hg1702/privacy-corpus-filtering-demo

Multilingual Privacy-Aware Corpus Filtering

This project trains embedding-based classifiers that identify whether short text snippets are suitable for inclusion in a training corpus.

The task has three labels:

• —safe: general text with no personal information.
• —personal_data: direct identifiers such as names, emails, phone numbers, addresses, usernames, or account-like details.
• —sensitive_or_private_context: private or sensitive context such as health, family, workplace, finances, or personal life.

Research Question

Can multilingual sentence embeddings help classify whether a text snippet is safe, contains personal information, or is borderline/private before it enters a training corpus?

Why Embeddings?

Regex can detect structured identifiers such as emails and phone numbers, but it cannot reliably detect semantic privacy risk. Sentence embeddings represent the meaning of text, which allows lightweight classifiers to learn patterns such as personal stories, health disclosures, family conflicts, or workplace issues even when no direct identifier is present.

Dataset

The custom dataset is stored in data/privacy_corpus_filtering_dataset.jsonl.

It contains synthetic multilingual examples in English, Swedish, German, French, and Spanish. All examples are fictional and created for this assignment.

Dataset columns:

• —text
• —label
• —language
• —source_type

Planned Hugging Face dataset link:

https://huggingface.co/datasets/<your-hf-username>/privacy-corpus-filtering

Models

The training script evaluates two multilingual sentence embedding models:

• —sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
• —intfloat/multilingual-e5-small

For each embedding model, the following classifiers are trained:

• —Logistic Regression
• —Linear SVM
• —Random Forest
• —k-Nearest Neighbors

Planned Hugging Face model link:

https://huggingface.co/<your-hf-username>/privacy-corpus-filtering-classifier

Demo

The demo is a Gradio app in app.py. It predicts the privacy label for a text snippet and shows the nearest training examples for basic interpretability.

Planned Hugging Face Space link:

https://huggingface.co/spaces/<your-hf-username>/privacy-corpus-filtering-demo

Setup

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Train and Evaluate

python src/train.py
python src/export_best_model.py

Training outputs are written to outputs/. The exported demo model is written to models/.

Run the Demo Locally

python app.py

Repository Structure

.
+-- app.py
+-- data/
|   +-- README.md
|   +-- privacy_corpus_filtering_dataset.jsonl
+-- reports/
|   +-- report_draft.md
+-- requirements.txt
+-- src/
    +-- export_best_model.py
    +-- predict.py
    +-- train.py

Limitations

The dataset is small and synthetic, so the model should be understood as a course project prototype rather than a production privacy system. A real deployment would require larger multilingual data, stricter annotation guidelines, legal review, and human-in-the-loop evaluation.