There are two multilingual models currently available. We do not plan to release
more single-language models, but we may release BERT-Large
versions of these
two in the future:
BERT-Base, Multilingual
: 102 languages, 12-layer, 768-hidden, 12-heads, 110M parametersBERT-Base, Chinese
: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters
See the list of languages that the Multilingual model supports. The Multilingual model does include Chinese (and English), but if your fine-tuning data is Chinese-only, then the Chinese model will likely produce better results.
To evaluate these systems, we use the XNLI dataset dataset, which is a version of MultiNLI where the dev and test sets have been translated (by humans) into 15 languages. Note that the training set was machine translated (we used the translations provided by XNLI, not Google NMT). For clarity, we only report on 6 languages below:
System | English | Chinese | Spanish | German | Arabic | Urdu |
---|---|---|---|---|---|---|
XNLI Baseline - Translate Train | 73.7 | 67.0 | 68.8 | 66.5 | 65.8 | 56.6 |
XNLI Baseline - Translate Test | 73.7 | 68.3 | 70.7 | 68.7 | 66.8 | 59.3 |
BERT -Translate Train | 81.4 | 74.2 | 77.3 | 75.2 | 70.5 | 61.7 |
BERT - Translate Test | 81.4 | 70.1 | 74.9 | 74.4 | 70.4 | 62.1 |
BERT - Zero Shot | 81.4 | 63.8 | 74.3 | 70.5 | 62.1 | 58.3 |
The first two rows are baselines from the XNLI paper and the last three rows are our results with BERT.
Translate Train means that the MultiNLI training set was machine translated from English into the foreign language. So training and evaluation were both done in the foreign language. Unfortunately, training was done on machine-translated data, so it is impossible to quantify how much of the lower accuracy (compared to English) is due to the quality of the machine translation vs. the quality of the pre-trained model.
Translate Test means that the XNLI test set was machine translated from the foreign language into English. So training and evaluation were both done on English. However, test evaluation was done on machine-translated English, so the accuracy depends on the quality of the machine translation system.
Zero Shot means that the Multilingual BERT system was fine-tuned on English MultiNLI, and then evaluated on the foreign language XNLI test. In this case, machine translation was not involved at all in either the pre-training or fine-tuning.
Note that the English result is worse than the 84.2 MultiNLI baseline because this training used Multilingual BERT rather than English-only BERT. This implies that for high-resource languages, the Multilingual model is somewhat worse than a single-language model. However, it is not feasible for us to train and maintain dozens of single-language model. Therefore, if your goal is to maximize performance with a language other than English or Chinese, you might find it beneficial to run pre-training for additional steps starting from our Multilingual model on data from your language of interest.
Here is a comparison of training Chinese models with the Multilingual
BERT-Base
and Chinese-only BERT-Base
:
System | Chinese |
---|---|
XNLI Baseline | 67.0 |
BERT Multilingual Model | 74.2 |
BERT Chinese-only Model | 77.2 |
Similar to English, the single-language model does 3% better than the Multilingual model.
The multilingual model does not require any special consideration or API
changes. We did update the implementation of BasicTokenizer
in
tokenization.py
to support Chinese character tokenization, so please update if
you forked it. However, we did not change the tokenization API.
To test the new models, we did modify run_classifier.py
to add support for the
XNLI dataset. This is a 15-language
version of MultiNLI where the dev/test sets have been human-translated, and the
training set has been machine-translated.
To run the fine-tuning code, please download the
XNLI dev/test set and the
XNLI machine-translated training set
and then unpack both .zip files into some directory $XNLI_DIR
.
To run fine-tuning on XNLI. The language is hard-coded into run_classifier.py
(Chinese by default), so please modify XnliProcessor
if you want to run on
another language.
This is a large dataset, so this will training will take a few hours on a GPU
(or about 30 minutes on a Cloud TPU). To run an experiment quickly for
debugging, just set num_train_epochs
to a small value like 0.1
.
export BERT_BASE_DIR=/path/to/bert/chinese_L-12_H-768_A-12 # or multilingual_L-12_H-768_A-12
export XNLI_DIR=/path/to/xnli
python run_classifier.py \
--task_name=XNLI \
--do_train=true \
--do_eval=true \
--data_dir=$XNLI_DIR \
--vocab_file=$BERT_BASE_DIR/vocab.txt \
--bert_config_file=$BERT_BASE_DIR/bert_config.json \
--init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
--max_seq_length=128 \
--train_batch_size=32 \
--learning_rate=5e-5 \
--num_train_epochs=2.0 \
--output_dir=/tmp/xnli_output/
With the Chinese-only model, the results should look something like this:
***** Eval results *****
eval_accuracy = 0.774116
eval_loss = 0.83554
global_step = 24543
loss = 0.74603
The languages chosen were the top 100 languages with the largest Wikipedias. The entire Wikipedia dump for each language (excluding user and talk pages) was taken as the training data for each language
However, the size of the Wikipedia for a given language varies greatly, and therefore low-resource languages may be "under-represented" in terms of the neural network model (under the assumption that languages are "competing" for limited model capacity to some extent).
However, the size of a Wikipedia also correlates with the number of speakers of a language, and we also don't want to overfit the model by performing thousands of epochs over a tiny Wikipedia for a particular language.
To balance these two factors, we performed exponentially smoothed weighting of the data during pre-training data creation (and WordPiece vocab creation). In other words, let's say that the probability of a language is P(L), e.g., P(English) = 0.21 means that after concatenating all of the Wikipedias together, 21% of our data is English. We exponentiate each probability by some factor S and then re-normalize, and sample from that distribution. In our case we use S=0.7. So, high-resource languages like English will be under-sampled, and low-resource languages like Icelandic will be over-sampled. E.g., in the original distribution English would be sampled 1000x more than Icelandic, but after smoothing it's only sampled 100x more.
For tokenization, we use a 110k shared WordPiece vocabulary. The word counts are weighted the same way as the data, so low-resource languages are upweighted by some factor. We intentionally do not use any marker to denote the input language (so that zero-shot training can work).
Because Chinese does not have whitespace characters, we add spaces around every character in the CJK Unicode range before applying WordPiece. This means that Chinese is effectively character-tokenized. Note that the CJK Unicode block only includes Chinese-origin characters and does not include Hangul Korean or Katakana/Hiragana Japanese, which are tokenized with whitespace+WordPiece like all other languages.
For all other languages, we apply the same recipe as English: (a) lower casing+accent removal, (b) punctuation splitting, (c) whitespace tokenization. We understand that accent markers have substantial meaning in some languages, but felt that the benefits of reducing the effective vocabulary make up for this. Generally the strong contextual models of BERT should make up for any ambiguity introduced by stripping accent markers.
The multilingual model supports the following languages. These languages were chosen because they are the top 100 languages with the largest Wikipedias:
- Afrikaans
- Albanian
- Arabic
- Aragonese
- Armenian
- Asturian
- Azerbaijani
- Bashkir
- Basque
- Bavarian
- Belarusian
- Bengali
- Bishnupriya Manipuri
- Bosnian
- Breton
- Bulgarian
- Burmese
- Catalan
- Cebuano
- Chechen
- Chinese (Simplified)
- Chinese (Traditional)
- Chuvash
- Croatian
- Czech
- Danish
- Dutch
- English
- Estonian
- Finnish
- French
- Galician
- Georgian
- German
- Greek
- Gujarati
- Haitian
- Hebrew
- Hindi
- Hungarian
- Icelandic
- Ido
- Indonesian
- Irish
- Italian
- Japanese
- Javanese
- Kannada
- Kazakh
- Kirghiz
- Korean
- Latin
- Latvian
- Lithuanian
- Lombard
- Low Saxon
- Luxembourgish
- Macedonian
- Malagasy
- Malay
- Malayalam
- Marathi
- Minangkabau
- Nepali
- Newar
- Norwegian (Bokmal)
- Norwegian (Nynorsk)
- Occitan
- Persian (Farsi)
- Piedmontese
- Polish
- Portuguese
- Punjabi
- Romanian
- Russian
- Scots
- Serbian
- Serbo-Croatian
- Sicilian
- Slovak
- Slovenian
- South Azerbaijani
- Spanish
- Sundanese
- Swahili
- Swedish
- Tagalog
- Tajik
- Tamil
- Tatar
- Telugu
- Turkish
- Ukrainian
- Urdu
- Uzbek
- Vietnamese
- Volapük
- Waray-Waray
- Welsh
- West
- Western Punjabi
- Yoruba
The only language which we had to unfortunately exclude was Thai, since it is the only language (other than Chinese) that does not use whitespace to delimit words, and it has too many characters-per-word to use character-based tokenization. Our WordPiece algorithm is quadratic with respect to the size of the input token so very long character strings do not work with it.