julianrisch
commited on
Commit
•
84cc8cd
1
Parent(s):
dd60ab6
Update germanquad.py
Browse files- germanquad.py +11 -30
germanquad.py
CHANGED
@@ -10,7 +10,7 @@ import datasets
|
|
10 |
logger = datasets.logging.get_logger(__name__)
|
11 |
|
12 |
|
13 |
-
_CITATION = """
|
14 |
@misc{möller2021germanquad,
|
15 |
title={GermanQuAD and GermanDPR: Improving Non-English Question Answering and Passage Retrieval},
|
16 |
author={Timo Möller and Julian Risch and Malte Pietsch},
|
@@ -21,7 +21,7 @@ _CITATION = """\\\\\\\\
|
|
21 |
}
|
22 |
"""
|
23 |
|
24 |
-
_DESCRIPTION = """
|
25 |
In order to raise the bar for non-English QA, we are releasing a high-quality, human-labeled German QA dataset consisting of 13 722 questions, incl. a three-way annotated test set.
|
26 |
The creation of GermanQuAD is inspired by insights from existing datasets as well as our labeling experience from several industry projects. We combine the strengths of SQuAD, such as high out-of-domain performance, with self-sufficient questions that contain all relevant information for open-domain QA as in the NaturalQuestions dataset. Our training and test datasets do not overlap like other popular datasets and include complex questions that cannot be answered with a single entity or only a few words.
|
27 |
"""
|
@@ -57,29 +57,15 @@ class GermanDPR(datasets.GeneratorBasedBuilder):
|
|
57 |
description=_DESCRIPTION,
|
58 |
features=datasets.Features(
|
59 |
{
|
|
|
|
|
60 |
"question": datasets.Value("string"),
|
61 |
-
"answers": datasets.features.Sequence(
|
62 |
-
"positive_ctxs": datasets.features.Sequence(
|
63 |
{
|
64 |
-
"
|
65 |
-
"
|
66 |
-
"passage_id": datasets.Value("string"),
|
67 |
}
|
68 |
-
)
|
69 |
-
"negative_ctxs": datasets.features.Sequence(
|
70 |
-
{
|
71 |
-
"title": datasets.Value("string"),
|
72 |
-
"text": datasets.Value("string"),
|
73 |
-
"passage_id": datasets.Value("string"),
|
74 |
-
}
|
75 |
-
),
|
76 |
-
"hard_negative_ctxs": datasets.features.Sequence(
|
77 |
-
{
|
78 |
-
"title": datasets.Value("string"),
|
79 |
-
"text": datasets.Value("string"),
|
80 |
-
"passage_id": datasets.Value("string"),
|
81 |
-
}
|
82 |
-
),
|
83 |
}
|
84 |
),
|
85 |
# No default supervised_keys (as we have to pass both question
|
@@ -109,9 +95,7 @@ class GermanDPR(datasets.GeneratorBasedBuilder):
|
|
109 |
for qa in paragraph["qas"]:
|
110 |
question = qa["question"]
|
111 |
id_ = qa["id"]
|
112 |
-
|
113 |
-
answer_starts = [answer["answer_start"] for answer in qa["answers"]]
|
114 |
-
answers = [answer["text"] for answer in qa["answers"]]
|
115 |
|
116 |
# Features currently used are "context", "question", and "answers".
|
117 |
# Others are extracted here for the ease of future expansions.
|
@@ -119,9 +103,6 @@ class GermanDPR(datasets.GeneratorBasedBuilder):
|
|
119 |
"context": context,
|
120 |
"question": question,
|
121 |
"id": id_,
|
122 |
-
"answers":
|
123 |
-
"answer_start": answer_starts,
|
124 |
-
"text": answers,
|
125 |
-
},
|
126 |
}
|
127 |
-
|
|
|
10 |
logger = datasets.logging.get_logger(__name__)
|
11 |
|
12 |
|
13 |
+
_CITATION = """
|
14 |
@misc{möller2021germanquad,
|
15 |
title={GermanQuAD and GermanDPR: Improving Non-English Question Answering and Passage Retrieval},
|
16 |
author={Timo Möller and Julian Risch and Malte Pietsch},
|
|
|
21 |
}
|
22 |
"""
|
23 |
|
24 |
+
_DESCRIPTION = """
|
25 |
In order to raise the bar for non-English QA, we are releasing a high-quality, human-labeled German QA dataset consisting of 13 722 questions, incl. a three-way annotated test set.
|
26 |
The creation of GermanQuAD is inspired by insights from existing datasets as well as our labeling experience from several industry projects. We combine the strengths of SQuAD, such as high out-of-domain performance, with self-sufficient questions that contain all relevant information for open-domain QA as in the NaturalQuestions dataset. Our training and test datasets do not overlap like other popular datasets and include complex questions that cannot be answered with a single entity or only a few words.
|
27 |
"""
|
|
|
57 |
description=_DESCRIPTION,
|
58 |
features=datasets.Features(
|
59 |
{
|
60 |
+
"id": datasets.Value("int32"),
|
61 |
+
"context": datasets.Value("string"),
|
62 |
"question": datasets.Value("string"),
|
63 |
+
"answers": datasets.features.Sequence(
|
|
|
64 |
{
|
65 |
+
"text": datasets.Value("string"),
|
66 |
+
"answer_start": datasets.Value("int32"),
|
|
|
67 |
}
|
68 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
}
|
70 |
),
|
71 |
# No default supervised_keys (as we have to pass both question
|
|
|
95 |
for qa in paragraph["qas"]:
|
96 |
question = qa["question"]
|
97 |
id_ = qa["id"]
|
98 |
+
answers = [{"answer_start": answer["answer_start"], "text": answer["text"]} for answer in qa["answers"]]
|
|
|
|
|
99 |
|
100 |
# Features currently used are "context", "question", and "answers".
|
101 |
# Others are extracted here for the ease of future expansions.
|
|
|
103 |
"context": context,
|
104 |
"question": question,
|
105 |
"id": id_,
|
106 |
+
"answers": answers,
|
|
|
|
|
|
|
107 |
}
|
108 |
+
|