ocr-corrector-flan-t5-base

OCR Corrector — flan-t5-base

Fine-tuned google/flan-t5-base (250M parameters) that corrects OCR errors in electrical engineering academic text.

This is the model released with the paper "Automatic OCR Error Correction in Electrical Engineering Academic Papers Using Sequence-to-Sequence Models" (IEEE LA-CCI 2026).

• Code and dataset: https://github.com/Cronosteak/ocr-corrector-papers
• License: MIT

Usage

The model expects the prefix correct OCR: on every input.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "Cronosteak/ocr-corrector-flan-t5-base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

text = "correct OCR: Thc powcr systcm opcrates at 60 Hz"
inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True)
outputs = model.generate(**inputs, max_new_tokens=256, num_beams=4)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training data

6,440 synthetic OCR–ground-truth pairs built without manual annotation: open-access electrical engineering abstracts were collected from the OpenAlex API, then corrupted with synthetic OCR noise (character confusions, spurious spaces/hyphens, deleted spaces, duplicated characters) at four rates: r ∈ {0.02, 0.04, 0.06, 0.08}.

Splits are per-document (80/10/10) to avoid leakage between train and test.

Results

Held-out test set of 676 pairs:

A 37.5% relative CER reduction and an 80.9% relative WER reduction over the uncorrected input. Per-noise-rate metrics are in eval_results.json.

Ablations showed that multi-level noise diversity is an important contributor to generalization: restricting training to r=0.04 alone drops exact match from 44.23% to 35.21% (that variant also trains on fewer pairs, so diversity and dataset size are not fully disentangled).

Training procedure

Limitations

• Trained on synthetic OCR noise, not on real Tesseract output. Real-world error distributions may differ.
• Domain is electrical engineering and adjacent fields (computer science, engineering) in English; performance on other domains or languages is untested.
• Residual errors concentrate on lowercase common words at the highest noise rate (r=0.08), where corrupted word shapes can be decoded into plausible but wrong alternatives. Because it is a generative model, it can rewrite text that was already correct — review output before using it in a critical pipeline.
• Inputs are truncated at 256 tokens; long passages must be chunked.

Citation

@inproceedings{espinosa2026ocr,
  title     = {Automatic OCR Error Correction in Electrical Engineering
               Academic Papers Using Sequence-to-Sequence Models},
  author    = {Espinosa Ortiz, Breinner Farid and
               Forero Mendoza, Leonardo Alfredo and
               Pacheco, Marco Aur{\'e}lio C. and
               Batista, Evelyn C. S.},
  booktitle = {2026 IEEE Latin American Conference on Computational
               Intelligence (LA-CCI)},
  year      = {2026},
  publisher = {IEEE}
}