bert-large-finnish-cased-v1 for toxicity detection
This is the bert-base-finnish-cased-v1 model, fine-tuned using the Finnish jigsaw_toxicity_pred_fi dataset. The model is trained to predict probabilities for 6 different toxicity labels introduced in the dataset card.
Overview
Language model: bert-base-finnish-v1
Language: Finnish
Downstream-task: Multi-label toxicity detection (multi-label text classification)
Training data: jigsaw_toxicity_pred_fi
Eval data: jigsaw_toxicity_pred_fi
Citing
If you use this model please cite us using the following bibtex.
@inproceedings{eskelinen-etal-2023-toxicity,
title = "Toxicity Detection in {F}innish Using Machine Translation",
author = "Eskelinen, Anni and
Silvala, Laura and
Ginter, Filip and
Pyysalo, Sampo and
Laippala, Veronika",
booktitle = "Proceedings of the 24th Nordic Conference on Computational Linguistics (NoDaLiDa)",
month = may,
year = "2023",
address = "T{\'o}rshavn, Faroe Islands",
publisher = "University of Tartu Library",
url = "https://aclanthology.org/2023.nodalida-1.68",
pages = "685--697",
abstract = "Due to the popularity of social media platforms and the sheer amount of user-generated content online, the automatic detection of toxic language has become crucial in the creation of a friendly and safe digital space. Previous work has been mostly focusing on English leaving many lower-resource languages behind. In this paper, we present novel resources for toxicity detection in Finnish by introducing two new datasets, a machine translated toxicity dataset for Finnish based on the widely used English Jigsaw dataset and a smaller test set of Suomi24 discussion forum comments originally written in Finnish and manually annotated following the definitions of the labels that were used to annotate the Jigsaw dataset. We show that machine translating the training data to Finnish provides better toxicity detection results than using the original English training data and zero-shot cross-lingual transfer with XLM-R, even with our newly annotated dataset from Suomi24.",
}
Usage
the model can be used through a huggingface pipeline:
model = transformers.AutoModelForSequenceClassification.from_pretrained("TurkuNLP/bert-large-finnish-cased-toxicity")
tokenizer = transformers.AutoTokenizer.from_pretrained("TurkuNLP/bert-large-finnish-cased-v1")
pipe = transformers.pipeline(task="text-classification", model=model, tokenizer=tokenizer, function_to_apply="sigmoid", top_k=None)
Hyperparameters
batch_size = 12
epochs = 10 (trained for 4)
base_LM_model = "bert-large-finnish-cased-v1"
max_seq_len = 512
learning_rate = 2e-5
Performance
F1-micro = 0.66
F1-macro = 0.57
Precision (micro) = 0.58
Recall (micro) = 0.76