masaenger commited on
Commit
fbd0581
1 Parent(s): 397ba86

Update bc5cdr based on git version 3276470

Browse files
Files changed (3) hide show
  1. README.md +61 -0
  2. bc5cdr.py +373 -0
  3. bigbiohub.py +592 -0
README.md ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ ---
3
+ language:
4
+ - en
5
+ bigbio_language:
6
+ - English
7
+ license: other
8
+ multilinguality: monolingual
9
+ bigbio_license_shortname: PUBLIC_DOMAIN_MARK_1p0
10
+ pretty_name: BC5CDR
11
+ homepage: http://www.biocreative.org/tasks/biocreative-v/track-3-cdr/
12
+ bigbio_pubmed: True
13
+ bigbio_public: True
14
+ bigbio_tasks:
15
+ - NAMED_ENTITY_RECOGNITION
16
+ - NAMED_ENTITY_DISAMBIGUATION
17
+ - RELATION_EXTRACTION
18
+ ---
19
+
20
+
21
+ # Dataset Card for BC5CDR
22
+
23
+ ## Dataset Description
24
+
25
+ - **Homepage:** http://www.biocreative.org/tasks/biocreative-v/track-3-cdr/
26
+ - **Pubmed:** True
27
+ - **Public:** True
28
+ - **Tasks:** NER,NED,RE
29
+
30
+
31
+ The BioCreative V Chemical Disease Relation (CDR) dataset is a large annotated text corpus of human annotations of all chemicals, diseases and their interactions in 1,500 PubMed articles.
32
+
33
+
34
+
35
+ ## Citation Information
36
+
37
+ ```
38
+ @article{DBLP:journals/biodb/LiSJSWLDMWL16,
39
+ author = {Jiao Li and
40
+ Yueping Sun and
41
+ Robin J. Johnson and
42
+ Daniela Sciaky and
43
+ Chih{-}Hsuan Wei and
44
+ Robert Leaman and
45
+ Allan Peter Davis and
46
+ Carolyn J. Mattingly and
47
+ Thomas C. Wiegers and
48
+ Zhiyong Lu},
49
+ title = {BioCreative {V} {CDR} task corpus: a resource for chemical disease
50
+ relation extraction},
51
+ journal = {Database J. Biol. Databases Curation},
52
+ volume = {2016},
53
+ year = {2016},
54
+ url = {https://doi.org/10.1093/database/baw068},
55
+ doi = {10.1093/database/baw068},
56
+ timestamp = {Thu, 13 Aug 2020 12:41:41 +0200},
57
+ biburl = {https://dblp.org/rec/journals/biodb/LiSJSWLDMWL16.bib},
58
+ bibsource = {dblp computer science bibliography, https://dblp.org}
59
+ }
60
+
61
+ ```
bc5cdr.py ADDED
@@ -0,0 +1,373 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """
16
+ To this end, we set up a challenge task through BioCreative V to automatically
17
+ extract CDRs from the literature. More specifically, we designed two challenge
18
+ tasks: disease named entity recognition (DNER) and chemical-induced disease
19
+ (CID) relation extraction. To assist system development and assessment, we
20
+ created a large annotated text corpus that consists of human annotations of
21
+ all chemicals, diseases and their interactions in 1,500 PubMed articles.
22
+
23
+ -- 'Overview of the BioCreative V Chemical Disease Relation (CDR) Task'
24
+ """
25
+ import collections
26
+ import itertools
27
+ import os
28
+
29
+ import datasets
30
+ from bioc import biocxml
31
+
32
+ from .bigbiohub import kb_features
33
+ from .bigbiohub import BigBioConfig
34
+ from .bigbiohub import Tasks
35
+ from .bigbiohub import get_texts_and_offsets_from_bioc_ann
36
+
37
+
38
+ _LANGUAGES = ['English']
39
+ _PUBMED = True
40
+ _LOCAL = False
41
+ _CITATION = """\
42
+ @article{DBLP:journals/biodb/LiSJSWLDMWL16,
43
+ author = {Jiao Li and
44
+ Yueping Sun and
45
+ Robin J. Johnson and
46
+ Daniela Sciaky and
47
+ Chih{-}Hsuan Wei and
48
+ Robert Leaman and
49
+ Allan Peter Davis and
50
+ Carolyn J. Mattingly and
51
+ Thomas C. Wiegers and
52
+ Zhiyong Lu},
53
+ title = {BioCreative {V} {CDR} task corpus: a resource for chemical disease
54
+ relation extraction},
55
+ journal = {Database J. Biol. Databases Curation},
56
+ volume = {2016},
57
+ year = {2016},
58
+ url = {https://doi.org/10.1093/database/baw068},
59
+ doi = {10.1093/database/baw068},
60
+ timestamp = {Thu, 13 Aug 2020 12:41:41 +0200},
61
+ biburl = {https://dblp.org/rec/journals/biodb/LiSJSWLDMWL16.bib},
62
+ bibsource = {dblp computer science bibliography, https://dblp.org}
63
+ }
64
+ """
65
+
66
+ _DATASETNAME = "bc5cdr"
67
+ _DISPLAYNAME = "BC5CDR"
68
+
69
+ _DESCRIPTION = """\
70
+ The BioCreative V Chemical Disease Relation (CDR) dataset is a large annotated \
71
+ text corpus of human annotations of all chemicals, diseases and their \
72
+ interactions in 1,500 PubMed articles.
73
+ """
74
+
75
+ _HOMEPAGE = "https://biocreative.bioinformatics.udel.edu/tasks/biocreative-v/track-3-cdr/"
76
+
77
+ _LICENSE = 'Public Domain Mark 1.0'
78
+
79
+ _URLs = {
80
+ "source": "https://github.com/JHnlp/BioCreative-V-CDR-Corpus/raw/master/CDR_Data.zip",
81
+ "bigbio_kb": "https://github.com/JHnlp/BioCreative-V-CDR-Corpus/raw/master/CDR_Data.zip",
82
+ }
83
+
84
+ _SUPPORTED_TASKS = [
85
+ Tasks.NAMED_ENTITY_RECOGNITION,
86
+ Tasks.NAMED_ENTITY_DISAMBIGUATION,
87
+ Tasks.RELATION_EXTRACTION,
88
+ ]
89
+ _SOURCE_VERSION = "01.05.16"
90
+ _BIGBIO_VERSION = "1.0.0"
91
+
92
+
93
+ class Bc5cdrDataset(datasets.GeneratorBasedBuilder):
94
+ """
95
+ BioCreative V Chemical Disease Relation (CDR) Task.
96
+ """
97
+
98
+ DEFAULT_CONFIG_NAME = "bc5cdr_source"
99
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
100
+ BIGBIO_VERSION = datasets.Version(_BIGBIO_VERSION)
101
+
102
+ BUILDER_CONFIGS = [
103
+ BigBioConfig(
104
+ name="bc5cdr_source",
105
+ version=SOURCE_VERSION,
106
+ description="BC5CDR source schema",
107
+ schema="source",
108
+ subset_id="bc5cdr",
109
+ ),
110
+ BigBioConfig(
111
+ name="bc5cdr_bigbio_kb",
112
+ version=BIGBIO_VERSION,
113
+ description="BC5CDR simplified BigBio schema",
114
+ schema="bigbio_kb",
115
+ subset_id="bc5cdr",
116
+ ),
117
+ ]
118
+
119
+ def _info(self):
120
+
121
+ if self.config.schema == "source":
122
+ # this is a variation on the BioC format
123
+ features = datasets.Features(
124
+ {
125
+ "passages": [
126
+ {
127
+ "document_id": datasets.Value("string"),
128
+ "type": datasets.Value("string"),
129
+ "text": datasets.Value("string"),
130
+ "entities": [
131
+ {
132
+ "id": datasets.Value("string"),
133
+ "offsets": [[datasets.Value("int32")]],
134
+ "text": [datasets.Value("string")],
135
+ "type": datasets.Value("string"),
136
+ "normalized": [
137
+ {
138
+ "db_name": datasets.Value("string"),
139
+ "db_id": datasets.Value("string"),
140
+ }
141
+ ],
142
+ }
143
+ ],
144
+ "relations": [
145
+ {
146
+ "id": datasets.Value("string"),
147
+ "type": datasets.Value("string"),
148
+ "arg1_id": datasets.Value("string"),
149
+ "arg2_id": datasets.Value("string"),
150
+ }
151
+ ],
152
+ }
153
+ ]
154
+ }
155
+ )
156
+
157
+ elif self.config.schema == "bigbio_kb":
158
+ features = kb_features
159
+
160
+ return datasets.DatasetInfo(
161
+ description=_DESCRIPTION,
162
+ features=features,
163
+ supervised_keys=None,
164
+ homepage=_HOMEPAGE,
165
+ license=str(_LICENSE),
166
+ citation=_CITATION,
167
+ )
168
+
169
+ def _split_generators(self, dl_manager):
170
+ """Returns SplitGenerators."""
171
+ my_urls = _URLs[self.config.schema]
172
+ data_dir = dl_manager.download_and_extract(my_urls)
173
+ return [
174
+ datasets.SplitGenerator(
175
+ name=datasets.Split.TRAIN,
176
+ # These kwargs will be passed to _generate_examples
177
+ gen_kwargs={
178
+ "filepath": os.path.join(
179
+ data_dir, "CDR_Data/CDR.Corpus.v010516/CDR_TrainingSet.BioC.xml"
180
+ ),
181
+ "split": "train",
182
+ },
183
+ ),
184
+ datasets.SplitGenerator(
185
+ name=datasets.Split.TEST,
186
+ # These kwargs will be passed to _generate_examples
187
+ gen_kwargs={
188
+ "filepath": os.path.join(
189
+ data_dir, "CDR_Data/CDR.Corpus.v010516/CDR_TestSet.BioC.xml"
190
+ ),
191
+ "split": "test",
192
+ },
193
+ ),
194
+ datasets.SplitGenerator(
195
+ name=datasets.Split.VALIDATION,
196
+ # These kwargs will be passed to _generate_examples
197
+ gen_kwargs={
198
+ "filepath": os.path.join(
199
+ data_dir,
200
+ "CDR_Data/CDR.Corpus.v010516/CDR_DevelopmentSet.BioC.xml",
201
+ ),
202
+ "split": "dev",
203
+ },
204
+ ),
205
+ ]
206
+
207
+ def _get_bioc_entity(self, span, doc_text, db_id_key="MESH"):
208
+ """Parse BioC entity annotation.
209
+
210
+ Parameters
211
+ ----------
212
+ span : BioCAnnotation
213
+ BioC entity annotation
214
+ doc_text : string
215
+ document text, required to construct text spans
216
+ db_id_key : str, optional
217
+ database name used for normalization, by default "MESH"
218
+
219
+ Returns
220
+ -------
221
+ dict
222
+ entity information
223
+ """
224
+ # offsets = [(loc.offset, loc.offset + loc.length) for loc in span.locations]
225
+ # texts = [doc_text[i:j] for i, j in offsets]
226
+ offsets, texts = get_texts_and_offsets_from_bioc_ann(span)
227
+ db_ids = span.infons[db_id_key] if db_id_key else "-1"
228
+
229
+ # some entities are not linked and
230
+ # some entities are linked to multiple normalized ids
231
+ if db_ids == "-1":
232
+ db_ids_list = []
233
+ else:
234
+ db_ids_list = db_ids.split("|")
235
+
236
+ normalized = [{"db_name": db_id_key, "db_id": db_id} for db_id in db_ids_list]
237
+
238
+ return {
239
+ "id": span.id,
240
+ "offsets": offsets,
241
+ "text": texts,
242
+ "type": span.infons["type"],
243
+ "normalized": normalized,
244
+ }
245
+
246
+ def _get_relations(self, relations, entities):
247
+ """
248
+ BC5CDR provides abstract-level annotations for entity-linked relation
249
+ pairs rather than materializing links between all surface form
250
+ mentions of relations. An example from train id=2670794, the relation
251
+ - (chemical, disease) (D014148, D004211)
252
+ is materialized as 6 mentions of entity pairs
253
+ - 2x ('tranexamic acid', 'intravascular coagulation')
254
+ - 4x ('AMCA', 'intravascular coagulation')
255
+ """
256
+ # index entities by normalized id
257
+ index = collections.defaultdict(list)
258
+ for ent in entities:
259
+ for norm in ent["normalized"]:
260
+ index[norm["db_id"]].append(ent)
261
+ index = dict(index)
262
+
263
+ # transform doc-level relations to mention-level
264
+ rela_mentions = []
265
+ for rela in relations:
266
+ arg1 = rela.infons["Chemical"]
267
+ arg2 = rela.infons["Disease"]
268
+ # all mention pairs
269
+ all_pairs = itertools.product(index[arg1], index[arg2])
270
+ for a, b in all_pairs:
271
+ # create relations linked by entity ids
272
+ rela_mentions.append(
273
+ {
274
+ "id": None,
275
+ "type": rela.infons["relation"],
276
+ "arg1_id": a["id"],
277
+ "arg2_id": b["id"],
278
+ "normalized": [],
279
+ }
280
+ )
281
+ return rela_mentions
282
+
283
+ def _get_document_text(self, xdoc):
284
+ """Build document text for unit testing entity span offsets."""
285
+ text = ""
286
+ for passage in xdoc.passages:
287
+ pad = passage.offset - len(text)
288
+ text += (" " * pad) + passage.text
289
+ return text
290
+
291
+ def _generate_examples(
292
+ self,
293
+ filepath,
294
+ split, # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
295
+ ):
296
+ """Yields examples as (key, example) tuples."""
297
+ if self.config.schema == "source":
298
+ reader = biocxml.BioCXMLDocumentReader(str(filepath))
299
+
300
+ for uid, xdoc in enumerate(reader):
301
+ doc_text = self._get_document_text(xdoc)
302
+ yield uid, {
303
+ "passages": [
304
+ {
305
+ "document_id": xdoc.id,
306
+ "type": passage.infons["type"],
307
+ "text": passage.text,
308
+ "entities": [
309
+ self._get_bioc_entity(span, doc_text)
310
+ for span in passage.annotations
311
+ ],
312
+ "relations": [
313
+ {
314
+ "id": rel.id,
315
+ "type": rel.infons["relation"],
316
+ "arg1_id": rel.infons["Chemical"],
317
+ "arg2_id": rel.infons["Disease"],
318
+ }
319
+ for rel in xdoc.relations
320
+ ],
321
+ }
322
+ for passage in xdoc.passages
323
+ ]
324
+ }
325
+
326
+ elif self.config.schema == "bigbio_kb":
327
+ reader = biocxml.BioCXMLDocumentReader(str(filepath))
328
+ uid = 0 # global unique id
329
+
330
+ for i, xdoc in enumerate(reader):
331
+ data = {
332
+ "id": uid,
333
+ "document_id": xdoc.id,
334
+ "passages": [],
335
+ "entities": [],
336
+ "relations": [],
337
+ "events": [],
338
+ "coreferences": [],
339
+ }
340
+ uid += 1
341
+ doc_text = self._get_document_text(xdoc)
342
+
343
+ char_start = 0
344
+ # passages must not overlap and spans must cover the entire document
345
+ for passage in xdoc.passages:
346
+ offsets = [[char_start, char_start + len(passage.text)]]
347
+ char_start = char_start + len(passage.text) + 1
348
+ data["passages"].append(
349
+ {
350
+ "id": uid,
351
+ "type": passage.infons["type"],
352
+ "text": [passage.text],
353
+ "offsets": offsets,
354
+ }
355
+ )
356
+ uid += 1
357
+
358
+ # entities
359
+ for passage in xdoc.passages:
360
+ for span in passage.annotations:
361
+ ent = self._get_bioc_entity(span, doc_text, db_id_key="MESH")
362
+ ent["id"] = uid # override BioC default id
363
+ data["entities"].append(ent)
364
+ uid += 1
365
+
366
+ # relations
367
+ relations = self._get_relations(xdoc.relations, data["entities"])
368
+ for rela in relations:
369
+ rela["id"] = uid # assign unique id
370
+ data["relations"].append(rela)
371
+ uid += 1
372
+
373
+ yield i, data
bigbiohub.py ADDED
@@ -0,0 +1,592 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from collections import defaultdict
2
+ from dataclasses import dataclass
3
+ from enum import Enum
4
+ import logging
5
+ from pathlib import Path
6
+ from types import SimpleNamespace
7
+ from typing import TYPE_CHECKING, Dict, Iterable, List, Tuple
8
+
9
+ import datasets
10
+
11
+ if TYPE_CHECKING:
12
+ import bioc
13
+
14
+ logger = logging.getLogger(__name__)
15
+
16
+
17
+ BigBioValues = SimpleNamespace(NULL="<BB_NULL_STR>")
18
+
19
+
20
+ @dataclass
21
+ class BigBioConfig(datasets.BuilderConfig):
22
+ """BuilderConfig for BigBio."""
23
+
24
+ name: str = None
25
+ version: datasets.Version = None
26
+ description: str = None
27
+ schema: str = None
28
+ subset_id: str = None
29
+
30
+
31
+ class Tasks(Enum):
32
+ NAMED_ENTITY_RECOGNITION = "NER"
33
+ NAMED_ENTITY_DISAMBIGUATION = "NED"
34
+ EVENT_EXTRACTION = "EE"
35
+ RELATION_EXTRACTION = "RE"
36
+ COREFERENCE_RESOLUTION = "COREF"
37
+ QUESTION_ANSWERING = "QA"
38
+ TEXTUAL_ENTAILMENT = "TE"
39
+ SEMANTIC_SIMILARITY = "STS"
40
+ TEXT_PAIRS_CLASSIFICATION = "TXT2CLASS"
41
+ PARAPHRASING = "PARA"
42
+ TRANSLATION = "TRANSL"
43
+ SUMMARIZATION = "SUM"
44
+ TEXT_CLASSIFICATION = "TXTCLASS"
45
+
46
+
47
+ entailment_features = datasets.Features(
48
+ {
49
+ "id": datasets.Value("string"),
50
+ "premise": datasets.Value("string"),
51
+ "hypothesis": datasets.Value("string"),
52
+ "label": datasets.Value("string"),
53
+ }
54
+ )
55
+
56
+ pairs_features = datasets.Features(
57
+ {
58
+ "id": datasets.Value("string"),
59
+ "document_id": datasets.Value("string"),
60
+ "text_1": datasets.Value("string"),
61
+ "text_2": datasets.Value("string"),
62
+ "label": datasets.Value("string"),
63
+ }
64
+ )
65
+
66
+ qa_features = datasets.Features(
67
+ {
68
+ "id": datasets.Value("string"),
69
+ "question_id": datasets.Value("string"),
70
+ "document_id": datasets.Value("string"),
71
+ "question": datasets.Value("string"),
72
+ "type": datasets.Value("string"),
73
+ "choices": [datasets.Value("string")],
74
+ "context": datasets.Value("string"),
75
+ "answer": datasets.Sequence(datasets.Value("string")),
76
+ }
77
+ )
78
+
79
+ text_features = datasets.Features(
80
+ {
81
+ "id": datasets.Value("string"),
82
+ "document_id": datasets.Value("string"),
83
+ "text": datasets.Value("string"),
84
+ "labels": [datasets.Value("string")],
85
+ }
86
+ )
87
+
88
+ text2text_features = datasets.Features(
89
+ {
90
+ "id": datasets.Value("string"),
91
+ "document_id": datasets.Value("string"),
92
+ "text_1": datasets.Value("string"),
93
+ "text_2": datasets.Value("string"),
94
+ "text_1_name": datasets.Value("string"),
95
+ "text_2_name": datasets.Value("string"),
96
+ }
97
+ )
98
+
99
+ kb_features = datasets.Features(
100
+ {
101
+ "id": datasets.Value("string"),
102
+ "document_id": datasets.Value("string"),
103
+ "passages": [
104
+ {
105
+ "id": datasets.Value("string"),
106
+ "type": datasets.Value("string"),
107
+ "text": datasets.Sequence(datasets.Value("string")),
108
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
109
+ }
110
+ ],
111
+ "entities": [
112
+ {
113
+ "id": datasets.Value("string"),
114
+ "type": datasets.Value("string"),
115
+ "text": datasets.Sequence(datasets.Value("string")),
116
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
117
+ "normalized": [
118
+ {
119
+ "db_name": datasets.Value("string"),
120
+ "db_id": datasets.Value("string"),
121
+ }
122
+ ],
123
+ }
124
+ ],
125
+ "events": [
126
+ {
127
+ "id": datasets.Value("string"),
128
+ "type": datasets.Value("string"),
129
+ # refers to the text_bound_annotation of the trigger
130
+ "trigger": {
131
+ "text": datasets.Sequence(datasets.Value("string")),
132
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
133
+ },
134
+ "arguments": [
135
+ {
136
+ "role": datasets.Value("string"),
137
+ "ref_id": datasets.Value("string"),
138
+ }
139
+ ],
140
+ }
141
+ ],
142
+ "coreferences": [
143
+ {
144
+ "id": datasets.Value("string"),
145
+ "entity_ids": datasets.Sequence(datasets.Value("string")),
146
+ }
147
+ ],
148
+ "relations": [
149
+ {
150
+ "id": datasets.Value("string"),
151
+ "type": datasets.Value("string"),
152
+ "arg1_id": datasets.Value("string"),
153
+ "arg2_id": datasets.Value("string"),
154
+ "normalized": [
155
+ {
156
+ "db_name": datasets.Value("string"),
157
+ "db_id": datasets.Value("string"),
158
+ }
159
+ ],
160
+ }
161
+ ],
162
+ }
163
+ )
164
+
165
+
166
+ TASK_TO_SCHEMA = {
167
+ Tasks.NAMED_ENTITY_RECOGNITION.name: "KB",
168
+ Tasks.NAMED_ENTITY_DISAMBIGUATION.name: "KB",
169
+ Tasks.EVENT_EXTRACTION.name: "KB",
170
+ Tasks.RELATION_EXTRACTION.name: "KB",
171
+ Tasks.COREFERENCE_RESOLUTION.name: "KB",
172
+ Tasks.QUESTION_ANSWERING.name: "QA",
173
+ Tasks.TEXTUAL_ENTAILMENT.name: "TE",
174
+ Tasks.SEMANTIC_SIMILARITY.name: "PAIRS",
175
+ Tasks.TEXT_PAIRS_CLASSIFICATION.name: "PAIRS",
176
+ Tasks.PARAPHRASING.name: "T2T",
177
+ Tasks.TRANSLATION.name: "T2T",
178
+ Tasks.SUMMARIZATION.name: "T2T",
179
+ Tasks.TEXT_CLASSIFICATION.name: "TEXT",
180
+ }
181
+
182
+ SCHEMA_TO_TASKS = defaultdict(set)
183
+ for task, schema in TASK_TO_SCHEMA.items():
184
+ SCHEMA_TO_TASKS[schema].add(task)
185
+ SCHEMA_TO_TASKS = dict(SCHEMA_TO_TASKS)
186
+
187
+ VALID_TASKS = set(TASK_TO_SCHEMA.keys())
188
+ VALID_SCHEMAS = set(TASK_TO_SCHEMA.values())
189
+
190
+ SCHEMA_TO_FEATURES = {
191
+ "KB": kb_features,
192
+ "QA": qa_features,
193
+ "TE": entailment_features,
194
+ "T2T": text2text_features,
195
+ "TEXT": text_features,
196
+ "PAIRS": pairs_features,
197
+ }
198
+
199
+
200
+ def get_texts_and_offsets_from_bioc_ann(ann: "bioc.BioCAnnotation") -> Tuple:
201
+
202
+ offsets = [(loc.offset, loc.offset + loc.length) for loc in ann.locations]
203
+
204
+ text = ann.text
205
+
206
+ if len(offsets) > 1:
207
+ i = 0
208
+ texts = []
209
+ for start, end in offsets:
210
+ chunk_len = end - start
211
+ texts.append(text[i : chunk_len + i])
212
+ i += chunk_len
213
+ while i < len(text) and text[i] == " ":
214
+ i += 1
215
+ else:
216
+ texts = [text]
217
+
218
+ return offsets, texts
219
+
220
+
221
+ def remove_prefix(a: str, prefix: str) -> str:
222
+ if a.startswith(prefix):
223
+ a = a[len(prefix) :]
224
+ return a
225
+
226
+
227
+ def parse_brat_file(
228
+ txt_file: Path,
229
+ annotation_file_suffixes: List[str] = None,
230
+ parse_notes: bool = False,
231
+ ) -> Dict:
232
+ """
233
+ Parse a brat file into the schema defined below.
234
+ `txt_file` should be the path to the brat '.txt' file you want to parse, e.g. 'data/1234.txt'
235
+ Assumes that the annotations are contained in one or more of the corresponding '.a1', '.a2' or '.ann' files,
236
+ e.g. 'data/1234.ann' or 'data/1234.a1' and 'data/1234.a2'.
237
+ Will include annotator notes, when `parse_notes == True`.
238
+ brat_features = datasets.Features(
239
+ {
240
+ "id": datasets.Value("string"),
241
+ "document_id": datasets.Value("string"),
242
+ "text": datasets.Value("string"),
243
+ "text_bound_annotations": [ # T line in brat, e.g. type or event trigger
244
+ {
245
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
246
+ "text": datasets.Sequence(datasets.Value("string")),
247
+ "type": datasets.Value("string"),
248
+ "id": datasets.Value("string"),
249
+ }
250
+ ],
251
+ "events": [ # E line in brat
252
+ {
253
+ "trigger": datasets.Value(
254
+ "string"
255
+ ), # refers to the text_bound_annotation of the trigger,
256
+ "id": datasets.Value("string"),
257
+ "type": datasets.Value("string"),
258
+ "arguments": datasets.Sequence(
259
+ {
260
+ "role": datasets.Value("string"),
261
+ "ref_id": datasets.Value("string"),
262
+ }
263
+ ),
264
+ }
265
+ ],
266
+ "relations": [ # R line in brat
267
+ {
268
+ "id": datasets.Value("string"),
269
+ "head": {
270
+ "ref_id": datasets.Value("string"),
271
+ "role": datasets.Value("string"),
272
+ },
273
+ "tail": {
274
+ "ref_id": datasets.Value("string"),
275
+ "role": datasets.Value("string"),
276
+ },
277
+ "type": datasets.Value("string"),
278
+ }
279
+ ],
280
+ "equivalences": [ # Equiv line in brat
281
+ {
282
+ "id": datasets.Value("string"),
283
+ "ref_ids": datasets.Sequence(datasets.Value("string")),
284
+ }
285
+ ],
286
+ "attributes": [ # M or A lines in brat
287
+ {
288
+ "id": datasets.Value("string"),
289
+ "type": datasets.Value("string"),
290
+ "ref_id": datasets.Value("string"),
291
+ "value": datasets.Value("string"),
292
+ }
293
+ ],
294
+ "normalizations": [ # N lines in brat
295
+ {
296
+ "id": datasets.Value("string"),
297
+ "type": datasets.Value("string"),
298
+ "ref_id": datasets.Value("string"),
299
+ "resource_name": datasets.Value(
300
+ "string"
301
+ ), # Name of the resource, e.g. "Wikipedia"
302
+ "cuid": datasets.Value(
303
+ "string"
304
+ ), # ID in the resource, e.g. 534366
305
+ "text": datasets.Value(
306
+ "string"
307
+ ), # Human readable description/name of the entity, e.g. "Barack Obama"
308
+ }
309
+ ],
310
+ ### OPTIONAL: Only included when `parse_notes == True`
311
+ "notes": [ # # lines in brat
312
+ {
313
+ "id": datasets.Value("string"),
314
+ "type": datasets.Value("string"),
315
+ "ref_id": datasets.Value("string"),
316
+ "text": datasets.Value("string"),
317
+ }
318
+ ],
319
+ },
320
+ )
321
+ """
322
+
323
+ example = {}
324
+ example["document_id"] = txt_file.with_suffix("").name
325
+ with txt_file.open() as f:
326
+ example["text"] = f.read()
327
+
328
+ # If no specific suffixes of the to-be-read annotation files are given - take standard suffixes
329
+ # for event extraction
330
+ if annotation_file_suffixes is None:
331
+ annotation_file_suffixes = [".a1", ".a2", ".ann"]
332
+
333
+ if len(annotation_file_suffixes) == 0:
334
+ raise AssertionError(
335
+ "At least one suffix for the to-be-read annotation files should be given!"
336
+ )
337
+
338
+ ann_lines = []
339
+ for suffix in annotation_file_suffixes:
340
+ annotation_file = txt_file.with_suffix(suffix)
341
+ try:
342
+ with annotation_file.open() as f:
343
+ ann_lines.extend(f.readlines())
344
+ except Exception:
345
+ continue
346
+
347
+ example["text_bound_annotations"] = []
348
+ example["events"] = []
349
+ example["relations"] = []
350
+ example["equivalences"] = []
351
+ example["attributes"] = []
352
+ example["normalizations"] = []
353
+
354
+ if parse_notes:
355
+ example["notes"] = []
356
+
357
+ for line in ann_lines:
358
+ line = line.strip()
359
+ if not line:
360
+ continue
361
+
362
+ if line.startswith("T"): # Text bound
363
+ ann = {}
364
+ fields = line.split("\t")
365
+
366
+ ann["id"] = fields[0]
367
+ ann["type"] = fields[1].split()[0]
368
+ ann["offsets"] = []
369
+ span_str = remove_prefix(fields[1], (ann["type"] + " "))
370
+ text = fields[2]
371
+ for span in span_str.split(";"):
372
+ start, end = span.split()
373
+ ann["offsets"].append([int(start), int(end)])
374
+
375
+ # Heuristically split text of discontiguous entities into chunks
376
+ ann["text"] = []
377
+ if len(ann["offsets"]) > 1:
378
+ i = 0
379
+ for start, end in ann["offsets"]:
380
+ chunk_len = end - start
381
+ ann["text"].append(text[i : chunk_len + i])
382
+ i += chunk_len
383
+ while i < len(text) and text[i] == " ":
384
+ i += 1
385
+ else:
386
+ ann["text"] = [text]
387
+
388
+ example["text_bound_annotations"].append(ann)
389
+
390
+ elif line.startswith("E"):
391
+ ann = {}
392
+ fields = line.split("\t")
393
+
394
+ ann["id"] = fields[0]
395
+
396
+ ann["type"], ann["trigger"] = fields[1].split()[0].split(":")
397
+
398
+ ann["arguments"] = []
399
+ for role_ref_id in fields[1].split()[1:]:
400
+ argument = {
401
+ "role": (role_ref_id.split(":"))[0],
402
+ "ref_id": (role_ref_id.split(":"))[1],
403
+ }
404
+ ann["arguments"].append(argument)
405
+
406
+ example["events"].append(ann)
407
+
408
+ elif line.startswith("R"):
409
+ ann = {}
410
+ fields = line.split("\t")
411
+
412
+ ann["id"] = fields[0]
413
+ ann["type"] = fields[1].split()[0]
414
+
415
+ ann["head"] = {
416
+ "role": fields[1].split()[1].split(":")[0],
417
+ "ref_id": fields[1].split()[1].split(":")[1],
418
+ }
419
+ ann["tail"] = {
420
+ "role": fields[1].split()[2].split(":")[0],
421
+ "ref_id": fields[1].split()[2].split(":")[1],
422
+ }
423
+
424
+ example["relations"].append(ann)
425
+
426
+ # '*' seems to be the legacy way to mark equivalences,
427
+ # but I couldn't find any info on the current way
428
+ # this might have to be adapted dependent on the brat version
429
+ # of the annotation
430
+ elif line.startswith("*"):
431
+ ann = {}
432
+ fields = line.split("\t")
433
+
434
+ ann["id"] = fields[0]
435
+ ann["ref_ids"] = fields[1].split()[1:]
436
+
437
+ example["equivalences"].append(ann)
438
+
439
+ elif line.startswith("A") or line.startswith("M"):
440
+ ann = {}
441
+ fields = line.split("\t")
442
+
443
+ ann["id"] = fields[0]
444
+
445
+ info = fields[1].split()
446
+ ann["type"] = info[0]
447
+ ann["ref_id"] = info[1]
448
+
449
+ if len(info) > 2:
450
+ ann["value"] = info[2]
451
+ else:
452
+ ann["value"] = ""
453
+
454
+ example["attributes"].append(ann)
455
+
456
+ elif line.startswith("N"):
457
+ ann = {}
458
+ fields = line.split("\t")
459
+
460
+ ann["id"] = fields[0]
461
+ ann["text"] = fields[2]
462
+
463
+ info = fields[1].split()
464
+
465
+ ann["type"] = info[0]
466
+ ann["ref_id"] = info[1]
467
+ ann["resource_name"] = info[2].split(":")[0]
468
+ ann["cuid"] = info[2].split(":")[1]
469
+ example["normalizations"].append(ann)
470
+
471
+ elif parse_notes and line.startswith("#"):
472
+ ann = {}
473
+ fields = line.split("\t")
474
+
475
+ ann["id"] = fields[0]
476
+ ann["text"] = fields[2] if len(fields) == 3 else BigBioValues.NULL
477
+
478
+ info = fields[1].split()
479
+
480
+ ann["type"] = info[0]
481
+ ann["ref_id"] = info[1]
482
+ example["notes"].append(ann)
483
+
484
+ return example
485
+
486
+
487
+ def brat_parse_to_bigbio_kb(brat_parse: Dict) -> Dict:
488
+ """
489
+ Transform a brat parse (conforming to the standard brat schema) obtained with
490
+ `parse_brat_file` into a dictionary conforming to the `bigbio-kb` schema (as defined in ../schemas/kb.py)
491
+ :param brat_parse:
492
+ """
493
+
494
+ unified_example = {}
495
+
496
+ # Prefix all ids with document id to ensure global uniqueness,
497
+ # because brat ids are only unique within their document
498
+ id_prefix = brat_parse["document_id"] + "_"
499
+
500
+ # identical
501
+ unified_example["document_id"] = brat_parse["document_id"]
502
+ unified_example["passages"] = [
503
+ {
504
+ "id": id_prefix + "_text",
505
+ "type": "abstract",
506
+ "text": [brat_parse["text"]],
507
+ "offsets": [[0, len(brat_parse["text"])]],
508
+ }
509
+ ]
510
+
511
+ # get normalizations
512
+ ref_id_to_normalizations = defaultdict(list)
513
+ for normalization in brat_parse["normalizations"]:
514
+ ref_id_to_normalizations[normalization["ref_id"]].append(
515
+ {
516
+ "db_name": normalization["resource_name"],
517
+ "db_id": normalization["cuid"],
518
+ }
519
+ )
520
+
521
+ # separate entities and event triggers
522
+ unified_example["events"] = []
523
+ non_event_ann = brat_parse["text_bound_annotations"].copy()
524
+ for event in brat_parse["events"]:
525
+ event = event.copy()
526
+ event["id"] = id_prefix + event["id"]
527
+ trigger = next(
528
+ tr
529
+ for tr in brat_parse["text_bound_annotations"]
530
+ if tr["id"] == event["trigger"]
531
+ )
532
+ if trigger in non_event_ann:
533
+ non_event_ann.remove(trigger)
534
+ event["trigger"] = {
535
+ "text": trigger["text"].copy(),
536
+ "offsets": trigger["offsets"].copy(),
537
+ }
538
+ for argument in event["arguments"]:
539
+ argument["ref_id"] = id_prefix + argument["ref_id"]
540
+
541
+ unified_example["events"].append(event)
542
+
543
+ unified_example["entities"] = []
544
+ anno_ids = [ref_id["id"] for ref_id in non_event_ann]
545
+ for ann in non_event_ann:
546
+ entity_ann = ann.copy()
547
+ entity_ann["id"] = id_prefix + entity_ann["id"]
548
+ entity_ann["normalized"] = ref_id_to_normalizations[ann["id"]]
549
+ unified_example["entities"].append(entity_ann)
550
+
551
+ # massage relations
552
+ unified_example["relations"] = []
553
+ skipped_relations = set()
554
+ for ann in brat_parse["relations"]:
555
+ if (
556
+ ann["head"]["ref_id"] not in anno_ids
557
+ or ann["tail"]["ref_id"] not in anno_ids
558
+ ):
559
+ skipped_relations.add(ann["id"])
560
+ continue
561
+ unified_example["relations"].append(
562
+ {
563
+ "arg1_id": id_prefix + ann["head"]["ref_id"],
564
+ "arg2_id": id_prefix + ann["tail"]["ref_id"],
565
+ "id": id_prefix + ann["id"],
566
+ "type": ann["type"],
567
+ "normalized": [],
568
+ }
569
+ )
570
+ if len(skipped_relations) > 0:
571
+ example_id = brat_parse["document_id"]
572
+ logger.info(
573
+ f"Example:{example_id}: The `bigbio_kb` schema allows `relations` only between entities."
574
+ f" Skip (for now): "
575
+ f"{list(skipped_relations)}"
576
+ )
577
+
578
+ # get coreferences
579
+ unified_example["coreferences"] = []
580
+ for i, ann in enumerate(brat_parse["equivalences"], start=1):
581
+ is_entity_cluster = True
582
+ for ref_id in ann["ref_ids"]:
583
+ if not ref_id.startswith("T"): # not textbound -> no entity
584
+ is_entity_cluster = False
585
+ elif ref_id not in anno_ids: # event trigger -> no entity
586
+ is_entity_cluster = False
587
+ if is_entity_cluster:
588
+ entity_ids = [id_prefix + i for i in ann["ref_ids"]]
589
+ unified_example["coreferences"].append(
590
+ {"id": id_prefix + str(i), "entity_ids": entity_ids}
591
+ )
592
+ return unified_example