sloberta-frenk-hate

Text classification model based on EMBEDDIA/sloberta and fine-tuned on the FRANK 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 models and with the same dataset. 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

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": 14,
        "learning_rate": 1e-5,
        "train_batch_size": 21,
        }

model = ClassificationModel(
    "bert", "5roop/bcms-bertic-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])