[PYTHON] Learn Japanese document categorization using spaCy / GiNZA (failure)

This article is the 9th day article of Unique Vision Co., Ltd. Advent Calendar 2019.

Introduction

This article is a sequel to Learn document categorization with spaCy CLI. Last time, we dealt with English documents, but in this article, we will study categorization using spaCy's CLI for Japanese documents.

Prerequisites

Like last time, the execution environment uses the GPU runtime of google Colab, and the GPU is Tesla P100. Also, the version of the library is

is. GiNZA is installed according to the following procedure.

$ pip install "https://github.com/megagonlabs/ginza/releases/download/latest/ginza-latest.tar.gz"

This time, we will run it on Google Colab, so we will also execute the following magic.

import pkg_resources, imp
imp.reload(pkg_resources)

Use the Livedoor News Corpus for the dataset. Download the data in advance as follows.

$ wget https://www.rondhuit.com/download/ldcc-20140209.tar.gz
$ tar -xvf ldcc-20140209.tar.gz

Data preparation

The data structure is the same as last time, so the save process remains the same.

import spacy
import srsly
from spacy.gold import docs_to_json

def save_to_json(model, data, targets, target_names, output_file, n_texts=0):
  def get_categories(target):
    return dict([(key, int(target == i)) for i, key in enumerate(target_names)])

  nlp = spacy.load(model)
  nlp.disable_pipes(*nlp.pipe_names)
  sentencizer = nlp.create_pipe("sentencizer")
  nlp.add_pipe(sentencizer, first=True)

  docs = []
  count = 0
  for i, doc in enumerate(nlp.pipe(data)):
    doc.cats = get_categories(targets[i])
    docs.append(doc)

    if n_texts > 0 and count == n_texts:
      break
    count += 1

  srsly.write_json(output_file, [docs_to_json(docs)])
  return count

Load the data so that you can use this function.

from pathlib import Path

data = []
targets = []
target_names = []

for target, target_name in enumerate([p for p in Path('text').iterdir() if p.is_dir()]):
  target_names.append(target_name.name)
  for news in target_name.iterdir():
    if 'LICENSE' in news.name:
      continue
    with open(news) as f:
      s = '\n'.join(f.read().splitlines()[2:])
    data.append(s)
    targets.append(target)

Save it in the file as follows.

from sklearn.model_selection import train_test_split

X_train, X_dev, y_train, y_dev = train_test_split(data, targets, test_size=0.20, random_state=42)

save_to_json(
    'ja_ginza',
    X_train,
    y_train,
    target_names,
    'train.json'
)

save_to_json(
    'ja_ginza',
    X_dev,
    y_dev,
    target_names,
    'dev.json'
)

Note that unlike the previous article, we specified ja_ginza for the model. If you create a JSON file by this method, non-ASCII characters will be Unicode escaped and output, but there is no problem because it will make it easier to read.

Learning

If you learn in the same way as last time, you will get the following error.

$ !time python -m spacy train ja output train.json dev.json -v ja_ginza -p textcat -ta simple_cnn -g 0
Training pipeline: ['textcat']
Starting with blank model 'ja'
Loading vector from model 'ja_ginza'
Traceback (most recent call last):
  File "/usr/lib/python3.6/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/usr/local/lib/python3.6/dist-packages/spacy/__main__.py", line 33, in <module>
    plac.call(commands[command], sys.argv[1:])
  File "/usr/local/lib/python3.6/dist-packages/plac_core.py", line 328, in call
    cmd, result = parser.consume(arglist)
  File "/usr/local/lib/python3.6/dist-packages/plac_core.py", line 207, in consume
    return cmd, self.func(*(args + varargs + extraopts), **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/spacy/cli/train.py", line 213, in train
    _load_vectors(nlp, vectors)
  File "/usr/local/lib/python3.6/dist-packages/spacy/cli/train.py", line 530, in _load_vectors
    util.load_model(vectors, vocab=nlp.vocab)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 162, in load_model
    return load_model_from_link(name, **overrides)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 179, in load_model_from_link
    return cls.load(**overrides)
  File "/usr/local/lib/python3.6/dist-packages/spacy/data/ja_ginza/__init__.py", line 12, in load
    return load_model_from_init_py(__file__, **overrides)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 228, in load_model_from_init_py
    return load_model_from_path(data_path, meta, **overrides)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 211, in load_model_from_path
    return nlp.from_disk(model_path)
  File "/usr/local/lib/python3.6/dist-packages/spacy/language.py", line 941, in from_disk
    util.from_disk(path, deserializers, exclude)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 654, in from_disk
    reader(path / key)
  File "/usr/local/lib/python3.6/dist-packages/spacy/language.py", line 936, in <lambda>
    p, exclude=["vocab"]
  File "nn_parser.pyx", line 665, in spacy.syntax.nn_parser.Parser.from_disk
  File "nn_parser.pyx", line 77, in spacy.syntax.nn_parser.Parser.Model
  File "/usr/local/lib/python3.6/dist-packages/spacy/_ml.py", line 323, in Tok2Vec
    return _legacy_tok2vec.Tok2Vec(width, embed_size, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/spacy/ml/_legacy_tok2vec.py", line 44, in Tok2Vec
    glove = StaticVectors(pretrained_vectors, width, column=cols.index(ID))
  File "/usr/local/lib/python3.6/dist-packages/thinc/neural/_classes/static_vectors.py", line 43, in __init__
    vectors = self.get_vectors()
  File "/usr/local/lib/python3.6/dist-packages/thinc/neural/_classes/static_vectors.py", line 55, in get_vectors
    return get_vectors(self.ops, self.lang)
  File "/usr/local/lib/python3.6/dist-packages/thinc/extra/load_nlp.py", line 26, in get_vectors
    nlp = get_spacy(lang)
  File "/usr/local/lib/python3.6/dist-packages/thinc/extra/load_nlp.py", line 14, in get_spacy
    SPACY_MODELS[lang] = spacy.load(lang, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/spacy/__init__.py", line 30, in load
    return util.load_model(name, **overrides)
  File "/usr/local/lib/python3.6/dist-packages/spacy/util.py", line 169, in load_model
    raise IOError(Errors.E050.format(name=name))
OSError: [E050] Can't find model 'ja_nopn.vectors'. It doesn't seem to be a shortcut link, a Python package or a valid path to a data directory.

real	0m1.610s
user	0m1.299s
sys	0m0.394s

When I read the error message, it seems that there is a problem with _load_vectors, so I will try to reproduce the situation with the following code.

from spacy.util import get_lang_class, load_model
lang_cls = get_lang_class('ja')
nlp = lang_cls()
load_model('ja_ginza', vocab=nlp.vocab)

Then this will succeed. I'm a little exhausted here, so I'll start learning from an empty model.

Learning from an empty model

Just remove -v ja_ginza from the above command.

$ !time python -m spacy train ja output train.json dev.json -p textcat -ta simple_cnn -g 0
Training pipeline: ['textcat']
Starting with blank model 'ja'
Counting training words (limit=0)
tcmalloc: large alloc 2128887808 bytes == 0x629bc000 @  0x7f7744def1e7 0x5acd6b 0x7f773a41a5db 0x7f773a41abf0 0x7f773a41ae36 0x7f773a4185c1 0x50ac25 0x50c5b9 0x7f76e032ab20 0x7f76e032f98f 0x7f76e03226c5 0x7f76e0371c47 0x7f76e31d372a 0x7f76e035efce 0x7f76e31d372a 0x7f76e038b1e7 0x7f76e31d372a 0x7f76e0352148 0x7f76e035c24b 0x59509c 0x54a8a5 0x551b81 0x5aa6ec 0x50abb3 0x50d390 0x508245 0x589471 0x5a067e 0x50d966 0x508245 0x50a080
Textcat evaluation score: F1-score macro-averaged across the labels 'peachy,
smax, it-life-hack, sports-watch, movie-enter, livedoor-homme, dokujo-tsushin,
kaden-channel, topic-news'

Itn  Textcat Loss  Textcat  Token %  CPU WPS  GPU WPS
---  ------------  -------  -------  -------  -------
  1      1312.086   84.850   99.995    32519   119886
  2       181.481   89.898   99.995    32879   119157
  3       119.681   91.490   99.995    32741   122003
  4        87.429   92.677   99.995    32618   119988
  5        66.618   92.674   99.995    32005   122189
  6        45.137   92.484   99.995    32293   113857

Apparently I can learn, but it is very heavy compared to en. It is a level of weight that does not feel like letting you learn casually.

Summary

I tried categorizing Japanese documents using the CLI of spaCy. The bottom line is failure.

First of all, I think we have to load the base model and solve the heavy problem.

Recommended Posts

Learn Japanese document categorization using spaCy / GiNZA (failure)
Learn document categorization with spaCy CLI
[PyTorch] Introduction to Japanese document classification using BERT
Perform entity analysis using spaCy / GiNZA in Python
[Failure] Development of Japanese TTS using Tacotron2 ~ 2nd work ~