This paper presents a multilingual text detoxification method based on a chain-of-thoughts prompting approach for the PAN at CLEF 2025. Multilingual text detoxification aims to transform toxic texts into neutral versions while preserving the original meaning and grammatical structure. Our method leverages large language models (LLMs) to classify toxic sentences and detoxify them through carefully designed prompts. We evaluate our approach on the PAN 2025 multilingual text detoxification task, demonstrating its potential and stability in handling various languages.
Our method consists of three key stages: 1) Extracting toxic text features; 2) Classifying toxic text by features; 3) Text detoxification via text features and classification. All three steps are implemented through carefully designed prompts fed into the large language model.
In this stage, we adopt a similar approach to that proposed by Dementieva et al. [
Please analyze the provided sentence using the structure below to identify elements of toxicity and suggest improvements, when I tell you, use words from the keywords list (can be more than one word!): keywords = [Neutral, Informative, Casual, Assertive, Dismissive, Condescending, Friendly, Commanding, Instructive Derogatory, Confrontational, Insulting, Vulgar, Formal, Informal, Offensive, Technical, Playful, Positive, Frustration, Analytical, Professional, Hostile, Hatred, Helpful, Angry, Friendly, Arrogant] Analysis Structure (do not use “ and [] and "" in your answer and do not suggest improvement!): {
Sentence: {{SENTENCE}}, Toxicity Level: Specify here (Low/Medium/High), Tone: the overall tone of the sentence- choose from keywords, Language: Language style—choose from keywords, Implied Sentiment: the overall sentiment- choose from keywords, Context: Brief description of how context contributes to toxicity, Negative Connotations: List specific negative words/phrases here,
Intent: Describe the perceived intent behind the sentence.
Please cluster the provided sentence using the structure below, when I tell you, use group_index from the groups. The sentences will clustered into 3 groups while each group has its own characterizes.
Cluster 0 is more Offensive, Hostile and Vulgar; Cluster 1 is more Condescending, Derogatory and Hostile; Cluster 2 is more Informal, Casual, Dismissive.
Analysis Structure (do not use “ and [] and ”“ in your answer and do not suggest improvement!): {
Sentence: {{SENTENCE}}, Toxicity Level: {{TL}}, Tone: {{TONE}}, Language:{{Language}}, Implied Sentiment: {{IS}},
Cluster: <the cluster index of the sentence>
text features into the model. With its powerful understanding and analytical capabilities, the LLM
can uncover latent semantic connections between text features, thereby achieving more precise and
semantically logical clustering of texts. Figure 2 illustrates the structure of this prompt.
}
2.3. Text Detoxification via Text Features and Classification
To achieve text detoxification, we extended the approach of Dementieva et al. [
Please detoxify the provided sentence using the structure below without changing the real meaning of the sentence. The sentences are clustered into 3 groups while each group has its own characterizes.
Cluster 0 is more Offensive, Hostile and Vulgar; Cluster 1 is more Condescending, Derogatory and Hostile; Cluster 2 is more Informal, Casual, Dismissive.
For each sentence and cluster that I give, make the sentence non-toxic by making it Neutral/Informal/Casual without changing the meaning.
Analysis Structure (do not use ” and [] and “” in your answer and do not suggest improvement!): {
Sentence: {{Sentence}}, Toxicity level: {{TL}}, Cluster: {{Cluster}},
Fixed sentence: <the non-toxic sentence after making it Neutral/Informal/Casual without changing the meaning, use the same language as the original sentence>; }
PAN 2025 multilingual text detoxification task has been improved and expanded in terms of data 1, evaluation metrics, and task settings. In this year’s competition, six additional languages have been introduced, none of which provide parallel datasets.
In this competition, we utilized the DeepSeek-R1[
The oficial provided four metrics
3 . Each metric component lies in the range [
The oficial provides an evaluation using the LLM-as-a-Judge approach. Specifically, they have fine-tuned the Llama-3.1-8B-Instruct 5 model on the manual annotations from the TextDetox2024 6 dataset.
Table 1 shows the final results of the LLM-as-a-Judge 7 evaluation. The golden annotation model performed the best, with an average score of 0.828, especially achieving a high score of 0.904 in Japanese. The Team cake (our method) model ranked second with an average score of 0.674. Although this score is lower than that of golden annotation, it outperformed all other baseline models. This indicates that the Team cake model has considerable potential and stability in multilingual evaluation tasks, although there is still room for improvement in certain languages. Overall, the golden annotation model demonstrated very stable performance across all languages, while the Team cake model showed competitive strength comparable to the golden annotation in the majority of languages.
In this paper, we proposed a multilingual text detoxification method based on a chain-of-thoughts prompting approach. Our method efectively utilizes large language models to extract toxic text features, classify toxic sentences, and detoxify them while preserving their original meaning. The experimental results on the PAN 2025 multilingual text detoxification task show that our method has considerable potential and stability, outperforming other baseline models. Although there is still room for improvement in certain languages, our approach demonstrates its efectiveness in handling multilingual text detoxification tasks. Future work will focus on further enhancing the performance of our method and exploring more advanced techniques to improve text detoxification across diferent languages. 4https://huggingface.co/s-nlp/mt0-xl-detox-orpo 5https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct 6https://github.com/textdetox/textdetox_clef_2024/tree/main/human_evaluation_results 7https://pan.webis.de/clef25/pan25-web/text-detoxification.html#results
This work is supported by the Social Science Foundation of Guangdong Province, China (No.GD24CZY02)
During the preparation of this work, we used a Large Language Model; specifically, we employed DeepSeek-R1, provided by Volcengine, for the following purposes: extracting toxic text features, classifying toxic texts, and performing text detoxification. We also used basic AI-powered text-editing tools to check the grammar and spelling of the manuscript content. After using these tools, we reviewed and edited the content as needed and take full responsibility for the publication’s content.