Datasets:

Languages:
English
License:
gabrielaltay commited on
Commit
cc0c03a
•
1 Parent(s): fb40aa1

initial commit

Browse files
Files changed (3) hide show
  1. README.md +40 -0
  2. bigbiohub.py +153 -0
  3. mednli.py +201 -0
README.md CHANGED
@@ -1,3 +1,43 @@
1
  ---
 
2
  license: other
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language: en
3
  license: other
4
+ multilinguality: monolingual
5
+ pretty_name: MedNLI
6
+ paperswithcode_id: mednli
7
  ---
8
+
9
+
10
+ # Dataset Card for MedNLI
11
+
12
+ ## Dataset Description
13
+
14
+ - **Homepage:** https://physionet.org/content/mednli/1.0.0/
15
+ - **Pubmed:** False
16
+ - **Public:** False
17
+ - **Tasks:** Textual Entailment
18
+
19
+
20
+ State of the art models using deep neural networks have become very good in learning an accurate
21
+ mapping from inputs to outputs. However, they still lack generalization capabilities in conditions
22
+ that differ from the ones encountered during training. This is even more challenging in specialized,
23
+ and knowledge intensive domains, where training data is limited. To address this gap, we introduce
24
+ MedNLI - a dataset annotated by doctors, performing a natural language inference task (NLI),
25
+ grounded in the medical history of patients. As the source of premise sentences, we used the
26
+ MIMIC-III. More specifically, to minimize the risks to patient privacy, we worked with clinical
27
+ notes corresponding to the deceased patients. The clinicians in our team suggested the Past Medical
28
+ History to be the most informative section of a clinical note, from which useful inferences can be
29
+ drawn about the patient.
30
+
31
+
32
+ ## Citation Information
33
+
34
+ ```
35
+ @misc{https://doi.org/10.13026/c2rs98,
36
+ title = {MedNLI — A Natural Language Inference Dataset For The Clinical Domain},
37
+ author = {Shivade, Chaitanya},
38
+ year = 2017,
39
+ publisher = {physionet.org},
40
+ doi = {10.13026/C2RS98},
41
+ url = {https://physionet.org/content/mednli/}
42
+ }
43
+ ```
bigbiohub.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from dataclasses import dataclass
2
+ from enum import Enum
3
+ import datasets
4
+ from types import SimpleNamespace
5
+
6
+
7
+ BigBioValues = SimpleNamespace(NULL="<BB_NULL_STR>")
8
+
9
+
10
+ @dataclass
11
+ class BigBioConfig(datasets.BuilderConfig):
12
+ """BuilderConfig for BigBio."""
13
+
14
+ name: str = None
15
+ version: datasets.Version = None
16
+ description: str = None
17
+ schema: str = None
18
+ subset_id: str = None
19
+
20
+
21
+ class Tasks(Enum):
22
+ NAMED_ENTITY_RECOGNITION = "NER"
23
+ NAMED_ENTITY_DISAMBIGUATION = "NED"
24
+ EVENT_EXTRACTION = "EE"
25
+ RELATION_EXTRACTION = "RE"
26
+ COREFERENCE_RESOLUTION = "COREF"
27
+ QUESTION_ANSWERING = "QA"
28
+ TEXTUAL_ENTAILMENT = "TE"
29
+ SEMANTIC_SIMILARITY = "STS"
30
+ TEXT_PAIRS_CLASSIFICATION = "TXT2CLASS"
31
+ PARAPHRASING = "PARA"
32
+ TRANSLATION = "TRANSL"
33
+ SUMMARIZATION = "SUM"
34
+ TEXT_CLASSIFICATION = "TXTCLASS"
35
+
36
+
37
+ entailment_features = datasets.Features(
38
+ {
39
+ "id": datasets.Value("string"),
40
+ "premise": datasets.Value("string"),
41
+ "hypothesis": datasets.Value("string"),
42
+ "label": datasets.Value("string"),
43
+ }
44
+ )
45
+
46
+ pairs_features = datasets.Features(
47
+ {
48
+ "id": datasets.Value("string"),
49
+ "document_id": datasets.Value("string"),
50
+ "text_1": datasets.Value("string"),
51
+ "text_2": datasets.Value("string"),
52
+ "label": datasets.Value("string"),
53
+ }
54
+ )
55
+
56
+ qa_features = datasets.Features(
57
+ {
58
+ "id": datasets.Value("string"),
59
+ "question_id": datasets.Value("string"),
60
+ "document_id": datasets.Value("string"),
61
+ "question": datasets.Value("string"),
62
+ "type": datasets.Value("string"),
63
+ "choices": [datasets.Value("string")],
64
+ "context": datasets.Value("string"),
65
+ "answer": datasets.Sequence(datasets.Value("string")),
66
+ }
67
+ )
68
+
69
+ text_features = datasets.Features(
70
+ {
71
+ "id": datasets.Value("string"),
72
+ "document_id": datasets.Value("string"),
73
+ "text": datasets.Value("string"),
74
+ "labels": [datasets.Value("string")],
75
+ }
76
+ )
77
+
78
+ text2text_features = datasets.Features(
79
+ {
80
+ "id": datasets.Value("string"),
81
+ "document_id": datasets.Value("string"),
82
+ "text_1": datasets.Value("string"),
83
+ "text_2": datasets.Value("string"),
84
+ "text_1_name": datasets.Value("string"),
85
+ "text_2_name": datasets.Value("string"),
86
+ }
87
+ )
88
+
89
+ kb_features = datasets.Features(
90
+ {
91
+ "id": datasets.Value("string"),
92
+ "document_id": datasets.Value("string"),
93
+ "passages": [
94
+ {
95
+ "id": datasets.Value("string"),
96
+ "type": datasets.Value("string"),
97
+ "text": datasets.Sequence(datasets.Value("string")),
98
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
99
+ }
100
+ ],
101
+ "entities": [
102
+ {
103
+ "id": datasets.Value("string"),
104
+ "type": datasets.Value("string"),
105
+ "text": datasets.Sequence(datasets.Value("string")),
106
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
107
+ "normalized": [
108
+ {
109
+ "db_name": datasets.Value("string"),
110
+ "db_id": datasets.Value("string"),
111
+ }
112
+ ],
113
+ }
114
+ ],
115
+ "events": [
116
+ {
117
+ "id": datasets.Value("string"),
118
+ "type": datasets.Value("string"),
119
+ # refers to the text_bound_annotation of the trigger
120
+ "trigger": {
121
+ "text": datasets.Sequence(datasets.Value("string")),
122
+ "offsets": datasets.Sequence([datasets.Value("int32")]),
123
+ },
124
+ "arguments": [
125
+ {
126
+ "role": datasets.Value("string"),
127
+ "ref_id": datasets.Value("string"),
128
+ }
129
+ ],
130
+ }
131
+ ],
132
+ "coreferences": [
133
+ {
134
+ "id": datasets.Value("string"),
135
+ "entity_ids": datasets.Sequence(datasets.Value("string")),
136
+ }
137
+ ],
138
+ "relations": [
139
+ {
140
+ "id": datasets.Value("string"),
141
+ "type": datasets.Value("string"),
142
+ "arg1_id": datasets.Value("string"),
143
+ "arg2_id": datasets.Value("string"),
144
+ "normalized": [
145
+ {
146
+ "db_name": datasets.Value("string"),
147
+ "db_id": datasets.Value("string"),
148
+ }
149
+ ],
150
+ }
151
+ ],
152
+ }
153
+ )
mednli.py ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ """
17
+ State of the art models using deep neural networks have become very good in learning an accurate
18
+ mapping from inputs to outputs. However, they still lack generalization capabilities in conditions
19
+ that differ from the ones encountered during training. This is even more challenging in specialized,
20
+ and knowledge intensive domains, where training data is limited. To address this gap, we introduce
21
+ MedNLI - a dataset annotated by doctors, performing a natural language inference task (NLI),
22
+ grounded in the medical history of patients. As the source of premise sentences, we used the
23
+ MIMIC-III. More specifically, to minimize the risks to patient privacy, we worked with clinical
24
+ notes corresponding to the deceased patients. The clinicians in our team suggested the Past Medical
25
+ History to be the most informative section of a clinical note, from which useful inferences can be
26
+ drawn about the patient.
27
+
28
+ The files comprising this dataset must be on the users local machine in a single directory that is
29
+ passed to `datasets.load_datset` via the `data_dir` kwarg. This loader script will read the archive
30
+ files directly (i.e. the user should not uncompress, untar or unzip any of the files). For example,
31
+ if `data_dir` is `"mednli"` it should contain the following files:
32
+
33
+ mednli
34
+ ├── mednli-a-natural-language-inference-dataset-for-the-clinical-domain-1.0.0.zip
35
+ """
36
+
37
+ import json
38
+ import os
39
+ from typing import Dict, List, Tuple
40
+
41
+ import datasets
42
+
43
+ from .bigbiohub import entailment_features
44
+ from .bigbiohub import BigBioConfig
45
+ from .bigbiohub import Tasks
46
+
47
+
48
+ _LANGUAGES = ["English"]
49
+ _PUBMED = False
50
+ _LOCAL = True
51
+ _CITATION = """\
52
+ @misc{https://doi.org/10.13026/c2rs98,
53
+ title = {MedNLI — A Natural Language Inference Dataset For The Clinical Domain},
54
+ author = {Shivade, Chaitanya},
55
+ year = 2017,
56
+ publisher = {physionet.org},
57
+ doi = {10.13026/C2RS98},
58
+ url = {https://physionet.org/content/mednli/}
59
+ }
60
+ """
61
+
62
+
63
+ _DATASETNAME = "mednli"
64
+ _DISPLAYNAME = "MedNLI"
65
+
66
+ _DESCRIPTION = """\
67
+ State of the art models using deep neural networks have become very good in learning an accurate
68
+ mapping from inputs to outputs. However, they still lack generalization capabilities in conditions
69
+ that differ from the ones encountered during training. This is even more challenging in specialized,
70
+ and knowledge intensive domains, where training data is limited. To address this gap, we introduce
71
+ MedNLI - a dataset annotated by doctors, performing a natural language inference task (NLI),
72
+ grounded in the medical history of patients. As the source of premise sentences, we used the
73
+ MIMIC-III. More specifically, to minimize the risks to patient privacy, we worked with clinical
74
+ notes corresponding to the deceased patients. The clinicians in our team suggested the Past Medical
75
+ History to be the most informative section of a clinical note, from which useful inferences can be
76
+ drawn about the patient.
77
+ """
78
+
79
+
80
+ _HOMEPAGE = "https://physionet.org/content/mednli/1.0.0/"
81
+
82
+ _LICENSE = "PHYSIONET_LICENSE_1p5"
83
+
84
+ _URLS = {}
85
+
86
+ _SUPPORTED_TASKS = [Tasks.TEXTUAL_ENTAILMENT]
87
+
88
+ _SOURCE_VERSION = "1.0.0"
89
+ _BIGBIO_VERSION = "1.0.0"
90
+
91
+
92
+ class MedNLIDataset(datasets.GeneratorBasedBuilder):
93
+ """MedNLI"""
94
+
95
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
96
+ BIGBIO_VERSION = datasets.Version(_BIGBIO_VERSION)
97
+
98
+ BUILDER_CONFIGS = [
99
+ BigBioConfig(
100
+ name="mednli_source",
101
+ version=SOURCE_VERSION,
102
+ description="MedNLI source schema",
103
+ schema="source",
104
+ subset_id="mednli",
105
+ ),
106
+ BigBioConfig(
107
+ name="mednli_bigbio_te",
108
+ version=BIGBIO_VERSION,
109
+ description="MedNLI BigBio schema",
110
+ schema="bigbio_te",
111
+ subset_id="mednli",
112
+ ),
113
+ ]
114
+
115
+ DEFAULT_CONFIG_NAME = "mednli_source"
116
+
117
+ def _info(self) -> datasets.DatasetInfo:
118
+
119
+ if self.config.schema == "source":
120
+ features = datasets.Features(
121
+ {
122
+ "pairID": datasets.Value("string"),
123
+ "gold_label": datasets.Value("string"),
124
+ "sentence1": datasets.Value("string"),
125
+ "sentence2": datasets.Value("string"),
126
+ "sentence1_parse": datasets.Value("string"),
127
+ "sentence2_parse": datasets.Value("string"),
128
+ "sentence1_binary_parse": datasets.Value("string"),
129
+ "sentence2_binary_parse": datasets.Value("string"),
130
+ }
131
+ )
132
+
133
+ elif self.config.schema == "bigbio_te":
134
+ features = entailment_features
135
+
136
+ return datasets.DatasetInfo(
137
+ description=_DESCRIPTION,
138
+ features=features,
139
+ homepage=_HOMEPAGE,
140
+ license=str(_LICENSE),
141
+ citation=_CITATION,
142
+ )
143
+
144
+ def _split_generators(self, dl_manager) -> List[datasets.SplitGenerator]:
145
+ if self.config.data_dir is None:
146
+ raise ValueError(
147
+ "This is a local dataset. Please pass the data_dir kwarg to load_dataset."
148
+ )
149
+ else:
150
+ extract_dir = dl_manager.extract(
151
+ os.path.join(
152
+ self.config.data_dir,
153
+ "mednli-a-natural-language-inference-dataset-for-the-clinical-domain-1.0.0.zip",
154
+ )
155
+ )
156
+ data_dir = os.path.join(
157
+ extract_dir,
158
+ "mednli-a-natural-language-inference-dataset-for-the-clinical-domain-1.0.0",
159
+ )
160
+
161
+ return [
162
+ datasets.SplitGenerator(
163
+ name=datasets.Split.TRAIN,
164
+ gen_kwargs={
165
+ "filepath": os.path.join(data_dir, "mli_train_v1.jsonl"),
166
+ "split": "train",
167
+ },
168
+ ),
169
+ datasets.SplitGenerator(
170
+ name=datasets.Split.TEST,
171
+ gen_kwargs={
172
+ "filepath": os.path.join(data_dir, "mli_test_v1.jsonl"),
173
+ "split": "test",
174
+ },
175
+ ),
176
+ datasets.SplitGenerator(
177
+ name=datasets.Split.VALIDATION,
178
+ gen_kwargs={
179
+ "filepath": os.path.join(data_dir, "mli_dev_v1.jsonl"),
180
+ "split": "dev",
181
+ },
182
+ ),
183
+ ]
184
+
185
+ def _generate_examples(self, filepath, split: str) -> Tuple[int, Dict]:
186
+ with open(filepath, "r") as f:
187
+ if self.config.schema == "source":
188
+ for line in f:
189
+ json_line = json.loads(line)
190
+ yield json_line["pairID"], json_line
191
+
192
+ elif self.config.schema == "bigbio_te":
193
+ for line in f:
194
+ json_line = json.loads(line)
195
+ entailment_example = {
196
+ "id": json_line["pairID"],
197
+ "premise": json_line["sentence1"],
198
+ "hypothesis": json_line["sentence2"],
199
+ "label": json_line["gold_label"],
200
+ }
201
+ yield json_line["pairID"], entailment_example