Datasets:
Tasks:
Token Classification
Modalities:
Text
Formats:
parquet
Sub-tasks:
named-entity-recognition
Size:
1M - 10M
ArXiv:
License:
Update files from the datasets library (from 1.6.0)
Browse filesRelease notes: https://github.com/huggingface/datasets/releases/tag/1.6.0
- README.md +5 -2
- dataset_infos.json +0 -0
- wikiann.py +78 -2
README.md
CHANGED
@@ -429,7 +429,10 @@ WikiANN (sometimes called PAN-X) is a multilingual named entity recognition data
|
|
429 |
|
430 |
### Data Fields
|
431 |
|
432 |
-
|
|
|
|
|
|
|
433 |
|
434 |
### Data Splits
|
435 |
|
@@ -535,4 +538,4 @@ while the 176 languages supported in this version are associated with the follow
|
|
535 |
|
536 |
### Contributions
|
537 |
|
538 |
-
Thanks to [@lewtun](https://github.com/lewtun) for adding this dataset.
|
|
|
429 |
|
430 |
### Data Fields
|
431 |
|
432 |
+
- `tokens`: a `list` of `string` features.
|
433 |
+
- `langs`: a `list` of `string` features that correspond to the language of each token.
|
434 |
+
- `ner_tags`: a `list` of classification labels, with possible values including `O` (0), `B-PER` (1), `I-PER` (2), `B-ORG` (3), `I-ORG` (4), `B-LOC` (5), `I-LOC` (6).
|
435 |
+
- `spans`: a `list` of `string` features, that is the list of named entities in the input text formatted as ``<TAG>: <mention>``
|
436 |
|
437 |
### Data Splits
|
438 |
|
|
|
538 |
|
539 |
### Contributions
|
540 |
|
541 |
+
Thanks to [@lewtun](https://github.com/lewtun) and [@rabeehk](https://github.com/rabeehk) for adding this dataset.
|
dataset_infos.json
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
wikiann.py
CHANGED
@@ -15,7 +15,6 @@
|
|
15 |
|
16 |
"""The WikiANN dataset for multilingual named entity recognition"""
|
17 |
|
18 |
-
from __future__ import absolute_import, division, print_function
|
19 |
|
20 |
import os
|
21 |
|
@@ -241,6 +240,61 @@ class Wikiann(datasets.GeneratorBasedBuilder):
|
|
241 |
WikiannConfig(name=lang, description=f"WikiANN NER examples in language {lang}") for lang in _LANGS
|
242 |
]
|
243 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
244 |
def _info(self):
|
245 |
features = datasets.Features(
|
246 |
{
|
@@ -259,6 +313,7 @@ class Wikiann(datasets.GeneratorBasedBuilder):
|
|
259 |
)
|
260 |
),
|
261 |
"langs": datasets.Sequence(datasets.Value("string")),
|
|
|
262 |
}
|
263 |
)
|
264 |
return datasets.DatasetInfo(
|
@@ -290,6 +345,26 @@ class Wikiann(datasets.GeneratorBasedBuilder):
|
|
290 |
]
|
291 |
|
292 |
def _generate_examples(self, filepath):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
guid_index = 1
|
294 |
with open(filepath, encoding="utf-8") as f:
|
295 |
tokens = []
|
@@ -298,7 +373,8 @@ class Wikiann(datasets.GeneratorBasedBuilder):
|
|
298 |
for line in f:
|
299 |
if line == "" or line == "\n":
|
300 |
if tokens:
|
301 |
-
|
|
|
302 |
guid_index += 1
|
303 |
tokens = []
|
304 |
ner_tags = []
|
|
|
15 |
|
16 |
"""The WikiANN dataset for multilingual named entity recognition"""
|
17 |
|
|
|
18 |
|
19 |
import os
|
20 |
|
|
|
240 |
WikiannConfig(name=lang, description=f"WikiANN NER examples in language {lang}") for lang in _LANGS
|
241 |
]
|
242 |
|
243 |
+
def _tags_to_spans(self, tags):
|
244 |
+
"""Convert tags to spans."""
|
245 |
+
spans = set()
|
246 |
+
span_start = 0
|
247 |
+
span_end = 0
|
248 |
+
active_conll_tag = None
|
249 |
+
for index, string_tag in enumerate(tags):
|
250 |
+
# Actual BIO tag.
|
251 |
+
bio_tag = string_tag[0]
|
252 |
+
assert bio_tag in ["B", "I", "O"], "Invalid Tag"
|
253 |
+
conll_tag = string_tag[2:]
|
254 |
+
if bio_tag == "O":
|
255 |
+
# The span has ended.
|
256 |
+
if active_conll_tag:
|
257 |
+
spans.add((active_conll_tag, (span_start, span_end)))
|
258 |
+
active_conll_tag = None
|
259 |
+
# We don't care about tags we are
|
260 |
+
# told to ignore, so we do nothing.
|
261 |
+
continue
|
262 |
+
elif bio_tag == "B":
|
263 |
+
# We are entering a new span; reset indices and active tag to new span.
|
264 |
+
if active_conll_tag:
|
265 |
+
spans.add((active_conll_tag, (span_start, span_end)))
|
266 |
+
active_conll_tag = conll_tag
|
267 |
+
span_start = index
|
268 |
+
span_end = index
|
269 |
+
elif bio_tag == "I" and conll_tag == active_conll_tag:
|
270 |
+
# We're inside a span.
|
271 |
+
span_end += 1
|
272 |
+
else:
|
273 |
+
# This is the case the bio label is an "I", but either:
|
274 |
+
# 1) the span hasn't started - i.e. an ill formed span.
|
275 |
+
# 2) We have IOB1 tagging scheme.
|
276 |
+
# We'll process the previous span if it exists, but also include this
|
277 |
+
# span. This is important, because otherwise, a model may get a perfect
|
278 |
+
# F1 score whilst still including false positive ill-formed spans.
|
279 |
+
if active_conll_tag:
|
280 |
+
spans.add((active_conll_tag, (span_start, span_end)))
|
281 |
+
active_conll_tag = conll_tag
|
282 |
+
span_start = index
|
283 |
+
span_end = index
|
284 |
+
# Last token might have been a part of a valid span.
|
285 |
+
if active_conll_tag:
|
286 |
+
spans.add((active_conll_tag, (span_start, span_end)))
|
287 |
+
# Return sorted list of spans
|
288 |
+
return sorted(list(spans), key=lambda x: x[1][0])
|
289 |
+
|
290 |
+
def _get_spans(self, tokens, tags):
|
291 |
+
"""Convert tags to textspans."""
|
292 |
+
spans = self._tags_to_spans(tags)
|
293 |
+
text_spans = [x[0] + ": " + " ".join([tokens[i] for i in range(x[1][0], x[1][1] + 1)]) for x in spans]
|
294 |
+
if not text_spans:
|
295 |
+
text_spans = ["None"]
|
296 |
+
return text_spans
|
297 |
+
|
298 |
def _info(self):
|
299 |
features = datasets.Features(
|
300 |
{
|
|
|
313 |
)
|
314 |
),
|
315 |
"langs": datasets.Sequence(datasets.Value("string")),
|
316 |
+
"spans": datasets.Sequence(datasets.Value("string")),
|
317 |
}
|
318 |
)
|
319 |
return datasets.DatasetInfo(
|
|
|
345 |
]
|
346 |
|
347 |
def _generate_examples(self, filepath):
|
348 |
+
"""Reads line by line format of the NER dataset and generates examples.
|
349 |
+
Input Format:
|
350 |
+
en:rick B-PER
|
351 |
+
en:and O
|
352 |
+
en:morty B-PER
|
353 |
+
en:are O
|
354 |
+
en:cool O
|
355 |
+
en:. O
|
356 |
+
Output Format:
|
357 |
+
{
|
358 |
+
'tokens': ["rick", "and", "morty", "are", "cool", "."],
|
359 |
+
'ner_tags': ["B-PER", "O" , "B-PER", "O", "O", "O"],
|
360 |
+
'langs': ["en", "en", "en", "en", "en", "en"]
|
361 |
+
'spans': ["PER: rick", "PER: morty"]
|
362 |
+
}
|
363 |
+
Args:
|
364 |
+
filepath: Path to file with line by line NER format.
|
365 |
+
Returns:
|
366 |
+
Examples with the format listed above.
|
367 |
+
"""
|
368 |
guid_index = 1
|
369 |
with open(filepath, encoding="utf-8") as f:
|
370 |
tokens = []
|
|
|
373 |
for line in f:
|
374 |
if line == "" or line == "\n":
|
375 |
if tokens:
|
376 |
+
spans = self._get_spans(tokens, ner_tags)
|
377 |
+
yield guid_index, {"tokens": tokens, "ner_tags": ner_tags, "langs": langs, "spans": spans}
|
378 |
guid_index += 1
|
379 |
tokens = []
|
380 |
ner_tags = []
|