XLM-RoBERTa (base) language-detection model (modern and medieval)

This model is a fine-tuned version of xlm-roberta-base on the monasterium.net dataset.

Model description

On the top of this XLM-RoBERTa transformer model is a classification head. Please refer this model together with to the XLM-RoBERTa (base-sized model) card or the paper Unsupervised Cross-lingual Representation Learning at Scale by Conneau et al. for additional information.

Intended uses & limitations

You can directly use this model as a language detector, i.e. for sequence classification tasks. Currently, it supports the following 41 languages, modern and medieval:

Modern: Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Irish (ga), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru), Turkish (tr), Basque (eu), Catalan (ca), Albanian (sq), Serbian (se), Ukrainian (uk), Norwegian (no), Arabic (ar), Chinese (zh), Hebrew (he)

Medieval: Middle High German (mhd), Latin (la), Middle Low German (gml), Old French (fro), Old Church Slavonic (chu), Early New High German (fnhd), Ancient and Medieval Greek (grc)

Training and evaluation data

The model was fine-tuned using the Monasterium and Wikipedia datasets, which consist of text sequences in 41 languages. The training set contains 80k samples, while the validation and test sets contain 16k. The average accuracy on the test set is 99.59% (this matches the average macro/weighted F1-score, the test set being perfectly balanced).

Training procedure

Fine-tuning was done via the Trainer API with WeightedLossTrainer.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 20
eval_batch_size: 20
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3

mixed_precision_training: Native AMP

Training results

Training Loss	Validation Loss	F1
0.000300	0.048985	0.991585
0.000100	0.033340	0.994663
0.000000	0.032938	0.995979

Using example

#Install packages
!pip install transformers --quiet

#Import libraries
import torch
from transformers import pipeline

#Define pipeline
classificator = pipeline("text-classification", model="ERCDiDip/langdetect")

#Use pipeline
classificator("clemens etc dilecto filio scolastico ecclesie wetflari ensi treveren dioc salutem etc significarunt nobis dilecti filii commendator et fratres hospitalis beate marie theotonicorum")

Updates

25th November 2022: Adding Ancient and Medieval Greek (grc)

Framework versions

Transformers 4.24.0
Pytorch 1.13.0
Datasets 2.6.1
Tokenizers 0.13.3

Citation

Please cite the following papers when using this model.

@misc{ercdidip2022,
  title={langdetect (Revision 0215f72)},
  author={Kovács, Tamás, Atzenhofer-Baumgartner, Florian, Aoun, Sandy, Nicolaou, Anguelos, Luger, Daniel, Decker, Franziska, Lamminger, Florian and Vogeler, Georg},
  year         = { 2022 },
  url          = { https://huggingface.co/ERCDiDip/40_langdetect_v01 },
  doi          = { 10.57967/hf/0135 },
  publisher    = { Hugging Face }
}

This model is part of the From Digital to Distant Diplomatics (DiDip) ERC project funded by the European Research Council.