In 2017 and 2019 a competion on Post-OCR Text Correction was organized. This repository contains the 'working' notebooks for reproducing the best results results of the competition and possibly improving them. The code in the notebooks use functionality from the ocrpostcorrection package.
git clone https://github.com/jvdzwaan/ocrpostcorrection.git
cd ocrpostcorrection
pip install -e .This repository contains two sets of notebooks:
localnotebooks to be run locally, e.g., for generating datasetscolabnotebooks to be run on machines with a GPU, e.g., for training neural networks
ocrpostcorrection
├── LICENSE
├── README.md
├── colab <- Notebooks to be run on GPU
│ ├── icdar-task1-hf-evaluation.ipynb <- Evaluate Huggingface BERT model for error detection
│ ├── icdar-task1-hf-train.ipynb <- Train Huggingface BERT model for error detection
│ ├── icdar-task2-seq2seq-evaluation.ipynb <- Evaluate performance of error correction model
│ └── icdar-task2-train-seq2seq.ipynb <- Train error correction model
└── local <- Notebooks to be run locally
├── data <- Data generated and/or used by local notebooks
├── evalTool_ICDAR2017.py <- ICDAR competition evaluation script
├── icdar-create-hf-dataset.ipynb <- Create Huggingface dataset from the icdar data
├── icdar-task2-create-dataset.ipynb <- Create error correction dataset from the icdar data
├── icdar-task2-results-analysis.ipynb <- Preliminary analysis of error correction results
└── perfect_task1+2_output_analysis.ipynb <- Analysis of evalTool script for measuring performance