This repository contains the data, the codes (implemented in tensorflow) and the supplementary material for the CoNLL 2018 paper "Upcycle Your OCR: Reusing OCRs for Post-OCR Text Correction in Romanised Sanskrit"
nmt/nmt_data contains the data files.
- Training files (train_BPE.src and train_BPE.trg) contain the BPE encoded input strings which were the output of the OCR system.
- Testing files (test_BPE.src and test_BPE.trg) contain the BPE encoded input strings for testing. These strings are taken from Gita and Sahasranama manuscripts.
- Validation files (valid_BPE.src and valid_BPE.trg) contain the BPE encoded strings for validation.
- Vocabulary file vocab_BPE.src contains the shared vocabulary obtained after BPE. This vocab file can contain specific tokens which are only to be copied. In our experiement, we used the complete shared vocabulary as tokens which need to be copied and/or generated.
Test while training -
python2.7 -m nmt.nmt --copynet=True --share_vocab=True --attention=scaled_luong --src=src --tgt=trg --vocab_prefix=nmt/nmt_data/vocab_BPE --train_prefix=nmt/nmt_data/train_BPE --dev_prefix=nmt/nmt_data/valid_BPE --test_prefix=nmt/nmt_data/test_BPE --out_dir=nmt/copynet_models --num_train_steps=12000 --steps_per_stats=100 --encoder_type=bi --num_layers=4 --num_units=128 --dropout=0.4 --metrics=bleu --check_special_token=FalseOnly training -
python2.7 -m nmt.nmt --copynet=True --share_vocab=True --attention=scaled_luong --src=src --tgt=trg --vocab_prefix=nmt/nmt_data/vocab_BPE --train_prefix=nmt/nmt_data/train_BPE --dev_prefix=nmt/nmt_data/valid_BPE --out_dir=nmt/copynet_models --num_train_steps=12000 --steps_per_stats=100 --encoder_type=bi --num_layers=4 --num_units=128 --dropout=0.4 --metrics=bleu --check_special_token=FalseTest after training -
python2.7 -m nmt.nmt --out_dir=nmt/copynet_models --inference_input_file=nmt/my_infer_file.vi --inference_output_file=nmt/copynet_models/output_inferRequirements - tesnorflow 1.5
-
After running CopyNet model for 50,000 steps, the output will be obtained in output_infer file. For each line of output in output_infer file and corresponding ground truth in test_BPE file, obtain the longest common subsequence (LCS). Calculate CRR by summing the LCS length for each line divided by total characters present in test_BPE and that will yield a CRR of 94.02%.
-
In our paper, we have used a mixture of synthetic data as training set obtained from various settings which yielded a CRR of 97.01%. Here, one of the such settings are presented which yields 94.02% CRR. This is representative since all the other models yield CRRs which maintain the original order of them presented in the paper.
-
valid_BPE.src and valid_BPE.trg contain one pair of sentence. This is pre-processing step. Required number (20% of the training set) can be samppled from training files and placed in the validation files for experiments.
CopyNet Paper: Incorporating Copying Mechanism in Sequence-to-Sequence Learning.
This CopyNet implementation is taken from https://github.com/lspvic/CopyNet
CopyNet mechanism is wrapped with an exsiting RNN cell and used as an normal RNN cell.
Official nmt is also modified to enable CopyNet mechanism.
Since in copynet scenarios the target sequence contains words from source sentences, the best choice is to use a shared vocabulary for source vocabulary and target vcabulary. And we also use a parameter generated vocabulary size, namely, the number of target vocabulary excluding words from source sequences (copied words), to indicate that the first N(=generated vocabulary size) words in shared vocabulary are in generate mode and target word indexes larger than N are copied.
In this codebase, vocab_size and gen_vocab_size are variables representing shared vocabulary size and generated vocabulalry size.
Full nmt usages are in nmt.
--copynet argument added to nmt command line to enable copy mechanism.
--share_vocab argument must be set.
--gen_vocab_size argument represents the size of generated vocabulary (excluding copy words from target vocabulary), if is not set, it equals the size of whole vocabulary.
python nmt.nmt.nmt.py --copynet --share_vocab --gen_vocab_size=2345 ...other_nmt_arguments