Telecom-Specific Multilingual PII Detection & Anonymization
Overview
This service provides high-accuracy Personal Identifiable Information (PII) detection and masking specifically optimized for conversational data in the telecommunications sector. The service utilizes Generative LLMs (e.g., EuroLLM-22B, Mistral, Salamandra) for context-aware anonymization. The architecture is based on decoder-only LLM architectures (EuroLLM, Mistral, Salamandra). Training uses Parameter-Efficient Fine-Tuning (LoRA/PEFT) for large-scale generative models. A synthetic data generation pipeline produces domain-specific PII training data (handling telecom jargon and product- and service-specific conversations) from human-annotated real transcript seeds.
Outcome
A specialized PII detection model that enables secure data sharing and model training on anonymized transcripts. Outcomes include model weights, performance benchmarks, and standardized annotation guidelines for telecom data.
Service Card
| Category | AI products |
| Domain | Conversational Analytics, Cyber Security |
| Vertical | Digital Society |
| Target users | European Telecom Providers, Data Protection Officers |
| Service type | Fine-tuned Model (NER & Generative Masking) |
| Outputs | Specialized PII detection checkpoints (LoRA adapters or full weights) for training; token-level PII classification and masked/anonymized text output for inference. |
| Requirements | Hardware: NVIDIA GPU node (H100 80GB preferred). Memory: 50GB—100GB disk space per node for 22B/40B model storage. Software: Support for PEFT/LoRA libraries and containerized inference environments. |
| Limitations | Models must be deployed in-house by partners to ensure GDPR compliance before anonymization. |
| Availability | Q3/2027 |
| Licensing | Apache 2.0 (Large Models) |
| Visual content | assets/service-visuals/telecom-specific-multilingual-pii-detection-anonymization.png |
| Visual caption |