Anonyma

Anonyma will be a multilingual text anonymization tool designed to identify and mask sensitive information in textual data.

It will leverage transformer-based models to replace personally identifiable information (PII) with anonymized placeholders, ensuring privacy without compromising text usability. The project focuses on custom fine-tuning using LoRA (Low-Rank Adaptation) for efficiency and performance.

---

Planned Features

• Multilingual Support: Process text in multiple languages.
• Entity Recognition: Detect and anonymize entities such as names, dates, and locations.
• Customizable: Allow users to define which entities to anonymize.

---

Possible Models

To effectively handle both entity recognition and text transformation, Sequence-to-Sequence (Seq2Seq) transformer models are recommended:

1. T5 (Text-to-Text Transfer Transformer)
   • Description: A flexible model that can handle various NLP tasks by framing them as text-to-text problems. Suitable for multilingual anonymization with fine-tuning.
2. mBART (Multilingual BART)
   • Description: A multilingual variant of BART, designed for tasks involving multiple languages. Ideal for anonymizing text across diverse languages.
3. Non-Instruction Transformer SLMs (Qwen2.5, Phi-Series)
   • Description: Encoder-decoder Small Language Models optimized for NLP tasks without instruction tuning. These models provide consistent task-focused outputs without misinterpreting prompts.

---

Possible Datasets

To train and evaluate Anonyma, the following datasets from Hugging Face can be used:

1. PII Masking 200k

   • Description: Contains over 200,000 examples for training models to mask PII in English, French, German, and Italian.
   • PII Masking 200k Dataset

2. PII Masking 300k

   • Description: An expanded version with 300,000 examples across multiple languages.
   • PII Masking 300k Dataset