Automatic Classifier of Lombard Orthography Variants

This repository contains tools for automatically classifying Lombard language text into different orthographic variants. The classifier supports both traditional machine learning (sklearn) and deep learning (PyTorch) approaches.

Supported Orthographic Variants

The classifier identifies the following Lombard orthographic variants:

• MILCLASS - Classical Milanese orthography
• LOCC - Unified Western Lombard
• LORUNIF - Unified Eastern Lombard
• SL - Scriver Lombard
• NOL - New Lombard Orthography
• CRES - Cremonese orthography
• BREMOD - Modern Brescian orthography
• BERGDUC - Bergamasque orthography

Repository Structure

ortho_classifier/
├── README.md                    # This file
├── data/
│   ├── gold/                    # Labeled training data
│   │   ├── train.jsonl         # 8,950 training examples
│   │   ├── valid.jsonl         # 1,118 validation examples
│   │   └── test.jsonl          # 1,118 test examples
│   └── no-tag/                  # Automatically tagged data
│       └── no-tag.jsonl        # 94,520 examples
├── models/                      # Pre-trained models
└── scripts/
    ├── classifiers.py          # Core classifier class definitions
    ├── train_classifier.py     # Training script
    ├── apply_classifier.py     # Classification script
    └── usage.md                # Reference

Dataset Statistics

Gold Standard Dataset

Split	Examples
Training	8,950
Validation	1,118
Test	1,118
Total	11,186

Tag Distribution (Training Set)

Variant	Count	Percentage
MILCLASS	3,606	40.3%
LOCC	2,907	32.5%
LORUNIF	1,901	21.2%
SL	174	1.9%
NOL	109	1.2%
CRES	98	1.1%
BREMOD	94	1.1%
BERGDUC	59	0.7%
LSI	2	0.0%
Total	8,950	100%

Performance

Model	MILCLASS	LOCC	LORUNIF	SL	NOL	CRES	BREMOD	BERGDUC	Overall	Avg Class
Log. byte	97.31	92.89	89.08	100.0	75.00	94.74	53.85	83.33	93.38	85.78
Log. byte+char+word	97.31	94.74	94.76	100.0	75.00	100.0	46.15	66.67	95.08	84.33
Log. char	95.07	93.42	84.28	100.0	75.00	94.74	53.85	83.33	91.67	84.96
Log. word	94.62	94.21	88.65	93.75	37.50	94.74	69.23	66.67	92.39	79.92
SVM byte	97.76	94.21	99.13	87.50	75.00	94.74	30.77	50.00	95.43	78.64
SVM byte+char+word	97.76	96.58	99.13	93.75	50.00	94.74	23.08	50.00	96.06	75.63
SVM char	98.21	94.74	98.25	93.75	50.00	94.74	30.77	50.00	95.52	76.31
SVM word	95.52	95.53	95.20	93.75	12.50	89.47	30.77	50.00	93.73	70.34
NB byte	98.88	91.05	93.45	0.00	0.00	0.00	0.00	0.00	89.62	35.42
NB byte+char+word	98.21	93.42	96.07	0.00	0.00	0.00	0.00	0.00	90.69	35.96
NB char	98.65	91.05	94.32	0.00	0.00	0.00	0.00	0.00	89.70	35.50
NB word	97.09	92.37	94.76	12.50	0.00	15.79	0.00	0.00	90.06	39.06
RF byte	98.43	92.37	96.94	68.75	0.00	68.42	0.00	0.00	92.75	53.11
RF byte+char+word	98.65	92.11	95.63	75.00	0.00	68.42	7.69	0.00	92.66	54.69
RF char	98.65	93.16	95.20	50.00	0.00	68.42	7.69	0.00	92.57	51.64
RF word	97.31	87.89	93.89	68.75	0.00	63.16	15.38	0.00	90.24	53.30
CNN byte	97.98	95.53	97.38	75.00	12.50	78.95	7.69	16.67	94.27	60.21
CNN char	98.21	94.47	98.69	87.50	12.50	63.16	7.69	16.67	94.18	59.86
Deep CNN byte	83.86	94.74	92.14	75.00	12.50	73.68	15.38	0.00	87.20	55.91
Deep CNN char	97.09	92.37	96.94	0.00	12.50	57.89	7.69	0.00	91.23	45.56
LSTM byte	97.53	91.05	95.63	0.00	0.00	63.16	0.00	0.00	90.60	43.42
LSTM char	98.21	91.05	95.63	0.00	0.00	68.42	0.00	16.67	91.05	46.25
Transformer byte	96.64	85.00	94.76	68.75	0.00	5.26	0.00	0.00	88.00	43.80
Transformer char	96.41	88.95	93.89	0.00	0.00	31.58	0.00	0.00	88.54	38.85
Best accuracy	98.88	96.58	99.13	100.0	75.00	100.0	69.23	83.33	96.06	85.78
Worst accuracy	83.86	85.00	84.28	0.00	0.00	0.00	0.00	0.00	87.20	35.42
Accuracy range	15.02	11.58	14.85	100.0	75.00	100.0	69.23	83.33	8.86	50.36

Upcoming Features

• Multilabel Classification

• Tagset Expansion (Non-Lombard; Non-Language; Macro-Diaclectal only)

• Dataset Expansion (Non-Lombard; Non-Language)

• Othography Converter

Citation

If you use these classifiers or the data, please cite:

Edoardo Signoroni and Pavel Rychlý, LombardoGraphia: Automatic Classification of Lombard Orthography Variants, Accepted at LREC 2026

@misc{signoroni2026lombardographiaautomaticclassificationlombard,
      title={LombardoGraphia: Automatic Classification of Lombard Orthography Variants}, 
      author={Edoardo Signoroni and Pavel Rychlý},
      year={2026},
      eprint={2603.28418},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.28418}, 
}

Contacts

For questions or issues, please contact e.signoroni@mail.muni.cz.

License

The dataset is licensed under CC-BY-SA 4.0 The software is licesed under GPL-3.0

Acknowledgments

This work has been supported by the Ministry of Education, Youth and Sports of the Czech Republic, Project No. LM2023062 LINDAT/CLARIAH-CZ. We wish to thank the volunteers and contributors of Lombard Wikipedia for the creation of the data and the useful discussions and information.