and Stan Szpa
Page 189 and 190:
Robust processing of noisy web-coll
Page 191 and 192:
1. first strip of the papery skin o
Page 193 and 194:
References Jon Chamberlain, Massimo
Page 195 and 196:
ments between its LSRs. Additionall
Page 197 and 198:
Figure 3: In the sense comparison v
Page 199 and 200:
Using Information Retrieval Technol
Page 201 and 202:
against every tree. Others avoid th
Page 203 and 204:
A 0.780715 NN Figure 4: A sp
Page 205 and 206:
# #Documents Index Size 1 29,704 1.
Page 207 and 208:
References P. Bański, C. Belica, H
Page 209 and 210:
(2012), in their corpus study of va
Page 211 and 212:
DET NSUBJ ADV OBJA AMOD Seine Tante
Page 213 and 214:
A Computational Semantic Analysis o
Page 215 and 216:
(6) a. Is Spain very large? b. Is B
Page 217 and 218:
References Alan C. Bale. 2008. A Un
Page 219 and 220:
219 Proceedings of KONVENS 2012 (Ma
Page 221 and 222:
221 Proceedings of KONVENS 2012 (Ma
Page 223 and 224:
Using subcategorization frames to i
Page 225 and 226:
5.2 Word clustering based on the Le
Page 227 and 228:
erage of english grammars. In Proce
Page 229 and 230:
cases not disambiguated using langu
Page 231 and 232:
Not surprisingly, the results show
Page 233 and 234:
Automatic Identification of Languag
Page 235 and 236:
Differences also appear at the synt
Page 237 and 238:
Acknowledgments The authors express
Page 239 and 240:
2 Related Work Previous work on POS
Page 241 and 242:
The XY tag is used for all non-word
Page 243 and 244:
Comparing Variety Corpora with Vis-
Page 245 and 246:
INPUT variety corpus comparability
Page 247 and 248:
References Andrea Abel and Stefanie
Page 249 and 250:
Features Context features (token wi
Page 251 and 252:
0.7 0.7 0.6 0.6 0.5 0.5 F-measure 0
Page 253 and 254:
ties into fine-grained classes. In
Page 255 and 256:
verbials (e.g., therefore, still).
Page 257 and 258:
word raw freq. non conn. conn. (of
Page 259 and 260:
Corpus-based Acquisition of German
Page 261 and 262:
Events Objects IPrec CPrec IPrec CP
Page 263 and 264:
References Marco Baroni, Silvia Ber
Page 265 and 266:
Preface The present proceedings con
Page 267 and 268:
Workshop Program Friday, September
Page 269 and 270:
Unsupervised Sentiment Analysis wit
Page 271 and 272:
pling (Geman and Geman, 1984) to es
Page 273 and 274:
References Anders Björkelund, Bern
Page 275 and 276:
Class # Events in % neutral 45’01
Page 277 and 278:
3.1.1 Rules Let t 1..n be tokens of
Page 279 and 280:
Human 6 System 6 Human 4 System 4 A
Page 281 and 282:
Codebuch Akteursanalyse. rev. 2011.
Page 283 and 284:
of our work in the whole applicatio
Page 285 and 286:
3.4 Negation and graduation Valence
Page 287 and 288:
local context. For instance, in une
Page 289 and 290:
n’était pas très favorable au G
Page 291 and 292:
Domain-specific variation of sentim
Page 293 and 294:
query building blocks comments real
Page 295 and 296:
References Monica Bednarek. 2006. E
Page 297 and 298:
3 Basic Methodology of Annotation T
Page 299 and 300:
(3) A. V případě jeho kandidatur
Page 301 and 302:
4.3 Auxiliary data - Mall.cz We hav
Page 303 and 304:
Overlap f-score Source 0.750 Polar
Page 305 and 306:
A Phrase-Based Opinion List for the
Page 307 and 308:
the Stuttgart-Tübingen Tagset (STT
Page 309 and 310:
As for noun-based opinion phrases a
Page 311 and 312:
More than half of the words and phr
Page 313 and 314:
treebank. In Proceedings of the Wor
Page 315 and 316:
Apart from these, we distinguish op
Page 317 and 318:
close variants)”. In fact importa
Page 319 and 320:
What is the Meaning of 5 *’s? An
Page 321 and 322:
For each category, we follow (Const
Page 323 and 324:
Positivity Term 2.87692577130 elske
Page 325 and 326:
In recent years, the use of rating
Page 327 and 328:
KONVENS 2012 LThist 2012: First Int
Page 329 and 330:
Organisation Committee: Thierry Dec
Page 331 and 332:
Page 333 and 334:
Rule-Based Normalisation of Histori
Page 335 and 336:
isation rules were applied to the i
Page 337 and 338:
ments in the gold standard have bee
Page 339 and 340:
cation in the contemporary SUC samp
Page 341 and 342:
proach. References Maria Ågren, Ro
Page 343 and 344:
2 Related Work There has been consi
Page 345 and 346:
vided to ensure that they are norma
Page 347 and 348:
alects, or even languages. Norma ca
Page 349 and 350:
ization (rather than modernization)
Page 351 and 352:
The Sketch Engine as infrastructure
Page 353 and 354:
2.1 The Corpus of English Dialogues
Page 355 and 356:
Other notable contrasts are that 17
Page 357 and 358:
Evaluating a Post-editing Approach
Page 359 and 360:
Figure 1: Example of a marriage lic
Page 361 and 362:
all possible concatenations of indi
Page 363 and 364:
Table 3: Post-editing results. Time
Page 365 and 366:
okstaffua, bokstaffwa, bokstafwa, b
Page 367 and 368:
Rule Cost v↔u 0.1 v↔w 0.1 u↔o
Page 369 and 370:
graphic word boundaries and conside
Page 371 and 372:
ation. Gulordava and Baroni (2011),
Page 373 and 374:
ut adding a statistical measure to
Page 375 and 376:
e.g. philosophical and political pr
Page 377 and 378:
of the EACL 2009 Workshop on GEMS:
Page 379 and 380:
Most of the research on genre class
Page 381 and 382:
Precision Recall F 1 Unigrams 0.551
Page 383 and 384:
The DTA ‘base format’: A TEI-Su
Page 385 and 386:
Lib are complementary in the sense
Page 387 and 388:
inline metadata solutions for edito
Page 389 and 390:
ing. The fourth level contains an e
Page 391 and 392:
the work package ‘Resources and T
Page 393 and 394:
is mainly used with parallel corpor
Page 395 and 396:
Corrections by the editor were omit
Page 397 and 398:
Accuracy N % All Words 781/1000 78.
Page 399 and 400:
$$ $$ $$ $$ $$ $$ $$ $$ $$ Figure 4
Page 401 and 402:
Digitised Historical Text: Does it
Page 403 and 404:
3 Related Work Previous research on
Page 405 and 406:
5.1 End-of-line Soft Hyphen Deletio
Page 407 and 408:
quite difficult to get an idea of h
Page 409 and 410:
References Ahmad Abdulkader and Mat
Page 411 and 412:
cision, recall, and F1. Section 4 d
Page 413 and 414:
not corrected corrected p P R F1 P
Page 415 and 416:
From Semi-Automatic to Automatic Af
Page 417 and 418:
Figure 2: ER-Diagramm of the Morphi
Page 419 and 420:
Figure 4: Morphilizer’s implement
Page 421 and 422:
Figure 6: Architecture of Morphorm
Page 423 and 424:
References Harald R. Baayen, R. Pie
Page 425 and 426:
edition has been adopted. This is d
Page 427 and 428:
marginalia, emendations or punctuat
Page 429 and 430:
Word Lemma Class Subclass Type Tens
Page 431 and 432:
Figure 3. Sample screenshot of comp
Page 433 and 434:
KONVENS 2012 LexSem 2012: Workshop
Page 435 and 436:
English) closely and discusses the
Page 437 and 438:
Page 439 and 440:
verbs were collected mostly from an
Page 441 and 442:
3 Theoretical Background Let us now
Page 443 and 444:
are sensible to semantic properties
Page 445 and 446:
Cançado, Márcia and Amaral, Luana
Page 447 and 448:
morphological tagger for Russian la
Page 449 and 450:
whom/which something is taken. (7cz
Page 451 and 452:
Verb class same frame different fra
Page 453 and 454:
Construction Grammar are introduced
Page 455 and 456:
posed to Hon åker på måndag ‘S
Page 457 and 458:
nitions of Agent, for example. Inst
Page 459 and 460:
Framenet++, VR dnr 2010-6013), by t
Page 461 and 462:
Källström and Inger Lindberg, edi
Page 463 and 464:
scoring function σ : W ×L×W →
Page 465 and 466:
esults in the triple 〈heutige Son
Page 467 and 468:
Unlexicalized links 7,833,635 n (i)
Page 469 and 470:
The difference in accuracies betwee
Page 471 and 472:
Towards high-accuracy bilingual phr
Page 473 and 474:
over a set of candidate translation
Page 475 and 476:
each soft restriction soft m set on
Page 477 and 478:
(> 98% in scalable precision) and a
Page 479 and 480:
References Colin Bannard and Chris
Page 481 and 482:
Preface The need for automatically
Page 483 and 484:
15:50-16:15 Coffee break 16:15-16:4
Page 485 and 486:
Standards for the technical infrast
Page 487 and 488:
⎡ ⎤ phrase CATEGORY np ⎡ ⎤
Page 489 and 490:
The representation of the feature v
Page 491 and 492:
ement, while the interpretation as
Page 493 and 494:
hand, checking the validity of FSR
Page 495 and 496:
Standardizing metadata descriptions
Page 497 and 498:
2 LMF and semantic interoperability
Page 499 and 500:
can register and create DCs. Users
Page 501 and 502:
all kinds of lexemes, (see Cruse (1
Page 503 and 504:
constraints, because many aspects o
Page 505 and 506:
Menzo Windhouwer. 2012. RELcat: a R
Page 507 and 508:
the meeting in Rome as a starting p
Page 509 and 510:
As these documents provide the refe
Page 511 and 512:
Two multi-layer annotated corpora r
Page 513 and 514:
As audio data is not directly store
Page 515 and 516:
Jörg Mayer. 1995. Transcription of
Page 517 and 518:
Towards standardized lexical semant
Page 519:
Getting involved into language reso
show all

Proceedings - Österreichische Gesellschaft für Artificial Intelligence

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

Delete template?

Save as template?