Introduction

This project aims to evaluate the performance of an OCR system applied to historical French documents from the 18th century (Dictionnaire de Trévoux Tomme 5).

The main challenges include:

  • Old orthography (e.g “portoient”, “appelloit”).

  • Non-standard punctuation and spelling.

  • Degraded scan quality.

  • Historical vocabulary.

OCR Pipeline Description

We propose a pipeline combining:

Alternative text for the image
1. Mistral OCR for raw text extraction:

  • Input: scanned historical document images.

  • Output: raw OCR transcription.

  • Preserves basic layout (lines, spacing when available).

  • Handles complex structures (columns, footnotes, dictionary entries).

2. Cleaning and normalization:

  • Removes OCR noise (artifacts, markdown, broken characters).

  • Standardizes whitespace and line breaks.

  • Unicode normalization and special character handling (e.g., ſ → s, æ → ae).

  • Output: cleaned text ready for correction.

3. LLM-based correction using Mistral Large:

  • Input: cleaned OCR text.

  • Corrects OCR errors while preserving historical spelling.

  • Fixes misrecognitions and word splits.

  • Preserves archaic French forms (portoient, appelloit, …).

  • Output: corrected transcription.

4. Evaluation using CER and WER metrics:

  • Reference: gold standard text.

  • CER: character-level error measurement.

  • WER: word-level error measurement.

  • Comparison: OCR vs gold, corrected vs gold.