Train models to restore punctuation and capitalization for texts in different languages. Based on NeMo project.

At first you need to download the dataset with the texts in desired language. You can use the Tatoeba dataset for that purpose. It has fairly enough punctuation marks.

Use the following code to get Tatoeba dataset:

python3 ./get_tatoeba.py --data_dir ./dataset --lang rus

Dataset will be downloaded into dataset folder in sentences.csv file. After that the sentences with your language code will be written in sentences.txt file.

At the next step we need to prepare the dataset for training.

• Clean sentences
  • Remove the punctuation you don't want to restore
    • Desired chars to restore are in the CHARS constant in the prepare_data.py script. Update it if needed.
  • Standartize the punctuation
    • Remove some fancy unicode punctuation with the usual one
• Encode sentences
  • Translate the sentences into something like OU OO ?O
    • It's special markup for our model
    • Details are here
• Split data on train and dev sets

To do that run the following code:

python3 ./prepare_data.py --data_dir ./dataset --num_samples 10000 --percent_dev 0.2

And finally we will train the model. You can train the model locally or use the following notebook (it's a fixed NeMo's Jupyter nb).

• train.ipynb

• Choose base BERT model in PRETRAINED_BERT_MODEL

  • DeepPavlov/distilrubert-tiny-cased-conversational-v1 used in this example notebook.

• Try to train in Colab

• To increase the quality please make sure that your dataset is balanced in terms of punctuation ratio.
• Adjust the training parameters.
• Try to choose different pretrained models (you can choose any BERT model from huggingface).

queries = [
        'меня зовут сергей а как тебя',
        'подскажи пожалуйста сегодня вторник или среда',
        'закрой за мной дверь я ухожу'
    ]

inference_results = model.add_punctuation_capitalization(queries)

for query, result in zip(queries, inference_results):
    print(f'Query   : {query}')
    print(f'Combined: {result.strip()}\n')

Query   : меня зовут сергей а как тебя
Combined: Меня зовут Сергей. А как тебя?

Query   : подскажи пожалуйста сегодня вторник или среда
Combined: Подскажи, пожалуйста, сегодня вторник или среда.

Query   : закрой за мной дверь я ухожу
Combined: Закрой за мной дверь. Я ухожу.