Update README.md

README.md

@@ -10,11 +10,11 @@ widget:
 
 # POET: A French Extended Part-of-Speech Tagger
 
-- Corpus: [UD_FRENCH_TREEBANKS](https://universaldependencies.org/treebanks/fr_gsd/index.html)
+- Corpora: [UD_FRENCH_TREEBANKS](https://universaldependencies.org/treebanks/fr_gsd/index.html)
 - Model: [Flair](https://www.aclweb.org/anthology/C18-1139.pdf)
 - Embeddings: [FastText](https://fasttext.cc/)
-- Additionnel: [LSTM-CRF](https://arxiv.org/abs/1011.4088)
-- Nombre d'Epochs: 115
+- Sequence Labelling: [LSTM-CRF](https://arxiv.org/abs/1011.4088)
+- Number of Epochs: 115
 
 **People Involved**
 
@@ -54,12 +54,14 @@ Output:
 
 `UD_FRENCH_GSD_Plus` is a part-of-speech tagging corpora based on [UD_French-GSD](https://universaldependencies.org/treebanks/fr_gsd/index.html) which was originally created in 2015 and is based on the [universal dependency treebank v2.0](https://github.com/ryanmcd/uni-dep-tb).
 
-Originally, the corpora consists of 400,399 words (16,341 sentences) and had 17 different classes. Now, after applying our tags augmentation we obtain 60 different classes which add semantic information such as the gender, number, mood, person, tense or verb form given in the different CoNLL-03 fields from the original corpora.
+Originally, the corpora consists of 400,399 words (16,341 sentences) and had 17 different classes. Now, after applying our tags augmentation we obtain 60 different classes which add linguistic and semantic information such as the gender, number, mood, person, tense or verb form given in the different CoNLL-03 fields from the original corpora.
 
 We based our tags on the level of details given by the [LIA_TAGG](http://pageperso.lif.univ-mrs.fr/frederic.bechet/download.html) statistical POS tagger written by [Frédéric Béchet](http://pageperso.lif.univ-mrs.fr/frederic.bechet/index-english.html) in 2001.
 
 The corpora used for this model is available on [Github](https://github.com/qanastek/UD_FRENCH_GSD_PLUS) at the [CoNLL-U format](https://universaldependencies.org/format.html).
 
+Training data are fed to the model as free language and doesn't pass a normalization phase. Thus, it's made the model case and punctuation sensitive.
+
 ## Original Tags
 
 ```plain

train.py

@@ -0,0 +1,54 @@
+import os
+import argparse
+from datetime import datetime
+
+from flair.data import Corpus
+from flair.models import SequenceTagger
+from flair.trainers import ModelTrainer
+from flair.datasets import UniversalDependenciesCorpus
+from flair.embeddings import WordEmbeddings, StackedEmbeddings
+
+parser = argparse.ArgumentParser(description='Flair Training Part-of-speech tagging')
+parser.add_argument('-output', type=str, default="models/", help='The output directory')
+parser.add_argument('-epochs', type=int, default=1, help='Number of Epochs')
+args = parser.parse_args()
+
+output = os.path.join(args.output, "UPOS_UD_FRENCH_PLUS_" + str(args.epochs) + "_" + datetime.today().strftime('%Y-%m-%d-%H:%M:%S'))
+print(output)
+
+# corpus: Corpus = UD_FRENCH()
+corpus: Corpus = UniversalDependenciesCorpus(
+    data_folder='UD_FRENCH_PLUS',
+    train_file="fr_gsd-ud-train.conllu",
+    test_file="fr_gsd-ud-test.conllu",
+    dev_file="fr_gsd-ud-dev.conllu",
+)
+# print(corpus)
+
+tag_type = 'upos'
+
+tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
+# print(tag_dictionary)
+
+embedding_types = [
+    WordEmbeddings('fr'),
+]
+
+embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
+
+tagger: SequenceTagger = SequenceTagger(
+    hidden_size=256,
+    embeddings=embeddings,
+    tag_dictionary=tag_dictionary,
+    tag_type=tag_type,
+    use_crf=True
+)
+
+trainer: ModelTrainer = ModelTrainer(tagger, corpus)
+
+trainer.train(
+    output,
+    learning_rate=0.1,
+    mini_batch_size=128,
+    max_epochs=args.epochs
+)