This model was trained to do a token classification of online comments to determine whether the token contains a vulgarity or not (swear words, insult, ...).

This model is based don GBERT from deepset (https://huggingface.co/deepset/gbert-base) which was mainly trained on wikipedia. To this model we added a freshly initialized token classification header, which had to be trained on our labeled data.

For the training a dataset of 4500 comments german comments label on toxicity was used. This dataset is not publicly available, but can be requested form TU-Wien (https://doi.org/10.5281/zenodo.10996203).

The dataset contains additional tags, which are

We decided to use the Vulgarity tag to mark the words which are considered to be an insult. 1306 Comments contained a Vulgarity, but 452 did not belong to a toxic considered comment. These comments are split into 1484 number of sentences containing vulgarities. Data prepared to have sentence by sentence data set tagged with vulgarity token. [‘O’,’Vul’] (1484 sentences). A 80/10/10 train/validation/test split was used.

Out of 4500 comments 1306 contained a vulgarity tags. In order to identify an optimally performing model for classifying toxic speech, a large set of models was trained and evaluated. Hyperparameter:

The best model used 2 frozen layers and was evaluated on the training set with the following metrics:

Here is how to use this model to get the features of a given text in PyTorch:

[{'entity_group': 'Vul', 'score': 0.9548946, 'word': 'Bescheuert - Gen', 'start': 26, 'end': 40}]