Datasets:
Commit
·
30f6ac6
1
Parent(s):
bf50f5a
docs: Add readme
Browse files
README.md
CHANGED
@@ -1,3 +1,146 @@
|
|
1 |
---
|
2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
pretty_name: ScandiQA
|
3 |
+
language:
|
4 |
+
- da
|
5 |
+
- sv
|
6 |
+
- no
|
7 |
+
license:
|
8 |
+
- cc-by-sa-4.0
|
9 |
+
multilinguality:
|
10 |
+
- multilingual
|
11 |
+
size_categories:
|
12 |
+
- 1K<n<10K
|
13 |
+
source_datasets:
|
14 |
+
- mkqa|natural_questions
|
15 |
+
task_categories:
|
16 |
+
- question-answering
|
17 |
+
task_ids:
|
18 |
+
- extractive-qa
|
19 |
---
|
20 |
+
|
21 |
+
# Dataset Card for ScandiQA
|
22 |
+
|
23 |
+
## Dataset Description
|
24 |
+
|
25 |
+
- **Repository:** <https://github.com/alexandrainst/scandi-qa>
|
26 |
+
- **Point of Contact:** [Dan Saattrup Nielsen](mailto:[email protected])
|
27 |
+
- **Size of downloaded dataset files:** 69 MB
|
28 |
+
- **Size of the generated dataset:** 67 MB
|
29 |
+
- **Total amount of disk used:** 136 MB
|
30 |
+
|
31 |
+
### Dataset Summary
|
32 |
+
|
33 |
+
ScandiQA is a dataset of questions and answers in the Danish, Norwegian, and Swedish
|
34 |
+
languages. All samples come from the Natural Questions (NQ) dataset, which is a large
|
35 |
+
question answering dataset from Google searches. The Scandinavian questions and answers
|
36 |
+
come from the MKQA dataset, where 10,000 NQ samples were manually translated into,
|
37 |
+
among others, Danish, Norwegian, and Swedish. However, this did not include a
|
38 |
+
translated context, hindering the training of extractive question answering models.
|
39 |
+
|
40 |
+
We merged the NQ dataset with the MKQA dataset, and extracted contexts as either "long
|
41 |
+
answers" from the NQ dataset, being the paragraph in which the answer was found, or
|
42 |
+
otherwise we extract the context by locating the paragraphs which have the largest
|
43 |
+
cosine similarity to the question, and which contains the desired answer.
|
44 |
+
|
45 |
+
Further, many answers in the MKQA dataset were "language normalised": for instance, all
|
46 |
+
date answers were converted to the format "YYYY-MM-DD", meaning that in most cases
|
47 |
+
these answers are not appearing in any paragraphs. We solve this by extending the MKQA
|
48 |
+
answers with plausible "answer candidates", being slight perturbations or translations
|
49 |
+
of the answer.
|
50 |
+
|
51 |
+
With the contexts extracted, we translated these to Danish, Swedish and Norwegian using
|
52 |
+
the DeepL translation service for Danish and Swedish, and the Google Translation
|
53 |
+
service for Norwegian. After translation we ensured that the Scandinavian answers do
|
54 |
+
indeed occur in the translated contexts.
|
55 |
+
|
56 |
+
As we are filtering the MKQA samples at both the "merging stage" and the "translation
|
57 |
+
stage", we are not able to fully convert the 10,000 samples to the Scandinavian
|
58 |
+
languages, and instead get roughly 8,000 samples per language. These have further been
|
59 |
+
split into a training, validation and test split, with the former two containing
|
60 |
+
roughly 750 samples. The splits have been created in such a way that the proportion of
|
61 |
+
samples without an answer is roughly the same in each split.
|
62 |
+
|
63 |
+
|
64 |
+
### Supported Tasks and Leaderboards
|
65 |
+
|
66 |
+
Training machine learning models for extractive question answering is the intended task
|
67 |
+
for this dataset. No leaderboard is active at this point.
|
68 |
+
|
69 |
+
|
70 |
+
### Languages
|
71 |
+
|
72 |
+
The dataset is available in Danish (`da`), Swedish (`sv`) and Norwegian (`no`).
|
73 |
+
|
74 |
+
|
75 |
+
## Dataset Structure
|
76 |
+
|
77 |
+
### Data Instances
|
78 |
+
|
79 |
+
- **Size of downloaded dataset files:** 69 MB
|
80 |
+
- **Size of the generated dataset:** 67 MB
|
81 |
+
- **Total amount of disk used:** 136 MB
|
82 |
+
|
83 |
+
An example from the `train` split of the `da` subset looks as follows.
|
84 |
+
```
|
85 |
+
{
|
86 |
+
'example_id': 123,
|
87 |
+
'question': 'Er dette en test?',
|
88 |
+
'answer': 'Dette er en test',
|
89 |
+
'answer_start': 0,
|
90 |
+
'context': 'Dette er en testkontekst.',
|
91 |
+
'answer_en': 'This is a test',
|
92 |
+
'answer_start_en': 0,
|
93 |
+
'context_en': "This is a test",
|
94 |
+
'title_en': 'Train test'
|
95 |
+
}
|
96 |
+
```
|
97 |
+
|
98 |
+
### Data Fields
|
99 |
+
|
100 |
+
The data fields are the same among all splits.
|
101 |
+
|
102 |
+
- `example_id`: an `int64` feature.
|
103 |
+
- `question`: a `string` feature.
|
104 |
+
- `answer`: a `string` feature.
|
105 |
+
- `answer_start`: an `int64` feature.
|
106 |
+
- `context`: a `string` feature.
|
107 |
+
- `answer_en`: a `string` feature.
|
108 |
+
- `answer_start_en`: an `int64` feature.
|
109 |
+
- `context_en`: a `string` feature.
|
110 |
+
- `title_en`: a `string` feature.
|
111 |
+
|
112 |
+
### Data Splits
|
113 |
+
|
114 |
+
| name | train | validation | test |
|
115 |
+
|----------|------:|-----------:|-----:|
|
116 |
+
| da | 6311 | 749 | 750 |
|
117 |
+
| sv | 6299 | 750 | 749 |
|
118 |
+
| no | 6314 | 749 | 750 |
|
119 |
+
|
120 |
+
|
121 |
+
## Dataset Creation
|
122 |
+
|
123 |
+
### Curation Rationale
|
124 |
+
|
125 |
+
The Scandinavian languages does not have any gold standard question answering dataset.
|
126 |
+
This is not quite gold standard, but the fact both the questions and answers are all
|
127 |
+
manually translated, it is a solid silver standard dataset.
|
128 |
+
|
129 |
+
### Source Data
|
130 |
+
|
131 |
+
The original data was collected from the [MKQA](https://github.com/apple/ml-mkqa/) and
|
132 |
+
[Natural Questions](https://ai.google.com/research/NaturalQuestions) datasets from
|
133 |
+
Apple and Google, respectively.
|
134 |
+
|
135 |
+
|
136 |
+
## Additional Information
|
137 |
+
|
138 |
+
### Dataset Curators
|
139 |
+
|
140 |
+
[Dan Saattrup Nielsen](https://saattrupdan.github.io/) from the [The Alexandra
|
141 |
+
Institute](https://alexandra.dk/) curated this dataset.
|
142 |
+
|
143 |
+
### Licensing Information
|
144 |
+
|
145 |
+
The dataset is licensed under the [CC BY-SA 4.0
|
146 |
+
license](https://creativecommons.org/licenses/by-sa/4.0/).
|