Add support for Japanese BERT models by cl-tohoku

docs/source/pretrained_models.rst

@@ -61,6 +61,24 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese``                                     | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece.                                                 |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                            |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-whole-word-masking``                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                        |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                            |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-char``                                | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized into characters.                                                          |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-japanese-char-whole-word-masking``             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters.                                 |
+|                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                             |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |

transformers/__init__.py

@@ -37,6 +37,7 @@ if is_sklearn_available():
 from .tokenization_utils import (PreTrainedTokenizer)
 from .tokenization_auto import AutoTokenizer
 from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
+from .tokenization_bert_japanese import BertJapaneseTokenizer, MecabTokenizer, CharacterTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
 from .tokenization_gpt2 import GPT2Tokenizer

transformers/configuration_bert.py

@@ -42,6 +42,10 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
+    'bert-base-japanese': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-config.json",
+    'bert-base-japanese-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-whole-word-masking-config.json",
+    'bert-base-japanese-char': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-config.json",
+    'bert-base-japanese-char-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-whole-word-masking-config.json"
 }
 
 

transformers/modeling_bert.py

@@ -48,6 +48,10 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
+    'bert-base-japanese': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-pytorch_model.bin",
+    'bert-base-japanese-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-whole-word-masking-pytorch_model.bin",
+    'bert-base-japanese-char': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-pytorch_model.bin",
+    'bert-base-japanese-char-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-whole-word-masking-pytorch_model.bin"
 }
 
 
@@ -1233,9 +1237,9 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
         input_ids = tokenizer.encode(input_text)
-        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]
         start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
-        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
         print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
         # a nice puppet
 

transformers/modeling_tf_bert.py

@@ -48,6 +48,10 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-tf_model.h5",
+    'bert-base-japanese': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-tf_model.h5",
+    'bert-base-japanese-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-whole-word-masking-tf_model.h5",
+    'bert-base-japanese-char': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-tf_model.h5",
+    'bert-base-japanese-char-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-whole-word-masking-tf_model.h5"
 }
 
 
@@ -129,7 +133,7 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
                 linear tensor, float32 with shape [batch_size, length, vocab_size].
         Raises:
             ValueError: if mode is not valid.
-        
+
         Shared weights logic adapted from
             https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
         """
@@ -148,7 +152,7 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
             input_shape = shape_list(input_ids)
         else:
             input_shape = shape_list(inputs_embeds)[:-1]
-        
+
         seq_length = input_shape[1]
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
@@ -246,7 +250,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
         context_layer = tf.matmul(attention_probs, value_layer)
 
         context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
-        context_layer = tf.reshape(context_layer, 
+        context_layer = tf.reshape(context_layer,
                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
 
         outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
@@ -591,7 +595,7 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
             `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
 
     Parameters:
-        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
             Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
@@ -605,13 +609,13 @@ BERT_INPUTS_DOCSTRING = r"""
             (a) For sequence pairs:
 
                 ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-                
+
                 ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
 
             (b) For single sequences:
 
                 ``tokens:         [CLS] the dog is hairy . [SEP]``
-                
+
                 ``token_type_ids:   0   0   0   0  0     0   0``
 
             Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on

transformers/tokenization_auto.py

@@ -19,6 +19,7 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 import logging
 
 from .tokenization_bert import BertTokenizer
+from .tokenization_bert_japanese import BertJapaneseTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer
 from .tokenization_ctrl import CTRLTokenizer
@@ -118,6 +119,8 @@ class AutoTokenizer(object):
             return CamembertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'bert-japanese' in pretrained_model_name_or_path:
+            return BertJapaneseTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'openai-gpt' in pretrained_model_name_or_path:

transformers/tokenization_bert_japanese.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import collections
+import logging
+import os
+import unicodedata
+from io import open
+
+from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer, load_vocab
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'bert-base-japanese': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-vocab.txt",
+        'bert-base-japanese-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-whole-word-masking-vocab.txt",
+        'bert-base-japanese-char': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-vocab.txt",
+        'bert-base-japanese-char-whole-word-masking': "https://www.nlp.ecei.tohoku.ac.jp/~m-suzuki/bert-japanese/bert-base-japanese-char-whole-word-masking-vocab.txt"
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'bert-base-japanese': 512,
+    'bert-base-japanese-whole-word-masking': 512,
+    'bert-base-japanese-char': 512,
+    'bert-base-japanese-char-whole-word-masking': 512
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    'bert-base-japanese': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'wordpiece'
+    },
+    'bert-base-japanese-whole-word-masking':{
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'wordpiece'
+    },
+    'bert-base-japanese-char': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'character'
+    },
+    'bert-base-japanese-char-whole-word-masking': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'character'
+    }
+}
+
+
+class BertJapaneseTokenizer(BertTokenizer):
+    """BERT tokenizer for Japanese text"""
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, do_lower_case=False,
+                 do_word_tokenize=True, do_subword_tokenize=True,
+                 word_tokenizer_type='basic', subword_tokenizer_type='wordpiece',
+                 never_split=None, unk_token='[UNK]', sep_token='[SEP]',
+                 pad_token='[PAD]', cls_token='[CLS]', mask_token='[MASK]', **kwargs):
+        """Constructs a MecabBertTokenizer.
+
+        Args:
+            **vocab_file**: Path to a one-wordpiece-per-line vocabulary file.
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input.
+                Only has an effect when do_basic_tokenize=True.
+            **do_word_tokenize**: (`optional`) boolean (default True)
+                Whether to do word tokenization.
+            **do_subword_tokenize**: (`optional`) boolean (default True)
+                Whether to do subword tokenization.
+            **word_tokenizer_type**: (`optional`) string (default "basic")
+                Type of word tokenizer.
+            **subword_tokenizer_type**: (`optional`) string (default "wordpiece")
+                Type of subword tokenizer.
+        """
+        super(BertTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
+                                            pad_token=pad_token, cls_token=cls_token,
+                                            mask_token=mask_token, **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict(
+            [(ids, tok) for tok, ids in self.vocab.items()])
+
+        self.do_word_tokenize = do_word_tokenize
+        if do_word_tokenize:
+            if word_tokenizer_type == 'basic':
+                self.word_tokenizer = BasicTokenizer(do_lower_case=do_lower_case,
+                                                     never_split=never_split,
+                                                     tokenize_chinese_chars=False)
+            elif word_tokenizer_type == 'mecab':
+                self.word_tokenizer = MecabTokenizer(do_lower_case=do_lower_case,
+                                                     never_split=never_split)
+            else:
+                raise ValueError(
+                    "Invalid word_tokenizer_type '{}' is specified.".format(word_tokenizer_type))
+
+        self.do_subword_tokenize = do_subword_tokenize
+        if do_subword_tokenize:
+            if subword_tokenizer_type == 'wordpiece':
+                self.subword_tokenizer = WordpieceTokenizer(vocab=self.vocab,
+                                                            unk_token=self.unk_token)
+            elif subword_tokenizer_type == 'character':
+                self.subword_tokenizer = CharacterTokenizer(vocab=self.vocab,
+                                                            unk_token=self.unk_token)
+            else:
+                raise ValueError(
+                    "Invalid subword_tokenizer_type '{}' is specified.".format(subword_tokenizer_type))
+
+
+    def _tokenize(self, text):
+        if self.do_word_tokenize:
+            tokens = self.word_tokenizer.tokenize(text,
+                                                  never_split=self.all_special_tokens)
+        else:
+            tokens = [text]
+
+        if self.do_subword_tokenize:
+            split_tokens = [sub_token for token in tokens
+                            for sub_token in self.subword_tokenizer.tokenize(token)]
+        else:
+            split_tokens = tokens
+
+        return split_tokens
+
+
+class MecabTokenizer(object):
+    """Runs basic tokenization with MeCab morphological parser."""
+
+    def __init__(self, do_lower_case=False, never_split=None, normalize_text=True):
+        """Constructs a MecabTokenizer.
+
+        Args:
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input.
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+            **normalize_text**: (`optional`) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+        """
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split if never_split is not None else []
+        self.normalize_text = normalize_text
+
+        import MeCab
+        self.mecab = MeCab.Tagger()
+
+    def tokenize(self, text, never_split=None, **kwargs):
+        """Tokenizes a piece of text."""
+        if self.normalize_text:
+            text = unicodedata.normalize('NFKC', text)
+
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        tokens = []
+
+        cursor = 0
+        for line in self.mecab.parse(text).split('\n'):
+            if line == 'EOS':
+                break
+
+            token, _ = line.split('\t')
+            token_start = text.index(token, cursor)
+            token_end = token_start + len(token)
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+
+            tokens.append(token)
+            cursor = token_end
+
+        return tokens
+
+
+class CharacterTokenizer(object):
+    """Runs Character tokenziation."""
+
+    def __init__(self, vocab, unk_token, normalize_text=True):
+        """Constructs a CharacterTokenizer.
+
+        Args:
+            **vocab**:
+                Vocabulary object.
+            **unk_token**: str
+                A special symbol for out-of-vocabulary token.
+            **normalize_text**: (`optional`) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+        """
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.normalize_text = normalize_text
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into characters.
+
+        For example:
+            input = "apple"
+            output = ["a", "p", "p", "l", "e"]
+        Args:
+            text: A single token or whitespace separated tokens.
+                This should have already been passed through `BasicTokenizer`.
+        Returns:
+            A list of characters.
+        """
+        if self.normalize_text:
+            text = unicodedata.normalize('NFKC', text)
+
+        output_tokens = []
+        for i, char in enumerate(text):
+            if char not in self.vocab:
+                output_tokens.append(self.unk_token)
+                continue
+
+            output_tokens.append(char)
+
+        return output_tokens