README-distrib 13.8 KB
Newer Older
1 2 3 4 5 6
-------------------------------------------------------
Deep Sequoia corpus v7.0
-------------------------------------------------------
november 2015

The corpus contains French sentences, from Europarl, Est Republicain newspaper, 
7
French Wikipedia and European Medicine Agency, with the following manual annotations: 
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

- parts-of-speech and morphological features 
- grammatical compound words (merged as one token)
- surface syntax (dependencies and constituents)
- deep syntax (dependencies only)


1. Licence
2. History of the corpus
3. References
4. Content of the release, and formats
5. Appendix

** contact : sequoia@inria.fr

** web site : deep-sequoia.inria.fr

** annotation guide : http://passage.inria.fr/deepwiki/node/19 (in French)

------------------------------------------------------
1. Licence
------------------------------------------------------
The corpus is freely available under the free licence LGPL-LR 
(Lesser General Public License For Linguistic Resources)
32
 cf. http://deep-sequoia.inria.fr/lgpl-lr/
33 34 35 36 37 38 39 40

------------------------------------------------------
2. History of the corpus
------------------------------------------------------

The Sequoia corpus was first manually annotated for part-of-speech and phrase-structure, and automatically converted to surface syntactic dependency trees.
(Candito and Seddah, 2012a).
The phrase-structure annotation follows mainly the French Treebank guidelines
41
(http://www.llf.cnrs.fr/Gens/Abeille/French-Treebank-fr.php),
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
modified in the context of conversion to dependencies:
   - prepositions that dominate a infinitival VP do project a PP
   - any sentence introduced by a complementizer (CS tag) is grouped into a Sint constituent


A further step of manual annotation was carried out, aiming at correcting the governor of extracted elements:
in case of long-distance dependency, the automatic conversion from constituents to dependencies picks out
a wrong governor for the extracted element.
These were manually corrected, leading to a few non-projective links.
(Candito and Seddah, 2012b).

Then a collaboration started in 2013 between the Alpage and Sémagramme teams, 
to obtain DEEP SYNTACTIC DEPENDENCIES on top of surface dependencies.
The main characteristics of the deep syntactic annotation scheme are:
(i) explicitation of subjects of non finite verbs and of adjectives
(ii) neutralization of diathesis alternations
(iii) distribution of dependents over coordinated governors

This led to a first release of the Deep Sequoia corpus (v 1.0)
(Candito et al., 2014; Perrier et al., 2014). 
Annotating the corpus for deep syntax has sometimes led to correct some surface dependencies.

A further step of systematic search for inconsistencies was carried out,
65
using the Grew system (http://grew.loria.fr).
66 67 68 69
This led to the current release (7.0).
(Note: the current release number (7.0) was chosen to get same version numbers for the surface and the deep syntactic annotations of the corpus)

The deep sequoia corpus and the surface sequoia corpus contain the same 3099 sentences,
70
but note that the original surface corpus (versions prior to 6.0) contained 101 more sentences, that turned out to be duplicates and were thus 
71 72 73 74 75 76 77
subsequently removed (from the EMEA-test part of the corpus).
See the appendix for the ids of the removed sentences.

------------------------------------------------------
3. References
------------------------------------------------------

78
** Deep syntactic annotation: 
79 80 81 82 83 84

- Marie Candito, Guy Perrier, Bruno Guillaume, Corentin Ribeyre, Karën Fort, Djamé Seddah and Éric de la Clergerie. (2014) Deep Syntax Annotation of the Sequoia French Treebank. Proc. of LREC 2014, Reykjavic, Iceland.

- Guy Perrier, Marie Candito, Bruno Guillaume, Corentin Ribeyre, Karën Fort and Djamé Seddah. (2014) Un schéma d’annotation en dépendances syntaxiques profondes pour le français. Proc. of TALN 2014, Marseille, France.


85
** Original paper (surface syntactic annotation):
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122

Candito M. and Seddah D., 2012a : "Le corpus Sequoia : annotation syntaxique et exploitation pour l’adaptation d’analyseur par pont lexical", Actes de TALN'2012, Grenoble, France

Candito M. and Seddah D., 2012b : "Effectively long-distance dependencies in French : annotation and parsing evaluation", Proceedings of TLT'11, 2012, Lisbon, Portugal)

------------------------------------------------------
4. Content and formats
------------------------------------------------------

The corpus contains 3099 sentences.

Number of sentences for each sub-domain :
----------------------------------------

561 sentences	Europarl	 file= Europar.550+fct.mrg	
529 sentences	EstRepublicain   file= annodis.er+fct.mrg 
996 sentences	French Wikipedia file= frwiki_50.1000+fct.mrg
574 sentences	EMEA (dev)  	 file= emea-fr-dev+fct.mrg
544 sentences	EMEA (test) 	 file= emea-fr-test+fct.mrg, among which 101 were removed (because duplicates) in surface version 6.0 and 1.0 deep version.

Tokenization :
--------------
Current tokenization contains some multi-word expressions.
More precisely, only grammatical (aka functional) multi-word were recognized as such, and treated as one token (components separated with an underscore, as in "parce_que").
Other multi-word expressions (e.g. nominal) are not represented (i.e. "pomme de terre" appears as three tokens).

The release contains :

******************
sequoia.surf.conll
******************
CoNLL format for surface syntax, augmented with "comments" lines (starting with a #)
A few arcs are non projective.
This arise in some cases of extraction (that can be traces by searching the "fctpath" attribute).

The first token of each sentence contains a sentid attribute for the sentence id.

bguillaum's avatar
bguillaum committed
123 124 125 126 127 128 129 130 131 132 133 134
For instance: 
<------------------------------------ annodis.er_00449 [surf] ----------------------------------->
# sentence-text: Une prochaine réunion est prévue mardi
1	Une	un	D	DET	g=f|n=s|s=ind|sentid=annodis.er_00449	3	det	_	_
2	prochaine	prochain	A	ADJ	g=f|n=s|s=qual	3	mod	_	_
3	réunion	réunion	N	NC	g=f|n=s|s=c	5	suj	_	_
4	est	être	V	V	m=ind|n=s|p=3|t=pst	5	aux.pass	_	_
5	prévue	prévoir	V	VPP	g=f|m=part|n=s|t=past	0	root	_	_
6	mardi	mardi	N	NC	g=m|n=s|s=c	5	mod	_	_
<------------------------------------------------------------------------------------------------>
Graphical representation of the 3 formats is available at: http://talc2.loria.fr/deep-sequoia/sequoia-7.0/html/annodis.er_00449.html

135 136 137 138 139 140
******************
sequoia.deep.conll
******************
Extended CoNLL format for deep syntax:
The CoNLL format still contains one token per line, but a given token may have several governors:
this is represented by column 7 containing several labels, and column 8 containing several governor ids.
bguillaum's avatar
bguillaum committed
141 142 143
The number of labels in column 7 is the same as the number of governors in column 8 and they are interpreted in parallel:
first label in column 7 corresponds to first governor in column 8, etc.

144 145 146 147
For instance, the noun "réunion" in the example below is both the (canonical) object of token 5 (prévue), 
and the subject of token 2 (prochaine).

Further the semantically void tokens are present in the deep format for readability (and parsability):
148
they are not strictly speaking belonging to the deep representation, and appear with a "0" in column 7 and "void" in column 8,
149 150
and a void=y feature (e.g token 4 below).
 
bguillaum's avatar
bguillaum committed
151 152 153 154 155
<------------------------------------ annodis.er_00449 [deep] ----------------------------------->
# sentence-text: Une prochaine réunion est prévue mardi
1	Une	un	D	DET	g=f|n=s|s=ind|sentid=annodis.er_00449	3	det	_	_
2	prochaine	prochain	A	ADJ	g=f|n=s|s=qual	3	mod	_	_
3	réunion	réunion	N	NC	g=f|n=s|s=c	5|2	obj|suj	_	_
156
4	est	être	V	V	dl=être|m=ind|n=s|p=3|t=pst|void=y	0	void	_	_
bguillaum's avatar
bguillaum committed
157 158 159 160 161 162 163 164
5	prévue	prévoir	V	VPP	diat=passif|dl=prévoir|dm=ind|g=f|m=part|n=s|t=past	0	root	_	_
6	mardi	mardi	N	NC	g=m|n=s|s=c	5	mod	_	_
<------------------------------------------------------------------------------------------------>

***************************
sequoia.deep_and_surf.conll
***************************
A compact representation of the two previous formats:
165
- relations subject to diathesis alternations contain a double label (ex: "suj:obj" stands for final subject and canonical object)
bguillaum's avatar
bguillaum committed
166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
- relations that are only in the deep representation are prefixed by "D:" (ex: "D:suj:suj")
- relations that are only in the surf representation are prefixed by "S:" (ex: "S:aux.pass")
- a special feature "void=y" is used for tokens removed in the deep syntax (ex: token 4 below)

The surf format is obtained by removing relations prefixed by "D:" and by taking the first part of double labels.
The deep format is obtained by removing relations prefixed by "S:" and by taking the second part of double labels.

For instance: 
<------------------------------- annodis.er_00449 [deep_and_surf] ------------------------------->
# sentence-text: Une prochaine réunion est prévue mardi
1	Une	un	D	DET	g=f|n=s|s=ind|sentid=annodis.er_00449	3	det	_	_
2	prochaine	prochain	A	ADJ	g=f|n=s|s=qual	3	mod	_	_
3	réunion	réunion	N	NC	g=f|n=s|s=c	5|2	suj:obj|D:suj:suj	_	_
4	est	être	V	V	dl=être|m=ind|n=s|p=3|t=pst|void=y	5	S:aux.pass	_	_
5	prévue	prévoir	V	VPP	diat=passif|dl=prévoir|dm=ind|g=f|m=part|n=s|t=past	0	root	_	_
6	mardi	mardi	N	NC	g=m|n=s|s=c	5	mod	_	_
<------------------------------------------------------------------------------------------------>
183

bguillaum's avatar
bguillaum committed
184
******************
185
sequoia.surf.const
bguillaum's avatar
bguillaum committed
186
******************
187 188 189 190 191 192 193 194 195 196 197 198 199
Constituency trees, in bracketed format.

NB: before deep syntactic annotation of the sentences,
the dependency trees were obtained by automatic conversion from the constituency trees.
But when the sentences were annotated with deep dependency syntax,
some "surface" annotation was corrected.
HENCE: the isomorphism between the surface dependency trees and the constituency trees no longer holds.
(Constituency trees are isomorphic to surface dependencies up to version 5.2)

Grammatical functions for dependents of verbs are appended to the non terminal symbol, after a "-"
(example NP-SUJ stands for a subject NP).


bguillaum's avatar
bguillaum committed
200
*********************
201
sequoia.predeep.const
bguillaum's avatar
bguillaum committed
202
*********************
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
A few manual deep syntactic annotations were added to the bracketed format.
NB: this is only a very small part of the phenomena handled in the deep dependency format.

With respect to FTB bracketed format, some grammatical functions have been specialized
P_OBJ has been split to P_OBJ.O and P_OBJ.AGT
MOD has been split to MOD, MOD.APP, MOD.INC, MOD.CLEFT

- non referential "il" : 
(CLS-SUJ##_@@void=y il) is used for a non referential "il", entering an impersonal alternation 
(meaning the verb alternates with a diathesis in which the subject is referential : Il arrive trois personnes <=> Trois personnes arrivent)

(CLS-SUJ##_@@void=y,intrinsimp=y Il) is used for a non referential "il", without syntactic alternation
(e.g.  "il faut 3 personnes" does not alternate with a construction of the same verb with referential subject ("*3 personnes faut")


- causative constructions
The canonical grammatical functions in case of causative verb complex have been marked
E.g : 

( (SENT (NP-SUJ##ARGC (DET Le) (NC conteur)) (VN (V a) (VPP fait) (VINF@@diat=causi jouer)) (NP-OBJ##SUJ (DET les) (NC enfants)) (PONCT .)))

The final subject is canonical "argc" (causer argument), and the final object (les enfants) is the canonical subject of the "jouer" verb.


 
----------------------------------
5. Appendix
----------------------------------

** Data split (TALN 2012 experiments)

The "neutral" domain is made of EstRepublicain + Europarl + FrWiki,
and the split into dev and test sets is the following :

head -265 annodis.er+fct.mrg >> sequoia-neutre-dev+fct.mrg
head -280 Europar.550+fct.mrg >> sequoia-neutre-dev+fct.mrg
head -498 frwiki_50.1000+fct.mrg >> sequoia-neutre-dev+fct.mrg

tail -264 annodis.er+fct.mrg >> sequoia-neutre-test+fct.mrg
tail -281 Europar.550+fct.mrg >> sequoia-neutre-test+fct.mrg
tail -498 frwiki_50.1000+fct.mrg >> sequoia-neutre-test+fct.mrg

** Appendix : duplicate sentences removed in version 6.0

< emea-fr-test_00301
< emea-fr-test_00302
< emea-fr-test_00303
< emea-fr-test_00304
< emea-fr-test_00305
< emea-fr-test_00306
< emea-fr-test_00307
< emea-fr-test_00308
< emea-fr-test_00309
< emea-fr-test_00310
< emea-fr-test_00311
< emea-fr-test_00312
< emea-fr-test_00313
< emea-fr-test_00314
< emea-fr-test_00315
< emea-fr-test_00316
< emea-fr-test_00317
< emea-fr-test_00318
< emea-fr-test_00319
< emea-fr-test_00320
< emea-fr-test_00321
< emea-fr-test_00322
< emea-fr-test_00323
< emea-fr-test_00324
< emea-fr-test_00325
< emea-fr-test_00326
< emea-fr-test_00327
< emea-fr-test_00328
< emea-fr-test_00329
< emea-fr-test_00330
< emea-fr-test_00331
< emea-fr-test_00332
< emea-fr-test_00333
< emea-fr-test_00334
< emea-fr-test_00335
< emea-fr-test_00336
< emea-fr-test_00337
< emea-fr-test_00338
< emea-fr-test_00339
< emea-fr-test_00340
< emea-fr-test_00341
< emea-fr-test_00342
< emea-fr-test_00343
< emea-fr-test_00344
< emea-fr-test_00345
< emea-fr-test_00346
< emea-fr-test_00347
< emea-fr-test_00348
< emea-fr-test_00349
< emea-fr-test_00350
< emea-fr-test_00351
< emea-fr-test_00352
< emea-fr-test_00353
< emea-fr-test_00354
< emea-fr-test_00355
< emea-fr-test_00356
< emea-fr-test_00357
< emea-fr-test_00358
< emea-fr-test_00359
< emea-fr-test_00360
< emea-fr-test_00361
< emea-fr-test_00362
< emea-fr-test_00363
< emea-fr-test_00364
< emea-fr-test_00365
< emea-fr-test_00366
< emea-fr-test_00367
< emea-fr-test_00368
< emea-fr-test_00369
< emea-fr-test_00370
< emea-fr-test_00371
< emea-fr-test_00372
< emea-fr-test_00373
< emea-fr-test_00374
< emea-fr-test_00375
< emea-fr-test_00376
< emea-fr-test_00377
< emea-fr-test_00378
< emea-fr-test_00379
< emea-fr-test_00380
< emea-fr-test_00381
< emea-fr-test_00382
< emea-fr-test_00383
< emea-fr-test_00384
< emea-fr-test_00385
< emea-fr-test_00386
< emea-fr-test_00387
< emea-fr-test_00388
< emea-fr-test_00389
< emea-fr-test_00390
< emea-fr-test_00391
< emea-fr-test_00392
< emea-fr-test_00393
< emea-fr-test_00394
< emea-fr-test_00395
< emea-fr-test_00396
< emea-fr-test_00397
< emea-fr-test_00398
< emea-fr-test_00399
< emea-fr-test_00400