Datasets:

ArXiv:
License:
holylovenia commited on
Commit
e9785dd
1 Parent(s): 6e43292

Upload tatoeba.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. tatoeba.py +192 -0
tatoeba.py ADDED
@@ -0,0 +1,192 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from pathlib import Path
2
+ from typing import Dict, List, Tuple
3
+
4
+ import datasets
5
+ from datasets.download.download_manager import DownloadManager
6
+
7
+ from seacrowd.utils import schemas
8
+ from seacrowd.utils.configs import SEACrowdConfig
9
+ from seacrowd.utils.constants import Licenses, Tasks
10
+
11
+ _CITATION = """\
12
+ @article{tatoeba,
13
+ title = {Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond},
14
+ author = {Mikel, Artetxe and Holger, Schwenk,},
15
+ journal = {arXiv:1812.10464v2},
16
+ year = {2018}
17
+ }
18
+ """
19
+
20
+ _LOCAL = False
21
+ _LANGUAGES = ["ind", "vie", "tgl", "jav", "tha", "eng"]
22
+ _DATASETNAME = "tatoeba"
23
+ _DESCRIPTION = """\
24
+ This dataset is a subset of the Tatoeba corpus containing language pairs for Indonesian, Vietnamese, Tagalog, Javanese, and Thai.
25
+ The original dataset description can be found below:
26
+
27
+ This data is extracted from the Tatoeba corpus, dated Saturday 2018/11/17.
28
+ For each languages, we have selected 1000 English sentences and their translations, if available. Please check
29
+ this paper for a description of the languages, their families and scripts as well as baseline results.
30
+ Please note that the English sentences are not identical for all language pairs. This means that the results are
31
+ not directly comparable across languages. In particular, the sentences tend to have less variety for several
32
+ low-resource languages, e.g. "Tom needed water", "Tom needs water", "Tom is getting water", ...
33
+ """
34
+
35
+ _HOMEPAGE = "https://github.com/facebookresearch/LASER/blob/main/data/tatoeba/v1/README.md"
36
+ _LICENSE = Licenses.APACHE_2_0.value
37
+ _URL = "https://github.com/facebookresearch/LASER/raw/main/data/tatoeba/v1/"
38
+
39
+ _SUPPORTED_TASKS = [Tasks.MACHINE_TRANSLATION]
40
+ _SOURCE_VERSION = "1.0.0"
41
+ _SEACROWD_VERSION = "2024.06.20"
42
+
43
+
44
+ class TatoebaDataset(datasets.GeneratorBasedBuilder):
45
+ """Tatoeba subset for Indonesian, Vietnamese, Tagalog, Javanese, and Thai."""
46
+
47
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
48
+ SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
49
+
50
+ SEACROWD_SCHEMA_NAME = "t2t"
51
+
52
+ # Add configurations for loading a dataset per language.
53
+ dataset_names = sorted([f"tatoeba_{lang}_eng" for lang in _LANGUAGES[:-1]]) + sorted([f"tatoeba_eng_{lang}" for lang in _LANGUAGES[:-1]])
54
+ BUILDER_CONFIGS = []
55
+ for name in dataset_names:
56
+ source_config = SEACrowdConfig(
57
+ name=f"{name}_source",
58
+ version=SOURCE_VERSION,
59
+ description=f"{_DATASETNAME} source schema",
60
+ schema="source",
61
+ subset_id=name,
62
+ )
63
+ BUILDER_CONFIGS.append(source_config)
64
+ seacrowd_config = SEACrowdConfig(
65
+ name=f"{name}_seacrowd_{SEACROWD_SCHEMA_NAME}",
66
+ version=SEACROWD_VERSION,
67
+ description=f"{_DATASETNAME} SEACrowd schema",
68
+ schema=f"seacrowd_{SEACROWD_SCHEMA_NAME}",
69
+ subset_id=name,
70
+ )
71
+ BUILDER_CONFIGS.append(seacrowd_config)
72
+
73
+ # Add configuration that allows loading all datasets at once.
74
+ BUILDER_CONFIGS.extend(
75
+ [
76
+ # tatoeba_source
77
+ SEACrowdConfig(
78
+ name=f"{_DATASETNAME}_source",
79
+ version=SOURCE_VERSION,
80
+ description=f"{_DATASETNAME} source schema (all)",
81
+ schema="source",
82
+ subset_id=_DATASETNAME,
83
+ ),
84
+ # tatoeba_seacrowd_t2t
85
+ SEACrowdConfig(
86
+ name=f"{_DATASETNAME}_seacrowd_{SEACROWD_SCHEMA_NAME}",
87
+ version=SEACROWD_VERSION,
88
+ description=f"{_DATASETNAME} SEACrowd schema (all)",
89
+ schema=f"seacrowd_{SEACROWD_SCHEMA_NAME}",
90
+ subset_id=_DATASETNAME,
91
+ ),
92
+ ]
93
+ )
94
+
95
+ # Choose first language as default
96
+ DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_source"
97
+
98
+ def _info(self) -> datasets.DatasetInfo:
99
+ if self.config.schema == "source":
100
+ features = datasets.Features(
101
+ {
102
+ "source_sentence": datasets.Value("string"),
103
+ "target_sentence": datasets.Value("string"),
104
+ "source_lang": datasets.Value("string"),
105
+ "target_lang": datasets.Value("string"),
106
+ }
107
+ )
108
+ elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
109
+ features = schemas.text2text_features
110
+ return datasets.DatasetInfo(
111
+ description=_DESCRIPTION,
112
+ features=features,
113
+ homepage=_HOMEPAGE,
114
+ license=_LICENSE,
115
+ citation=_CITATION,
116
+ )
117
+
118
+ def _split_generators(self, dl_manager: DownloadManager) -> List[datasets.SplitGenerator]:
119
+ """Return SplitGenerators."""
120
+ language_pairs = []
121
+ tatoeba_source_data = []
122
+ tatoeba_eng_data = []
123
+
124
+ lang_1 = self.config.name.split("_")[1]
125
+ lang_2 = self.config.name.split("_")[2]
126
+ if lang_1 == "eng":
127
+ lang = lang_2
128
+ else:
129
+ lang = lang_1
130
+
131
+ if lang in _LANGUAGES:
132
+ # Load data per language
133
+ tatoeba_source_data.append(dl_manager.download_and_extract(_URL + f"tatoeba.{lang}-eng.{lang_1}"))
134
+ tatoeba_eng_data.append(dl_manager.download_and_extract(_URL + f"tatoeba.{lang}-eng.{lang_2}"))
135
+ language_pairs.append((lang_1, lang_2))
136
+ else:
137
+ # Load examples from all languages at once
138
+ # We just want to run this part when tatoeba_source / tatoeba_seacrowd_t2t was chosen.
139
+ for lang in _LANGUAGES[:-1]:
140
+ tatoeba_source_data.append(dl_manager.download_and_extract(_URL + f"tatoeba.{lang}-eng.{lang}"))
141
+ tatoeba_eng_data.append(dl_manager.download_and_extract(_URL + f"tatoeba.{lang}-eng.eng"))
142
+ language_pairs.append((lang, "eng"))
143
+ return [
144
+ datasets.SplitGenerator(
145
+ name=datasets.Split.VALIDATION,
146
+ gen_kwargs={
147
+ "filepaths": (tatoeba_source_data, tatoeba_eng_data),
148
+ "split": "dev",
149
+ "language_pairs": language_pairs,
150
+ },
151
+ )
152
+ ]
153
+
154
+ def _generate_examples(self, filepaths: Tuple[List[Path], List[Path]], split: str, language_pairs: List[str]) -> Tuple[int, Dict]:
155
+ """Yield examples as (key, example) tuples"""
156
+ source_files, target_files = filepaths
157
+ source_sents = []
158
+ target_sents = []
159
+ source_langs = []
160
+ target_langs = []
161
+
162
+ for source_file, target_file, (lang_1, lang_2) in zip(source_files, target_files, language_pairs):
163
+ with open(source_file, encoding="utf-8") as f1:
164
+ for row in f1:
165
+ source_sents.append(row.strip())
166
+ source_langs.append(lang_1)
167
+ with open(target_file, encoding="utf-8") as f2:
168
+ for row in f2:
169
+ target_sents.append(row.strip())
170
+ target_langs.append(lang_2)
171
+
172
+ for idx, (source, target, lang_src, lang_tgt) in enumerate(zip(source_sents, target_sents, source_langs, target_langs)):
173
+ if self.config.schema == "source":
174
+ example = {
175
+ "source_sentence": source,
176
+ "target_sentence": target,
177
+ # The source_lang in the HuggingFace source seems incorrect
178
+ # I am overriding it with the actual language code.
179
+ "source_lang": lang_src,
180
+ "target_lang": lang_tgt,
181
+ }
182
+ elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
183
+ example = {
184
+ "id": str(idx),
185
+ "text_1": source,
186
+ "text_2": target,
187
+ # The source_lang in the HuggingFace source seems incorrect
188
+ # I am overriding it with the actual language code.
189
+ "text_1_name": lang_src,
190
+ "text_2_name": lang_tgt,
191
+ }
192
+ yield idx, example