/rpunct

📝An easy-to-use package to restore punctuation of the text.

Primary LanguagePython

bbc-punctuator

This is the BBC Punctuator, developed by BBC R&D's Speech-to-Text team. BBC Punctuator is a machine learning model for restoring punctuation and capitalisation into plaintext, such as raw transcripts produced by STT systems.

The model uses HuggingFace's bert-base-uncased model weights that have been fine-tuned for Punctuation restoration for one epoch over a new BBC built composite punctuator training dataset, composed of BBC News articles (written between 2020-2), BBC News transcripts, and subtitles from a range of BBC TV programmes [model=composite20v3-1e]. The parameters learned over this training process are stored in bbc-data (bbc/speech/transcription/bbc-punctuator:0.0.1), which when referenced from this code make a complete system for punctuating text.

The system utilises the CPU by default but can be instructed to run on a GPU.

List of text featured available for recovery:

  • Capitalisation:
    • Capitalisaton
    • Acronyms (upper-casing)
    • Mixed-casing
  • Punctuation:
    • Period: .
    • Comma: ,
    • Question Mark: ?
    • Exclamation: !
    • Colon: :
    • Semi-colon: ;
    • Apostrophe: '
    • Hypen: -
  • Numbers:
    • Digitising natural language numbers
    • Currencies: £, $, €, ¥
    • Percentages: %
    • Ordinals
    • Decimals
    • Dates (years/decades)

Usage:

  1. Pull the model files from bbc-data:
    bbc-data pull bbc/rpunct:0.0.1

  2. Install the BBC Punctuator system from Git:
    git clone git@github.com:bbc/rpunct.git
    cd rpunct
    pip install -r requirements.txt

  3. Move the model files to the reference location within bbc-punctuator:
    mv $(bbc-data path bbc/rpunct) model-files

  4. Run the BBC Punctuator by pointing the run file to the model files (from inside the cloned repo):

    1. To punctuate a plaintext (.txt) file from the command line:
      python run.py rpunct -m model-files -i input_file_path.txt -o output_file_path.txt

    2. To use bbc-punctuator within a python script (from within the rpunct directory):

      from rpunct.rpunct_recoverer import RPunctRecoverer

punctuator = RPunctRecoverer()
text = "hello and welcome to the bbc punctuator \
this is an example piece of plaintext \
that exhibits no punctuation or capitalisation"

punctuated_text = punctuator.process(text)

Performance

Metric Prec Recall F1
abbreviations 0.9377 0.8605 0.8974
capitalisation 0.9213 0.9434 0.9322
commas 0.6977 0.7029 0.7003
hyphens 0.7198 0.8953 0.7980
numbers 0.7499 0.7747 0.7621
sentences 0.8351 0.8762 0.8552
fullstops 0.8110 0.8710 0.8399
questions 0.7537 0.4770 0.5843

Authors