Cronosteak/ocr-corrector-flan-t5-base

OCR Corrector Papers

Automatic OCR error correction for electrical engineering academic papers, using sequence-to-sequence models fine-tuned on data built automatically from OpenAlex.

This is the official code release for the paper:

Automatic OCR Error Correction in Electrical Engineering Academic Papers Using Sequence-to-Sequence Models Breinner Farid Espinosa Ortiz, Leonardo Alfredo Forero Mendoza, Marco Aurélio C. Pacheco, Evelyn C. S. Batista 2026 IEEE Latin American Conference on Computational Intelligence (LA-CCI)

Model weights: Cronosteak/ocr-corrector-flan-t5-base on the Hugging Face Hub.

---

Results

Fine-tuned google/flan-t5-base (250M parameters) on 6,440 synthetic OCR–ground-truth pairs, evaluated on a held-out 676-pair test set.

Method	CER (%) ↓	WER (%) ↓	BLEU ↑	EM (%) ↑
Noisy input (no correction)	3.28	21.99	63.95	0.00
Spellchecker baseline	3.35	15.87	73.80	5.47
flan-t5-base (ours)	2.05	4.21	91.94	44.23

That is a 37.5% relative CER reduction and an 80.9% relative WER reduction over the uncorrected input.

Note that the dictionary-based spellchecker fails to improve CER over doing nothing: it introduces spurious character-level edits on domain terms and proper nouns that the noisy input had preserved correctly.

Ablation study

Variant	CER (%)	WER (%)	BLEU	EM (%)
Full model (ours)	2.05	4.21	91.94	44.23
w/o warmup	2.09	4.41	91.74	40.53
w/o weight decay	1.92	4.15	92.19	41.72
Single noise (r=0.04 only)	2.40	5.33	89.89	35.21

Training across multiple noise levels is an important contributor to generalization — restricting training to one noise rate drops exact match by 9 percentage points, though that variant also trains on fewer pairs (1,306 vs 5,178), so diversity and dataset size are not fully disentangled.

All raw numbers above are reproducible from the JSON files in results/.

---

Repository structure

ocr-corrector-papers/
├── src/
│   ├── pipeline/           # Data pipeline
│   │   ├── fetch_openalex.py    # Query OpenAlex for open-access papers
│   │   ├── download_pdfs.py     # Download PDFs
│   │   ├── ocr_extract.py       # Tesseract OCR extraction
│   │   ├── align_text.py        # Align noisy OCR with ground truth
│   │   └── build_dataset.py     # Synthetic noise injection + train/val/test splits
│   ├── model/
│   │   ├── train.py             # Fine-tuning loop
│   │   ├── postprocess.py       # Full evaluation: metrics, baselines, plots
│   │   ├── evaluate.py          # Lightweight metrics-only entry point
│   │   ├── baselines.py         # Spellchecker baseline (used by postprocess)
│   │   ├── error_analysis.py    # Token-level error categorization
│   │   ├── compare_ablations.py # Ablation comparison tables
│   │   └── predict.py           # Inference on new text
│   ├── utils/                   # Metrics, logging, text cleaning, timing
│   └── api/                     # Optional FastAPI REST service
├── configs/                # Training and ablation YAML configs
├── data/pairs/             # Pre-generated 6,440-pair dataset (versioned)
├── results/                # Evaluation artifacts for every trained variant
├── slurm_scripts/          # SLURM job scripts used to produce the results
├── notebooks/              # Exploration and results notebooks
├── tests/                  # Unit tests
└── docker/                 # Dockerfile for reproducible runs

Heavy artifacts (raw PDFs, OCR text, model checkpoints) are not versioned. The dataset needed to reproduce every number in the paper is — see data/pairs/.

---

Installation

pip install -r requirements.txt
cp .env.example .env    # then edit the paths/credentials

System dependencies for the OCR stage: tesseract-ocr and poppler-utils.

sudo apt install tesseract-ocr poppler-utils

A Dockerfile is provided if you prefer a pinned environment.

---

Reproducing the paper

The dataset is already committed, so you can skip straight to training:

# 1. Train the model (config: configs/train_config.yaml)
python -m src.model.train

# 2. Full evaluation: CER/WER/BLEU/EM, per-noise-rate breakdown, both
#    baselines (noisy input + spellchecker), plots and qualitative examples.
#    Writes eval_results.json, predictions.json and the paper figures.
python -m src.model.postprocess --model models/ocr-corrector \
                                --data data/pairs/synthetic_test.json

# 3. Token-level error categorization (Figure 6 in the paper)
python -m src.model.error_analysis --predictions models/ocr-corrector/predictions.json

src/model/evaluate.py is a lighter standalone entry point if you only want the headline metrics printed as JSON:

python -m src.model.evaluate --model models/ocr-corrector

Ablations

slurm_scripts/run_ablations.sh submits the three variants via SLURM (sbatch), which is how they were run for the paper. Edit REPO_PATH in the .srm files to point at your checkout first.

./slurm_scripts/run_ablations.sh          # SLURM clusters
python -m src.model.compare_ablations     # -> models/ablation_results.md

Without SLURM, train each variant directly and then compare:

python -m src.model.train --config configs/ablation_no_warmup.yaml
python -m src.model.train --config configs/ablation_no_wd.yaml
python -m src.model.train --config configs/ablation_single_noise.yaml
python -m src.model.compare_ablations

To rebuild the dataset from scratch instead — note this re-downloads and re-OCRs several hundred PDFs and takes roughly 3 hours, dominated by the OCR stage:

python -m src.pipeline.fetch_openalex    # OpenAlex query -> data/works.json
python -m src.pipeline.download_pdfs     # -> data/raw/*.pdf
python -m src.pipeline.ocr_extract       # -> data/ocr/*.txt
python -m src.pipeline.build_dataset     # -> data/pairs/*.json

Stage timings from our run are recorded in data/pipeline_stats.json.

---

Usage

Correct a piece of OCR text:

python -m src.model.predict --input "Thc powcr systcm opcrates at 60 Hz"

Or serve it over HTTP:

uvicorn src.api.server:app --reload

---

Configuration

File	Purpose
configs/openalex_query.yaml	OpenAlex search filters (field, year range, OA status)
configs/train_config.yaml	Base model, hyperparameters, splits
configs/ablation_no_warmup.yaml	Ablation: warmup steps set to 0
configs/ablation_no_wd.yaml	Ablation: weight decay set to 0
configs/ablation_single_noise.yaml	Ablation: single noise rate (r=0.04)

---

Tests

pytest tests/

---

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}
}

License

Released under the MIT License.