README.md
8.7 KB · 312 lines · markdown Raw
1 ---
2 language:
3 - multilingual
4 - af
5 - sq
6 - ar
7 - an
8 - hy
9 - ast
10 - az
11 - ba
12 - eu
13 - bar
14 - be
15 - bn
16 - inc
17 - bs
18 - br
19 - bg
20 - my
21 - ca
22 - ceb
23 - ce
24 - zh
25 - cv
26 - hr
27 - cs
28 - da
29 - nl
30 - en
31 - et
32 - fi
33 - fr
34 - gl
35 - ka
36 - de
37 - el
38 - gu
39 - ht
40 - he
41 - hi
42 - hu
43 - is
44 - io
45 - id
46 - ga
47 - it
48 - ja
49 - jv
50 - kn
51 - kk
52 - ky
53 - ko
54 - la
55 - lv
56 - lt
57 - roa
58 - nds
59 - lm
60 - mk
61 - mg
62 - ms
63 - ml
64 - mr
65 - min
66 - ne
67 - new
68 - nb
69 - nn
70 - oc
71 - fa
72 - pms
73 - pl
74 - pt
75 - pa
76 - ro
77 - ru
78 - sco
79 - sr
80 - hr
81 - scn
82 - sk
83 - sl
84 - aze
85 - es
86 - su
87 - sw
88 - sv
89 - tl
90 - tg
91 - ta
92 - tt
93 - te
94 - tr
95 - uk
96 - ud
97 - uz
98 - vi
99 - vo
100 - war
101 - cy
102 - fry
103 - pnb
104 - yo
105 license: apache-2.0
106 datasets:
107 - wikipedia
108 ---
109
110 # BERT multilingual base model (uncased)
111
112 Pretrained model on the top 102 languages with the largest Wikipedia using a masked language modeling (MLM) objective.
113 It was introduced in [this paper](https://arxiv.org/abs/1810.04805) and first released in
114 [this repository](https://github.com/google-research/bert). This model is uncased: it does not make a difference
115 between english and English.
116
117 Disclaimer: The team releasing BERT did not write a model card for this model so this model card has been written by
118 the Hugging Face team.
119
120 ## Model description
121
122 BERT is a transformers model pretrained on a large corpus of multilingual data in a self-supervised fashion. This means
123 it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of
124 publicly available data) with an automatic process to generate inputs and labels from those texts. More precisely, it
125 was pretrained with two objectives:
126
127 - Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run
128 the entire masked sentence through the model and has to predict the masked words. This is different from traditional
129 recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like
130 GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the
131 sentence.
132 - Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes
133 they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to
134 predict if the two sentences were following each other or not.
135
136 This way, the model learns an inner representation of the languages in the training set that can then be used to
137 extract features useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a
138 standard classifier using the features produced by the BERT model as inputs.
139
140 ## Intended uses & limitations
141
142 You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to
143 be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=bert) to look for
144 fine-tuned versions on a task that interests you.
145
146 Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)
147 to make decisions, such as sequence classification, token classification or question answering. For tasks such as text
148 generation you should look at model like GPT2.
149
150 ### How to use
151
152 You can use this model directly with a pipeline for masked language modeling:
153
154 ```python
155 >>> from transformers import pipeline
156 >>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-uncased')
157 >>> unmasker("Hello I'm a [MASK] model.")
158
159 [{'sequence': "[CLS] hello i'm a top model. [SEP]",
160 'score': 0.1507750153541565,
161 'token': 11397,
162 'token_str': 'top'},
163 {'sequence': "[CLS] hello i'm a fashion model. [SEP]",
164 'score': 0.13075384497642517,
165 'token': 23589,
166 'token_str': 'fashion'},
167 {'sequence': "[CLS] hello i'm a good model. [SEP]",
168 'score': 0.036272723227739334,
169 'token': 12050,
170 'token_str': 'good'},
171 {'sequence': "[CLS] hello i'm a new model. [SEP]",
172 'score': 0.035954564809799194,
173 'token': 10246,
174 'token_str': 'new'},
175 {'sequence': "[CLS] hello i'm a great model. [SEP]",
176 'score': 0.028643041849136353,
177 'token': 11838,
178 'token_str': 'great'}]
179 ```
180
181 Here is how to use this model to get the features of a given text in PyTorch:
182
183 ```python
184 from transformers import BertTokenizer, BertModel
185 tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-uncased')
186 model = BertModel.from_pretrained("bert-base-multilingual-uncased")
187 text = "Replace me by any text you'd like."
188 encoded_input = tokenizer(text, return_tensors='pt')
189 output = model(**encoded_input)
190 ```
191
192 and in TensorFlow:
193
194 ```python
195 from transformers import BertTokenizer, TFBertModel
196 tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-uncased')
197 model = TFBertModel.from_pretrained("bert-base-multilingual-uncased")
198 text = "Replace me by any text you'd like."
199 encoded_input = tokenizer(text, return_tensors='tf')
200 output = model(encoded_input)
201 ```
202
203 ### Limitations and bias
204
205 Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
206 predictions:
207
208 ```python
209 >>> from transformers import pipeline
210 >>> unmasker = pipeline('fill-mask', model='bert-base-multilingual-uncased')
211 >>> unmasker("The man worked as a [MASK].")
212
213 [{'sequence': '[CLS] the man worked as a teacher. [SEP]',
214 'score': 0.07943806052207947,
215 'token': 21733,
216 'token_str': 'teacher'},
217 {'sequence': '[CLS] the man worked as a lawyer. [SEP]',
218 'score': 0.0629938617348671,
219 'token': 34249,
220 'token_str': 'lawyer'},
221 {'sequence': '[CLS] the man worked as a farmer. [SEP]',
222 'score': 0.03367974981665611,
223 'token': 36799,
224 'token_str': 'farmer'},
225 {'sequence': '[CLS] the man worked as a journalist. [SEP]',
226 'score': 0.03172805905342102,
227 'token': 19477,
228 'token_str': 'journalist'},
229 {'sequence': '[CLS] the man worked as a carpenter. [SEP]',
230 'score': 0.031021825969219208,
231 'token': 33241,
232 'token_str': 'carpenter'}]
233
234 >>> unmasker("The Black woman worked as a [MASK].")
235
236 [{'sequence': '[CLS] the black woman worked as a nurse. [SEP]',
237 'score': 0.07045423984527588,
238 'token': 52428,
239 'token_str': 'nurse'},
240 {'sequence': '[CLS] the black woman worked as a teacher. [SEP]',
241 'score': 0.05178029090166092,
242 'token': 21733,
243 'token_str': 'teacher'},
244 {'sequence': '[CLS] the black woman worked as a lawyer. [SEP]',
245 'score': 0.032601192593574524,
246 'token': 34249,
247 'token_str': 'lawyer'},
248 {'sequence': '[CLS] the black woman worked as a slave. [SEP]',
249 'score': 0.030507225543260574,
250 'token': 31173,
251 'token_str': 'slave'},
252 {'sequence': '[CLS] the black woman worked as a woman. [SEP]',
253 'score': 0.027691684663295746,
254 'token': 14050,
255 'token_str': 'woman'}]
256 ```
257
258 This bias will also affect all fine-tuned versions of this model.
259
260 ## Training data
261
262 The BERT model was pretrained on the 102 languages with the largest Wikipedias. You can find the complete list
263 [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
264
265 ## Training procedure
266
267 ### Preprocessing
268
269 The texts are lowercased and tokenized using WordPiece and a shared vocabulary size of 110,000. The languages with a
270 larger Wikipedia are under-sampled and the ones with lower resources are oversampled. For languages like Chinese,
271 Japanese Kanji and Korean Hanja that don't have space, a CJK Unicode block is added around every character.
272
273 The inputs of the model are then of the form:
274
275 ```
276 [CLS] Sentence A [SEP] Sentence B [SEP]
277 ```
278
279 With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in
280 the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a
281 consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two
282 "sentences" has a combined length of less than 512 tokens.
283
284 The details of the masking procedure for each sentence are the following:
285 - 15% of the tokens are masked.
286 - In 80% of the cases, the masked tokens are replaced by `[MASK]`.
287 - In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
288 - In the 10% remaining cases, the masked tokens are left as is.
289
290
291 ### BibTeX entry and citation info
292
293 ```bibtex
294 @article{DBLP:journals/corr/abs-1810-04805,
295 author = {Jacob Devlin and
296 Ming{-}Wei Chang and
297 Kenton Lee and
298 Kristina Toutanova},
299 title = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language
300 Understanding},
301 journal = {CoRR},
302 volume = {abs/1810.04805},
303 year = {2018},
304 url = {http://arxiv.org/abs/1810.04805},
305 archivePrefix = {arXiv},
306 eprint = {1810.04805},
307 timestamp = {Tue, 30 Oct 2018 20:39:56 +0100},
308 biburl = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib},
309 bibsource = {dblp computer science bibliography, https://dblp.org}
310 }
311 ```
312