mhvdr/neolatiner-model

NeoLatiNER

NeoLatiNER is an XLM-RoBERTa-large model fine-tuned for named-entity recognition in OCR-derived Early Modern Latin texts about natural philosophy. It predicts persons (PRS), groups (GRP), and geographical locations (GEO) using IOB2 labels.

Person-name recognition was the primary research target. The GRP and GEO classes were retained for reuse but were not evaluated in the same depth.

Usage

Install transformers and PyTorch, then run:

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="mhvdr/neolatiner-model",
    aggregation_strategy="simple",
)

entities = ner("Cartesius in Gallia natus est.")
for entity in entities:
    print(entity["word"], entity["entity_group"], entity["score"])

The underlying labels are O, B-PRS, I-PRS, B-GRP, I-GRP, B-GEO, and I-GEO. With aggregation_strategy="simple", the pipeline combines adjacent IOB2 predictions and returns entity groups such as PRS, GRP, and GEO.

For lower-level control:

from transformers import AutoModelForTokenClassification, AutoTokenizer

repo_id = "mhvdr/neolatiner-model"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForTokenClassification.from_pretrained(repo_id)

Long documents must be split into passages that fit XLM-R's input limit. Preserve enough surrounding context for ambiguous names and retain confidence scores for review.

Evaluation

The released checkpoint was evaluated on the held-out test split of mhvdr/neolatiner-dataset using strict IOB2 entity matching. The test split contains 1,460 records and 639 annotated entities.

Entity	Precision	Recall	F1	Support
GEO	0.6533	0.7784	0.7104	167
GRP	0.4937	0.4588	0.4756	85
PRS	0.7541	0.7209	0.7371	387
Micro average	0.6914	0.7011	0.6962	639

Training data

The model builds on XLM-R and Latin NER material from the Herodotos Project and the Corpus Burgundiae Medii Aevi, followed by fine-tuning on two rounds of manually annotated, in-domain Early Modern Latin excerpts. The final in-domain dataset is available at mhvdr/neolatiner-dataset.

Limitations

The target texts contain OCR errors, spelling variation, inflection, abbreviated names, and inconsistent capitalization. Predictions require domain-aware review and post-processing. The annotation sampling strategy emphasized excerpts likely to contain entities and low-confidence predictions, so the data is not a representative random sample of Early Modern Latin. The model was designed for research use, not high-stakes or fully automated decisions.

Citation

Please cite the associated NeoLatiNER publication. Final bibliographic metadata and DOI will be added at publication.