Repositorio del proyecto desarrollado para la tarea PAN 2025 de Multilingual Text Detoxification. Su objetivo es generar versiones no tóxicas de textos ofensivos en varios idiomas manteniendo el contenido original. Incluye experimentos con modelos de lenguaje, evaluación y análisis de resultados.
This project focuses on text detoxification, transforming toxic or harmful texts into neutral versions that preserve their original meaning but remove any form of toxicity.
There are several possible approaches to achieve this goal:
- Masking and filling: identifying toxic words or expressions and replacing them with neutral alternatives.
- Text generation: generating a new text based on the toxic one, maintaining the same semantic content while modifying specific words and expressions.
- Hybrid approach (explained later in section Hybrid Approach).
Each approach has its own strengths and limitations. The following sections explore these ideas in detail.
This approach will likely involve two or more models.
First, a classifier will label each token as either toxic or non-toxic.
This can be done using two of the provided datasets:
datasets/multilingual_toxic_lexicon and datasets/multilingual_toxic_spans.
These datasets link tokens and expressions directly with their toxicity labels.
Once toxic tokens are detected, they will be masked, and a filling model will generate new words to replace them appropriately.
Advantages:
- Only the classifier needs to be trained; the filling model can often work without fine-tuning.
- Uses two out of the four available datasets.
Disadvantages:
- Limited flexibility - only modifies the tokens labeled as toxic.
- If additional words need to change for grammatical or semantic coherence, this method cannot handle it effectively.
In this approach, a text generation model will be fine-tuned (using techniques such as LoRA or QLoRA) with the dataset datasets/toxic_nontoxic.
During fine-tuning, toxic phrases will be used as inputs, and the model will learn to minimize the difference between the output embeddings and those of the corresponding non-toxic phrases.
This allows the model to generate non-toxic text conditioned on a toxic input.
Advantages:
- Highly flexible - works at the semantic level rather than purely lexical.
Disadvantages:
- Fine-tuning can be computationally expensive, especially with large models.
- Uses only one dataset.