Commit
·
8a1283f
1
Parent(s):
147f6f9
upgrade to pie-datasets 0.3.0
Browse files- README.md +80 -0
- brat.py +85 -77
- requirements.txt +1 -0
README.md
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# PIE Dataset Card for "conll2003"
|
2 |
+
|
3 |
+
This is a [PyTorch-IE](https://github.com/ChristophAlt/pytorch-ie) wrapper for the
|
4 |
+
[BRAT Huggingface dataset loading script](https://huggingface.co/datasets/DFKI-SLT/brat).
|
5 |
+
|
6 |
+
## Data Schema
|
7 |
+
|
8 |
+
The document type for this dataset is `BratDocument` or `BratDocumentWithMergedSpans`, depending on if the
|
9 |
+
data was loaded with `merge_fragmented_spans=True` (default: `False`). They define the following data fields:
|
10 |
+
|
11 |
+
- `text` (str)
|
12 |
+
- `id` (str, optional)
|
13 |
+
- `metadata` (dictionary, optional)
|
14 |
+
|
15 |
+
and the following annotation layers:
|
16 |
+
|
17 |
+
- `spans` (annotation type: `LabeledMultiSpan` in the case of `BratDocument` and `LabeledSpan` and in the case of `BratDocumentWithMergedSpans`, target: `text`)
|
18 |
+
- `relations` (annotation type: `BinaryRelation`, target: `spans`)
|
19 |
+
- `span_attributes` (annotation type: `Attribute`, target: `spans`)
|
20 |
+
- `relation_attributes` (annotation type: `Attribute`, target: `relations`)
|
21 |
+
|
22 |
+
The `Attribute` annotation type is defined as follows:
|
23 |
+
|
24 |
+
- `annotation` (type: `Annotation`): the annotation to which the attribute is attached
|
25 |
+
- `label` (type: `str`)
|
26 |
+
- `value` (type: `str`, optional)
|
27 |
+
- `score` (type: `float`, optional, not included in comparison)
|
28 |
+
|
29 |
+
See [here](https://github.com/ChristophAlt/pytorch-ie/blob/main/src/pytorch_ie/annotations.py) for the remaining annotation type definitions.
|
30 |
+
|
31 |
+
## Document Converters
|
32 |
+
|
33 |
+
The dataset provides no predefined document converters because the BRAT format is very flexible and can be used
|
34 |
+
for many different tasks. You can add your own document converter by doing the following:
|
35 |
+
|
36 |
+
```python
|
37 |
+
import dataclasses
|
38 |
+
from typing import Optional
|
39 |
+
|
40 |
+
from pytorch_ie.core import AnnotationList, annotation_field
|
41 |
+
from pytorch_ie.documents import TextBasedDocument
|
42 |
+
from pytorch_ie.annotations import LabeledSpan
|
43 |
+
|
44 |
+
from pie_datasets import DatasetDict
|
45 |
+
|
46 |
+
# define your document class
|
47 |
+
@dataclasses.dataclass
|
48 |
+
class MyDocument(TextBasedDocument):
|
49 |
+
my_field: Optional[str] = None
|
50 |
+
my_span_annotations: AnnotationList[LabeledSpan] = annotation_field(target="text")
|
51 |
+
|
52 |
+
# define your document converter
|
53 |
+
def my_converter(document: BratDocumentWithMergedSpans) -> MyDocument:
|
54 |
+
# create your document with the data from the original document.
|
55 |
+
# The fields "text", "id" and "metadata" are derived from the TextBasedDocument.
|
56 |
+
my_document = MyDocument(id=document.id, text=document.text, metadata=document.metadata, my_field="my_value")
|
57 |
+
|
58 |
+
# create a new span annotation
|
59 |
+
new_span = LabeledSpan(label="my_label", start=2, end=10)
|
60 |
+
# add the new span annotation to your document
|
61 |
+
my_document.my_span_annotations.append(new_span)
|
62 |
+
|
63 |
+
# add annotations from the document to your document
|
64 |
+
for span in document.spans:
|
65 |
+
# we need to copy the span because an annotation can only be attached to one document
|
66 |
+
my_document.my_span_annotations.append(span.copy())
|
67 |
+
|
68 |
+
return my_document
|
69 |
+
|
70 |
+
|
71 |
+
# load the dataset. We use the "merge_fragmented_spans" dataset variant here
|
72 |
+
# because it provides documents of type BratDocumentWithMergedSpans.
|
73 |
+
dataset = DatasetDict.load_dataset("pie/brat", name="merge_fragmented_spans", data_dir="path/to/brat/data")
|
74 |
+
|
75 |
+
# attach your document converter to the dataset
|
76 |
+
dataset.register_document_converter(my_converter)
|
77 |
+
|
78 |
+
# convert the dataset
|
79 |
+
converted_dataset = dataset.to_document_type(MyDocument)
|
80 |
+
```
|
brat.py
CHANGED
@@ -1,16 +1,14 @@
|
|
1 |
import dataclasses
|
2 |
import logging
|
|
|
3 |
from typing import Any, Dict, List, Optional, Tuple, Union
|
4 |
|
5 |
import datasets
|
6 |
-
import
|
7 |
-
from pytorch_ie.
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
_post_init_single_label,
|
12 |
-
)
|
13 |
-
from pytorch_ie.core import Annotation, AnnotationList, Document, annotation_field
|
14 |
|
15 |
logger = logging.getLogger(__name__)
|
16 |
|
@@ -28,41 +26,32 @@ def ld2dl(
|
|
28 |
|
29 |
@dataclasses.dataclass(eq=True, frozen=True)
|
30 |
class Attribute(Annotation):
|
31 |
-
|
32 |
label: str
|
33 |
value: Optional[str] = None
|
34 |
-
score: float =
|
35 |
-
|
36 |
-
def __post_init__(self) -> None:
|
37 |
-
_post_init_single_label(self)
|
38 |
|
39 |
|
40 |
@dataclasses.dataclass
|
41 |
-
class BratDocument(
|
42 |
-
text: str
|
43 |
-
id: Optional[str] = None
|
44 |
-
metadata: Dict[str, Any] = dataclasses.field(default_factory=dict)
|
45 |
spans: AnnotationList[LabeledMultiSpan] = annotation_field(target="text")
|
46 |
relations: AnnotationList[BinaryRelation] = annotation_field(target="spans")
|
47 |
-
|
48 |
-
|
49 |
|
50 |
|
51 |
@dataclasses.dataclass
|
52 |
-
class BratDocumentWithMergedSpans(
|
53 |
-
text: str
|
54 |
-
id: Optional[str] = None
|
55 |
-
metadata: Dict[str, Any] = dataclasses.field(default_factory=dict)
|
56 |
spans: AnnotationList[LabeledSpan] = annotation_field(target="text")
|
57 |
relations: AnnotationList[BinaryRelation] = annotation_field(target="spans")
|
58 |
-
|
59 |
-
|
60 |
|
61 |
|
62 |
def example_to_document(
|
63 |
-
example: Dict[str, Any],
|
64 |
) -> BratDocument:
|
65 |
-
if
|
66 |
doc = BratDocumentWithMergedSpans(text=example["context"], id=example["file_name"])
|
67 |
else:
|
68 |
doc = BratDocument(text=example["context"], id=example["file_name"])
|
@@ -85,14 +74,19 @@ def example_to_document(
|
|
85 |
f"joined span parts do not match stripped span text field content. "
|
86 |
f'joined_span_texts_stripped: "{joined_span_texts_stripped}" != stripped "text": "{span_text_stripped}"'
|
87 |
)
|
88 |
-
if
|
89 |
if len(starts) > 1:
|
90 |
# check if the text in between the fragments holds only space
|
91 |
-
merged_content_texts = [
|
92 |
-
|
|
|
|
|
|
|
|
|
93 |
if len(merged_content_texts_not_empty) > 0:
|
94 |
logger.warning(
|
95 |
-
f"document '{doc.id}' contains a non-contiguous span with text content in between
|
|
|
96 |
f"newly covered text parts: {merged_content_texts_not_empty}, "
|
97 |
f"merged span text: '{doc.text[starts[0]:ends[-1]]}', "
|
98 |
f"annotation: {span_dict}"
|
@@ -130,28 +124,29 @@ def example_to_document(
|
|
130 |
if len(events) > 0:
|
131 |
raise NotImplementedError("converting events is not yet implemented")
|
132 |
|
133 |
-
|
134 |
-
|
135 |
-
for
|
136 |
-
target_id =
|
137 |
if target_id in spans:
|
138 |
target_layer_name = "spans"
|
139 |
-
|
140 |
elif target_id in relations:
|
141 |
target_layer_name = "relations"
|
142 |
-
|
143 |
else:
|
144 |
-
raise Exception("only span and relation
|
145 |
-
|
146 |
-
|
147 |
-
label=
|
148 |
-
value=
|
149 |
)
|
150 |
-
|
151 |
-
|
152 |
|
153 |
-
doc.
|
154 |
-
doc.
|
|
|
155 |
|
156 |
normalizations = dl2ld(example["normalizations"])
|
157 |
if len(normalizations) > 0:
|
@@ -199,7 +194,8 @@ def document_to_example(
|
|
199 |
prev_ann_dict = span_dicts[span]
|
200 |
ann_dict = span_dict
|
201 |
logger.warning(
|
202 |
-
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']}
|
|
|
203 |
)
|
204 |
span_dicts[span] = span_dict
|
205 |
example["spans"] = ld2dl(list(span_dicts.values()), keys=["id", "type", "locations", "text"])
|
@@ -221,7 +217,8 @@ def document_to_example(
|
|
221 |
prev_ann_dict = relation_dicts[rel]
|
222 |
ann_dict = relation_dict
|
223 |
logger.warning(
|
224 |
-
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']}
|
|
|
225 |
)
|
226 |
relation_dicts[rel] = relation_dict
|
227 |
|
@@ -230,31 +227,41 @@ def document_to_example(
|
|
230 |
example["equivalence_relations"] = ld2dl([], keys=["type", "targets"])
|
231 |
example["events"] = ld2dl([], keys=["id", "type", "trigger", "arguments"])
|
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 |
example["attributions"] = ld2dl(
|
257 |
-
list(
|
258 |
)
|
259 |
example["normalizations"] = ld2dl(
|
260 |
[], keys=["id", "type", "target", "resource_id", "entity_id"]
|
@@ -267,31 +274,32 @@ def document_to_example(
|
|
267 |
class BratConfig(datasets.BuilderConfig):
|
268 |
"""BuilderConfig for BratDatasetLoader."""
|
269 |
|
270 |
-
def __init__(self,
|
271 |
"""BuilderConfig for DocRED.
|
|
|
272 |
Args:
|
273 |
**kwargs: keyword arguments forwarded to super.
|
274 |
"""
|
275 |
super().__init__(**kwargs)
|
276 |
-
self.
|
277 |
|
278 |
|
279 |
-
class BratDatasetLoader(
|
280 |
# this requires https://github.com/ChristophAlt/pytorch-ie/pull/288
|
281 |
DOCUMENT_TYPES = {
|
282 |
"default": BratDocument,
|
283 |
-
"
|
284 |
}
|
285 |
|
286 |
DEFAULT_CONFIG_NAME = "default"
|
287 |
BUILDER_CONFIGS = [
|
288 |
BratConfig(name="default"),
|
289 |
-
BratConfig(name="
|
290 |
]
|
291 |
|
292 |
BASE_DATASET_PATH = "DFKI-SLT/brat"
|
293 |
|
294 |
def _generate_document(self, example, **kwargs):
|
295 |
return example_to_document(
|
296 |
-
example,
|
297 |
)
|
|
|
1 |
import dataclasses
|
2 |
import logging
|
3 |
+
from collections import defaultdict
|
4 |
from typing import Any, Dict, List, Optional, Tuple, Union
|
5 |
|
6 |
import datasets
|
7 |
+
from pytorch_ie.annotations import BinaryRelation, LabeledMultiSpan, LabeledSpan
|
8 |
+
from pytorch_ie.core import Annotation, AnnotationList, annotation_field
|
9 |
+
from pytorch_ie.documents import TextBasedDocument
|
10 |
+
|
11 |
+
from pie_datasets import GeneratorBasedBuilder
|
|
|
|
|
|
|
12 |
|
13 |
logger = logging.getLogger(__name__)
|
14 |
|
|
|
26 |
|
27 |
@dataclasses.dataclass(eq=True, frozen=True)
|
28 |
class Attribute(Annotation):
|
29 |
+
annotation: Annotation
|
30 |
label: str
|
31 |
value: Optional[str] = None
|
32 |
+
score: Optional[float] = dataclasses.field(default=None, compare=False)
|
|
|
|
|
|
|
33 |
|
34 |
|
35 |
@dataclasses.dataclass
|
36 |
+
class BratDocument(TextBasedDocument):
|
|
|
|
|
|
|
37 |
spans: AnnotationList[LabeledMultiSpan] = annotation_field(target="text")
|
38 |
relations: AnnotationList[BinaryRelation] = annotation_field(target="spans")
|
39 |
+
span_attributes: AnnotationList[Attribute] = annotation_field(target="spans")
|
40 |
+
relation_attributes: AnnotationList[Attribute] = annotation_field(target="relations")
|
41 |
|
42 |
|
43 |
@dataclasses.dataclass
|
44 |
+
class BratDocumentWithMergedSpans(TextBasedDocument):
|
|
|
|
|
|
|
45 |
spans: AnnotationList[LabeledSpan] = annotation_field(target="text")
|
46 |
relations: AnnotationList[BinaryRelation] = annotation_field(target="spans")
|
47 |
+
span_attributes: AnnotationList[Attribute] = annotation_field(target="spans")
|
48 |
+
relation_attributes: AnnotationList[Attribute] = annotation_field(target="relations")
|
49 |
|
50 |
|
51 |
def example_to_document(
|
52 |
+
example: Dict[str, Any], merge_fragmented_spans: bool = False
|
53 |
) -> BratDocument:
|
54 |
+
if merge_fragmented_spans:
|
55 |
doc = BratDocumentWithMergedSpans(text=example["context"], id=example["file_name"])
|
56 |
else:
|
57 |
doc = BratDocument(text=example["context"], id=example["file_name"])
|
|
|
74 |
f"joined span parts do not match stripped span text field content. "
|
75 |
f'joined_span_texts_stripped: "{joined_span_texts_stripped}" != stripped "text": "{span_text_stripped}"'
|
76 |
)
|
77 |
+
if merge_fragmented_spans:
|
78 |
if len(starts) > 1:
|
79 |
# check if the text in between the fragments holds only space
|
80 |
+
merged_content_texts = [
|
81 |
+
doc.text[start:end] for start, end in zip(ends[:-1], starts[1:])
|
82 |
+
]
|
83 |
+
merged_content_texts_not_empty = [
|
84 |
+
text.strip() for text in merged_content_texts if text.strip() != ""
|
85 |
+
]
|
86 |
if len(merged_content_texts_not_empty) > 0:
|
87 |
logger.warning(
|
88 |
+
f"document '{doc.id}' contains a non-contiguous span with text content in between "
|
89 |
+
f"(will be merged into a single span): "
|
90 |
f"newly covered text parts: {merged_content_texts_not_empty}, "
|
91 |
f"merged span text: '{doc.text[starts[0]:ends[-1]]}', "
|
92 |
f"annotation: {span_dict}"
|
|
|
124 |
if len(events) > 0:
|
125 |
raise NotImplementedError("converting events is not yet implemented")
|
126 |
|
127 |
+
attribute_annotations: Dict[str, Dict[str, Attribute]] = defaultdict(dict)
|
128 |
+
attribute_ids = []
|
129 |
+
for attribute_dict in dl2ld(example["attributions"]):
|
130 |
+
target_id = attribute_dict["target"]
|
131 |
if target_id in spans:
|
132 |
target_layer_name = "spans"
|
133 |
+
annotation = spans[target_id]
|
134 |
elif target_id in relations:
|
135 |
target_layer_name = "relations"
|
136 |
+
annotation = relations[target_id]
|
137 |
else:
|
138 |
+
raise Exception("only span and relation attributes are supported yet")
|
139 |
+
attribute = Attribute(
|
140 |
+
annotation=annotation,
|
141 |
+
label=attribute_dict["type"],
|
142 |
+
value=attribute_dict["value"],
|
143 |
)
|
144 |
+
attribute_annotations[target_layer_name][attribute_dict["id"]] = attribute
|
145 |
+
attribute_ids.append((target_layer_name, attribute_dict["id"]))
|
146 |
|
147 |
+
doc.span_attributes.extend(attribute_annotations["spans"].values())
|
148 |
+
doc.relation_attributes.extend(attribute_annotations["relations"].values())
|
149 |
+
doc.metadata["attribute_ids"] = attribute_ids
|
150 |
|
151 |
normalizations = dl2ld(example["normalizations"])
|
152 |
if len(normalizations) > 0:
|
|
|
194 |
prev_ann_dict = span_dicts[span]
|
195 |
ann_dict = span_dict
|
196 |
logger.warning(
|
197 |
+
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
198 |
+
f"are identical"
|
199 |
)
|
200 |
span_dicts[span] = span_dict
|
201 |
example["spans"] = ld2dl(list(span_dicts.values()), keys=["id", "type", "locations", "text"])
|
|
|
217 |
prev_ann_dict = relation_dicts[rel]
|
218 |
ann_dict = relation_dict
|
219 |
logger.warning(
|
220 |
+
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
221 |
+
f"are identical"
|
222 |
)
|
223 |
relation_dicts[rel] = relation_dict
|
224 |
|
|
|
227 |
example["equivalence_relations"] = ld2dl([], keys=["type", "targets"])
|
228 |
example["events"] = ld2dl([], keys=["id", "type", "trigger", "arguments"])
|
229 |
|
230 |
+
annotation_dicts = {
|
231 |
+
"spans": span_dicts,
|
232 |
+
"relations": relation_dicts,
|
233 |
+
}
|
234 |
+
all_attribute_annotations = {
|
235 |
+
"spans": document.span_attributes,
|
236 |
+
"relations": document.relation_attributes,
|
237 |
+
}
|
238 |
+
attribute_dicts: Dict[Annotation, Dict[str, Any]] = dict()
|
239 |
+
attribute_ids_per_target = defaultdict(list)
|
240 |
+
for target_layer, attribute_id in document.metadata["attribute_ids"]:
|
241 |
+
attribute_ids_per_target[target_layer].append(attribute_id)
|
242 |
+
|
243 |
+
for target_layer, attribute_ids in attribute_ids_per_target.items():
|
244 |
+
attribute_annotations = all_attribute_annotations[target_layer]
|
245 |
+
assert len(attribute_ids) == len(attribute_annotations)
|
246 |
+
for i, attribute_annotation in enumerate(attribute_annotations):
|
247 |
+
target_id = annotation_dicts[target_layer][attribute_annotation.annotation]["id"]
|
248 |
+
attribute_dict = {
|
249 |
+
"id": attribute_ids_per_target[target_layer][i],
|
250 |
+
"type": attribute_annotation.label,
|
251 |
+
"target": target_id,
|
252 |
+
"value": attribute_annotation.value,
|
253 |
+
}
|
254 |
+
if attribute_annotation in attribute_dicts:
|
255 |
+
prev_ann_dict = attribute_dicts[attribute_annotation]
|
256 |
+
ann_dict = attribute_annotation
|
257 |
+
logger.warning(
|
258 |
+
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
259 |
+
f"are identical"
|
260 |
+
)
|
261 |
+
attribute_dicts[attribute_annotation] = attribute_dict
|
262 |
|
263 |
example["attributions"] = ld2dl(
|
264 |
+
list(attribute_dicts.values()), keys=["id", "type", "target", "value"]
|
265 |
)
|
266 |
example["normalizations"] = ld2dl(
|
267 |
[], keys=["id", "type", "target", "resource_id", "entity_id"]
|
|
|
274 |
class BratConfig(datasets.BuilderConfig):
|
275 |
"""BuilderConfig for BratDatasetLoader."""
|
276 |
|
277 |
+
def __init__(self, merge_fragmented_spans: bool = False, **kwargs):
|
278 |
"""BuilderConfig for DocRED.
|
279 |
+
|
280 |
Args:
|
281 |
**kwargs: keyword arguments forwarded to super.
|
282 |
"""
|
283 |
super().__init__(**kwargs)
|
284 |
+
self.merge_fragmented_spans = merge_fragmented_spans
|
285 |
|
286 |
|
287 |
+
class BratDatasetLoader(GeneratorBasedBuilder):
|
288 |
# this requires https://github.com/ChristophAlt/pytorch-ie/pull/288
|
289 |
DOCUMENT_TYPES = {
|
290 |
"default": BratDocument,
|
291 |
+
"merge_fragmented_spans": BratDocumentWithMergedSpans,
|
292 |
}
|
293 |
|
294 |
DEFAULT_CONFIG_NAME = "default"
|
295 |
BUILDER_CONFIGS = [
|
296 |
BratConfig(name="default"),
|
297 |
+
BratConfig(name="merge_fragmented_spans", merge_fragmented_spans=True),
|
298 |
]
|
299 |
|
300 |
BASE_DATASET_PATH = "DFKI-SLT/brat"
|
301 |
|
302 |
def _generate_document(self, example, **kwargs):
|
303 |
return example_to_document(
|
304 |
+
example, merge_fragmented_spans=self.config.merge_fragmented_spans
|
305 |
)
|
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
pie-datasets>=0.3.0
|