Replace mecab-python3 with fugashi for Japanese tokenization (#6086)

* Replace mecab-python3 with fugashi

This replaces mecab-python3 with fugashi for Japanese tokenization. I am
the maintainer of both projects.

Both projects are MeCab wrappers, so the underlying C++ code is the
same. fugashi is the newer wrapper and doesn't use SWIG, so for basic
use of the MeCab API it's easier to use.

This code insures the use of a version of ipadic installed via pip,
which should make versioning and tracking down issues easier.

fugashi has wheels for Windows, OSX, and Linux, which will help with
issues with installing old versions of mecab-python3 on Windows.
Compared to mecab-python3, because fugashi doesn't use SWIG, it doesn't
require a C++ runtime to be installed on Windows.

In adding this change I removed some code dealing with `cursor`,
`token_start`, and `token_end` variables. These variables didn't seem to
be used for anything, it is unclear to me why they were there.

I ran the tests and they passed, though I couldn't figure out how to run
the slow tests (`--runslow` gave an error) and didn't try testing with
Tensorflow.

* Style fix

* Remove unused variable

Forgot to delete this...

* Adapt doc with install instructions

* Fix typo
Co-authored-by: sgugger <sylvain.gugger@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

.circleci/config.yml

@@ -56,7 +56,7 @@ jobs:
             RUN_CUSTOM_TOKENIZERS: yes
         steps:
             - checkout
-            - run: sudo pip install .[mecab,testing]
+            - run: sudo pip install .[ja,testing]
             - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
             - store_artifacts:
                 path: ~/transformers/output.txt

docs/source/pretrained_models.rst

@@ -74,14 +74,16 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/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.                                                          |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
+|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
+|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/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.                                                          |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
+|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
+|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

setup.cfg

@@ -10,10 +10,10 @@ known_third_party =
     faiss
     fastprogress
     fire
+    fugashi
     git
     h5py
     matplotlib
-    MeCab
     nlp
     nltk
     numpy

setup.py

@@ -65,7 +65,7 @@ if stale_egg_info.exists():
 
 extras = {}
 
-extras["mecab"] = ["mecab-python3<1"]
+extras["ja"] = ["fugashi>=1.0", "ipadic>=1.0,<2.0"]
 extras["sklearn"] = ["scikit-learn"]
 
 # keras2onnx and onnxconverter-common version is specific through a commit until 1.7.0 lands on pypi
@@ -97,7 +97,7 @@ extras["quality"] = [
     "isort @ git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort",
     "flake8",
 ]
-extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3<1", "scikit-learn", "tensorflow", "torch"]
+extras["dev"] = extras["testing"] + extras["quality"] + extras["ja"] + ["scikit-learn", "tensorflow", "torch"]
 
 setup(
     name="transformers",

src/transformers/tokenization_bert_japanese.py

@@ -185,9 +185,14 @@ class MecabTokenizer:
         self.never_split = never_split if never_split is not None else []
         self.normalize_text = normalize_text
 
-        import MeCab
+        import fugashi
+        import ipadic
 
-        self.mecab = MeCab.Tagger(mecab_option) if mecab_option is not None else MeCab.Tagger()
+        # Use ipadic by default (later options can override it)
+        mecab_option = mecab_option or ""
+        mecab_option = ipadic.MECAB_ARGS + " " + mecab_option
+
+        self.mecab = fugashi.GenericTagger(mecab_option)
 
     def tokenize(self, text, never_split=None, **kwargs):
         """Tokenizes a piece of text."""
@@ -197,21 +202,13 @@ class MecabTokenizer:
         never_split = self.never_split + (never_split if never_split is not None else [])
         tokens = []
 
-        mecab_output = self.mecab.parse(text)
-
-        cursor = 0
-        for line in mecab_output.split("\n"):
-            if line == "EOS":
-                break
+        for word in self.mecab(text):
+            token = word.surface
 
-            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