IBS Software built a bilingual NER model for cargo logistics on Amazon Bedrock

The AWS Machine Learning Blog has published a technical breakdown of how IBS Software built a bilingual named entity recognition (NER) system for cargo logistics, using token-based knowledge distillation on Amazon Bedrock.

What NER is and why logistics needs it

Named Entity Recognition (NER) is a natural language processing task in which a model finds and classifies specific entities in text: company names, addresses, container numbers, dates, and port names. In cargo logistics, documents such as waybills, customs declarations, and instructions often arrive in different languages, and extracting this data by hand is labor-intensive and error-prone.

How token-based knowledge distillation works

Knowledge distillation is a technique in which a smaller model learns to reproduce the behavior of a larger one. The smaller model keeps most of the accuracy at significantly lower compute cost. The token-based approach applies distillation at the level of individual text tokens. That matters especially for NER, where labeling every word in a sentence correctly is critical.

• Technology — a bilingual NER model for cargo logistics
• Training method — token-based knowledge distillation
• Platform — Amazon Bedrock, using AWS knowledge distillation capabilities
• Case study author — IBS Software

By using knowledge distillation instead of deploying a full-size large language model, IBS Software gets acceptable entity recognition accuracy at a lower inference cost. That is critical for logistics companies that process large volumes of documents every day.

What this means

The IBS Software case shows a practical pattern for industries with highly specialized, multilingual data: you don't necessarily need to run a heavy general-purpose model in production. A compact distilled model trained for a specific task such as NER can handle it more efficiently and at lower cost.