Edit model card

Clone from https://huggingface.co/papluca/xlm-roberta-base-language-detection

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

Model description

This model is an XLM-RoBERTa transformer model with a classification head on top (i.e. a linear layer on top of the pooled output). For additional information please refer to the xlm-roberta-base model card or to the paper Unsupervised Cross-lingual Representation Learning at Scale by Conneau et al.

Intended uses & limitations

You can directly use this model as a language detector, i.e. for sequence classification tasks. Currently, it supports the following 20 languages:

arabic (ar), bulgarian (bg), german (de), modern greek (el), english (en), spanish (es), french (fr), hindi (hi), italian (it), japanese (ja), dutch (nl), polish (pl), portuguese (pt), russian (ru), swahili (sw), thai (th), turkish (tr), urdu (ur), vietnamese (vi), and chinese (zh)

Training and evaluation data

The model was fine-tuned on the Language Identification dataset, which consists of text sequences in 20 languages. The training set contains 70k samples, while the validation and test sets 10k each. The average accuracy on the test set is 99.6% (this matches the average macro/weighted F1-score being the test set perfectly balanced). A more detailed evaluation is provided by the following table.

Language Precision Recall F1-score support
ar 0.998 0.996 0.997 500
bg 0.998 0.964 0.981 500
de 0.998 0.996 0.997 500
el 0.996 1.000 0.998 500
en 1.000 1.000 1.000 500
es 0.967 1.000 0.983 500
fr 1.000 1.000 1.000 500
hi 0.994 0.992 0.993 500
it 1.000 0.992 0.996 500
ja 0.996 0.996 0.996 500
nl 1.000 1.000 1.000 500
pl 1.000 1.000 1.000 500
pt 0.988 1.000 0.994 500
ru 1.000 0.994 0.997 500
sw 1.000 1.000 1.000 500
th 1.000 0.998 0.999 500
tr 0.994 0.992 0.993 500
ur 1.000 1.000 1.000 500
vi 0.992 1.000 0.996 500
zh 1.000 1.000 1.000 500

Benchmarks

As a baseline to compare xlm-roberta-base-language-detection against, we have used the Python langid library. Since it comes pre-trained on 97 languages, we have used its .set_languages() method to constrain the language set to our 20 languages. The average accuracy of langid on the test set is 98.5%. More details are provided by the table below.

Language Precision Recall F1-score support
ar 0.990 0.970 0.980 500
bg 0.998 0.964 0.981 500
de 0.992 0.944 0.967 500
el 1.000 0.998 0.999 500
en 1.000 1.000 1.000 500
es 1.000 0.968 0.984 500
fr 0.996 1.000 0.998 500
hi 0.949 0.976 0.963 500
it 0.990 0.980 0.985 500
ja 0.927 0.988 0.956 500
nl 0.980 1.000 0.990 500
pl 0.986 0.996 0.991 500
pt 0.950 0.996 0.973 500
ru 0.996 0.974 0.985 500
sw 1.000 1.000 1.000 500
th 1.000 0.996 0.998 500
tr 0.990 0.968 0.979 500
ur 0.998 0.996 0.997 500
vi 0.971 0.990 0.980 500
zh 1.000 1.000 1.000 500

Training procedure

Fine-tuning was done via the Trainer API.

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 2e-05
  • train_batch_size: 64
  • eval_batch_size: 128
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • num_epochs: 2
  • mixed_precision_training: Native AMP

Training results

The validation results on the valid split of the Language Identification dataset are summarised here below.

Training Loss Epoch Step Validation Loss Accuracy F1
0.2492 1.0 1094 0.0149 0.9969 0.9969
0.0101 2.0 2188 0.0103 0.9977 0.9977

In short, it achieves the following results on the validation set:

  • Loss: 0.0101
  • Accuracy: 0.9977
  • F1: 0.9977

Framework versions

  • Transformers 4.12.5
  • Pytorch 1.10.0+cu111
  • Datasets 1.15.1
  • Tokenizers 0.10.3
Downloads last month
139
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for eleldar/language-detection

Adapters
1 model