README.md · classla/sloberta-frenk-hate at refs/pr/1

metadata

language: sl
license: cc-by-sa-4.0
tags:
  - text-classification
  - hate-speech
widget:
  - text: Silva, ti si grda in neprijazna

Text classification model based on EMBEDDIA/sloberta and fine-tuned on the FRENK dataset comprising of LGBT and migrant hatespeech. Only the slovenian subset of the data was used for fine-tuning and the dataset has been relabeled for binary classification (offensive or acceptable).

Fine-tuning hyperparameters

Fine-tuning was performed with simpletransformers. Beforehand a brief hyperparameter optimisation was performed and the presumed optimal hyperparameters are:

model_args = {
        "num_train_epochs": 14,
        "learning_rate": 1e-5,
        "train_batch_size": 21,
        }

Performance

The same pipeline was run with two other transformer models and fasttext for comparison. Accuracy and macro F1 score were recorded for each of the 6 fine-tuning sessions and post festum analyzed.

model	average accuracy	average macro F1
sloberta-frenk-hate	0.7785	0.7764
EMBEDDIA/crosloengual-bert	0.7616	0.7585
xlm-roberta-base	0.686	0.6827
fasttext	0.709	0.701

From recorded accuracies and macro F1 scores p-values were also calculated:

Comparison with crosloengual-bert:

test	accuracy p-value	macro F1 p-value
Wilcoxon	0.00781	0.00781
Mann Whithney U test	0.00163	0.00108
Student t-test	0.000101	3.95e-05

Comparison with xlm-roberta-base:

test	accuracy p-value	macro F1 p-value
Wilcoxon	0.00781	0.00781
Mann Whithney U test	0.00108	0.00108
Student t-test	9.46e-11	6.94e-11

Use examples

from simpletransformers.classification import ClassificationModel
model_args = {
        "num_train_epochs": 6,
        "learning_rate": 3e-6,
        "train_batch_size": 69}

model = ClassificationModel(
    "camembert", "5roop/sloberta-frenk-hate", use_cuda=True,
    args=model_args
    
)

predictions, logit_output = model.predict(["Silva, ti si grda in neprijazna", "Naša hiša ima dimnik"])
predictions
### Output:
### array([1, 0])

Citation

If you use the model, please cite the following paper on which the original model is based:

@article{DBLP:journals/corr/abs-1907-11692,
  author    = {Yinhan Liu and
               Myle Ott and
               Naman Goyal and
               Jingfei Du and
               Mandar Joshi and
               Danqi Chen and
               Omer Levy and
               Mike Lewis and
               Luke Zettlemoyer and
               Veselin Stoyanov},
  title     = {RoBERTa: {A} Robustly Optimized {BERT} Pretraining Approach},
  journal   = {CoRR},
  volume    = {abs/1907.11692},
  year      = {2019},
  url       = {http://arxiv.org/abs/1907.11692},
  archivePrefix = {arXiv},
  eprint    = {1907.11692},
  timestamp = {Thu, 01 Aug 2019 08:59:33 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1907-11692.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

and the dataset used for fine-tuning:

@misc{ljubešić2019frenk,
      title={The FRENK Datasets of Socially Unacceptable Discourse in Slovene and English}, 
      author={Nikola Ljubešić and Darja Fišer and Tomaž Erjavec},
      year={2019},
      eprint={1906.02045},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/1906.02045}
}