Transformer Based Punctuation Restoration Models for Turkish

You can find the BERT model used in the paper Transformer Based Punctuation Restoration for Turkish. Aim of this work is correctly place pre-decided punctuation marks in a given text. We present three pre-trained transformer models to predict period(.), comma(,) and question(?) marks for the Turkish language.

Usage

Inference

Recommended usage is via HuggingFace. You can run an inference using the pre-trained BERT model with the following code:

from transformers import pipeline

pipe = pipeline(task="token-classification", model="uygarkurt/bert-restore-punctuation-turkish")

sample_text = "Türkiye toprakları üzerindeki ilk yerleşmeler Yontma Taş Devri'nde başlar Doğu Trakya'da Traklar olmak üzere Hititler Frigler Lidyalılar ve Dor istilası sonucu Yunanistan'dan kaçan Akalar tarafından kurulan İyon medeniyeti gibi çeşitli eski Anadolu medeniyetlerinin ardından Makedonya kralı Büyük İskender'in egemenliğiyle ve fetihleriyle birlikte Helenistik Dönem başladı"

out = pipe(sample_text)

To use a different pre-trained model you can just replace the model argument with one of the other available models we provided.

Data

Dataset is provided in data/ directory as train, validation and test splits.

Dataset can be summarized as below:

Available Models

We experimented with BERT, ELECTRA and ConvBERT. Pre-trained models can be accessed via Huggingface.

BERT: https://huggingface.co/uygarkurt/bert-restore-punctuation-turkish
ELECTRA: https://huggingface.co/uygarkurt/electra-restore-punctuation-turkish
ConvBERT: https://huggingface.co/uygarkurt/convbert-restore-punctuation-turkish

Results

Precision and Recall and F1 scores for each model and punctuation mark are summarized below.

Citation

@INPROCEEDINGS{10286690,
    author={Kurt, Uygar and Çayır, Aykut},
    booktitle={2023 8th International Conference on Computer Science and Engineering (UBMK)}, 
    title={Transformer Based Punctuation Restoration for Turkish}, 
    year={2023},
    volume={},
    number={},
    pages={169-174},
    doi={10.1109/UBMK59864.2023.10286690}
}