Edit model card

Model Details

This is a conditional language model based on gpt2-medium but with a vocabulary from t5-base, for compatibility with T5-based paraphrasers such as t5-paranmt-detox. The model is conditional on two styles, toxic and normal, and was fine-tuned on the dataset from the Jigsaw toxic comment classification challenge.

The model was trained for the paper Text Detoxification using Large Pre-trained Neural Models (Dale et al, 2021) that describes its possible usage in more detail.

An example of its use and the code for its training is given in https://github.com/skoltech-nlp/detox.

Model Description

Uses

The model is intended for usage as a discriminator in a text detoxification pipeline using the ParaGeDi approach (see the paper for more details). It can also be used for text generation conditional on toxic or non-toxic style, but we do not know how to condition it on the things other than toxicity, so we do not recommend this usage. Another possible use is as a toxicity classifier (using the Bayes rule), but the model is not expected to perform better than e.g. a BERT-based standard classifier.

Bias, Risks, and Limitations

The model inherits all the risks of its parent model, gpt2-medium. It also inherits all the biases of the Jigsaw dataset on which it was fine-tuned. The model is intended to be conditional on style, but in fact it does not clearly separate the concepts of style and content, so it might regard some texts as toxic or safe based not on the style, but on their topics or keywords.

Training Details

See the paper Text Detoxification using Large Pre-trained Neural Models and the associated code.

Evaluation

The model has not been evaluated on its own, only as a part as a ParaGeDi text detoxification pipeline (see the paper).

Citation

BibTeX:

@inproceedings{dale-etal-2021-text,
   title = "Text Detoxification using Large Pre-trained Neural Models",
   author = "Dale, David  and
     Voronov, Anton  and
     Dementieva, Daryna  and
     Logacheva, Varvara  and
     Kozlova, Olga  and
     Semenov, Nikita  and
     Panchenko, Alexander",
   booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
   month = nov,
   year = "2021",
   address = "Online and Punta Cana, Dominican Republic",
   publisher = "Association for Computational Linguistics",
   url = "https://aclanthology.org/2021.emnlp-main.629",
   pages = "7979--7996",
}
Downloads last month
111