naubull2 / vpunct

An English auto punctuator for voice recognized texts

Voice Input Punctuator

• Tries to automatically puctuate voice recognized text

Data

• Datasets are from babi tasks data and the intent classifier dataset

• Simple questions V2
• movie dialogue task
• NLU benchmark dataset

Requirements

• python3
  • spacy, sanic, keras

Howto

• You can either tag [.!?] based on pattern rules only, otherwise use a neural network model

python punctuate.py -m [neural|pattern]  // defaults to pattern, in which case you won't need keras
python punctuate.py -e   // evaluate on a movie dialogue dataset

Performance

The following evaluation is performed on the movie dialogue task corpus.

• Toal number of lines : 136771

• Number of lines with question marks : 45122

• Pattern Only

Accuracy  : 0.7358669903196561
Precision : 0.8589784517158818
Recall    : 0.23853109347989895
F1 Score  : 0.3733782002358982

• NNet Only

TBD

• Joint Model

TBD

References

• babi dataset
• NLU benchmark dataset