metadata

license: openrail++
task_categories:
  - text-generation
language:
  - en
tags:
  - detoxification
size_categories:
  - 10K<n<100K

ParaDetox: Text Detoxification with Parallel Data (English)

This repository contains information about ParaDetox dataset -- the first parallel corpus for the detoxification task -- as well as models and evaluation methodology for the detoxification of English texts. The original paper "ParaDetox: Detoxification with Parallel Data" was presented at ACL 2022 main conference.

📰 Updates

[2024] We have also created versions of ParaDetox in more languages. You can checkout a RuParaDetox dataset as well as a Multilingual TextDetox project that includes 9 languages.

Corresponding papers:

MultiParaDetox: Extending Text Detoxification with Parallel Data to New Languages (NAACL 2024)
Overview of the multilingual text detoxification task at pan 2024 (CLEF Shared Task 2024)

ParaDetox Collection Pipeline

The ParaDetox Dataset collection was done via Yandex.Toloka crowdsource platform. The collection was done in three steps:

Task 1: Generation of Paraphrases: The first crowdsourcing task asks users to eliminate toxicity in a given sentence while keeping the content.
Task 2: Content Preservation Check: We show users the generated paraphrases along with their original variants and ask them to indicate if they have close meanings.
Task 3: Toxicity Check: Finally, we check if the workers succeeded in removing toxicity.

All these steps were done to ensure high quality of the data and make the process of collection automated. For more details please refer to the original paper.

ParaDetox Dataset

As a result, we get paraphrases for 11,939 toxic sentences (on average 1.66 paraphrases per sentence), 19,766 paraphrases total.

In addition to all ParaDetox dataset, we also make public samples that were marked by annotators as "cannot rewrite" in Task 1 of crowdsource pipeline.

Detoxification evaluation

The automatic evaluation of the model were produced based on three parameters:

style transfer accuracy (STA): percentage of nontoxic outputs identified by a style classifier. We pretrained toxicity classifier on Jigsaw data and put it online in HuggingFace🤗 repo.
content preservation (SIM): cosine similarity between the embeddings of the original text and the output computed with the model of Wieting et al. (2019).
fluency (FL): percentage of fluent sentences identified by a RoBERTa-based classifier of linguistic acceptability trained on the CoLA dataset.

All code used for our experiments to evluate different detoxifcation models can be run via Colab notebook

Detoxification model

At-the-time SOTA for text detoxification task in English -- BART (base) model trained on ParaDetox dataset -- we release online in HuggingFace🤗 repo.

You can also check out our web-demo.

Citation

@inproceedings{logacheva-etal-2022-paradetox,
    title = "{P}ara{D}etox: Detoxification with Parallel Data",
    author = "Logacheva, Varvara  and
      Dementieva, Daryna  and
      Ustyantsev, Sergey  and
      Moskovskiy, Daniil  and
      Dale, David  and
      Krotova, Irina  and
      Semenov, Nikita  and
      Panchenko, Alexander",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.469",
    pages = "6804--6818",
    abstract = "We present a novel pipeline for the collection of parallel data for the detoxification task. We collect non-toxic paraphrases for over 10,000 English toxic sentences. We also show that this pipeline can be used to distill a large existing corpus of paraphrases to get toxic-neutral sentence pairs. We release two parallel corpora which can be used for the training of detoxification models. To the best of our knowledge, these are the first parallel datasets for this task.We describe our pipeline in detail to make it fast to set up for a new language or domain, thus contributing to faster and easier development of new parallel resources.We train several detoxification models on the collected data and compare them with several baselines and state-of-the-art unsupervised approaches. We conduct both automatic and manual evaluations. All models trained on parallel data outperform the state-of-the-art unsupervised models by a large margin. This suggests that our novel datasets can boost the performance of detoxification systems.",
}

and

@inproceedings{dementieva2021crowdsourcing,
    title = "Crowdsourcing of Parallel Corpora: the Case of Style Transfer for Detoxification",
    author = {Dementieva, Daryna
                 and Ustyantsev, Sergey
                 and Dale, David 
                 and Kozlova, Olga
                 and Semenov, Nikita
                 and Panchenko, Alexander
                 and Logacheva, Varvara},
    booktitle = "Proceedings of the 2nd Crowd Science Workshop: Trust, Ethics, and Excellence in Crowdsourced Data Management at Scale co-located with 47th International Conference on Very Large Data Bases (VLDB 2021 (https://vldb.org/2021/))",
    year = "2021",
    address = "Copenhagen, Denmark",
    publisher = "CEUR Workshop Proceedings",
    pages = "35--49",
    url={http://ceur-ws.org/Vol-2932/paper2.pdf}
}

Contacts

If you find some issue, do not hesitate to add it to Github Issues.

For any questions and get the TEST SET, please, contact: Daryna Dementieva ([email protected]), Daniil Moskovskiy ([email protected]), or Alexander Panchenko ([email protected])

Dataset Card corresponding contact: Daryna Dementieva ([email protected])