Toxic online language poses a severe threat to the safety and inclusivity of users on many online websites. This work is a part of PAN at CLEF 2025 shared task named Multilingual Text Detoxification (TextDetox) shared task 2025, which tries to convert toxic texts into non-toxic ones while preserving semantic meaning in diferent languages that range from being high-resource to underrepresented ones. The dataset used in this task consists of toxic sentences in 15 languages from around the globe such as English, Spanish, German, Chinese, Arabic, Hindi, Ukrainian, Russian, Amharic, Italian, French, Hebrew, Hinglish, Japanese and Tatar, as provided by the CLEF PAN-25 initiative. Our method, applicable to all 15 languages, the approach begins by identifying and masking toxic words in input sentences. The original and masked versions are then provided to large language models (LLMs) to generate detoxified outputs that retain the intended meaning while eliminating ofensive language. These experiments are evaluated based on style accuracy, semantic preservation, and fluency. The study show competitive results across multiple languages, highlighting the efectiveness of a hybrid approach in multilingual style transfer tasks. Our results further go toward inclusive and robust moderation tools that allow safer communication in multilingual digital spaces. All our codes can be seen on GitHub1."
Online interactions are becoming multilingual, and while this global reach is of immense value, it also
poses the challenge of moderating ofensive and toxic content [
The Multilingual Text Detoxification (TextDetox) 2025 [
One of the main challenges in this area is the variability in toxic phrasing across languages and dialects. The vast majority of languages lack labeled data for this issue, and supervised learning becomes unfeasible. In addition, translating detoxed content often loses its meaning or gets misinterpreted as culturally unpleasant.
To address these challenges, The proposed study used a hybrid approach that combines rule-based masking with model-guided generation. First, toxic spans were identified using keyword-based lexicons. Such lexicons were subsequently masked to reduce generation bias. By feeding the original and masked sentences into multilingual large language models, we ensured that the detoxified outputs were semantically aligned with the input and efectively removed ofensive content. This two-input method allowed the model to better identify contextual nuances, especially in code-mixed and low-resource languages, and improved the outputs.
The dataset, acquired via CLEF 2025 PAN [
The section outlines the specific design and development approach used for developing our multilingual text detoxification system. We explain the architecture, low-level techniques behind identifying and cleaning up toxic text to create detoxified versions of them, and how implementation strategies hold across languages. The objective of the study is to develop a generalizable system in support of several languages and cleaning up toxic input without sacrificing any original meaning or context.
The system is designed to accommodate a multilingual detoxification pipeline in which input text in toxic form is converted into its non-toxic equivalent for diferent languages. The system starts with a toxic sentence submitted by the user in one of the languages supported. This input is fed to a modular detoxification engine that uses several detoxification models concurrently. Each model processes the input independently and produces its detoxed equivalent. These outputs are then gathered and assessed to compare fluency, factuality consistency, and reduction of toxicity. The architecture was made such a way that it could be easily extended with more LLMs.
The framework has multiple functional layers. The Input Layer takes raw toxic sentences without preprocessing. In the Masking Layer, toxic words of the sentence are detected automatically and masked to maintain sentence structure with a focus on objectionable content. The Model Layer also takes the original and masked sentences as inputs to multilingual LLMs. The models give detoxified versions of text that try to remove toxicity while maintaining the original semantic intent.
The toxic dataset comprises 9,000 instances in the test set. For every toxic sentence, first of all, the toxic words are deleted, and they get replaced with the token ([MASK]). This gives a masked version of the original sentence, keeping the structure intact without ofensive content.
Both the original toxic sentence and its censored counterpart are fed into large language models
(LLMs) with few-shot prompting. The study used multilingual instruction-following LLMs such as
GPT4o-mini[
This hybrid approach produces efective detoxification across both high-resource and low-resource languages, without depending upon supervised fine-tuning or parallel corpora. Instead, it uses the generative capabilities of LLMs through few-shot prompting and contextual understanding. Refer to Fig.1 and Fig.2 for an overview of the masking and generation pipeline.
There are two major steps of our implementation as follows:
• Toxic Word Deletion(lexicon-based): This defines a class called Detoxification, which uses a
multilingual word list to filter out toxic words. The word list can either be loaded from a local file
or taken from the multilingual-toxic-lexicon dataset [
To evaluate the quality of detoxified outputs across languages, TextDetox 2025 shared task used three
major evaluation metrics : Style Transfer Accuracy, Content Preservation, and Fluency.
Style Transfer Accuracy assesses whether the output successfully transfers a toxic sentence into a
nontoxic sentence. For this, we employed a binary toxicity classifier on the basis of the XLM-Roberta-Large
model [
Content Preservation is the measure of how closely the semantic meaning of the original toxic sentence
is to the detoxified sentence. This is calculated as cosine similarity between LaBSE [
Where: =_(, , ) × (0.4 × (, )
+ 0.6 × (, )) × =
∑︀ _ _ + (_)
2 _ = __() _ = _ ≤ __ __ = __() (1) (2) (3) (4) (5)
For this shared task, the evaluation is also carried out in another way, namely LLM-as-a-Judge, a popular tool in recent times for the evaluation. In our submissions, this framework is used for the evaluation, using a LLaMA 3.1-8B-Instruct fine-tuned version, which is trained on human-annotated pairwise comparisons that have been taken from the TextDetox 2024 dataset. This fine-tuned model evaluates the results by comparing them in pairs, which would be more human-aligned judgments than traditional automatic metrics. For the fluency evaluation, the xCOMET model is used as always due to its strong correlation with human fluency assessments.
The sections demonstrate a detailed assessment of the multilingual text detoxification models on their performance in the 15 languages of the Multilingual ParaDetox dataset. The task ofer both quantitative outcomes based on traditional evaluation measures as well as qualitative examples of model behavior across varied linguistic and stylistic scenarios. The discussion further comments on model generalizability implications, and model efectiveness across low-resource languages.
baseline_mt0 baseline_gpt4 hybrid gpt-4o-mini baseline_o3mini baseline_gpt4o baseline_delete baseline_backtranslation baseline_duplicate Language(s) Average Score 9 supervised languages 9 supervised languages 9 supervised languages 9 supervised languages 9 supervised languages 9 supervised languages 9 supervised languages 9 supervised languages
Average Cross-lingual examples are presented in Table 5 to demonstrate the project’s capability to rephrase toxic inputs while preserving their original meaning and improving overall fluency. For highly toxic inputs, some outputs became vague or overly neutral. Using masked inputs helped the model focus better on toxic segments
This paper provides a study of multilingual text detoxification, introduced as part of the PAN 2025 shared task. Our hybrid method that integrates lexicon based toxic word masking with few-shot prompting of multilingual LLMs such as GPT-4o-mini showed robust detoxification performance on both high-resource domain as well as low-resource languages compared to baseline methods. Semantic consistency was preserved during the removal of explicit toxic content, as confirmed by strong results in style transfer accuracy, semantic preservation, and fluency. These outcomes were validated using both automated evaluation methods and assessments by large language models acting as judges.
The findings indicate that hybrid models using large generative models are able to generalize across widely varying linguistic patterns and cultural contexts even without large parallel corpora. This research provides the foundation for developing robust, scalable, and inclusive content moderation tools that can engage multilingual digital societies. Future work will focus on enhancing domain adaptation, managing implicit toxicity, and scaling the model to real-world applications.
All models were evaluated using three carefully selected metrics to comprehensively assess the efectiveness of detoxification: Style Transfer Accuracy (quantified with XLM-RoBERTa fine-tuned for toxicity detection), Content Preservation (cosine similarity of LaBSE embeddings), and Fluency (with the xCOMET model that agrees well with hu- man ratings on text quality) and also through LLM-as-a-Judge. A combined score function with these metrics enabled us to compare models fairly across languages.
The author(s) used Grammarly, ChatGPT, and Gemini for language editing and LaTeX table formatting. The content was reviewed and finalized by the author(s), who take full responsibility for it.