declan101 commited on
Commit
cfb61d6
·
verified ·
1 Parent(s): a0ff569

Upload W03-0419.txt

Browse files
Files changed (1) hide show
  1. W03-0419.txt +929 -0
W03-0419.txt ADDED
@@ -0,0 +1,929 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Introduction to the CoNLL-2003 Shared Task:
2
+ Language-Independent Named Entity Recognition
3
+ Erik F. Tjong Kim Sang and Fien De Meulder
4
+ CNTS - Language Technology Group
5
+ University of Antwerp
6
+ {erikt,fien.demeulder}@uia.ua.ac.be
7
+
8
+ Abstract
9
+ We describe the CoNLL-2003 shared task:
10
+ language-independent named entity recognition. We give background information on
11
+ the data sets (English and German) and
12
+ the evaluation method, present a general
13
+ overview of the systems that have taken
14
+ part in the task and discuss their performance.
15
+
16
+ of the 2003 shared task have been offered training
17
+ and test data for two other European languages:
18
+ English and German. They have used the data
19
+ for developing a named-entity recognition system
20
+ that includes a machine learning component. The
21
+ shared task organizers were especially interested in
22
+ approaches that made use of resources other than
23
+ the supplied training data, for example gazetteers
24
+ and unannotated data.
25
+
26
+ 2
27
+ 1
28
+
29
+ Introduction
30
+
31
+ Named entities are phrases that contain the names
32
+ of persons, organizations and locations. Example:
33
+ [ORG U.N. ] official [PER Ekeus ] heads for
34
+ [LOC Baghdad ] .
35
+ This sentence contains three named entities: Ekeus
36
+ is a person, U.N. is a organization and Baghdad is
37
+ a location. Named entity recognition is an important task of information extraction systems. There
38
+ has been a lot of work on named entity recognition,
39
+ especially for English (see Borthwick (1999) for an
40
+ overview). The Message Understanding Conferences
41
+ (MUC) have offered developers the opportunity to
42
+ evaluate systems for English on the same data in a
43
+ competition. They have also produced a scheme for
44
+ entity annotation (Chinchor et al., 1999). More recently, there have been other system development
45
+ competitions which dealt with different languages
46
+ (IREX and CoNLL-2002).
47
+ The shared task of CoNLL-2003 concerns
48
+ language-independent named entity recognition. We
49
+ will concentrate on four types of named entities:
50
+ persons, locations, organizations and names of
51
+ miscellaneous entities that do not belong to the previous three groups. The shared task of CoNLL-2002
52
+ dealt with named entity recognition for Spanish and
53
+ Dutch (Tjong Kim Sang, 2002). The participants
54
+
55
+ Data and Evaluation
56
+
57
+ In this section we discuss the sources of the data
58
+ that were used in this shared task, the preprocessing
59
+ steps we have performed on the data, the format of
60
+ the data and the method that was used for evaluating
61
+ the participating systems.
62
+ 2.1
63
+
64
+ Data
65
+
66
+ The CoNLL-2003 named entity data consists of eight
67
+ files covering two languages: English and German1 .
68
+ For each of the languages there is a training file, a development file, a test file and a large file with unannotated data. The learning methods were trained with
69
+ the training data. The development data could be
70
+ used for tuning the parameters of the learning methods. The challenge of this year’s shared task was
71
+ to incorporate the unannotated data in the learning
72
+ process in one way or another. When the best parameters were found, the method could be trained on
73
+ the training data and tested on the test data. The
74
+ results of the different learning methods on the test
75
+ sets are compared in the evaluation of the shared
76
+ task. The split between development data and test
77
+ data was chosen to avoid systems being tuned to the
78
+ test data.
79
+ The English data was taken from the Reuters Corpus2 . This corpus consists of Reuters news stories
80
+ 1
81
+
82
+ Data files (except the words) can be found on
83
+ http://lcg-www.uia.ac.be/conll2003/ner/
84
+ 2
85
+ http://www.reuters.com/researchandstandards/
86
+
87
+ English data
88
+ Training set
89
+ Development set
90
+ Test set
91
+
92
+ Articles
93
+ 946
94
+ 216
95
+ 231
96
+
97
+ Sentences
98
+ 14,987
99
+ 3,466
100
+ 3,684
101
+
102
+ Tokens
103
+ 203,621
104
+ 51,362
105
+ 46,435
106
+
107
+ English data
108
+ Training set
109
+ Development set
110
+ Test set
111
+
112
+ LOC
113
+ 7140
114
+ 1837
115
+ 1668
116
+
117
+ MISC
118
+ 3438
119
+ 922
120
+ 702
121
+
122
+ ORG
123
+ 6321
124
+ 1341
125
+ 1661
126
+
127
+ PER
128
+ 6600
129
+ 1842
130
+ 1617
131
+
132
+ German data
133
+ Training set
134
+ Development set
135
+ Test set
136
+
137
+ Articles
138
+ 553
139
+ 201
140
+ 155
141
+
142
+ Sentences
143
+ 12,705
144
+ 3,068
145
+ 3,160
146
+
147
+ Tokens
148
+ 206,931
149
+ 51,444
150
+ 51,943
151
+
152
+ German data
153
+ Training set
154
+ Development set
155
+ Test set
156
+
157
+ LOC
158
+ 4363
159
+ 1181
160
+ 1035
161
+
162
+ MISC
163
+ 2288
164
+ 1010
165
+ 670
166
+
167
+ ORG
168
+ 2427
169
+ 1241
170
+ 773
171
+
172
+ PER
173
+ 2773
174
+ 1401
175
+ 1195
176
+
177
+ Table 1: Number of articles, sentences and tokens in
178
+ each data file.
179
+
180
+ Table 2: Number of named entities per data file
181
+ 2.3
182
+
183
+ between August 1996 and August 1997. For the
184
+ training and development set, ten days’ worth of data
185
+ were taken from the files representing the end of August 1996. For the test set, the texts were from December 1996. The preprocessed raw data covers the
186
+ month of September 1996.
187
+ The text for the German data was taken from the
188
+ ECI Multilingual Text Corpus3 . This corpus consists
189
+ of texts in many languages. The portion of data that
190
+ was used for this task, was extracted from the German newspaper Frankfurter Rundshau. All three of
191
+ the training, development and test sets were taken
192
+ from articles written in one week at the end of August 1992. The raw data were taken from the months
193
+ of September to December 1992.
194
+ Table 1 contains an overview of the sizes of the
195
+ data files. The unannotated data contain 17 million
196
+ tokens (English) and 14 million tokens (German).
197
+ 2.2
198
+
199
+ Data preprocessing
200
+
201
+ The participants were given access to the corpus after some linguistic preprocessing had been done: for
202
+ all data, a tokenizer, part-of-speech tagger, and a
203
+ chunker were applied to the raw data. We created
204
+ two basic language-specific tokenizers for this shared
205
+ task. The English data was tagged and chunked by
206
+ the memory-based MBT tagger (Daelemans et al.,
207
+ 2002). The German data was lemmatized, tagged
208
+ and chunked by the decision tree tagger Treetagger
209
+ (Schmid, 1995).
210
+ Named entity tagging of English and German
211
+ training, development, and test data, was done by
212
+ hand at the University of Antwerp. Mostly, MUC
213
+ conventions were followed (Chinchor et al., 1999).
214
+ An extra named entity category called MISC was
215
+ added to denote all names which are not already in
216
+ the other categories. This includes adjectives, like
217
+ Italian, and events, like 1000 Lakes Rally, making it
218
+ a very diverse category.
219
+ 3
220
+
221
+ http://www.ldc.upenn.edu/
222
+
223
+ Data format
224
+
225
+ All data files contain one word per line with empty
226
+ lines representing sentence boundaries. At the end
227
+ of each line there is a tag which states whether the
228
+ current word is inside a named entity or not. The
229
+ tag also encodes the type of named entity. Here is
230
+ an example sentence:
231
+ U.N.
232
+ official
233
+ Ekeus
234
+ heads
235
+ for
236
+ Baghdad
237
+ .
238
+
239
+ NNP
240
+ NN
241
+ NNP
242
+ VBZ
243
+ IN
244
+ NNP
245
+ .
246
+
247
+ I-NP
248
+ I-NP
249
+ I-NP
250
+ I-VP
251
+ I-PP
252
+ I-NP
253
+ O
254
+
255
+ I-ORG
256
+ O
257
+ I-PER
258
+ O
259
+ O
260
+ I-LOC
261
+ O
262
+
263
+ Each line contains four fields: the word, its partof-speech tag, its chunk tag and its named entity
264
+ tag. Words tagged with O are outside of named entities and the I-XXX tag is used for words inside a
265
+ named entity of type XXX. Whenever two entities of
266
+ type XXX are immediately next to each other, the
267
+ first word of the second entity will be tagged B-XXX
268
+ in order to show that it starts another entity. The
269
+ data contains entities of four types: persons (PER),
270
+ organizations (ORG), locations (LOC) and miscellaneous names (MISC). This tagging scheme is the
271
+ IOB scheme originally put forward by Ramshaw and
272
+ Marcus (1995). We assume that named entities are
273
+ non-recursive and non-overlapping. When a named
274
+ entity is embedded in another named entity, usually
275
+ only the top level entity has been annotated.
276
+ Table 2 contains an overview of the number of
277
+ named entities in each data file.
278
+ 2.4
279
+
280
+ Evaluation
281
+
282
+ The performance in this task is measured with Fβ=1
283
+ rate:
284
+ Fβ =
285
+
286
+ (β 2 + 1) ∗ precision ∗ recall
287
+ (β 2 ∗ precision + recall)
288
+
289
+ (1)
290
+
291
+ Florian
292
+ Chieu
293
+ Klein
294
+ Zhang
295
+ Carreras (a)
296
+ Curran
297
+ Mayfield
298
+ Carreras (b)
299
+ McCallum
300
+ Bender
301
+ Munro
302
+ Wu
303
+ Whitelaw
304
+ Hendrickx
305
+ De Meulder
306
+ Hammerton
307
+
308
+ lex
309
+ +
310
+ +
311
+ +
312
+ +
313
+ +
314
+ +
315
+ +
316
+ +
317
+ +
318
+ +
319
+ +
320
+ +
321
+ +
322
+ +
323
+ +
324
+
325
+ pos
326
+ +
327
+ +
328
+ +
329
+ +
330
+ +
331
+ +
332
+ +
333
+ +
334
+ +
335
+ +
336
+ +
337
+ +
338
+ +
339
+ +
340
+
341
+ aff
342
+ +
343
+ +
344
+ +
345
+ +
346
+ +
347
+ +
348
+ +
349
+ +
350
+ +
351
+ +
352
+ +
353
+ +
354
+ +
355
+ -
356
+
357
+ pre
358
+ +
359
+ +
360
+ +
361
+ +
362
+ +
363
+ +
364
+ +
365
+ +
366
+ +
367
+ +
368
+ +
369
+ +
370
+ -
371
+
372
+ ort
373
+ +
374
+ +
375
+ +
376
+ +
377
+ +
378
+ +
379
+ +
380
+ +
381
+ +
382
+ +
383
+ +
384
+ +
385
+ -
386
+
387
+ gaz
388
+ +
389
+ +
390
+ +
391
+ +
392
+ +
393
+ +
394
+ +
395
+ +
396
+ +
397
+ +
398
+ +
399
+
400
+ chu
401
+ +
402
+ +
403
+ +
404
+ +
405
+ +
406
+ +
407
+ +
408
+ +
409
+ +
410
+
411
+ pat
412
+ +
413
+ +
414
+ +
415
+ +
416
+ +
417
+ -
418
+
419
+ cas
420
+ +
421
+ +
422
+ +
423
+ +
424
+ +
425
+ -
426
+
427
+ tri
428
+ +
429
+ +
430
+ +
431
+ +
432
+ -
433
+
434
+ bag
435
+ +
436
+ +
437
+ -
438
+
439
+ quo
440
+ +
441
+ +
442
+ -
443
+
444
+ doc
445
+ +
446
+ -
447
+
448
+ Table 3: Main features used by the the sixteen systems that participated in the CoNLL-2003 shared task
449
+ sorted by performance on the English test data. Aff: affix information (n-grams); bag: bag of words; cas:
450
+ global case information; chu: chunk tags; doc: global document information; gaz: gazetteers; lex: lexical
451
+ features; ort: orthographic information; pat: orthographic patterns (like Aa0); pos: part-of-speech tags; pre:
452
+ previously predicted NE tags; quo: flag signing that the word is between quotes; tri: trigger words.
453
+ with β=1 (Van Rijsbergen, 1975). Precision is the
454
+ percentage of named entities found by the learning
455
+ system that are correct. Recall is the percentage of
456
+ named entities present in the corpus that are found
457
+ by the system. A named entity is correct only if it
458
+ is an exact match of the corresponding entity in the
459
+ data file.
460
+
461
+ 3
462
+
463
+ Participating Systems
464
+
465
+ Sixteen systems have participated in the CoNLL2003 shared task. They employed a wide variety of
466
+ machine learning techniques as well as system combination. Most of the participants have attempted
467
+ to use information other than the available training data. This information included gazetteers and
468
+ unannotated data, and there was one participant
469
+ who used the output of externally trained named entity recognition systems.
470
+ 3.1
471
+
472
+ Learning techniques
473
+
474
+ The most frequently applied technique in the
475
+ CoNLL-2003 shared task is the Maximum Entropy
476
+ Model. Five systems used this statistical learning
477
+ method. Three systems used Maximum Entropy
478
+ Models in isolation (Bender et al., 2003; Chieu and
479
+ Ng, 2003; Curran and Clark, 2003). Two more
480
+ systems used them in combination with other techniques (Florian et al., 2003; Klein et al., 2003). Maximum Entropy Models seem to be a good choice for
481
+
482
+ this kind of task: the top three results for English
483
+ and the top two results for German were obtained
484
+ by participants who employed them in one way or
485
+ another.
486
+ Hidden Markov Models were employed by four of
487
+ the systems that took part in the shared task (Florian et al., 2003; Klein et al., 2003; Mayfield et al.,
488
+ 2003; Whitelaw and Patrick, 2003). However, they
489
+ were always used in combination with other learning
490
+ techniques. Klein et al. (2003) also applied the related Conditional Markov Models for combining classifiers.
491
+ Learning methods that were based on connectionist approaches were applied by four systems. Zhang
492
+ and Johnson (2003) used robust risk minimization,
493
+ which is a Winnow technique. Florian et al. (2003)
494
+ employed the same technique in a combination of
495
+ learners. Voted perceptrons were applied to the
496
+ shared task data by Carreras et al. (2003a) and
497
+ Hammerton used a recurrent neural network (Long
498
+ Short-Term Memory) for finding named entities.
499
+ Other learning approaches were employed less frequently. Two teams used AdaBoost.MH (Carreras
500
+ et al., 2003b; Wu et al., 2003) and two other groups
501
+ employed memory-based learning (De Meulder and
502
+ Daelemans, 2003; Hendrickx and Van den Bosch,
503
+ 2003). Transformation-based learning (Florian et
504
+ al., 2003), Support Vector Machines (Mayfield et al.,
505
+ 2003) and Conditional Random Fields (McCallum
506
+
507
+ and Li, 2003) were applied by one system each.
508
+ Combination of different learning systems has
509
+ proven to be a good method for obtaining excellent
510
+ results. Five participating groups have applied system combination. Florian et al. (2003) tested different methods for combining the results of four systems and found that robust risk minimization worked
511
+ best. Klein et al. (2003) employed a stacked learning system which contains Hidden Markov Models,
512
+ Maximum Entropy Models and Conditional Markov
513
+ Models. Mayfield et al. (2003) stacked two learners
514
+ and obtained better performance. Wu et al. (2003)
515
+ applied both stacking and voting to three learners.
516
+ Munro et al. (2003) employed both voting and bagging for combining classifiers.
517
+ 3.2
518
+
519
+ Features
520
+
521
+ The choice of the learning approach is important for
522
+ obtaining a good system for recognizing named entities. However, in the CoNLL-2002 shared task we
523
+ found out that choice of features is at least as important. An overview of some of the types of features
524
+ chosen by the shared task participants, can be found
525
+ in Table 3.
526
+ All participants used lexical features (words) except for Whitelaw and Patrick (2003) who implemented a character-based method. Most of the systems employed part-of-speech tags and two of them
527
+ have recomputed the English tags with better taggers (Hendrickx and Van den Bosch, 2003; Wu et al.,
528
+ 2003). Othographic information, affixes, gazetteers
529
+ and chunk information were also incorporated in
530
+ most systems although one group reports that the
531
+ available chunking information did not help (Wu et
532
+ al., 2003) Other features were used less frequently.
533
+ Table 3 does not reveal a single feature that would
534
+ be ideal for named entity recognition.
535
+ 3.3
536
+
537
+ External resources
538
+
539
+ Eleven of the sixteen participating teams have attempted to use information other than the training
540
+ data that was supplied for this shared task. All included gazetteers in their systems. Four groups examined the usability of unannotated data, either for
541
+ extracting training instances (Bender et al., 2003;
542
+ Hendrickx and Van den Bosch, 2003) or obtaining
543
+ extra named entities for gazetteers (De Meulder and
544
+ Daelemans, 2003; McCallum and Li, 2003). A reasonable number of groups have also employed unannotated data for obtaining capitalization features for
545
+ words. One participating team has used externally
546
+ trained named entity recognition systems for English
547
+ as a part in a combined system (Florian et al., 2003).
548
+ Table 4 shows the error reduction of the systems
549
+
550
+ Zhang
551
+ Florian
552
+ Chieu
553
+ Hammerton
554
+ Carreras (a)
555
+ Hendrickx
556
+ De Meulder
557
+ Bender
558
+ Curran
559
+ McCallum
560
+ Wu
561
+
562
+ G
563
+ +
564
+ +
565
+ +
566
+ +
567
+ +
568
+ +
569
+ +
570
+ +
571
+ +
572
+ +
573
+ +
574
+
575
+ U
576
+ +
577
+ +
578
+ +
579
+ +
580
+ -
581
+
582
+ E
583
+ +
584
+ -
585
+
586
+ English
587
+ 19%
588
+ 27%
589
+ 17%
590
+ 22%
591
+ 12%
592
+ 7%
593
+ 8%
594
+ 3%
595
+ 1%
596
+ ?
597
+ ?
598
+
599
+ German
600
+ 15%
601
+ 5%
602
+ 7%
603
+ 8%
604
+ 5%
605
+ 3%
606
+ 6%
607
+ ?
608
+ ?
609
+
610
+ Table 4: Error reduction for the two development data sets when using extra information like
611
+ gazetteers (G), unannotated data (U) or externally
612
+ developed named entity recognizers (E). The lines
613
+ have been sorted by the sum of the reduction percentages for the two languages.
614
+ with extra information compared to while using only
615
+ the available training data. The inclusion of extra named entity recognition systems seems to have
616
+ worked well (Florian et al., 2003). Generally the systems that only used gazetteers seem to gain more
617
+ than systems that have used unannotated data for
618
+ other purposes than obtaining capitalization information. However, the gain differences between the
619
+ two approaches are most obvious for English for
620
+ which better gazetteers are available. With the exception of the result of Zhang and Johnson (2003),
621
+ there is not much difference in the German results
622
+ between the gains obtained by using gazetteers and
623
+ those obtained by using unannotated data.
624
+ 3.4
625
+
626
+ Performances
627
+
628
+ A baseline rate was computed for the English and the
629
+ German test sets. It was produced by a system which
630
+ only identified entities which had a unique class in
631
+ the training data. If a phrase was part of more than
632
+ one entity, the system would select the longest one.
633
+ All systems that participated in the shared task have
634
+ outperformed the baseline system.
635
+ For all the Fβ=1 rates we have estimated significance boundaries by using bootstrap resampling
636
+ (Noreen, 1989). From each output file of a system,
637
+ 250 random samples of sentences have been chosen
638
+ and the distribution of the Fβ=1 rates in these samples is assumed to be the distribution of the performance of the system. We assume that performance
639
+ A is significantly different from performance B if A
640
+ is not within the center 90% of the distribution of B.
641
+ The performances of the sixteen systems on the
642
+
643
+ two test data sets can be found in Table 5. For English, the combined classifier of Florian et al. (2003)
644
+ achieved the highest overall Fβ=1 rate. However,
645
+ the difference between their performance and that
646
+ of the Maximum Entropy approach of Chieu and Ng
647
+ (2003) is not significant. An important feature of the
648
+ best system that other participants did not use, was
649
+ the inclusion of the output of two externally trained
650
+ named entity recognizers in the combination process.
651
+ Florian et al. (2003) have also obtained the highest
652
+ Fβ=1 rate for the German data. Here there is no significant difference between them and the systems of
653
+ Klein et al. (2003) and Zhang and Johnson (2003).
654
+ We have combined the results of the sixteen system in order to see if there was room for improvement. We converted the output of the systems to
655
+ the same IOB tagging representation and searched
656
+ for the set of systems from which the best tags for
657
+ the development data could be obtained with majority voting. The optimal set of systems was determined by performing a bidirectional hill-climbing
658
+ search (Caruana and Freitag, 1994) with beam size 9,
659
+ starting from zero features. A majority vote of five
660
+ systems (Chieu and Ng, 2003; Florian et al., 2003;
661
+ Klein et al., 2003; McCallum and Li, 2003; Whitelaw
662
+ and Patrick, 2003) performed best on the English
663
+ development data. Another combination of five systems (Carreras et al., 2003b; Mayfield et al., 2003;
664
+ McCallum and Li, 2003; Munro et al., 2003; Zhang
665
+ and Johnson, 2003) obtained the best result for the
666
+ German development data. We have performed a
667
+ majority vote with these sets of systems on the related test sets and obtained Fβ=1 rates of 90.30 for
668
+ English (14% error reduction compared with the best
669
+ system) and 74.17 for German (6% error reduction).
670
+
671
+ 4
672
+
673
+ Concluding Remarks
674
+
675
+ We have described the CoNLL-2003 shared task:
676
+ language-independent named entity recognition.
677
+ Sixteen systems have processed English and German
678
+ named entity data. The best performance for both
679
+ languages has been obtained by a combined learning system that used Maximum Entropy Models,
680
+ transformation-based learning, Hidden Markov Models as well as robust risk minimization (Florian et al.,
681
+ 2003). Apart from the training data, this system also
682
+ employed gazetteers and the output of two externally
683
+ trained named entity recognizers. The performance
684
+ of the system of Chieu et al. (2003) was not significantly different from the best performance for English and the method of Klein et al. (2003) and the
685
+ approach of Zhang and Johnson (2003) were not significantly worse than the best result for German.
686
+ Eleven teams have incorporated information other
687
+
688
+ than the training data in their system. Four of them
689
+ have obtained error reductions of 15% or more for
690
+ English and one has managed this for German. The
691
+ resources used by these systems, gazetteers and externally trained named entity systems, still require a
692
+ lot of manual work. Systems that employed unannotated data, obtained performance gains around 5%.
693
+ The search for an excellent method for taking advantage of the fast amount of available raw text, remains
694
+ open.
695
+
696
+ Acknowledgements
697
+ Tjong Kim Sang is financed by IWT STWW as a
698
+ researcher in the ATraNoS project. De Meulder is
699
+ supported by a BOF grant supplied by the University
700
+ of Antwerp.
701
+
702
+ References
703
+ Oliver Bender, Franz Josef Och, and Hermann Ney.
704
+ 2003. Maximum Entropy Models for Named Entity Recognition. In Proceedings of CoNLL-2003.
705
+ Andrew Borthwick. 1999. A Maximum Entropy Approach to Named Entity Recognition. PhD thesis,
706
+ New York University.
707
+ Xavier Carreras, Lluı́s Màrquez, and Lluı́s Padró.
708
+ 2003a. Learning a Perceptron-Based Named Entity Chunker via Online Recognition Feedback. In
709
+ Proceedings of CoNLL-2003.
710
+ Xavier Carreras, Lluı́s Màrquez, and Lluı́s Padró.
711
+ 2003b. A Simple Named Entity Extractor using
712
+ AdaBoost. In Proceedings of CoNLL-2003.
713
+ Rich Caruana and Dayne Freitag. 1994. Greedy Attribute Selection. In Proceedings of the Eleventh
714
+ International Conference on Machine Learning,
715
+ pages 28–36. New Brunswick, NJ, USA, Morgan
716
+ Kaufman.
717
+ Hai Leong Chieu and Hwee Tou Ng. 2003. Named
718
+ Entity Recognition with a Maximum Entropy Approach. In Proceedings of CoNLL-2003.
719
+ Nancy Chinchor, Erica Brown, Lisa Ferro, and Patty
720
+ Robinson. 1999. 1999 Named Entity Recognition
721
+ Task Definition. MITRE and SAIC.
722
+ James R. Curran and Stephen Clark. 2003. Language Independent NER using a Maximum Entropy Tagger. In Proceedings of CoNLL-2003.
723
+ Walter Daelemans, Jakub Zavrel, Ko van der Sloot,
724
+ and Antal van den Bosch. 2002. MBT: MemoryBased Tagger, version 1.0, Reference Guide. ILK
725
+ Technical Report ILK-0209, University of Tilburg,
726
+ The Netherlands.
727
+
728
+ Fien De Meulder and Walter Daelemans. 2003.
729
+ Memory-Based Named Entity Recognition using
730
+ Unannotated Data. In Proceedings of CoNLL2003.
731
+ Radu Florian, Abe Ittycheriah, Hongyan Jing, and
732
+ Tong Zhang. 2003. Named Entity Recognition
733
+ through Classifier Combination. In Proceedings of
734
+ CoNLL-2003.
735
+ James Hammerton. 2003. Named Entity Recognition with Long Short-Term Memory. In Proceedings of CoNLL-2003.
736
+ Iris Hendrickx and Antal van den Bosch. 2003.
737
+ Memory-based one-step named-entity recognition:
738
+ Effects of seed list features, classifier stacking, and
739
+ unannotated data. In Proceedings of CoNLL-2003.
740
+ Dan Klein, Joseph Smarr, Huy Nguyen, and Christopher D. Manning. 2003. Named Entity Recognition with Character-Level Models. In Proceedings
741
+ of CoNLL-2003.
742
+ James Mayfield, Paul McNamee, and Christine Piatko. 2003. Named Entity Recognition using Hundreds of Thousands of Features. In Proceedings of
743
+ CoNLL-2003.
744
+ Andrew McCallum and Wei Li. 2003. Early results
745
+ for Named Entity Recognition with Conditional
746
+ Random Fields, Feature Induction and WebEnhanced Lexicons. In Proceedings of CoNLL2003.
747
+ Robert Munro, Daren Ler, and Jon Patrick.
748
+ 2003. Meta-Learning Orthographic and Contextual Models for Language Independent Named Entity Recognition. In Proceedings of CoNLL-2003.
749
+ Eric W. Noreen. 1989. Computer-Intensive Methods
750
+ for Testing Hypotheses. John Wiley & Sons.
751
+
752
+ English test
753
+ Florian
754
+ Chieu
755
+ Klein
756
+ Zhang
757
+ Carreras (a)
758
+ Curran
759
+ Mayfield
760
+ Carreras (b)
761
+ McCallum
762
+ Bender
763
+ Munro
764
+ Wu
765
+ Whitelaw
766
+ Hendrickx
767
+ De Meulder
768
+ Hammerton
769
+ Baseline
770
+
771
+ Precision
772
+ 88.99%
773
+ 88.12%
774
+ 85.93%
775
+ 86.13%
776
+ 84.05%
777
+ 84.29%
778
+ 84.45%
779
+ 85.81%
780
+ 84.52%
781
+ 84.68%
782
+ 80.87%
783
+ 82.02%
784
+ 81.60%
785
+ 76.33%
786
+ 75.84%
787
+ 69.09%
788
+ 71.91%
789
+
790
+ Recall
791
+ 88.54%
792
+ 88.51%
793
+ 86.21%
794
+ 84.88%
795
+ 85.96%
796
+ 85.50%
797
+ 84.90%
798
+ 82.84%
799
+ 83.55%
800
+ 83.18%
801
+ 84.21%
802
+ 81.39%
803
+ 78.05%
804
+ 80.17%
805
+ 78.13%
806
+ 53.26%
807
+ 50.90%
808
+
809
+ Fβ=1
810
+ 88.76±0.7
811
+ 88.31±0.7
812
+ 86.07±0.8
813
+ 85.50±0.9
814
+ 85.00±0.8
815
+ 84.89±0.9
816
+ 84.67±1.0
817
+ 84.30±0.9
818
+ 84.04±0.9
819
+ 83.92±1.0
820
+ 82.50±1.0
821
+ 81.70±0.9
822
+ 79.78±1.0
823
+ 78.20±1.0
824
+ 76.97±1.2
825
+ 60.15±1.3
826
+ 59.61±1.2
827
+
828
+ German test
829
+ Florian
830
+ Klein
831
+ Zhang
832
+ Mayfield
833
+ Carreras (a)
834
+ Bender
835
+ Curran
836
+ McCallum
837
+ Munro
838
+ Carreras (b)
839
+ Wu
840
+ Chieu
841
+ Hendrickx
842
+ De Meulder
843
+ Whitelaw
844
+ Hammerton
845
+ Baseline
846
+
847
+ Precision
848
+ 83.87%
849
+ 80.38%
850
+ 82.00%
851
+ 75.97%
852
+ 75.47%
853
+ 74.82%
854
+ 75.61%
855
+ 75.97%
856
+ 69.37%
857
+ 77.83%
858
+ 75.20%
859
+ 76.83%
860
+ 71.15%
861
+ 63.93%
862
+ 71.05%
863
+ 63.49%
864
+ 31.86%
865
+
866
+ Recall
867
+ 63.71%
868
+ 65.04%
869
+ 63.03%
870
+ 64.82%
871
+ 63.82%
872
+ 63.82%
873
+ 62.46%
874
+ 61.72%
875
+ 66.21%
876
+ 58.02%
877
+ 59.35%
878
+ 57.34%
879
+ 56.55%
880
+ 51.86%
881
+ 44.11%
882
+ 38.25%
883
+ 28.89%
884
+
885
+ Fβ=1
886
+ 72.41±1.3
887
+ 71.90±1.2
888
+ 71.27±1.5
889
+ 69.96±1.4
890
+ 69.15±1.3
891
+ 68.88±1.3
892
+ 68.41±1.4
893
+ 68.11±1.4
894
+ 67.75±1.4
895
+ 66.48±1.5
896
+ 66.34±1.3
897
+ 65.67±1.4
898
+ 63.02±1.4
899
+ 57.27±1.6
900
+ 54.43±1.4
901
+ 47.74±1.5
902
+ 30.30±1.3
903
+
904
+ Lance A. Ramshaw and Mitchell P. Marcus.
905
+ 1995. Text Chunking Using Transformation-Based
906
+ Learning. In Proceedings of the Third ACL Workshop on Very Large Corpora, pages 82–94. Cambridge, MA, USA.
907
+
908
+ Table 5: Overall precision, recall and Fβ=1 rates obtained by the sixteen participating systems on the
909
+ test data sets for the two languages in the CoNLL2003 shared task.
910
+
911
+ Helmut Schmid. 1995. Improvements in Part-ofSpeech Tagging with an Application to German.
912
+ In Proceedings of EACL-SIGDAT 1995. Dublin,
913
+ Ireland.
914
+
915
+ Casey Whitelaw and Jon Patrick. 2003. Named Entity Recognition Using a Character-based Probabilistic Approach. In Proceedings of CoNLL-2003.
916
+
917
+ Erik F. Tjong Kim Sang. 2002. Introduction to the
918
+ CoNLL-2002 Shared Task: Language-Independent
919
+ Named Entity Recognition. In Proceedings of
920
+ CoNLL-2002, pages 155–158. Taipei, Taiwan.
921
+ C.J. van Rijsbergen. 1975. Information Retrieval.
922
+ Buttersworth.
923
+
924
+ Dekai Wu, Grace Ngai, and Marine Carpuat. 2003.
925
+ A Stacked, Voted, Stacked Model for Named Entity Recognition. In Proceedings of CoNLL-2003.
926
+ Tong Zhang and David Johnson. 2003. A Robust
927
+ Risk Minimization based Named Entity Recognition System. In Proceedings of CoNLL-2003.
928
+
929
+