punct_restore_fr

This model is a fine-tuned version of camembert-base on a raw, French opensubtitles dataset. It achieves the following results on the evaluation set:

Loss: 0.0301
Precision: 0.9601
Recall: 0.9527
F1: 0.9564
Accuracy: 0.9915

Model description

Classifies tokens based on beginning of French sentences (B-SENT) and everything else (O).

Intended uses & limitations

This model aims to help punctuation restoration on French YouTube auto-generated subtitles. In doing so, one can measure more in a corpus such as words per sentence, grammar structures per sentence, etc.

Training and evaluation data

1 million Open Subtitles (French) sentences. 80%/10%/10% training/validation/test split.

The sentences:

were lower-cased
had end punctuation (.?!) removed
were of length between 7 and 70 words
had beginning word of sentence tagged with B-SENT.
- All other words marked with O.

Token/tag pairs batched together in groups of 64. This helps show variety of positions for B-SENT and O tags. This also keeps training examples from just being one sentence. Otherwise, this leads to having the first word and only the first word in a sequence being labeled B-SENT.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 16
eval_batch_size: 8
seed: 1
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3

Training results

Framework versions

Transformers 4.8.1
Pytorch 1.9.0+cu102
Datasets 1.8.0
Tokenizers 0.10.3