Romanian DistilBERT

This repository contains the a Romanian cased version of DistilBERT (named DistilMulti-BERT-base-ro in the paper) that was obtained by distilling an ensemble of two teacher models: dumitrescustefan/bert-base-romanian-cased-v1 and readerbench/RoBERT-base.

The model was introduced in this paper. The adjacent code can be found here.

Usage

from transformers import AutoTokenizer, AutoModel

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained("racai/distilbert-multi-base-romanian-cased")
model = AutoModel.from_pretrained("racai/distilbert-multi-base-romanian-cased")

# tokenize a test sentence
input_ids = tokenizer.encode("Aceasta este o propoziție de test.", add_special_tokens=True, return_tensors="pt")

# run the tokens trough the model
outputs = model(input_ids)

print(outputs)

Model Size

The model is 35% smaller than bert-base-romanian-cased-v1 and 30% smaller than RoBERT-base.

Evaluation

We evaluated the model in comparison with its two teachers on 5 Romanian tasks:

• UPOS: Universal Part of Speech (F1-macro)
• XPOS: Extended Part of Speech (F1-macro)
• NER: Named Entity Recognition (F1-macro)
• SAPN: Sentiment Anlaysis - Positive vs Negative (Accuracy)
• SAR: Sentiment Analysis - Rating (F1-macro)
• DI: Dialect identification (F1-macro)
• STS: Semantic Textual Similarity (Pearson)

BibTeX entry and citation info

@article{avram2021distilling,
  title={Distilling the Knowledge of Romanian BERTs Using Multiple Teachers},
  author={Andrei-Marius Avram and Darius Catrina and Dumitru-Clementin Cercel and Mihai Dascălu and Traian Rebedea and Vasile Păiş and Dan Tufiş},
  journal={ArXiv},
  year={2021},
  volume={abs/2112.12650}
}