alexandrainst/da-hatespeech-classification-base

Danish BERT for hate speech classification

The BERT HateSpeech model classifies offensive Danish text into 4 categories:

• Særlig opmærksomhed (special attention, e.g. threat)
• Personangreb (personal attack)
• Sprogbrug (offensive language)
• Spam & indhold (spam) This model is intended to be used after the BERT HateSpeech detection model.

It is based on the pretrained Danish BERT model by BotXO which has been fine-tuned on social media data.

See the DaNLP documentation for more details.

Here is how to use the model:

from transformers import BertTokenizer, BertForSequenceClassification

model = BertForSequenceClassification.from_pretrained("alexandrainst/da-hatespeech-classification-base")
tokenizer = BertTokenizer.from_pretrained("alexandrainst/da-hatespeech-classification-base")

Training data

The data used for training has not been made publicly available. It consists of social media data manually annotated in collaboration with Danmarks Radio.