Dear mentors and contributors, I am very interested in the “ML-assisted Anonymization Layer and Targeted Pipeline Improvements for Greek Datasets” project and have been studying the proposal carefully. From a systems perspective, the core challenge appears to be designing a high-recall anonymization layer that remains robust under Greek morphological variation and OCR-induced noise, while integrating cleanly into the existing GlossAPI pipeline. Before drafting a proposal, I would appreciate clarification on a few technical points: 1. Are there existing annotated Greek NER datasets currently used or recommended for this project, or would part of the work involve curating or adapting such data? 2. Has there been prior evaluation of multilingual transformer-based NER models (e.g., XLM-R or similar) on your datasets, particularly under OCR noise conditions? 3. Do the datasets exhibit systematic OCR error patterns (e.g., accent loss, character confusion, token fragmentation) that should influence preprocessing design? 4. In terms of anonymization policy, is recall considered strictly more important than precision, or is there a defined tradeoff target? Additionally, I am curious whether LLM-based structured extraction approaches have been considered or experimentally evaluated (e.g., prompt-based entity extraction using multilingual instruction-tuned models), either as a primary method or as part of a hybrid architecture. I would be interested to understand whether such approaches have been explored and how they compare to token-classification models in terms of reliability and reproducibility within your pipeline constraints. My initial thinking is to explore a layered architecture combining: - Deterministic regex detection (emails, phone numbers), - Transformer-based NER for PERSON/ORG, - Noise-aware preprocessing, - Confidence-based masking and evaluation tooling, - Potentially a controlled LLM-assisted validation layer if appropriate. I would be glad to receive any guidance on preferred technical directions or constraints within the current GlossAPI codebase. Thank you for your time, and I look forward to following the discussion. Best regards, Emmanuel Adewumi
---- Λαμβάνετε αυτό το μήνυμα απο την λίστα: Λίστα αλληλογραφίας και συζητήσεων που απευθύνεται σε φοιτητές developers \& mentors έργων του Google Summer of Code - A discussion list for student developers and mentors of Google Summer of Code projects., https://lists.ellak.gr/gsoc-developers/listinfo.html Μπορείτε να απεγγραφείτε από τη λίστα στέλνοντας κενό μήνυμα ηλ. ταχυδρομείου στη διεύθυνση <gsoc-developers+unsubscribe [ at ] ellak [ dot ] gr>.