Code and checkpoints corresponding to the posts https://medium.com/@praneethbedapudi/deepcorrection-3-spell-correction-and-simple-grammar-correction-d033a52bc11d and https://medium.com/@praneethbedapudi/deepcorrection2-automatic-punctuation-restoration-ac4a837d92d9
Pre-trained models for punctuation correction (trained on google news, wikipedia and tatoeba) are available at https://drive.google.com/open?id=1Yd8cJaqfQkrJMbRVWIWtuyo4obTDYu-e
Demo of the punctuation model trained on google news corpus is available at http://bpraneeth.com/projects/deeppunct
This repo uses a seq2seq model written by me in keras with tensorflow backend. The multi-purpose seq2seq model can be found at https://github.com/bedapudi6788/txt2txt/
Usage:
from deepcorrect import DeepCorrect
corrector = DeepCorrect('params_path', 'checkpoint_path')
corrector.correct('hey')
'Hey!'
pip install deepcorrect
Max input and output lengths are 200
Segment text into sentences using https://github.com/bedapudi6788/deepsegment and run punctuation correction on each sentence seperately.
As explored in https://medium.com/@praneethbedapudi/deepcorrection-3-spell-correction-and-simple-grammar-correction-d033a52bc11d , the original aim of this repo is to do spell + simple grammar + punctuation correction. But, since, I realised that doing spell correction with seq2seq is a very dumb idea I didn't provide usable checkpoints for the same. Keep in mind that just for punctuation restoration, though seq2seq works well, it's very time consuming (since > len(input_sent) sequential predictions need to be made.).