Commit
•
1a4e9f5
1
Parent(s):
e911e49
Delete loading script
Browse files
qanta.py
DELETED
@@ -1,289 +0,0 @@
|
|
1 |
-
"""qanta dataset."""
|
2 |
-
|
3 |
-
|
4 |
-
import json
|
5 |
-
from typing import List, Tuple
|
6 |
-
|
7 |
-
import datasets
|
8 |
-
|
9 |
-
|
10 |
-
_CITATION = """
|
11 |
-
@article{Rodriguez2019QuizbowlTC,
|
12 |
-
title={Quizbowl: The Case for Incremental Question Answering},
|
13 |
-
author={Pedro Rodriguez and Shi Feng and Mohit Iyyer and He He and Jordan L. Boyd-Graber},
|
14 |
-
journal={ArXiv},
|
15 |
-
year={2019},
|
16 |
-
volume={abs/1904.04792}
|
17 |
-
}
|
18 |
-
"""
|
19 |
-
|
20 |
-
_DESCRIPTION = """
|
21 |
-
The Qanta dataset is a question answering dataset based on the academic trivia game Quizbowl.
|
22 |
-
"""
|
23 |
-
|
24 |
-
|
25 |
-
_QANTA_URL = "https://s3-us-west-2.amazonaws.com/pinafore-us-west-2/qanta-jmlr-datasets/qanta.mapped.2018.04.18.json"
|
26 |
-
_TRICK_URL = "https://s3-us-west-2.amazonaws.com/pinafore-us-west-2/trick-tacl-datasets/qanta.tacl-trick.json"
|
27 |
-
_VERSION = datasets.Version("2018.04.18")
|
28 |
-
_FIRST = "first"
|
29 |
-
_FULL = "full"
|
30 |
-
_SENTENCES = "sentences"
|
31 |
-
_RUNS = "runs"
|
32 |
-
# Order matters, the first one is default
|
33 |
-
_MODES = [_FULL, _FIRST, _SENTENCES, _RUNS]
|
34 |
-
_DEFAULT_CHAR_SKIP = 25
|
35 |
-
|
36 |
-
|
37 |
-
class QantaConfig(datasets.BuilderConfig):
|
38 |
-
"""BuilderConfig for Qanta."""
|
39 |
-
|
40 |
-
def __init__(self, mode: str, char_skip: int, **kwargs):
|
41 |
-
super(QantaConfig, self).__init__(version=_VERSION, **kwargs)
|
42 |
-
self.mode = mode
|
43 |
-
self.char_skip = char_skip
|
44 |
-
|
45 |
-
|
46 |
-
def create_char_runs(text: str, char_skip: int) -> List[Tuple[str, int]]:
|
47 |
-
"""
|
48 |
-
Returns runs of the question based on skipping char_skip characters at a time. Also returns the indices used
|
49 |
-
q: name this first united states president.
|
50 |
-
runs with char_skip=10:
|
51 |
-
['name this ',
|
52 |
-
'name this first unit',
|
53 |
-
'name this first united state p',
|
54 |
-
'name this first united state president.']
|
55 |
-
:param char_skip: Number of characters to skip each time
|
56 |
-
"""
|
57 |
-
char_indices = list(range(char_skip, len(text) + char_skip, char_skip))
|
58 |
-
return [(text[:idx], idx) for idx in char_indices]
|
59 |
-
|
60 |
-
|
61 |
-
def with_default(key, lookup, default):
|
62 |
-
if key in lookup:
|
63 |
-
value = lookup[key]
|
64 |
-
if value is None:
|
65 |
-
return default
|
66 |
-
else:
|
67 |
-
return value
|
68 |
-
else:
|
69 |
-
return default
|
70 |
-
|
71 |
-
|
72 |
-
def question_to_examples(question, mode: str, char_skip: int):
|
73 |
-
features = {
|
74 |
-
"qanta_id": question["qanta_id"],
|
75 |
-
"proto_id": with_default("proto_id", question, ""),
|
76 |
-
"qdb_id": with_default("qdb_id", question, -1),
|
77 |
-
# We refer to the actual answer as page, but this
|
78 |
-
# may be misleading externally, so rename here to
|
79 |
-
# be clearer
|
80 |
-
"page": question["page"],
|
81 |
-
"answer": question["page"],
|
82 |
-
"raw_answer": question["answer"],
|
83 |
-
"dataset": with_default("dataset", question, ""),
|
84 |
-
"full_question": question["text"],
|
85 |
-
"first_sentence": question["first_sentence"],
|
86 |
-
"tokenizations": question["tokenizations"],
|
87 |
-
"fold": question["fold"],
|
88 |
-
"gameplay": question["gameplay"],
|
89 |
-
"category": with_default("category", question, ""),
|
90 |
-
"subcategory": with_default("subcategory", question, ""),
|
91 |
-
"tournament": question["tournament"],
|
92 |
-
"difficulty": with_default("difficulty", question, ""),
|
93 |
-
"year": question["year"],
|
94 |
-
"char_idx": -1,
|
95 |
-
"sentence_idx": -1,
|
96 |
-
}
|
97 |
-
if mode == _FULL:
|
98 |
-
yield {
|
99 |
-
"text": question["text"],
|
100 |
-
"id": str(question["qanta_id"]) + "-full",
|
101 |
-
**features,
|
102 |
-
}
|
103 |
-
elif mode == _FIRST:
|
104 |
-
yield {
|
105 |
-
"text": question["first_sentence"],
|
106 |
-
"id": str(question["qanta_id"]) + "-first",
|
107 |
-
**features,
|
108 |
-
}
|
109 |
-
elif mode == _RUNS:
|
110 |
-
text = question["text"]
|
111 |
-
for text_run, char_idx in create_char_runs(text, char_skip):
|
112 |
-
yield {
|
113 |
-
"text": text_run,
|
114 |
-
"char_idx": char_idx,
|
115 |
-
"id": str(question["qanta_id"]) + "-char-" + str(char_idx),
|
116 |
-
**features,
|
117 |
-
}
|
118 |
-
elif mode == _SENTENCES:
|
119 |
-
for sentence_idx, (start, end) in enumerate(question["tokenizations"]):
|
120 |
-
sentence = question["text"][start:end]
|
121 |
-
yield {
|
122 |
-
"text": sentence,
|
123 |
-
"sentence_idx": sentence_idx,
|
124 |
-
"id": str(question["qanta_id"]) + "-sentence-" + str(sentence_idx),
|
125 |
-
**features,
|
126 |
-
}
|
127 |
-
else:
|
128 |
-
raise ValueError(f"Invalid mode: {mode}")
|
129 |
-
|
130 |
-
|
131 |
-
_FEATURES = {
|
132 |
-
# Generated ID based modes set, unique
|
133 |
-
"id": datasets.Value("string"),
|
134 |
-
# Dataset defined IDs
|
135 |
-
"qanta_id": datasets.Value("int32"),
|
136 |
-
"proto_id": datasets.Value("string"),
|
137 |
-
"qdb_id": datasets.Value("int32"),
|
138 |
-
"dataset": datasets.Value("string"),
|
139 |
-
# Inputs
|
140 |
-
"text": datasets.Value("string"),
|
141 |
-
"full_question": datasets.Value("string"),
|
142 |
-
"first_sentence": datasets.Value("string"),
|
143 |
-
"char_idx": datasets.Value("int32"),
|
144 |
-
"sentence_idx": datasets.Value("int32"),
|
145 |
-
# Character indices of sentences: List[Tuple[int, int]]
|
146 |
-
"tokenizations": datasets.features.Sequence(datasets.features.Sequence(datasets.Value("int32"), length=2)),
|
147 |
-
# Labels: Number is equal to number of unique pages across all folds
|
148 |
-
"answer": datasets.Value("string"),
|
149 |
-
"page": datasets.Value("string"),
|
150 |
-
"raw_answer": datasets.Value("string"),
|
151 |
-
# Meta Information
|
152 |
-
"fold": datasets.Value("string"),
|
153 |
-
"gameplay": datasets.Value("bool"),
|
154 |
-
"category": datasets.Value("string"),
|
155 |
-
"subcategory": datasets.Value("string"),
|
156 |
-
"tournament": datasets.Value("string"),
|
157 |
-
"difficulty": datasets.Value("string"),
|
158 |
-
"year": datasets.Value("int32"),
|
159 |
-
}
|
160 |
-
|
161 |
-
|
162 |
-
class Qanta(datasets.GeneratorBasedBuilder):
|
163 |
-
"""The Qanta dataset is a question answering dataset based on the academic trivia game Quizbowl."""
|
164 |
-
|
165 |
-
VERSION = _VERSION
|
166 |
-
BUILDER_CONFIGS = [
|
167 |
-
QantaConfig(
|
168 |
-
name=f"mode={mode},char_skip={_DEFAULT_CHAR_SKIP}",
|
169 |
-
description=f"Question format: {mode}, char_skip: {_DEFAULT_CHAR_SKIP}",
|
170 |
-
mode=mode,
|
171 |
-
char_skip=_DEFAULT_CHAR_SKIP,
|
172 |
-
)
|
173 |
-
for mode in _MODES
|
174 |
-
]
|
175 |
-
|
176 |
-
def _info(self):
|
177 |
-
return datasets.DatasetInfo(
|
178 |
-
# This is the description that will appear on the datasets page.
|
179 |
-
description=_DESCRIPTION,
|
180 |
-
# datasets.features.FeatureConnectors
|
181 |
-
features=datasets.Features(_FEATURES),
|
182 |
-
# Number of classes is a function of the dataset, ClassLabel doesn't support dynamic
|
183 |
-
# definition, so have to defer conversion to classes to later, so can't define
|
184 |
-
# supervied keys
|
185 |
-
supervised_keys=None,
|
186 |
-
# Homepage of the dataset for documentation
|
187 |
-
homepage="http://www.qanta.org/",
|
188 |
-
citation=_CITATION,
|
189 |
-
)
|
190 |
-
|
191 |
-
def _split_generators(self, dl_manager):
|
192 |
-
"""Returns SplitGenerators."""
|
193 |
-
qanta_path = dl_manager.download_and_extract(_QANTA_URL)
|
194 |
-
trick_path = dl_manager.download_and_extract(_TRICK_URL)
|
195 |
-
return [
|
196 |
-
datasets.SplitGenerator(
|
197 |
-
name=datasets.Split("guesstrain"),
|
198 |
-
gen_kwargs={
|
199 |
-
"qanta_filepath": qanta_path,
|
200 |
-
"trick_filepath": trick_path,
|
201 |
-
"fold": "guesstrain",
|
202 |
-
"mode": self.config.mode,
|
203 |
-
"char_skip": self.config.char_skip,
|
204 |
-
},
|
205 |
-
),
|
206 |
-
datasets.SplitGenerator(
|
207 |
-
name=datasets.Split("buzztrain"),
|
208 |
-
gen_kwargs={
|
209 |
-
"qanta_filepath": qanta_path,
|
210 |
-
"trick_filepath": trick_path,
|
211 |
-
"fold": "buzztrain",
|
212 |
-
"mode": self.config.mode,
|
213 |
-
"char_skip": self.config.char_skip,
|
214 |
-
},
|
215 |
-
),
|
216 |
-
datasets.SplitGenerator(
|
217 |
-
name=datasets.Split("guessdev"),
|
218 |
-
gen_kwargs={
|
219 |
-
"qanta_filepath": qanta_path,
|
220 |
-
"trick_filepath": trick_path,
|
221 |
-
"fold": "guessdev",
|
222 |
-
"mode": self.config.mode,
|
223 |
-
"char_skip": self.config.char_skip,
|
224 |
-
},
|
225 |
-
),
|
226 |
-
datasets.SplitGenerator(
|
227 |
-
name=datasets.Split("buzzdev"),
|
228 |
-
gen_kwargs={
|
229 |
-
"qanta_filepath": qanta_path,
|
230 |
-
"trick_filepath": trick_path,
|
231 |
-
"fold": "buzzdev",
|
232 |
-
"mode": self.config.mode,
|
233 |
-
"char_skip": self.config.char_skip,
|
234 |
-
},
|
235 |
-
),
|
236 |
-
datasets.SplitGenerator(
|
237 |
-
name=datasets.Split("guesstest"),
|
238 |
-
gen_kwargs={
|
239 |
-
"qanta_filepath": qanta_path,
|
240 |
-
"trick_filepath": trick_path,
|
241 |
-
"fold": "guesstest",
|
242 |
-
"mode": self.config.mode,
|
243 |
-
"char_skip": self.config.char_skip,
|
244 |
-
},
|
245 |
-
),
|
246 |
-
datasets.SplitGenerator(
|
247 |
-
name=datasets.Split("buzztest"),
|
248 |
-
gen_kwargs={
|
249 |
-
"qanta_filepath": qanta_path,
|
250 |
-
"trick_filepath": trick_path,
|
251 |
-
"fold": "buzztest",
|
252 |
-
"mode": self.config.mode,
|
253 |
-
"char_skip": self.config.char_skip,
|
254 |
-
},
|
255 |
-
),
|
256 |
-
datasets.SplitGenerator(
|
257 |
-
name=datasets.Split("adversarial"),
|
258 |
-
gen_kwargs={
|
259 |
-
"qanta_filepath": qanta_path,
|
260 |
-
"trick_filepath": trick_path,
|
261 |
-
"fold": "adversarial",
|
262 |
-
"mode": self.config.mode,
|
263 |
-
"char_skip": self.config.char_skip,
|
264 |
-
},
|
265 |
-
),
|
266 |
-
]
|
267 |
-
|
268 |
-
def _generate_examples(
|
269 |
-
self,
|
270 |
-
qanta_filepath: str,
|
271 |
-
trick_filepath: str,
|
272 |
-
fold: str,
|
273 |
-
mode: str,
|
274 |
-
char_skip: int,
|
275 |
-
):
|
276 |
-
"""Yields examples."""
|
277 |
-
if mode not in _MODES:
|
278 |
-
raise ValueError(f"Invalid mode: {mode}")
|
279 |
-
|
280 |
-
if fold == "adversarial":
|
281 |
-
path = trick_filepath
|
282 |
-
else:
|
283 |
-
path = qanta_filepath
|
284 |
-
with open(path, encoding="utf-8") as f:
|
285 |
-
questions = json.load(f)["questions"]
|
286 |
-
for q in questions:
|
287 |
-
if q["page"] is not None and q["fold"] == fold:
|
288 |
-
for example in question_to_examples(q, mode, char_skip):
|
289 |
-
yield example["id"], example
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|