Towards Prompt Engineering and Large Language Models for Post-OCR correction in handwritten texts

This experiment aims to perform post-OCR spelling corrections on handwritten text recognition competition datasets.

To evaluate the performance of LLMs in this task, prompts with complete sentences/sentences are created for the models to correct.

Create Conda environment:

conda create --name=llm_spelling_corrector python=3.10 -y
conda activate llm_spelling_corrector
pip install -r requirements.txt

Run experiments:

# Remember create .env file from .env.example
python3 predict.py \
  --datasets bressay rimes iam \
  --ocr_predictions bluche flor puigcerver \
  --output_folder corrections/gpt_4o_mini \
  --provider openai \ # gemini or ollama
  --prompt_type 1 \ # 1 for our prompt, 2 for previous work's prompt
  --llm gpt-4o-mini # gemini-2.0-flash-lite, phi4 etc.

Calculate metrics:

python3 metrics.py \
    --datasets bressay rimes iam \
    --ocr_predictions bluche flor puigcerver \
    --output_corrections corrections/gpt_4o_mini

Observations:

• To support other datasets modify the get_sentences function to identify the lines and create sentences to prompt (this is necessary to add more context to the llm In the end the lines are split again).
• The get_sentences function expects the dataset to be sorted to get the lines of the paragraphs in the correct order.
• To support other providers and models modify model.py file.

Citation:

@inproceedings{stil,
 author = {Sávio Araújo and Byron Bezerra and Arthur Sousa Neto},
 title = { Towards Prompt Engineering and Large Language Models for Post-OCR correction in handwritten texts},
 booktitle = {Anais do XVI Simpósio Brasileiro de Tecnologia da Informação e da Linguagem Humana},
 location = {Fortaleza/CE},
 year = {2025},
 keywords = {},
 issn = {0000-0000},
 pages = {572--579},
 publisher = {SBC},
 address = {Porto Alegre, RS, Brasil},
 doi = {10.5753/stil.2025.37859},
 url = {https://sol.sbc.org.br/index.php/stil/article/view/37859}
}