Latin Intertextuality Classifier (3-class)

This model is a fine-tuned version of ashleygong03/bamman-burns-latin-bert for sequence-pair classification of intertextual links between Jerome (Hieronymus) and other classical authors. It is intended to integrate with the LociSimiles Python package for Latin intertextuality workflows: https://julianschelb.github.io/locisimiles/api/

Labels

Given a pair of passages, the model predicts one of three classes:

The earlier -class-lat-intertext-v1 models solved the binary version of this task (match / no match). This model distinguishes the two positive types instead, so its outputs are not interchangeable with theirs.

Model Description

• Task: 3-class classification for detecting and typing intertextual links between classical Latin authors
• Model type: Sequence Classification
• Base model: ashleygong03/bamman-burns-latin-bert
• Max input tokens: 512
• Training data: one of five cross-validation splits of the Loci Similes benchmark
• Class balance: trained with class-balanced sampling, since real corpora are overwhelmingly negative
• Language: Latin
• License: Apache 2.0

Usage

When using standard tokenization for a sequence-pair classification task, the final input sequence follows the encoder-style pattern with special tokens:

<s> Jerome_phrase </s></s> Candidate_phrase </s>

Here is a complete example:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("julian-schelb/latin-bert-3class-lat-intertext-v1")
model = AutoModelForSequenceClassification.from_pretrained("julian-schelb/latin-bert-3class-lat-intertext-v1")
model.eval()

# Define your sentence pair
sentence1 = "omnia fert aetas, animum quoque; saepe ego longos cantando puerum memini me condere soles."
sentence2 = "saepe ego longos cantando puerum memini me condere soles."

# Tokenize the sentence pair for the model
inputs = tokenizer(
    sentence1,  # Hieronymus
    sentence2,  # Classical author
    add_special_tokens=True,
    truncation=True,
    padding="max_length",
    return_tensors="pt",
)

# Run the model in evaluation mode (no gradient calculation)
with torch.no_grad():
    probs = torch.softmax(model(**inputs).logits, dim=-1)[0]

# probs is indexed by class id: 0 = no_match, 1 = cit, 2 = cf
for idx, p in enumerate(probs):
    print(f"{model.config.id2label[idx]}: {p:.4f}")

Decision thresholds

Argmax over the three classes is a reasonable default, but the experiments this model comes from apply a per-class threshold instead, tuned one-vs-rest on the training split, falling back to no_match. For this checkpoint:

THRESHOLDS = {"cit": 0.4, "cf": 0.81}

# A positive class fires when its probability clears its own threshold.
# If both fire, the higher probability wins (ties go to `cit`).
def predict(probs):
    fired = {
        name: probs[model.config.label2id[name]]
        for name, threshold in THRESHOLDS.items()
        if probs[model.config.label2id[name]] >= threshold
    }
    return max(fired, key=fired.get) if fired else "no_match"

Real corpora are overwhelmingly negative, so thresholding trades recall for a large reduction in false positives. Note also that cf (loose thematic echo) is a much harder class than cit, since it carries no reliable lexical signal.

Citation

@misc{schelb2026locisimilesbenchmarkextracting,
      title={Loci Similes: A Benchmark for Extracting Intertextualities in Latin Literature},
      author={Julian Schelb and Michael Wittweiler and Marie Revellio and Barbara Feichtinger and Andreas Spitz},
      year={2026},
      eprint={2601.07533},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2601.07533},
}