Skip to content

Frequently Asked Questions

GDPR Pseudonymizer - Common questions and answers

Accuracy and Detection

What accuracy should I expect from the automatic detection?

The hybrid detection pipeline (NLP + regex + geography dictionary) automatically identifies approximately 60% of entities in French text (F1 59.97%). This is a pre-filtering step -- you review and confirm every entity during the mandatory validation workflow.

After human validation, accuracy is 100% because you control the final decision for every entity.

Why is the NER accuracy so low?

The spaCy fr_core_news_lg model was trained primarily on news text, not interview transcripts or business documents. Domain-specific language patterns (conversational registers, mixed formality) reduce out-of-the-box accuracy. A benchmark on 25 French documents with 1,855 entities measured 29.5% F1 for spaCy alone, improving to ~60% with the hybrid approach (F1 59.97%) after annotation cleanup, expanded regex patterns, and a French geography dictionary (Story 5.3).

Fine-tuning with real-world validation data is planned for v3.0 (targeting 70-85% F1).

Why is validation mandatory?

Because the AI detection misses entities. With ~80% recall, roughly one in five entities would go undetected without human review. For GDPR compliance, missing even one personal data entity could constitute a data breach. Mandatory validation ensures zero false negatives.

How do I process only specific entity types?

Use the --entity-types flag with a comma-separated list of types:

# Process only PERSON entities
gdpr-pseudo process doc.txt --entity-types PERSON

# Process PERSON and LOCATION (skip ORG)
gdpr-pseudo batch ./documents/ --entity-types PERSON,LOCATION

Valid types are PERSON, LOCATION, and ORG. When omitted, all entity types are processed. This flag works with both process and batch commands.

What is entity variant grouping?

Entity variant grouping automatically merges related forms of the same entity into a single validation item. For example, if a document contains "Marie Dubois", "Pr. Dubois", and "Dubois", the validation UI shows them as one item with "Also appears as:" listing the variant forms.

This reduces validation fatigue by eliminating redundant prompts. The grouping is type-aware:

  • PERSON: Title-stripped surname matching ("Dr. Dubois" = "Dubois"). Different first names stay separate ("Marie Dubois" vs "Jean Dubois").
  • LOCATION: French preposition stripping ("a Lyon" = "Lyon").
  • ORG: Case-insensitive matching ("ACME Corp" = "acme corp").

Document Formats and Languages

What document formats are supported?

Supported formats: - Plain text (.txt) - Markdown (.md) - PDF (.pdf) -- requires optional dependency - Microsoft Word (.docx) -- requires optional dependency - Excel (.xlsx) -- requires optional dependency (openpyxl) - CSV (.csv) -- no extra dependency needed

HTML and other formats are not yet supported. Convert those files to plain text before processing.

How do I process PDF, DOCX, Excel, or CSV files?

Install the optional format extras and process as usual:

# Install PDF/DOCX support
pip install gdpr-pseudonymizer[formats]

# Install Excel support
pip install gdpr-pseudonymizer[excel]

# Process a PDF
gdpr-pseudo process report.pdf

# Process a DOCX
gdpr-pseudo process interview.docx

# Process an Excel spreadsheet
gdpr-pseudo process data.xlsx

# Process a CSV file (no extra dependency needed)
gdpr-pseudo process data.csv

PDF and DOCX output is always plaintext (.txt). Excel and CSV inputs preserve their original format (.xlsx/.csv). Format preservation (PDF-to-PDF, DOCX-to-DOCX) is planned for a future release.

Excel notes: Formulas are read as cached values (not preserved in output). Each cell is processed independently through the NER pipeline. Legacy .xls format is not supported.

Can I use this for non-French documents?

No. v1.0 is designed exclusively for French-language text. The NLP model (fr_core_news_lg), regex patterns, and pseudonym libraries are all French-specific.

Multi-language support (English, Spanish, German) is planned for v3.0.

What languages are supported?

French only in v1.0. The tool uses spaCy's fr_core_news_lg model trained specifically on French text.

GDPR Compliance

Is this tool GDPR compliant?

GDPR Pseudonymizer supports GDPR compliance but does not guarantee it by itself. The tool implements:

  • Article 4(5): Pseudonymization with separate encrypted mapping storage
  • Article 25: Data protection by design (100% local processing, encrypted storage)
  • Article 32: Security measures (AES-256-SIV encryption, PBKDF2 key derivation)
  • Article 30: Processing records (comprehensive audit logging with export)

Important: Pseudonymized data is still personal data under GDPR (Recital 26). You remain the data controller and should consult your Data Protection Officer for specific compliance guidance.

See Methodology for the full GDPR compliance mapping.

Is pseudonymization the same as anonymization?

No. Pseudonymization replaces identifying information with pseudonyms but remains reversible (with the mapping table and passphrase). Pseudonymized data is still personal data under GDPR.

Anonymization irreversibly removes all identifying information so re-identification is impossible. GDPR does not apply to truly anonymous data.

GDPR Pseudonymizer performs pseudonymization, not anonymization. This is by design -- many use cases require reversibility (academic research follow-ups, legal proceedings, audit requirements).

Does my data leave my machine?

No. All processing happens locally. There are no cloud dependencies, API calls, telemetry, or network communication. After initial installation (downloading the spaCy model), the tool works completely offline.

GUI (Desktop Application)

How do I switch to dark mode?

Open View menu in the menu bar and select Dark theme. You can also change it in Settings > Appearance > Theme. Three themes are available: Light, Dark, and High Contrast. The preference is saved automatically.

Can I use the GUI without a mouse?

Yes. The GUI meets WCAG 2.1 Level AA accessibility standards with full keyboard navigation. Press Enter in the validation editor to start navigation mode, then use Tab/Shift+Tab to move between entities. See the GUI Guide for the complete keyboard shortcuts reference.

What's the difference between CLI and GUI?

Both use the same core processing engine. The CLI is text-based (terminal/command prompt) and ideal for scripting and automation. The GUI provides a visual interface with drag-and-drop, color-coded entity validation, batch progress dashboard, and database management. Choose based on your workflow preference.

Do I need Python for the GUI?

No. Standalone executables are available for Windows (.exe installer), macOS (.dmg), and Linux (.AppImage). Download from GitHub Releases. If you prefer PyPI: pip install gdpr-pseudonymizer[gui].

Comparison with Alternatives

How does this compare to manual redaction?

Aspect Manual Redaction GDPR Pseudonymizer
Time per document 15-30 minutes 2-5 minutes (AI pre-detection + validation)
Accuracy Human error prone 100% (mandatory validation)
Consistency Difficult across documents Guaranteed (same entity = same pseudonym)
Reversibility Not possible Yes (encrypted mapping table)
Audit trail Manual logging Automatic (operation logging, JSON/CSV export)
Cost for 50 documents 16-25 hours 2-3 hours

How does this compare to cloud pseudonymization services?

Aspect Cloud Services GDPR Pseudonymizer
Data location Third-party servers 100% local
Privacy risk Data exposure to provider Zero (no network)
Cost Per-document or subscription Free (open source)
Accuracy Varies (often no validation) 100% (mandatory validation)
Customization Limited Full (themed pseudonyms, config files)
Internet required Always Only for installation

How does this compare to fully automatic tools?

Aspect Automatic Tools GDPR Pseudonymizer
Human oversight None Mandatory
Accuracy Depends on NLP (often 70-90%) 100% (human validated)
Speed Faster (no validation) Slower but more accurate
GDPR defensibility Weaker (no human audit trail) Stronger (documented human review)
False negatives Possible Zero (by design)

Roadmap

What's planned for future versions?

v1.0 (Q1 2026): AI-assisted CLI with mandatory validation

v1.1 (Q1 2026): GDPR erasure, gender-aware pseudonyms, NER accuracy improvements, PDF/DOCX support, French docs

v2.0 (Q1 2026 — Current): Desktop GUI, standalone executables, WCAG AA accessibility, French UI, batch validation

v3.0 (2027+): NLP accuracy and automation - Fine-tuned French NER model (70-85% F1 target) - Optional --no-validate flag for high-confidence workflows - Multi-language support (English, Spanish, German)

Is there a GUI version?

Yes! v2.0 includes a full desktop GUI with drag-and-drop document processing, visual entity validation, batch processing dashboard, and standalone executables that don't require Python. See the GUI Guide for details.

Can I contribute?

Not yet. The project is in pre-release development. After v1.0 MVP launch, contributions will be welcome for bug reports, documentation improvements, translations, and feature suggestions.

GitHub: https://github.com/LioChanDaYo/RGPDpseudonymizer

Technical Questions

What NLP model is used?

spaCy fr_core_news_lg (version 3.8.0), a large French language model (~571MB). It was selected after benchmarking against Stanza, achieving 2.5x better F1 score on the test corpus.

How is the mapping database encrypted?

AES-256-SIV (deterministic authenticated encryption per RFC 5297) with PBKDF2-HMAC-SHA256 key derivation (210,000 iterations). Each database has a unique cryptographically random 32-byte salt. The passphrase is never stored on disk.

What happens if I forget my passphrase?

The mapping database cannot be decrypted without the correct passphrase. This means: - Existing pseudonym mappings are permanently inaccessible - You cannot reverse pseudonymization on previously processed documents - You must create a new database (poetry run gdpr-pseudo init --force)

Recommendation: Store your passphrase in a secure password manager.

Can I use different themes for the same project?

Technically yes, but it's not recommended. Different themes produce different pseudonyms for the same entity, creating inconsistencies. If you must switch themes, use a separate database for each theme.