ArneBinder
commited on
Commit
·
c73ae7d
1
Parent(s):
67a4e57
pie-datasets 0.4.0
Browse files- brat.py +3 -282
- requirements.txt +1 -1
brat.py
CHANGED
@@ -1,284 +1,5 @@
|
|
1 |
-
import
|
2 |
-
from collections import defaultdict
|
3 |
-
from typing import Any, Dict, List, Optional, Tuple, Union
|
4 |
|
5 |
-
import datasets
|
6 |
-
from pytorch_ie.annotations import BinaryRelation, LabeledMultiSpan, LabeledSpan
|
7 |
-
from pytorch_ie.core import Annotation
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
Attribute,
|
12 |
-
BratDocument,
|
13 |
-
BratDocumentWithMergedSpans,
|
14 |
-
)
|
15 |
-
|
16 |
-
logger = logging.getLogger(__name__)
|
17 |
-
|
18 |
-
|
19 |
-
def dl2ld(dict_of_lists: Dict[str, List[Any]]) -> List[Dict[str, Any]]:
|
20 |
-
return [dict(zip(dict_of_lists, t)) for t in zip(*dict_of_lists.values())]
|
21 |
-
|
22 |
-
|
23 |
-
def ld2dl(
|
24 |
-
list_fo_dicts: List[Dict[str, Any]], keys: Optional[List[str]] = None
|
25 |
-
) -> Dict[str, List[Any]]:
|
26 |
-
keys = keys or list(list_fo_dicts[0])
|
27 |
-
return {k: [dic[k] for dic in list_fo_dicts] for k in keys}
|
28 |
-
|
29 |
-
|
30 |
-
def example_to_document(
|
31 |
-
example: Dict[str, Any], merge_fragmented_spans: bool = False
|
32 |
-
) -> Union[BratDocument, BratDocumentWithMergedSpans]:
|
33 |
-
if merge_fragmented_spans:
|
34 |
-
doc = BratDocumentWithMergedSpans(text=example["context"], id=example["file_name"])
|
35 |
-
else:
|
36 |
-
doc = BratDocument(text=example["context"], id=example["file_name"])
|
37 |
-
|
38 |
-
spans: Dict[str, LabeledSpan] = dict()
|
39 |
-
span_locations: List[Tuple[Tuple[int, int]]] = []
|
40 |
-
span_texts: List[str] = []
|
41 |
-
for span_dict in dl2ld(example["spans"]):
|
42 |
-
starts: List[int] = span_dict["locations"]["start"]
|
43 |
-
ends: List[int] = span_dict["locations"]["end"]
|
44 |
-
slices = tuple(zip(starts, ends))
|
45 |
-
span_locations.append(slices)
|
46 |
-
span_texts.append(span_dict["text"])
|
47 |
-
# sanity check
|
48 |
-
span_text_parts = [doc.text[start:end] for start, end in slices]
|
49 |
-
joined_span_texts_stripped = " ".join(span_text_parts).strip()
|
50 |
-
span_text_stripped = span_dict["text"].strip()
|
51 |
-
if joined_span_texts_stripped != span_text_stripped:
|
52 |
-
logger.warning(
|
53 |
-
f"joined span parts do not match stripped span text field content. "
|
54 |
-
f'joined_span_texts_stripped: "{joined_span_texts_stripped}" != stripped "text": "{span_text_stripped}"'
|
55 |
-
)
|
56 |
-
if merge_fragmented_spans:
|
57 |
-
if len(starts) > 1:
|
58 |
-
# check if the text in between the fragments holds only space
|
59 |
-
merged_content_texts = [
|
60 |
-
doc.text[start:end] for start, end in zip(ends[:-1], starts[1:])
|
61 |
-
]
|
62 |
-
merged_content_texts_not_empty = [
|
63 |
-
text.strip() for text in merged_content_texts if text.strip() != ""
|
64 |
-
]
|
65 |
-
if len(merged_content_texts_not_empty) > 0:
|
66 |
-
logger.warning(
|
67 |
-
f"document '{doc.id}' contains a non-contiguous span with text content in between "
|
68 |
-
f"(will be merged into a single span): "
|
69 |
-
f"newly covered text parts: {merged_content_texts_not_empty}, "
|
70 |
-
f"merged span text: '{doc.text[starts[0]:ends[-1]]}', "
|
71 |
-
f"annotation: {span_dict}"
|
72 |
-
)
|
73 |
-
# just take everything
|
74 |
-
start = min(starts)
|
75 |
-
end = max(ends)
|
76 |
-
span = LabeledSpan(start=start, end=end, label=span_dict["type"])
|
77 |
-
else:
|
78 |
-
span = LabeledMultiSpan(slices=slices, label=span_dict["type"])
|
79 |
-
spans[span_dict["id"]] = span
|
80 |
-
|
81 |
-
doc.spans.extend(spans.values())
|
82 |
-
doc.metadata["span_ids"] = list(spans.keys())
|
83 |
-
doc.metadata["span_locations"] = span_locations
|
84 |
-
doc.metadata["span_texts"] = span_texts
|
85 |
-
|
86 |
-
relations: Dict[str, BinaryRelation] = dict()
|
87 |
-
for rel_dict in dl2ld(example["relations"]):
|
88 |
-
arguments = dict(zip(rel_dict["arguments"]["type"], rel_dict["arguments"]["target"]))
|
89 |
-
assert set(arguments) == {"Arg1", "Arg2"}
|
90 |
-
head = spans[arguments["Arg1"]]
|
91 |
-
tail = spans[arguments["Arg2"]]
|
92 |
-
rel = BinaryRelation(head=head, tail=tail, label=rel_dict["type"])
|
93 |
-
relations[rel_dict["id"]] = rel
|
94 |
-
|
95 |
-
doc.relations.extend(relations.values())
|
96 |
-
doc.metadata["relation_ids"] = list(relations.keys())
|
97 |
-
|
98 |
-
equivalence_relations = dl2ld(example["equivalence_relations"])
|
99 |
-
if len(equivalence_relations) > 0:
|
100 |
-
raise NotImplementedError("converting equivalence_relations is not yet implemented")
|
101 |
-
|
102 |
-
events = dl2ld(example["events"])
|
103 |
-
if len(events) > 0:
|
104 |
-
raise NotImplementedError("converting events is not yet implemented")
|
105 |
-
|
106 |
-
attribute_annotations: Dict[str, Dict[str, Attribute]] = defaultdict(dict)
|
107 |
-
attribute_ids = []
|
108 |
-
for attribute_dict in dl2ld(example["attributions"]):
|
109 |
-
target_id = attribute_dict["target"]
|
110 |
-
if target_id in spans:
|
111 |
-
target_layer_name = "spans"
|
112 |
-
annotation = spans[target_id]
|
113 |
-
elif target_id in relations:
|
114 |
-
target_layer_name = "relations"
|
115 |
-
annotation = relations[target_id]
|
116 |
-
else:
|
117 |
-
raise Exception("only span and relation attributes are supported yet")
|
118 |
-
attribute = Attribute(
|
119 |
-
annotation=annotation,
|
120 |
-
label=attribute_dict["type"],
|
121 |
-
value=attribute_dict["value"],
|
122 |
-
)
|
123 |
-
attribute_annotations[target_layer_name][attribute_dict["id"]] = attribute
|
124 |
-
attribute_ids.append((target_layer_name, attribute_dict["id"]))
|
125 |
-
|
126 |
-
doc.span_attributes.extend(attribute_annotations["spans"].values())
|
127 |
-
doc.relation_attributes.extend(attribute_annotations["relations"].values())
|
128 |
-
doc.metadata["attribute_ids"] = attribute_ids
|
129 |
-
|
130 |
-
normalizations = dl2ld(example["normalizations"])
|
131 |
-
if len(normalizations) > 0:
|
132 |
-
raise NotImplementedError("converting normalizations is not yet implemented")
|
133 |
-
|
134 |
-
notes = dl2ld(example["notes"])
|
135 |
-
if len(notes) > 0:
|
136 |
-
raise NotImplementedError("converting notes is not yet implemented")
|
137 |
-
|
138 |
-
return doc
|
139 |
-
|
140 |
-
|
141 |
-
def document_to_example(
|
142 |
-
document: Union[BratDocument, BratDocumentWithMergedSpans]
|
143 |
-
) -> Dict[str, Any]:
|
144 |
-
example = {
|
145 |
-
"context": document.text,
|
146 |
-
"file_name": document.id,
|
147 |
-
}
|
148 |
-
span_dicts: Dict[Union[LabeledSpan, LabeledMultiSpan], Dict[str, Any]] = dict()
|
149 |
-
assert len(document.metadata["span_locations"]) == len(document.spans)
|
150 |
-
assert len(document.metadata["span_texts"]) == len(document.spans)
|
151 |
-
assert len(document.metadata["span_ids"]) == len(document.spans)
|
152 |
-
for i, span in enumerate(document.spans):
|
153 |
-
locations = tuple((start, end) for start, end in document.metadata["span_locations"][i])
|
154 |
-
if isinstance(span, LabeledSpan):
|
155 |
-
assert locations[0][0] == span.start
|
156 |
-
assert locations[-1][1] == span.end
|
157 |
-
elif isinstance(span, LabeledMultiSpan):
|
158 |
-
assert span.slices == locations
|
159 |
-
else:
|
160 |
-
raise TypeError(f"span has unknown type [{type(span)}]: {span}")
|
161 |
-
|
162 |
-
starts, ends = zip(*locations)
|
163 |
-
span_dict = {
|
164 |
-
"id": document.metadata["span_ids"][i],
|
165 |
-
"locations": {
|
166 |
-
"start": list(starts),
|
167 |
-
"end": list(ends),
|
168 |
-
},
|
169 |
-
"text": document.metadata["span_texts"][i],
|
170 |
-
"type": span.label,
|
171 |
-
}
|
172 |
-
if span in span_dicts:
|
173 |
-
prev_ann_dict = span_dicts[span]
|
174 |
-
ann_dict = span_dict
|
175 |
-
logger.warning(
|
176 |
-
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
177 |
-
f"are identical"
|
178 |
-
)
|
179 |
-
span_dicts[span] = span_dict
|
180 |
-
example["spans"] = ld2dl(list(span_dicts.values()), keys=["id", "type", "locations", "text"])
|
181 |
-
|
182 |
-
relation_dicts: Dict[BinaryRelation, Dict[str, Any]] = dict()
|
183 |
-
assert len(document.metadata["relation_ids"]) == len(document.relations)
|
184 |
-
for i, rel in enumerate(document.relations):
|
185 |
-
arg1_id = span_dicts[rel.head]["id"]
|
186 |
-
arg2_id = span_dicts[rel.tail]["id"]
|
187 |
-
relation_dict = {
|
188 |
-
"id": document.metadata["relation_ids"][i],
|
189 |
-
"type": rel.label,
|
190 |
-
"arguments": {
|
191 |
-
"type": ["Arg1", "Arg2"],
|
192 |
-
"target": [arg1_id, arg2_id],
|
193 |
-
},
|
194 |
-
}
|
195 |
-
if rel in relation_dicts:
|
196 |
-
prev_ann_dict = relation_dicts[rel]
|
197 |
-
ann_dict = relation_dict
|
198 |
-
logger.warning(
|
199 |
-
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
200 |
-
f"are identical"
|
201 |
-
)
|
202 |
-
relation_dicts[rel] = relation_dict
|
203 |
-
|
204 |
-
example["relations"] = ld2dl(list(relation_dicts.values()), keys=["id", "type", "arguments"])
|
205 |
-
|
206 |
-
example["equivalence_relations"] = ld2dl([], keys=["type", "targets"])
|
207 |
-
example["events"] = ld2dl([], keys=["id", "type", "trigger", "arguments"])
|
208 |
-
|
209 |
-
annotation_dicts = {
|
210 |
-
"spans": span_dicts,
|
211 |
-
"relations": relation_dicts,
|
212 |
-
}
|
213 |
-
all_attribute_annotations = {
|
214 |
-
"spans": document.span_attributes,
|
215 |
-
"relations": document.relation_attributes,
|
216 |
-
}
|
217 |
-
attribute_dicts: Dict[Annotation, Dict[str, Any]] = dict()
|
218 |
-
attribute_ids_per_target = defaultdict(list)
|
219 |
-
for target_layer, attribute_id in document.metadata["attribute_ids"]:
|
220 |
-
attribute_ids_per_target[target_layer].append(attribute_id)
|
221 |
-
|
222 |
-
for target_layer, attribute_ids in attribute_ids_per_target.items():
|
223 |
-
attribute_annotations = all_attribute_annotations[target_layer]
|
224 |
-
assert len(attribute_ids) == len(attribute_annotations)
|
225 |
-
for i, attribute_annotation in enumerate(attribute_annotations):
|
226 |
-
target_id = annotation_dicts[target_layer][attribute_annotation.annotation]["id"]
|
227 |
-
attribute_dict = {
|
228 |
-
"id": attribute_ids_per_target[target_layer][i],
|
229 |
-
"type": attribute_annotation.label,
|
230 |
-
"target": target_id,
|
231 |
-
"value": attribute_annotation.value,
|
232 |
-
}
|
233 |
-
if attribute_annotation in attribute_dicts:
|
234 |
-
prev_ann_dict = attribute_dicts[attribute_annotation]
|
235 |
-
ann_dict = attribute_annotation
|
236 |
-
logger.warning(
|
237 |
-
f"document {document.id}: annotation exists twice: {prev_ann_dict['id']} and {ann_dict['id']} "
|
238 |
-
f"are identical"
|
239 |
-
)
|
240 |
-
attribute_dicts[attribute_annotation] = attribute_dict
|
241 |
-
|
242 |
-
example["attributions"] = ld2dl(
|
243 |
-
list(attribute_dicts.values()), keys=["id", "type", "target", "value"]
|
244 |
-
)
|
245 |
-
example["normalizations"] = ld2dl(
|
246 |
-
[], keys=["id", "type", "target", "resource_id", "entity_id"]
|
247 |
-
)
|
248 |
-
example["notes"] = ld2dl([], keys=["id", "type", "target", "note"])
|
249 |
-
|
250 |
-
return example
|
251 |
-
|
252 |
-
|
253 |
-
class BratConfig(datasets.BuilderConfig):
|
254 |
-
"""BuilderConfig for BratDatasetLoader."""
|
255 |
-
|
256 |
-
def __init__(self, merge_fragmented_spans: bool = False, **kwargs):
|
257 |
-
"""BuilderConfig for DocRED.
|
258 |
-
|
259 |
-
Args:
|
260 |
-
**kwargs: keyword arguments forwarded to super.
|
261 |
-
"""
|
262 |
-
super().__init__(**kwargs)
|
263 |
-
self.merge_fragmented_spans = merge_fragmented_spans
|
264 |
-
|
265 |
-
|
266 |
-
class BratDatasetLoader(GeneratorBasedBuilder):
|
267 |
-
# this requires https://github.com/ChristophAlt/pytorch-ie/pull/288
|
268 |
-
DOCUMENT_TYPES = {
|
269 |
-
"default": BratDocument,
|
270 |
-
"merge_fragmented_spans": BratDocumentWithMergedSpans,
|
271 |
-
}
|
272 |
-
|
273 |
-
DEFAULT_CONFIG_NAME = "default"
|
274 |
-
BUILDER_CONFIGS = [
|
275 |
-
BratConfig(name="default"),
|
276 |
-
BratConfig(name="merge_fragmented_spans", merge_fragmented_spans=True),
|
277 |
-
]
|
278 |
-
|
279 |
-
BASE_DATASET_PATH = "DFKI-SLT/brat"
|
280 |
-
|
281 |
-
def _generate_document(self, example, **kwargs):
|
282 |
-
return example_to_document(
|
283 |
-
example, merge_fragmented_spans=self.config.merge_fragmented_spans
|
284 |
-
)
|
|
|
1 |
+
from pie_datasets.builders import BratBuilder
|
|
|
|
|
2 |
|
|
|
|
|
|
|
3 |
|
4 |
+
class Brat(BratBuilder):
|
5 |
+
pass
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
requirements.txt
CHANGED
@@ -1 +1 @@
|
|
1 |
-
pie-datasets>=0.
|
|
|
1 |
+
pie-datasets>=0.4.0,<0.5.0
|