Abinaya Mahendiran
commited on
Commit
•
278a858
1
Parent(s):
305310d
Added data preparation script and updated data loader script
Browse files- data_preparation.py +112 -0
- squad_v2.py +38 -30
data_preparation.py
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
""" Script to prepare the SQuAD2.0 data to the GEM format
|
2 |
+
|
3 |
+
@author: AbinayaM02
|
4 |
+
"""
|
5 |
+
|
6 |
+
# Import libraries
|
7 |
+
import json
|
8 |
+
import pandas as pd
|
9 |
+
from sklearn.model_selection import train_test_split
|
10 |
+
|
11 |
+
# Function to generate gem id
|
12 |
+
def add_gem_id(data: dict, split: str) -> dict:
|
13 |
+
"""
|
14 |
+
Add gem id for each of the datapoint in the dataset.
|
15 |
+
|
16 |
+
Parameters:
|
17 |
+
-----------
|
18 |
+
data: dict,
|
19 |
+
data.
|
20 |
+
split: str,
|
21 |
+
split of data (train, test or validation).
|
22 |
+
|
23 |
+
Returns:
|
24 |
+
--------
|
25 |
+
dict
|
26 |
+
dictionary with updated id
|
27 |
+
"""
|
28 |
+
gem_id = -1
|
29 |
+
generated_data = {"data": []}
|
30 |
+
for example in data:
|
31 |
+
temp_dict = {}
|
32 |
+
title = example["title"]
|
33 |
+
for paragraph in example["paragraphs"]:
|
34 |
+
context = paragraph["context"] # do not strip leading blank spaces GH-2585
|
35 |
+
for qa in paragraph["qas"]:
|
36 |
+
question = qa["question"]
|
37 |
+
qa_id = qa["id"]
|
38 |
+
answer_starts = [answer["answer_start"] for answer in qa["answers"]]
|
39 |
+
answers = [answer["text"] for answer in qa["answers"]]
|
40 |
+
# Features currently used are "context", "question", and "answers".
|
41 |
+
# Others are extracted here for the ease of future expansions.
|
42 |
+
gem_id += 1
|
43 |
+
temp_dict["id"] = qa_id
|
44 |
+
temp_dict["gem_id"] = f"gem-squad_v2-{split}-{gem_id}"
|
45 |
+
temp_dict["title"] = title
|
46 |
+
temp_dict["context"] = context
|
47 |
+
temp_dict["question"] = question
|
48 |
+
temp_dict["answers"] = {
|
49 |
+
"answer_start": answer_starts,
|
50 |
+
"text": answers,
|
51 |
+
}
|
52 |
+
generated_data["data"].append(temp_dict)
|
53 |
+
return generated_data
|
54 |
+
|
55 |
+
|
56 |
+
# Function to split data
|
57 |
+
def split_data(file_name: str, data_type: str) -> (dict, dict):
|
58 |
+
"""
|
59 |
+
Method to split the data specific to SQuAD2.0
|
60 |
+
|
61 |
+
Parameters:
|
62 |
+
-----------
|
63 |
+
file_name: str,
|
64 |
+
name of the file.
|
65 |
+
data_type: str,
|
66 |
+
type of the data file.
|
67 |
+
|
68 |
+
Returns:
|
69 |
+
--------
|
70 |
+
(dict, dict)
|
71 |
+
split of data
|
72 |
+
"""
|
73 |
+
|
74 |
+
if data_type == "json":
|
75 |
+
with open(file_name, 'r') as json_file:
|
76 |
+
data = json.load(json_file)["data"]
|
77 |
+
json_file.close()
|
78 |
+
|
79 |
+
# split the data into train and test
|
80 |
+
train, test = train_test_split(data, train_size=0.7, random_state = 42)
|
81 |
+
return(train, test)
|
82 |
+
|
83 |
+
|
84 |
+
if __name__ == "__main__":
|
85 |
+
# split the train data
|
86 |
+
train, test = split_data("squad_data/train-v2.0.json", "json")
|
87 |
+
|
88 |
+
# add gem id and save the files
|
89 |
+
train = add_gem_id(train, "train")
|
90 |
+
test = add_gem_id(test, "test")
|
91 |
+
|
92 |
+
# save the train split
|
93 |
+
with open("train.json", "w") as train_file:
|
94 |
+
json.dump(train, train_file, indent = 2)
|
95 |
+
train_file.close()
|
96 |
+
|
97 |
+
# save the test split
|
98 |
+
with open("test.json", "w") as test_file:
|
99 |
+
json.dump(test, test_file, indent = 2)
|
100 |
+
test_file.close()
|
101 |
+
|
102 |
+
# load validation data
|
103 |
+
with open("squad_data/dev-v2.0.json", "r") as dev_file:
|
104 |
+
validation = json.load(dev_file)["data"]
|
105 |
+
dev_file.close()
|
106 |
+
|
107 |
+
# add gem id and save valid.json
|
108 |
+
validation = add_gem_id(validation, "validation")
|
109 |
+
with open("valid.json", "w") as val_file:
|
110 |
+
json.dump(validation, val_file, indent = 2)
|
111 |
+
val_file.close()
|
112 |
+
|
squad_v2.py
CHANGED
@@ -23,15 +23,15 @@ archivePrefix = {arXiv},
|
|
23 |
"""
|
24 |
|
25 |
_DESCRIPTION = """\
|
26 |
-
combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers
|
27 |
to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but
|
28 |
also determine when no answer is supported by the paragraph and abstain from answering.
|
29 |
"""
|
30 |
|
31 |
-
_URL = "https://rajpurkar.github.io/SQuAD-explorer/dataset/"
|
32 |
_URLS = {
|
33 |
-
"train":
|
34 |
-
"
|
|
|
35 |
}
|
36 |
|
37 |
|
@@ -64,6 +64,7 @@ class SquadV2(datasets.GeneratorBasedBuilder):
|
|
64 |
features=datasets.Features(
|
65 |
{
|
66 |
"gem_id": datasets.Value("string"),
|
|
|
67 |
"title": datasets.Value("string"),
|
68 |
"context": datasets.Value("string"),
|
69 |
"question": datasets.Value("string"),
|
@@ -99,35 +100,42 @@ class SquadV2(datasets.GeneratorBasedBuilder):
|
|
99 |
downloaded_files = dl_manager.download_and_extract(urls_to_download)
|
100 |
|
101 |
return [
|
102 |
-
datasets.SplitGenerator(
|
103 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
104 |
]
|
105 |
|
106 |
def _generate_examples(self, filepath, split):
|
107 |
"""Yields examples."""
|
108 |
# TODO(squad_v2): Yields (key, example) tuples from the dataset
|
109 |
with open(filepath, encoding="utf-8") as f:
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
for
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
# Features currently used are "context", "question", and "answers".
|
123 |
-
# Others are extracted here for the ease of future expansions.
|
124 |
-
yield id_, {
|
125 |
-
"title": title,
|
126 |
-
"context": context,
|
127 |
-
"question": question,
|
128 |
-
"gem_id": f"gem-{squad_v2}-{split}-{id_}",
|
129 |
-
"answers": {
|
130 |
-
"answer_start": answer_starts,
|
131 |
-
"text": answers,
|
132 |
-
},
|
133 |
-
}
|
|
|
23 |
"""
|
24 |
|
25 |
_DESCRIPTION = """\
|
26 |
+
SQuAD2.0 combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers
|
27 |
to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but
|
28 |
also determine when no answer is supported by the paragraph and abstain from answering.
|
29 |
"""
|
30 |
|
|
|
31 |
_URLS = {
|
32 |
+
"train": "train.json",
|
33 |
+
"test": "test.json",
|
34 |
+
"valid": "validation.json",
|
35 |
}
|
36 |
|
37 |
|
|
|
64 |
features=datasets.Features(
|
65 |
{
|
66 |
"gem_id": datasets.Value("string"),
|
67 |
+
"id": datasets.Value("string"),
|
68 |
"title": datasets.Value("string"),
|
69 |
"context": datasets.Value("string"),
|
70 |
"question": datasets.Value("string"),
|
|
|
100 |
downloaded_files = dl_manager.download_and_extract(urls_to_download)
|
101 |
|
102 |
return [
|
103 |
+
datasets.SplitGenerator(
|
104 |
+
name=datasets.Split.TRAIN,
|
105 |
+
gen_kwargs={
|
106 |
+
"filepath": downloaded_files["train"],
|
107 |
+
"split": "train",
|
108 |
+
},
|
109 |
+
),
|
110 |
+
datasets.SplitGenerator(
|
111 |
+
name=datasets.Split.VALIDATION,
|
112 |
+
gen_kwargs={
|
113 |
+
"filepath": downloaded_files["validation"],
|
114 |
+
"split": "validation",
|
115 |
+
},
|
116 |
+
),
|
117 |
+
datasets.SplitGenerator(
|
118 |
+
name=datasets.Split.TEST,
|
119 |
+
gen_kwargs={
|
120 |
+
"filepath": downloaded_files["test"],
|
121 |
+
"split": "test",
|
122 |
+
},
|
123 |
+
),
|
124 |
]
|
125 |
|
126 |
def _generate_examples(self, filepath, split):
|
127 |
"""Yields examples."""
|
128 |
# TODO(squad_v2): Yields (key, example) tuples from the dataset
|
129 |
with open(filepath, encoding="utf-8") as f:
|
130 |
+
for id_, row in enumerate(f):
|
131 |
+
data = json.loads(row)
|
132 |
+
# Features currently used are "context", "question", and "answers".
|
133 |
+
# Others are extracted here for the ease of future expansions.
|
134 |
+
yield id_, {
|
135 |
+
"id": data["id"],
|
136 |
+
"gem_id": data["gem_id"],
|
137 |
+
"title": data["title"],
|
138 |
+
"context": data["context"],
|
139 |
+
"question": data["question"],
|
140 |
+
"answers": data["answers"],
|
141 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|