The issue of managing toxic speech remains a crucial aspect of human communication and digital violence prevention [1], including the mitigation of toxic responses generated by Large Language Models (LLMs) [2]. The typical approach to dealing with abusive speech on social platforms involves message blocking [3]. To address this, numerous toxic and hate speech detection models have been developed for different languages, i.e. English [4], Spanish [5], Amharic [6], Code-Mixed Hindi [7], and many others [8]. However, the recent research indicates a necessity for more proactive moderation of abusive speech [9]. One such approach is text detoxification.
Within the baselines approaches for automatic text detoxification, multiple unsupervised baselines were created based on ideas of Delete-Retrieve-Generate [10], latent style spaces disentanglement [11], or conditional generation with Masked Language Modeling [12]. However, the latest state-of-the-art outcomes, particularly in English, were attained when parallel data and fine-tuning with text-to-text generation models were employed [13,14]. At the same time, the availability of such a corpus can be a challenge for new languages and cross-lingual transfer techniques should be applied [15]. In this shared task, we explored both setups-cross-lingual and multilingual one-providing new multilingual parallel text detoxification dataset for 9 languages [16]. The remainder of this paper is structured as follows. Section 2 gives an overview of the TextDetox shared task description. Section 3 provides the full overview of the new multilingual parallel text detoxification dataset collection per each language. In the following sections, the evaluation setups essentials are described-baselines in Section 4, automatic evaluation setup in Section 5, and human evaluation setup in Section 6. The submissions from participants are described in Section 7. Section 8 provides the details about final results-both automatic (Section 8.1) and human (Section 8.2) evaluation leaderboards. Finally, Section 9 concludes the paper.
All the resources produced from the task are listed at the shared task page 1 and are also mentioned in the corresponding sections.
Here, we provide the shared task main definitions-how we understand toxicity, text style transfer task, cross-lingual and multilingual setup, and the competition rules.
Definition of Toxicity While there can be different types of toxic language in conversations [17,18], i.e. sarcasm, hate speech, threats, in this work, we include samples with substrings that are commonly referred to as vulgar or profane language [19,13] while the whole main message can be both neutral and toxic, but not hateful with direct insult of individuals or groups of people.
In this work, we adapt the formal task definition of the text style transfer described in [20,21,13]:
Having a set of style 𝑆 and a corpus of texts 𝐷, a text style transfer (TST) model is a function 𝛼 : 𝑆 × 𝑆 × 𝐷 → 𝐷 that, given a source style 𝑠 𝑠𝑟𝑐 , a target style 𝑠 𝑡𝑔 , and an input text 𝑑 𝑠𝑟𝑐 , produces an output text 𝑑 𝑡𝑔 such that:
• The style of the text changes from the source style 𝑠 𝑠𝑟𝑐 to the target style 𝑠 𝑡𝑔 and is measured by a style classifier: 𝜎(𝑑 𝑠𝑟𝑐 ) ̸ = 𝜎(𝑑 𝑡𝑔 ), 𝜎(𝑑 𝑡𝑔 ) = 𝑠 𝑡𝑔 ;
• The content of the source text is saved in the target text as much as required for the task and estimated by a content similarity function: 𝛿(𝑑 𝑠𝑟𝑐 , 𝑑 𝑡𝑔 ) ≥ 𝑡 𝛿 ;
• The fluency of the target text achieves the required level according to the fluency estimator:
where 𝑡 𝛿 and 𝑡 𝜓 are the threshold values for the content preservation (𝛿) and fluency (𝜓) functions and can be adjusted to the specific task. In our task, the source style 𝑠 𝑠𝑟𝑐 is toxic and the target style 𝑠 𝑡𝑔 is non-toxic.
As parallel text detoxification corpora might not be available for any language, one of the important tasks is to explore cross-lingual text detoxification knowledge transfer. In this case, we assume that training data is available for the resource-rich language (i.e. English) and the task is to obtain a text detoxification system for a new language.
Multilingual Text Detoxification If parallel corpora available for multiple language, then both monolingual text detoxification models per language and multilingual model for all languages can be obtained.
The share task timeline was divided in to two phases-development and test.
For the first phase, only training parallel data for English and Russian from previous works [13,22] aiming to provide participants to explore cross-lingual transfer techniques.
Test Phase During the test phase, parallel text detoxification corpora were available for all target languages. Participants was invited to submit monolingual and multilingual solutions.
Leaderboards During both phases, the leaderboards based on automatic evaluation were available. We used Codalab platform [23] 2 (and TIRA [24] as a backup platform). However, despite having powerful models capable of classifying texts and embedding their meanings, human judgment remains superior for making final decisions in the text detoxification task [25]. Thus, based on a test part subset, we performed human evaluation of the participants submissions. The final leaderboard was based on the human judgements results.
For each of our 9 target languages, we prepared parallel text detoxification corpus. We asked experts and native speakers to contribute for corpora collection. Further, we describe the collection details per each language: English (Section 3.1), Russian (Section 3.2), Ukrainian (Section 3.3), Spanish (Section 3.4), German (Section 3.5), Hindi (Section 3.6), Amharic (Section 3.7), Arabic (Section 3.8), Chinese (Section 3.9). All the instructions per language are available online. 3 We also opensource the obtained resources for the public usage. 4For all the data, we adapt the concept of English ParaDetox [26] collection pipeline as it was designed to automate the data collection as well as verification with crowdsourcing. The pipeline consists of three tasks: Task 1: Rewrite text in a polite way Annotators need to provide the detoxified paraphrase of the text so it becomes non-toxic and the main content is saved or to skip paraphrasing if the text is not possible to rewrite in non-toxic way;
Task 2: Do these sentences mean the same? Check if the content is indeed the same between the original toxic text and its potential non-toxic paraphrase;
Task 3: Is this text offensive? Verification of the provided paraphrase if it is indeed non-toxic.
In the same manner, each language stakeholder asked the annotators to rewrite the toxic samples verifying the main three criteria: (i) the new paraphrase should be non-toxic; (ii) the content should be saved as much as possible; (iii) the resulted text should be fluent but may contain some minor mistakes (as the majority of the original toxic samples are examples from posts from social networks). The rewriting of the texts and verification of their quality could have been done either via crowdsourcing or via manual annotation. The main goal for each language was to obtain 1000 parallel pairs that were later splitted into dev and test sets.
Data Preprocessing For all languages, we maintain the length of samples as sentences of around 5-20 tokens. Also, if a text sample is from a social network, we anonymize any mentioning of usernames and links.
For English, we reused the data from English ParaDetox dataset [13] and additionally manually marked up approximately 400 pairs to form a validation dataset of 1000 examples.
For EnParaDetox, the original toxic texts were taken from Jigsaw toxicity identification challenge train dataset [27]. We have considered only texts labeled as toxic and severe toxic.
The training and validation sets of EnParaDetox were acquired through crowdsourcing via Toloka5 platform with fluent English speakers. Additionally, we employed annotators who are fluent in English and hold a Masters degree in Computer Science to compile additional samples to the test set.
The same as for English, there were previously available training and validation data from previous work [22,28]. We reused this data and manually annotated some additional toxic examples taken from various toxicity datasets.
The original toxic samples were taken from two binary toxicity classification Kaggle Toxic Comments datasets [29,30].
The training and validation sets of RuParaDetox were acquired through crowdsourcing via Toloka platform with fluent Russian speakers. Additionally, we employed annotators who are native in Russian and hold a Masters degree in Computer Science to compile additional samples to the test set.
We used the data presented in MultiParaDetox paper [28] providing the main details of data collection:
Input Toxicity Data For the Ukrainian language, there was no existing binary toxicity classification corpus. Therefore, we filtered explicitly toxic samples containing obscene lexicon from the predefined list [31] within the Ukrainian Tweets Corpus [32].
We adapt ParaDetox [26] collection pipeline and verified the data quality via crowdsourcing. We utilized the Toloka platform for crowdsourcing tasks in Ukrainian. The annotators, who were native Ukrainian speakers, underwent an examination before starting the tasks.
We used the data presented in MultiParaDetox paper [28] providing the main details of data collection:
Input Toxicity Data For Spanish, we selected samples for annotation from three datasets: hate speech detection ones [33,34] as well as filtered by keywords Spanish Tweets corpus [35].
We adapt ParaDetox [26] collection pipeline and verified the data quality via crowdsourcing. We utilized the Toloka platform for crowdsourcing tasks in Ukrainian. The annotators, who were native Spanish speakers, underwent an examination before starting the tasks.
German ParaDetox was collected with several annotators with manual quality verification:
The German language source data in this work is based on three datasets containing toxic, offensive, or hate speech comments on social media about primarily political events in Germany or the US. For the two datasets from the GermEval 2018 [36] and GermEval 2021 [37] shared tasks, we used data from both the test and the train split and filtered it as follows. For the GermEval 2018 data, we only used samples labeled with the coarse class "OFFENSE" whereas for the GermEval 2021 data -which contains different labels -we only used samples annotated with the "Sub1_Toxic" class. The third dataset [38] was filtered so only samples were kept where both expert annotators classified the samples as hate speech.
The data from the three datasets was merged and deduplicated via exact string matching. Furthermore, we removed all samples that included less than 5 or more than 30 white-space separated tokens.
To create the final parallel detoxified German dataset, we hired two native German annotators. Annotator A is a female born in 1994 who holds a Master of Arts degree in Social Sciences, and Annotator B is a male born in 1992 who holds a Master of Science degree in Computer Science. The data was distributed so that each sample was transcribed by only one of the annotators.
Hindi dataset was collected manually by a native-speaker annotator gaining data from multiple sources:
Input Toxicity Data We used the HASOC dataset created at FIRE 2019 [39] as source for Hindi language. Contents in this dataset are relevant within Indian subcontinent which are collected from various social media platforms prevalent in India. In this dataset, hostile posts are divided into HATE SPEECH, OFFENSIVE and PROFANE. For curation, posts containing OFFENSIVE and PROFANE contents in train and test splits were used. 1455 PROFANE posts (1237 train + 218 test) and 873 OFFENSIVE posts (676 train + 197 test) were chosen to prepare detoxifiable toxic data for our task.
Input Preprocessing On a total of 2328 samples, we first performed deduplication via exact string matching. Mentions, links and emojis were also removed as part of this step.
Annotation Task(s) The posts after input preprocessing were manually verified. Those with less than 5 white-space separated tokens were removed and which had more than 25 white-space separated tokens were re-framed to bring them down to this limit. Toxicity and meaning of the posts were unchanged during this re-framing. These posts were then bifurcated into detoxifiable and non-detoxifiable labels. The manual re-framing and bifurcation were carried by a NLP researcher with working experience on hate/toxic speech.
Out of 2328 samples, 1007 samples were marked as detoxifiable. From these detoxifiable samples, we carefully sampled 24 data points and detoxified them. These detoxified samples were evaluated by two experts who are native Hindi language speakers to provide precise samples to the annotators for detoxifying the whole dataset. Annotators were guided based on expert prepared samples and were asked to re-write toxic pairs in a non-toxic manner, keeping the meaning of the original post unchanged. Detoxification was carried out by two annotators and we provide their details in the corresponding subsection.
Annotators One male NLP researcher working in the field of hate/toxic speech and another female student enrolled in Bachelor's Degree and having working knowledge in Machine Learning, were employed to carry out the detoxification of whole dataset. Both annotators are Indian, native Hindi speakers and are well versed with the topicality covered in the dataset.
We compiled new Amharic ParaDetox datasets with the following annotation details, based on prior studies of hate and offensive language:
The Amharic ParaDetox dataset is derived from merging two pre-existing studies conducted on the X/Twitter datasets [6,40]. The dataset introduced by Ayele et al. [40] was initially annotated into categories of hate, offensive, normal, and unsure by three native speaker annotators, with the gold labels determined through a majority voting scheme. In contrast, the dataset presented by Ayele et al. [6] was annotated by two native speakers, with a third adjudicator annotator deciding the gold labels for instances where there was no majority consensus. We extracted a subset of these datasets labeled as offensive to create the new Amharic ParaDetox dataset and subsequently reworked this subset using new annotators to determine if the messages could be detoxified and to present non-toxic versions of each message.
The input toxicity data is entirely sourced from the two previous studies, namely Ayele et al. [6] and Ayele et al. [40], and has been adapted to meet the requirements of the ParaDetox task.
Annotation Task(s) We customized the Potato-POrtable Text Annotation TOol6 and utilized it for the annotation of Amharic ParaDetox dataset. Annotators were provided annotation guidelines, took hands-on practical training, completed independent training tasks before the main annotation task.
We conducted pilot annotation of 125 sample items with three native Amharic speaker annotators and evaluated the annotation quality with experts and annotators together in a group meeting to improve the understandings of annotators for the main task. The main annotation task comprises of 2,995 tweets, each annotated by one annotator. Annotators were asked to classify each tweet in to two broad categories, detoxifiable and non-detoxifiable. For the detoxifiable category, annotators are asked to detoxify and re-write the text. For non-detoxifiable tweets, annotators choose non-detoxifiable and select reason as; it is hate speech, it is normal speech or indeterminate to decide the label.
Annotators Annotators have previous hate speech annotation experiences and already holds Masters degree in Computer Science. Only two of the annotators were evolved in the main annotation task, where both of them are university lecturers and have basic knowledge of natural language processing tasks. One of the annotators is from Adama Scinece and Technology University with experience of 15 years of teaching Computer Science, who is female. The other annotator is a male, who has been teaching Computer science over 12 years in Kotebe University of Education.
Arabic ParaDetox was collected with several annotators with manual quality verification:
The Arabic ParaDetox dataset was created by combining parts of several existing datasets along with the Arabic-translated version of the Jigsaw dataset [27]. It includes the Levantine Twitter Dataset for Hate Speech and Abusive Language (L-HSAB) [41], which focuses on Levantine dialects, and the Tunisian Hate and Abusive Speech (T-HSAB) dataset [42], which targets Tunisian dialects. It also incorporates the OSACT dataset [43] and the Arabic Levantine Twitter Dataset for Misogynistic Language (LeT-Mi) [44], which specifically addresses gender-based abuse. These resources combine to form the Arabic ParaDetox dataset, aimed at aiding the development of toxicity classifiers capable of handling Arabic content across various dialects and contexts.
The detoxification process was conducted by three annotators, each with a PhD. The team includes two males and one female, all of whom have a strong interest in computational linguistics. These native Arabic speakers possess a deep understanding of the subjects encompassed within the dataset. Each text sample was transcribed by two of the annotators to ensure accuracy and consistency in the data.
We collected new Chinese ParaDetox datasets with the following annotation details:
Input Toxicity Data The Chinese ParaDetox dataset is derived from TOXICN [45], a recently released Chinese toxic language dataset. TOXICN was compiled from social media platforms and comprises 12,011 comments addressing several sensitive topics, including gender, race, region, and LGBTQ issues. From this dataset, we extracted a subset based on multiple criteria: the number of toxic words, the ratio of toxic words in the comments, the length of comments, and the toxic scores of comments.
We set thresholds for the criteria mentioned above: the number of toxic words ranged from 1 to 5, the ratio of toxic words in comments was less than 0.5, and the length of comments ranged from 3 to 50 words, ensuring suitability for annotators to rewrite them. Following these criteria, we extracted 1,516 samples from the training set and 231 samples from the test set.
Subsequently, we employed a pre-trained toxic classifier [45] to compute the toxic scores of the selected comments, using a threshold score of 0.978 to filter the candidates. Ultimately, we collected 1,149 samples from the training set and 231 samples from the test set, resulting in a total of 1,380 samples deemed suitable for annotation.
Annotation Tasks For data annotation and verification, we employed a specifically designed threetask pipeline: Task 1: Determine if the sentences are toxic or neutral. Annotators were required to choose one of three options: the given sentence is neutral, toxic but can be rewritten, or toxic and cannot be rewritten. The last option was included based on the observation that some toxic texts are impossible to rewrite in a non-toxic manner.
Task 2: Rewrite sentences in a non-toxic style. Annotators were instructed to create detoxified versions of the toxic sentences identified in Task 1. They were advised to retain the main content of the original sentences and rewrite the toxic words in a polite manner.
Task 3: Cross-check verification. The rewritten sentences from Task 2 were cross-distributed to different annotators for verification. The goal was to ensure the rewritten sentences were nontoxic and adhered to our guidelines. If annotators selected the "No" option, indicating the sentence did not meet the criteria, a further meta-rewrite process was initiated.
From the 1,380 toxic samples, 1,031 samples were successfully detoxified and verified, with 861 from the training set and 170 from the test set. The remaining 349 samples were either considered non-toxic or toxic but could not be rewritten.
Annotators For the detoxification process, we hired three native Chinese annotators. Two female annotators, both aged 22, hold Bachelor's degrees in Engineering, and a male annotator, aged 32, holds a Master's degree in Computer Science. All annotators are native Chinese speakers residing in mainland China, ensuring they deeply understand the Chinese language and the detoxification task.
The statistics of all ParaDetox datasets used in the shared task. The human detoxified references were collected either via crowdsourcing or locally hired native speaker. For English and Russian, the previously collected train data was available during all shared task's phases. For other languages, 1 000 samples per language were divided correspondingly into development and test parts.
Annotation The full picture of the collected ParaDetox data for all target languages is presented in Table 1. While the methods of collecting human annotations vary across languages-some data were gathered via crowdsourcing, others by hiring local native speakers-the quality of the texts was uniformly verified by experts to ensure three key attributes as introduced in [46,13]: (i) the style of new paraphrases is genuinely non-toxic, (ii) the main content is preserved, and (iii) the new texts are fluent.
For each language for the shared task's phases:
• During the development phase: 400 only toxic parts were available for participants to perform cross-lingual experiments.
• During the test phase: (i) 400 ParaDetox instances were fully released; (ii) participants should provide their final solutions for 600 toxic parts of the test dataset.
For English and Russian during all phases, additional training parallel datasets were available released from previous work [13,22,28]. You can find online fully released development part of the data 7 and the test part only toxic instances. 8
We provide four baselines for our shared task: (i) trivial Duplicate baseline; (ii) a rule-based Delete approach; (iii) Backtranslation pipeline that reduces the task to the monolingual one; (iv) finally, finetuned for the downstream task on the dev dataset mT5 instance. The code for all the baselines is available online.9
Duplicate Trivial baseline: the output sentence is a copy-paste of the input sentence. This baseline has 1.0 (or 100%) SIM score by definition.
Delete For the first unsupervised baseline, we perform an elimination of obscene and toxic substrings from a text according to the predefined lists of keywords. For the shared task, we collected and compiled together the lists of such toxic keywords for all target languages based on openly available sources (see Table 2). The amount of toxic keywords per language differs which displays the diversity of morphological forms and variations of toxicity expressions across languages. For participants and further public usage, we release our compiled list online. 10
The list of the original sources and the corresponding amount of obscene keywords used to compile multilingual toxic lexicon list for our Delete baseline. Backtranslation As for a more sophisticated unsupervised baseline, we perform translation of non-English texts in English with NLLB [19] instance 11 and then perform detoxification with the fine-tuned on English ParaDetox train part BART [13] instance. 12Fine-tuned mT5 Specifically for the test phase, we fine-tuned the multilingual text-to-text generation model mT5 [51]. We tuned the mT5-XL 13 on the released for the test phase parallel development part of the presented multilingual data.
We adopt the monolingual evaluation pipelines from [13,22] to our multilingual setup and provide the detailed description below. We evaluate the outputs based on three parameters-style of text, content preservation, and conformity to human references-combining them into the final Joint score. The evaluation script is available online. 14Style Transfer Accuracy (STA) ensures that the generated text is indeed more non-toxic. To prepare a model for this metric that covers our target languages, we subsampled 5 000 samples-2 500 toxic and 2 500 neutral-from toxicity classification corpora for each language (see references in Table 1) that were not used for ParaDetox data collection. We released 15 this compiled corpus for participants as an additional dataset for experiments and fine-tuned XLM-R [52] large instance for the binary toxicity classification task. The model is also available for the public usage 16 and is used in the shared task to estimate the level of non-toxicity in the texts.
Content Similarity (SIM) is the cosine similarity between LaBSE17 embeddings [53] of the source texts and the generated texts.
Fluency (ChrF1) is used to estimate the proximity of the detoxified texts to human references. While in several previous work language acceptability classifiers based on CoLa-like corpora were utilized for fluency estimation [13,15], the recent work [25] also showed that reference-based metrics achieved high correlations with human evaluation. Thus, we use an implementation of ChrF1 score from sacrebleu library [54].
Joint score (J) is the aggregation of the three above metrics. The metrics STA, SIM and ChrF1 are subsequently combined into the final J score used for the final ranking of approaches. Given an input toxic text 𝑥 𝑖 and its output detoxified version 𝑦 𝑖 , for a test set of 𝑛 samples:
where STA(𝑦 𝑖 ), SIM(𝑥 𝑖 , 𝑦 𝑖 ), ChrF1(𝑥 𝑖 , 𝑦 𝑖 ) ∈ [0, 1] for each text detoxification output 𝑦 𝑖 .
For the test set, we performed the human evaluation to obtain final judgements on the participants' systems. The details and instructions of the annotation setups are available for the public usage.18
We used Toloka 19 crowdsourcing platform for manual evaluation of automatic detoxification. For each project, the annotators who indicated the knowledge of the necessary language could access the tasks.
Quality Control First, the annotators underwent a rigorous selection process, starting with a training pool of 20 tasks with predefined answers and instructional feedback. Following this, annotators faced an examination pool comprising 10 tasks with known answers. Only those achieving a minimum accuracy of 65% in both the training and examination pools were permitted to proceed to the main evaluation tasks.
During the execution of the real tasks, annotators' performance was meticulously monitored using several measures. Response speed was a critical factor; annotators who responded rapidly (less than 15 seconds per page) to three consecutive task pages were permanently banned. Additionally, we evaluated the answers of the annotators to mixed tasks with known true answers. Those who demonstrated low performance, were defined as achieving less than 65% accuracy, were required to undergo re-examination. Furthermore, annotators who skipped three task pages were also banned from the study.
We selected 100 random original toxic samples per each language from the test part of our dataset. Each pair was annotated within a separate project dedicated to the respective language. These projects were categorized into three distinct types, which will be described in detail in the following section.
In general, the concept of the human evaluation mirrored the approach used in the automatic evaluation. Each project type focused on assessing one of the three key qualities of detoxification; style transfer accuracy, content similarity, or fluency. Below, we provide a detailed description of each project type and its specific focus within the evaluation process.
To measure style transfer accuracy, we employed a pairwise comparison between the original toxic text and the generated detoxified text. Annotators were tasked with determining which text was more toxic: the left text, the right text, or neither. An illustration of this task can be found in Figure 2.
To enhance realism, we randomized the sequence of original and detoxified texts. Annotators' votes were then converted into numerical values using the following logic: if the original toxic text was deemed more toxic, the vote was recorded as 1; if the detoxified text was considered more toxic, the vote was recorded as 0; if "none of the texts" was selected, the vote was recorded as 0.5. Try out tasks as a Toloker and review results
Look for possible problems that Tolokers might run into and fix them for best results. Check the instructions and the popular screen size. 83% of users use mobile devices when completing tasks
Which of the texts is more offensive?
Which of the texts is more offensive? getting kicked by gre practice tests made sure that i never make this same mistake again .
getting my ass kicked by gre practice tests made sure that i never make this same mistake again . Content Similarity For the content similarity metric, we presented pairs of texts in a fixed order (the toxic phrase followed by the detoxified phrase) and asked annotators to indicate whether the sense of the two texts was similar. Annotators' responses were binary, with "yes" indicating that the sense was similar and "no" indicating otherwise. The responses were mapped to numerical values, with a "yes" response being recorded as 1 and a "no" response being recorded as 0. An illustration of this task can be found in Figure 3.
Fluency For the fluency metric, sentences were evaluated individually rather than in pairs. Annotators were asked whether each text was intelligible and correct, with three possible responses: yes, partially, and no. Each response was mapped to a numeric value: "yes" was scored as 1, "partially" as 0.5, and "no" as 0. The fluency score for a pair of texts was determined by comparing the scores of the detoxified text to the original text. If the detoxified text received a higher or equal fluency score, the pair was scored as 1; otherwise, it was scored as 0. An illustration of this task can be found in Figure 4.
The final scores for each pair within each project were calculated by weighting them according to the skill of the crowdsourcing annotators. This skill was dynamically assigned based on the quality of their responses to both the tasks with known answers and the real tasks. The logic for calculating the Joint score in the manual evaluation mirrors that of the automatic evaluation. still defending their lying and selfish leader.
He deserves every beating and rape he will be subjected to.
Skip Do these sentences mean the same? language, the Joint score was determined by multiplying the three individual scores (style transfer accuracy, content similarity, and fluency) using the same formula as in the automatic evaluation.
We received 20 submissions for the development phase leaderboard and 31 submissions for the test phase leaderboard; the final manually evaluated leaderboard was based on 17 submissions who confirmed their participation in the competition. Here, we briefly describe the solutions of our final participants. Each team is presented with its respective leaderboard name (in some cases, additionally, by a username of the corresponding team member that did a submission) and approach used in brackets:
The participants achieved the resulting score with a few-short LLM inference by using a two stage process: first, 400 samples from EN and RU provided datasets were used to be detoxified by a proprietary LLM-Kimi.AI [56] which is a large language model chatbot developed by Moonshot AI, a Beijing-based startup. In the second step, the participants employed newly detoxified samples to construct a prompt where they were included as examples of the desired behavior and the model Kimi.AI, thus, was prompted to perform detoxification in target languages.
To get the results, Team SINAI employed the Tree-of-Thought prompting strategy based on the OpenAI's model GPT-3.5 [58]. Given a toxic sentence, the model was prompted to output three options of potential detoxified sentences.
Then the model was asked to decide in terms of offensiveness, content, and fluency which one out these sentences was detoxified the most appropriate way.
Team MarSanAI, Submission maryam.najafi (Mistral-7b with PPO) [59] This team offered a solution only for two languages: English and Russian. A reinforcement learning method was utilized to fine-tune an LLM-Mistral-7b [60]-coupled with a Proximal Policy Optimization (PPO) [61] using the implementation from HuggingFace TRL [62]; the reward was obtained using the provided toxicity classifier.
The solution consisted of two components: i) the supervised solution for the English and Russian languages; ii) the unsupervised solution for the other 7 languages. The supervised solution used T5 [64] and BART [65] as base models; the exponentially weighted moving average and ROUGE scores were used as loss functions for Russian and English, respectively. The unsupervised solution utilized hashing techniques, log odds ratio, and grammatical rules to identify and conceal toxic words across other 7 languages; additionally, it incorporated a mask prediction model to maintain the original sentences meaning intact.
Team VitalyProtasov (mT0-large) [66] In the proposed solution, the team used a text-to-text model-mT0-large [67]-which was trained on different combinations of languages. In addition, before training, certain filtering techniques were applied to the data.
The participant used the text-to-text mT0-XL [67] model that was fine-tuned in two stages. In the first stage, a model was fine-tuned on the parallel data of all languages; this model was used to generate synthetic parallel data from non-parallel samples. The resulting data was cleaned and filtered using a cosine distance between LaBSE embeddings and the toxicity scores by the provided classification models followed by a modification with improved delete approach. At the end, the synthetic and filtered "golden" data were merged into new training set to fine-tune a new instance of the text-to-text multilingual model.
Multilingual model mT0-XL [67] was as well used by this team. First, the model was fine-tuned for text generation using a combination of parallel and translated datasets. The model was further aligned with the Odds Ratio Preference Optimization (ORPO) [70]. During the inference stage, the best candidate generated by the model was chosen by calculating scores from STA and SIM models.
Team gleb.shnshn (zero-shot LLaMa-3) This solution was based on a modern open-source LLM-LlaMa3-70B [71]. The model was prompted using the zero-shot prompting method for the detoxification task.
In this solution, "uncensored" LLaMa3 [71] was introduced and initialized for every target language except Amharic. Using the recent alignment jailbreaking method by identifying "refusal" directions and subtracting them from model weights [73], they used LLaMa3-70B to get predictions using a few-shot prompting strategy. So, the model received 10 examples of detoxification via starting prompt. For the Amharic language, the text-to-text mT0-XL [67] model was used: the model was fine-tuned on the Amharic parallel dataset.
The team used a combination of different methods and models depending on the language. For the majority of languages, the participant used a text-to-text encoder-decoder NLLB translation model [19] to translate data from various languages into English. Then, the translated data was detoxified using the English BART-detox model [13]. After that, the resulting parallel synthetic data was translated back into the original languages. For Russian, the specifically Russian text-totext model-ruT5-base-detox [22]-was employed for the detoxification. In the case of Chinese, the participants, firstly, applied filtering of the training dataset, fine-tuned the pre-trained ruGPT3 [75] model, and applied the Delete method.
The participant employed a text-to-text model mT5 [51]. The provided multilingual parallel data from the development phase was used for fine-tuning.
Team NaiveNeuron, Submission erehulka (few-shot LLaMa-3) [77] The team used a text-to-text Llama3 [71] which was prompted using a few-shot method.
Team team0, Submission dkenco (few-shot Cotype-7b) In this case, the team put a stress solely on the English and Russian languages. Two language-specific approaches were used based on Cotype-7b model [78]. For English, there was employed a zero-shot prompting technique where the prompt included brief instructions for the text detoxification task. For the Russian language, the team used a few-shot approach: the system prompt included brief instructions for the task to be performed as well as five randomly picked samples from the parallel development set. During inference, for both languages, there were applied regular expressions intended as filters.
Team NLPunks, Submission bmmikheev (few-shot LlaMa-3) This team used a text-to-text Llama3-70B [71] by with a few-shot prompting method. For English and Russian, the generated output was evaluated manually. For other languages, GPT-3.5 [58] was used to evaluate outputs. For all languages, the system prompt was formulated in English.
The participant as well used a text-to-text Llama3-70b [71] with a few-shot prompting method.
Team MBZUAI-UnbabelDetox, Submission mkrisnai (few-shot GPT-3.5) In this team, a twostep prompting approach was utilized. At the first step, GPT-3.5 [58] was prompted with a few-shot method to produce synthetic detoxification data. Then, the resulting data was employed in the prompt to GPT-3.5 to perform detoxification.
The participant fine-tuned mT5-XL instances [51] on the provided development multilingual parallel dataset. Almost all of the participant used the current state-of-the-art Large Language Models (LLMs), among which are GPT-3.5 [58] and LLaMa-3 [71] models. To enhance the model's performance on the task of detoxification, most participants used the few-shot prompting method. Among smaller models, mT5 [51] and mT0 [67] were utilized: usually, these models were fine-tuned using ad hoc filtering and data augmentation techniques, for instance, as RAG and backtranslation. Additionally, region-specific LLMs were also employed-Cotype [78] and Kimi.AI [56].
Here, we provide the final results of the final test phase, of our tasks. The full detailed tables of results per each language and per each metric can be found in Appendix A.
We received 20 submissions for the development phase automatic leaderboard and 31 submissions for the test phase automatic leaderboard. Automatic evaluation leaderboards are publicly available online. 20The final leaderboard from the test automatic phase evaluation is presented in Table 3.
The leading solutions were consistent across most languages, except for Spanish, Chinese, and Hindi. However, with the automatic evaluation leaderboard publicly available to all participants, some teams focused on optimizing their models specifically for the evaluation metrics, leading to potential overfitting.
Most solutions surpassed the baseline for at least one language, and in some cases, participant systems approached the performance of human references. However, except for Hindi, no participant solution outperformed human references in the automatic evaluation across any language. Although the automatic evaluation scores for human references across most languages hovered around a J score of 0.7, the results for Chinese were notably poor, with the highest participant score being 0.178 and the best human reference score at 0.201. The results leads to a further investigations of the robustness of the automatic evaluation metrics.
The top three teams across the majority of the languages generally employed a similar strategy, fine-tuning the mT0-XL text-to-text model. Team SmurfCat is holding the best automatic evaluation scores for all the languages, which was achieved by additionally fine-tuning mT0-XL with a recent ORPO alignment method. The majority of the submissions were multilingual, designed to cover all languages within a single model. These models demonstrated consistent score distributions across languages, with notable declines in performance for Chinese and Hindi. An exception was user ansafronov, who achieved the top score specifically for Chinese.
Results of the automatic evaluation of the test phase. Scores are sorted by the average Joint score. The scores for each language are respective Joint scores. Baselines are highlighted with gray , Human References are highlighted with green . Three best scores for each language are highlighted with bold, the best score is underlined bold.
After participants confirmed their submissions via a form, we received 17 entries for the human evaluation phase. This evaluation was conducted on a subsample of 100 test set items through crowdsourcing.
The results of the human evaluation, organized by team and language, are publicly available. 21 The final leaderboard based on human evaluation is presented in Table 4.
Results of the human final evaluation of the test phase. Scores are sorted by the average Joint score. The scores for each language are respective Joint scores. Baselines are highlighted with gray , Human References are highlighted with green . Three best scores for each language are highlighted with bold, the best score is underlined bold. The human evaluation leaderboard saw significant changes compared to the automatic evaluation phase. Human references received high scores from the annotators, with J scores around 0.8 or higher. However, not all teams surpassed the mT5 and Delete baselines. Interestingly, the Delete baseline outperformed the mT5 text-to-text generation baseline in languages such as Arabic, Hindi, Ukrainian, Russian, and Amharic. This indicates that not all multilingual models are equally proficient in understanding and handling toxicity across different languages.
In the human evaluation phase, participants' solutions closely matched the human references, even surpassing the provided references from parallel datasets in some languages. The top solution, after manual evaluation, was presented by user SomethingAwful and was based on the "uncensored" LLaMa3-70B language model. Interestingly, SomethingAwful's solution did not achieve the highest scores across all nine languages but excelled in Spanish, German, and Russian. The leader of the automatic evaluation leaderboard, Team SmurfCat, secured second place. Participants nikita.sushko and VitalyProtasov switched places in the manual leaderboard.
Similar to the automatic leaderboard, human assessments revealed that certain models excelled in specific languages. For instance, nikita.sushko and VitalyProtasov achieved top results in Arabic and Hindi. Despite Team mkrisnai ranking 7th overall, their solution performed exceptionally well in English, Spanish, Russian, and Ukrainian. Additionally, Team Team cake secured the highest scores specifically for English and Chinese.
From the detailed results in Appendix A, it is evident that the solutions surpassed human references in English, Spanish, and German, often achieving near-perfect fluency. However, this success does not extend to other languages. These results highlight the impressive human-like text generation capabilities of modern LLMs, though they still struggle with handling toxicity and maintaining consistent controllable generation across languages. Future work should focus on developing more challenging tasks, particularly in cross-lingual contexts.
In Multilingual Text Detoxification task at PAN 2024, participants were tasked with transforming text style from toxic to non-toxic across nine languages: English, Spanish, German, Chinese, Arabic, Hindi, Ukrainian, Russian, and Amharic. The task was divided into two phases: (i) development phase focused on cross-lingual transfer approaches; (ii) test phase utilized parallel training data for all languages and encouraged multilingual solutions. Participants' submissions in both phases underwent evaluation using a set of automatic metrics, followed by human evaluation of the test subset to determine the final leaderboard rankings.
Participants employed modern state-of-the-art Large Language Models either by prompting them in few-shot formats or fine-tuning medium-sized models. For certain languages with sufficient training data, these models approached or even exceeded human reference provided in the shared task. However, this was primarily observed for resource-rich European languages. Opportunities for enhancement remain significant for less resource-rich languages and those with limited data, highlighting the need for further exploration in cross-lingual text detoxification and knowledge transfer.
We express our deepest gratitude to Toloka platform to support our shared task. Crowdsourced data collection and human evaluation were made possible through the provided research grant.
A. Panchenko) https://dardem.github.io (D. Dementieva); https://www.researchgate.net/profile/Daniil-Moskovskiy (D. Moskovskiy); https://github.com/bbkjunior/bbkjunior (N. Babakov); https://scholar.google.com/citations?user=g2m1wH4AAAAJ&hl=en (A. A. Ayele); https://www.linkedin.com/in/naquee-rizwan-a97abb159 (N. Rizwan); https://www.linkedin.com/in/flo-schneider-hh (F. Schneider); https://ethanscuter.github.io (X. Wang); https://seyyaw.github.io (S. M. Yimam); https://linkedin.com/in/ustalov (D. Ustalov); https://github.com/eistakovskii (E. Stakovskii); https://www.sharjah.ac.ae/en/academics/Colleges/CI/dept/cs/Pages/ppl_detail.aspx?mcid=4 (A. Elnagar); https://cse.iitkgp.ac.in/~animeshm (A. Mukherjee); https://alexanderpanchenko.github.io (A
Here, we provide the extended results-from both automatic and human evaluation setups-based on three evaluation parameters for all languages: English (Table 5), Spanish (Table 6), German (Table 7), Chinese (Table 8), Arabic (Table 9), Hindi (Table 10), Ukrainian (Table 11), Russian (Table 12), and Amharic (Table 13). In every table, the baselines are highlighted with gray ; Human References are highlighted with green ; the ordering is made by J score from Human Evaluation results. The automatic evaluation is based on the full test set of 600 samples per language; human evaluation was performed on 100 set of the test set per language.