Multilingual text detoxification is a style transfer task of creating neutral versions of toxic texts across multiple languages. In this paper, we use a mix of real and synthetic data to build a multilingual text detoxification model using a parallel corpus of toxic and non-toxic texts in 9 languages. We evaluate models trained on various combinations of the training data and determine the optimal training regime. Our proposed approach, which combines an ensemble model with a toxic word deletion baseline, achieves a top-3 score in automatic evaluations and a top-4 score in manual evaluations in the TextDetox 2024 shared task.
Text detoxification is a relatively new field, which started from a paper by dos Santos et. al. [ 4], where they utilized an encoder-decoder translation model, trained with cycle consistency loss to solve the task of unsupervised detoxification.
More recently, Laugier et. al. [5] proposed finetuning T5 model [ 6] on a detoxification task, using denoising and cyclic autoencoder loss. In RUSSE-2022 shared task [7] further explored non-English detoxification, with solutions ranging from using decoder-only networks with right prompts to finetuning an encoder-only tagger for toxic words and a style transfer encoder-decoder model for further detoxification [8].
In addition to these approaches, a paper by Dale et. al. [9], two algorithms were proposed. CondBERT approach, inspired by Wu et. al. [10], utilized a finetuned BERT model for replacing toxic tokens in the sequence to non-toxic. The second approach, ParaGedi, reframes text detoxification problem as a paraphrase and imposes constraints on toxic tokens used during the generation.
Authors of [11] proposed finetuning a multilingual mBART model on a big parallel corpus of English and Russian texts. Their work has shown, that reformulating the task of detoxification as a neural machine translation task boosts performance of the models, given enough data, outperforming CondBERT baseline. Also, they’ve proved that finetuning a pretrained multilingual model on any of the languages it knows, not on the main language of the model, is possible.
TextDetox 2024 shared task consisted of two phases. During the dev phase of the task, organizers provided a training set consisting of 1000 parallel toxic and neutral samples in Russian and English languages. During the test phase, organizers provided a training set, consisting of 400 parallel toxic and neutral samples in 9 languages: English, German, Spanish, Amharic, Arabic, Hindi, Chinese, Ukrainian and Russian.
Additionally, a non-parallel set of 2500 toxic and 2500 neutral sentences in the same 9 languages was provided, as well as a dataset of toxic lexicon, consisting of swear words in these languages. To assess the resulting models and given data, we calculated STA, SIM, chrF_1 and J metrics. STA metric measured the style transfer quality using the textdetox/xlmr-large-toxicity-classifier 4 [12] model. The SIM metric can be calculated by finding cosine similarity between the embeddings of sentencetransformers/LaBSE5 [13] model for the input and output (i.e., toxic and neutral sentences). chrF_1 [14] measures the similarity between model output and the references by using character n-grams. J metric is a multiplication of STA, SIM and chrF_1 metrics. The calculation of metrics was conducted using the evaluation script, provided by competition organizers, with toxic examples as the input and neutral examples as both references and output.
Upon examining the provided data, we found out that it’s quality varied significantly from language to language.
As shown in the Table 1, the quality of provided examples is suboptimal in Chinese and Hindi, as the "neutral" sentences have extremely low STA score. This indicates that only 25% of Chinese and 36% of Hindi neutral examples are actually non-toxic.
Furthermore, the neutral sentences in Amharic language are quite distinct from the toxic sentences, as evidenced by the SIM metric of 0.67. 4textdetox/xlmr-large-toxicity-classifier on HuggingFace https://huggingface.co/textdetox/xlmr-large-toxicity-classifier 5sentence-transformers/LaBSE on HuggingFace https://huggingface.co/sentence-transformers/LaBSE
To visualize this, we can sort the sentences in each language by their toxicity scores and plot STA and SIM scores on a graph by a language (Fig. 1a).
By applying a hard threshold of 0.6 for both similarity and toxicity metrics, we can filter out noisy data. However, this approach leads to a drastic reduction in the quantity of examples in Chinese and Hindi languages, with Chinese being left with only 84 examples and Hindi with 120 examples. In addition to that, we can also drop all examples longer than 512 symbols to ensure training stability (Fig. 1b).
Due to the limited amount of data available after removing non-detoxified pairs from the training data, we need to generate a new dataset. To achieve this, we employed the following algorithm: 1. Train a detox model on uncleaned dataset; 2. Run inference of this model on the toxic sentences from unpaired multilingual dataset; 3. Check if the toxicity classification model classifies the output as non-toxic, if the output of detoxification is still toxic, delete all toxic words from the data, using the toxic lexicon dataset; 4. Check if the toxicity classification model classifies the output as non-toxic; • If the output is toxic, do not add the sentence to the resulting dataset; • If the output is not toxic, add the sentence to the dataset.
For the toxicity classification model, we utilized the intfloat/multilingual-e5-large model [ 15], which
was trained on non-parallel data with an 80/20 train-test split. In contrast, for the detox model, we
employed the bigscience/mt0-xl model [
Although the resulting dataset is of lower quality than the real dataset, after applying the same cleaning procedure, its metrics become comparable to those of the cleaned original dataset (Fig. 2a, 2b, Table 2). By combining these two datasets, we obtained the training data for the final model.
There are various approaches to tackle the problem of text detoxification. One possible method is to employ encoder-only models, such as BERT, to identify toxic words in a sentence, mask them, and then treat the problem as a denoising task. However, given that we have a dataset consisting of parallel data (i.e., toxic and neutral versions of the same sentence), it is more intuitive to view this problem as a sequence-to-sequence task. Therefore, selecting a full transformer model is the obvious choice for this problem.
There are three primary families of multilingual encoder-decoder transformer models: mT5, UMT5,
and mT0. mT5 [
Our experiments show that fine-tuned mT0 models perform better in the task of text detoxification,
which led us to adopt the mT0 family as the foundation of our detoxification pipeline. Specifically,
we opted for the bigscience/mt0-xl6 [
During training, we explored ten diferent approaches to training models on synthetic data. We examined training models on real and synthetic data before and after cleaning, mixing the synthetic and real data before and after cleaning, and sequentially training on real + synthetic and synthetic + real data in a two-stage fashion, both before and after cleaning.
The models were trained using the following parameters: AdamW optimizer [20], inverse square root scheduler, learning rate (lr) = 8e-5, batch size (bs) = 4. The training was done in full precision. The models were trained for one epoch.
The best-performing model, according to evaluation set metrics (Table 3), is the model trained on a mix of synthetic and real data. We attribute this to the fact that adding synthetic data to the mix increases the STA metric, which is the hardest metric to optimize. Given enough training steps, the model learns more toxic words and becomes better at deleting them from the input data. Additionally, it is interesting to note that training on synthetic data boosts the STA metric and lowers the chrF_1 metric.
Two stage training yields middling results in both chrF_1 and STA, providing better scores than the worst models. The mixed training regime comes out on top, boasting both higher STA and chrF_1 then all other training regimes, although with slightly reduced SIM scores.
Cleaning the data significantly boosts both chrF_1 and STA metrics and moderately improves the SIM metric. The model trained on cleaned version of the real data, outperforms all models trained on non-cleaned data, even when we mix in the synthetic data.
Thus, the optimal approach for training detoxification models in this particular setting is to utilize the Cleaned Mixed training regime, which involves cleaning both synthetic and real datasets from the pairs where neutral outputs are still toxic or where the toxic and neutral sentences are dissimilar, and then mixing them together into one large training set on which the model is trained.
The bigscience/mt0-xl9 model, trained on a mix of synthetic and real data, was used for the final submission. The training parameters were as follows: AdamW optimizer, inverse square root scheduler, a learning rate of 8e-5, a batch size of 4. The model was trained in full precision for two epochs.
To ensure the model generated responses in the correct language, we used the following prompt: "Write a non-toxic version of the following text in ’language’: ’toxic sentence’." Without this prompt, the model tended to respond in a language diferent from the input. The final submission was based on a combination of answers from diferent models, taken from diferent training checkpoints.
Notably, the models sometimes failed to detoxify sentences and left out words that could be deleted simply by cutting them out. To address this, each output in the submission pipeline was additionally detoxed using the "delete" baseline method.
Our final model achieved third place in the automatic evaluation and fourth place in the manual human evaluation.
During the automatic evaluation, our model consistently ranked within the top three (Table 4), only being outperformed by other models in Spanish and Hindi. The model visibly struggled with scores on Chinese and Hindi datasets, where it performed much worse then in other languages. The reason behind this is that provided data after cleaning was insuficient for training a quality detoxification model and we had to rely on delete baseline for detoxification on Chinese language. We have tried to mitigate it by providing it synthetic data, but after cleaning it from non-detoxified samples, the amount of data was still insuficient for training a good detoxification model on these languages.
In the human evaluation, our model secured first place in Arabic detoxification and ranked among the top three models in Arabic, German, and Hindi (Table 5). Notably, our model outperformed human evaluators in Arabic and German languages in the human evaluation subset. You can see some examples of detoxification in the Table 6.
In conclusion, we have identified the optimal approach to training a multilingual seq2seq model for text detoxification tasks. When combined with the detoxification via toxic word deletion baseline, our resulting model achieved third place in the automatic evaluation stage of the PAN 2024 TextDetox competition. The model and dataset are available for download on HuggingFace. CLEF Association (CLEF 2024), Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2024. [4] C. N. dos Santos, I. Melnyk, I. Padhi, Fighting ofensive language on social media with unsupervised text style transfer, 2018. arXiv:1805.07685. [5] L. Laugier, J. Pavlopoulos, J. Sorensen, L. Dixon, Civil rephrases of toxic texts with self-supervised transformers, in: P. Merlo, J. Tiedemann, R. Tsarfaty (Eds.), Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, Association for Computational Linguistics, Online, 2021, pp. 1442–1461. URL: https://aclanthology.org/2021. eacl-main.124. doi:10.18653/v1/2021.eacl-main.124. [6] C. Rafel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, P. J. Liu, Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. arXiv:1910.10683. [7] D. Dementieva, V. Logacheva, I. Nikishina, A. Fenogenova, D. Dale, I. Krotova, N. Semenov, T. Shavrina, A. Panchenko, Russe-2022: Findings of the first russian detoxification shared task based on parallel corpora, 2022, pp. 114–131. doi:10.28995/2075-7182-2022-21-114-131. [8] I. Gusev, Russian texts detoxification with levenshtein editing, 2022. arXiv:2204.13638. [9] D. Dale, A. Voronov, D. Dementieva, V. Logacheva, O. Kozlova, N. Semenov, A. Panchenko, Text detoxification using large pre-trained neural models, 2021. arXiv:2109.08914. [10] X. Wu, S. Lv, L. Zang, J. Han, S. Hu, Conditional bert contextual augmentation, in: J. M. F.
Rodrigues, P. J. S. Cardoso, J. Monteiro, R. Lam, V. V. Krzhizhanovskaya, M. H. Lees, J. J. Dongarra, P. M. Sloot (Eds.), Computational Science – ICCS 2019, Springer International Publishing, Cham, 2019, pp. 84–95. [11] D. Moskovskiy, D. Dementieva, A. Panchenko, Exploring cross-lingual text detoxification with large multilingual language models., in: S. Louvan, A. Madotto, B. Madureira (Eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 346–354. URL: https://aclanthology.org/2022.acl-srw.26. doi:10.18653/v1/2022.acl-srw.26. [12] textdetox, xlmr-large-toxicity-classifier model on huggingface, https://huggingface.co/textdetox/ xlmr-large-toxicity-classifier, 2024. Accessed: 2024-05-15. [13] F. Feng, Y. Yang, D. Cer, N. Arivazhagan, W. Wang, Language-agnostic bert sentence embedding, 2022. arXiv:2007.01852. [14] M. Popović, chrF: character n-gram F-score for automatic MT evaluation, in: O. Bojar, R. Chatterjee, C. Federmann, B. Haddow, C. Hokamp, M. Huck, V. Logacheva, P. Pecina (Eds.), Proceedings of the Tenth Workshop on Statistical Machine Translation, Association for Computational Linguistics, Lisbon, Portugal, 2015, pp. 392–395. URL: https://aclanthology.org/W15-3049. doi:10.18653/v1/ W15-3049. [15] L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, F. Wei, Multilingual e5 text embeddings: A technical report, arXiv preprint arXiv:2402.05672 (2024). [16] H. W. Chung, X. Garcia, A. Roberts, Y. Tay, O. Firat, S. Narang, N. Constant, Unimax: Fairer and more efective language sampling for large-scale multilingual pretraining, in: The Eleventh International Conference on Learning Representations, 2023. URL: https://openreview.net/forum? id=kXwdL1cWOAi. [17] H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y. Tay, W. Fedus, Y. Li, X. Wang, M. Dehghani, S. Brahma, A. Webson, S. S. Gu, Z. Dai, M. Suzgun, X. Chen, A. Chowdhery, A. Castro-Ros, M. Pellat, K. Robinson, D. Valter, S. Narang, G. Mishra, A. Yu, V. Zhao, Y. Huang, A. Dai, H. Yu, S. Petrov, E. H. Chi, J. Dean, J. Devlin, A. Roberts, D. Zhou, Q. V. Le, J. Wei, Scaling instructionifnetuned language models, 2022. arXiv:2210.11416. [18] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, 2021. arXiv:2106.09685. [19] A. Üstün, V. Aryabumi, Z.-X. Yong, W.-Y. Ko, D. D’souza, G. Onilude, N. Bhandari, S. Singh, H.-L. Ooi, A. Kayid, F. Vargus, P. Blunsom, S. Longpre, N. Muennighof, M. Fadaee, J. Kreutzer, S. Hooker, Aya model: An instruction finetuned open-access multilingual language model, 2024.
URL: https://arxiv.org/abs/2402.07827. arXiv:2402.07827. [20] I. Loshchilov, F. Hutter, Decoupled weight decay regularization, 2019. arXiv:1711.05101.