Adapted from for HuggingFace/Transformers library


You must run the original ThaiTokenizer to have your tokenization match that of the original model. If you skip this step, you will not do much better than mBERT or random chance!

pip install pythainlp six sentencepiece==0.0.9
git clone
# download .vocab and .model files from ThAIKeras readme

Then set up ThaiTokenizer class - this is modified slightly to remove a TensorFlow dependency.

import collections
import unicodedata
import six

def convert_to_unicode(text):
  """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
  if six.PY3:
    if isinstance(text, str):
      return text
    elif isinstance(text, bytes):
      return text.decode("utf-8", "ignore")
      raise ValueError("Unsupported string type: %s" % (type(text)))
  elif six.PY2:
    if isinstance(text, str):
      return text.decode("utf-8", "ignore")
    elif isinstance(text, unicode):
      return text
      raise ValueError("Unsupported string type: %s" % (type(text)))
    raise ValueError("Not running on Python2 or Python 3?")

def load_vocab(vocab_file):
  vocab = collections.OrderedDict()
  index = 0
  with open(vocab_file, "r") as reader:
    while True:
      token = reader.readline()
      if token.split(): token = token.split()[0] # to support SentencePiece vocab file
      token = convert_to_unicode(token)
      if not token:
      token = token.strip()
      vocab[token] = index
      index += 1
  return vocab


from bert.bpe_helper import BPE
import sentencepiece as spm

def convert_by_vocab(vocab, items):
  output = []
  for item in items:
  return output

class ThaiTokenizer(object):
  """Tokenizes Thai texts."""

  def __init__(self, vocab_file, spm_file):
    self.vocab = load_vocab(vocab_file)
    self.inv_vocab = {v: k for k, v in self.vocab.items()}

    self.bpe = BPE(vocab_file)    
    self.s = spm.SentencePieceProcessor()

  def tokenize(self, text):
    bpe_tokens = self.bpe.encode(text).split(' ')
    spm_tokens = self.s.EncodeAsPieces(text)

    tokens = bpe_tokens if len(bpe_tokens) < len(spm_tokens) else spm_tokens

    split_tokens = []

    for token in tokens:
      new_token = token

      if token.startswith('_') and not token in self.vocab:
        new_token = token[1:]

      if not new_token in self.vocab:

    return split_tokens

  def convert_tokens_to_ids(self, tokens):
    return convert_by_vocab(self.vocab, tokens)

  def convert_ids_to_tokens(self, ids):
    return convert_by_vocab(self.inv_vocab, ids)

Then pre-tokenizing your own text:

from pythainlp import sent_tokenize
tokenizer = ThaiTokenizer(vocab_file='', spm_file='')  

og_text = "กรุงเทพมหานคร..."
split_sentences = ' '.join(sent_tokenize(txt))
split_words = ' '.join(tokenizer.tokenize(split_sentences))

> "▁ร้าน อาหาร ใหญ่มาก กก กก กก ▁ <unk> เลี้ยว..."

Original README follows:

Google's BERT is currently the state-of-the-art method of pre-training text representations which additionally provides multilingual models. Unfortunately, Thai is the only one in 103 languages that is excluded due to difficulties in word segmentation.

BERT-th presents the Thai-only pre-trained model based on the BERT-Base structure. It is now available to download.

BERT-th also includes relevant codes and scripts along with the pre-trained model, all of which are the modified versions of those in the original BERT project.


Data Source

Training data for BERT-th come from the latest article dump of Thai Wikipedia on November 2, 2018. The raw texts are extracted by using WikiExtractor.

Sentence Segmentation

Input data need to be segmented into separate sentences before further processing by BERT modules. Since Thai language has no explicit marker at the end of a sentence, it is quite problematic to pinpoint sentence boundaries. To the best of our knowledge, there is still no implementation of Thai sentence segmentation elsewhere. So, in this project, sentence segmentation is done by applying simple heuristics, considering spaces, sentence length and common conjunctions.

After preprocessing, the training corpus consists of approximately 2 million sentences and 40 million words (counting words after word segmentation by PyThaiNLP). The plain and segmented texts can be downloaded here.


BERT uses WordPiece as a tokenization mechanism. But it is Google internal, we cannot apply existing Thai word segmentation and then utilize WordPiece to learn the set of subword units. The best alternative is SentencePiece which implements BPE and needs no word segmentation.

In this project, we adopt a pre-trained Thai SentencePiece model from BPEmb. The model of 25000 vocabularies is chosen and the vocabulary file has to be augmented with BERT's special characters, including '[PAD]', '[CLS]', '[SEP]' and '[MASK]'. The model and vocabulary files can be downloaded here.

SentencePiece and from BPEmb are both used to tokenize data. ThaiTokenizer class has been added to BERT's for tokenizing Thai texts.


The data can be prepared before pre-training by using this script.

export BPE_DIR=/path/to/bpe
export TEXT_DIR=/path/to/text
export DATA_DIR=/path/to/data

python \
  --input_file=$TEXT_DIR/thaiwikitext_sentseg \
  --output_file=$DATA_DIR/tf_examples.tfrecord \
  --vocab_file=$BPE_DIR/ \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --masked_lm_prob=0.15 \
  --random_seed=12345 \
  --dupe_factor=5 \
  --thai_text=True \

Then, the following script can be run to learn a model from scratch.

export DATA_DIR=/path/to/data
export BERT_BASE_DIR=/path/to/bert_base

python \
  --input_file=$DATA_DIR/tf_examples.tfrecord \
  --output_dir=$BERT_BASE_DIR \
  --do_train=True \
  --do_eval=True \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --train_batch_size=32 \
  --max_seq_length=128 \
  --max_predictions_per_seq=20 \
  --num_train_steps=1000000 \
  --num_warmup_steps=100000 \
  --learning_rate=1e-4 \

We have trained the model for 1 million steps. On Tesla K80 GPU, it took around 20 days to complete. Though, we provide a snapshot at 0.8 million steps because it yields better results for downstream classification tasks.

Downstream Classification Tasks


XNLI is a dataset for evaluating a cross-lingual inferential classification task. The development and test sets contain 15 languages which data are thoroughly edited. The machine-translated versions of training data are also provided.

The Thai-only pre-trained BERT model can be applied to the XNLI task by using training data which are translated to Thai. Spaces between words in the training data need to be removed to make them consistent with inputs in the pre-training step. The processed files of XNLI related to Thai language can be downloaded here.

Afterwards, the XNLI task can be learned by using this script.

export BPE_DIR=/path/to/bpe
export XNLI_DIR=/path/to/xnli
export OUTPUT_DIR=/path/to/output
export BERT_BASE_DIR=/path/to/bert_base

python \
  --task_name=XNLI \
  --do_train=true \
  --do_eval=true \
  --data_dir=$XNLI_DIR \
  --vocab_file=$BPE_DIR/ \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/model.ckpt \
  --max_seq_length=128 \
  --train_batch_size=32 \
  --learning_rate=5e-5 \
  --num_train_epochs=2.0 \
  --output_dir=$OUTPUT_DIR \
  --xnli_language=th \

This table compares the Thai-only model with XNLI baselines and the Multilingual Cased model which is also trained by using translated data.

XNLI Baseline BERT
Translate Train Translate Test Multilingual Model Thai-only Model
62.8 64.4 66.1 68.9

Wongnai Review Dataset

Wongnai Review Dataset collects restaurant reviews and ratings from Wongnai website. The task is to classify a review into one of five ratings (1 to 5 stars). The dataset can be downloaded here and the following script can be run to use the Thai-only model for this task.

export BPE_DIR=/path/to/bpe
export WONGNAI_DIR=/path/to/wongnai
export OUTPUT_DIR=/path/to/output
export BERT_BASE_DIR=/path/to/bert_base

python \
  --task_name=wongnai \
  --do_train=true \
  --do_predict=true \
  --data_dir=$WONGNAI_DIR \
  --vocab_file=$BPE_DIR/ \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/model.ckpt \
  --max_seq_length=128 \
  --train_batch_size=32 \
  --learning_rate=5e-5 \
  --num_train_epochs=2.0 \
  --output_dir=$OUTPUT_DIR \

Without additional preprocessing and further fine-tuning, the Thai-only BERT model can achieve 0.56612 and 0.57057 for public and private test-set scores respectively.

Downloads last month
Hosted inference API

Unable to determine this model’s pipeline type. Check the docs .