receiving a few-shot natural-language training data comprising a new class for retraining an original Named Entity Recognition (NER) model, the NER model trained to identify a plurality of other classes using non-few-shot natural language training data;

generating, through a model inversion of the original NER model, synthetic training data that represents each of the plurality of other classes; and

forming a retrained NER model by retraining the original NER model to identify text in the plurality of other classes and the new class using the synthetic training data and the few-shot natural-language training data, wherein the retrained NER model is trained through a distillation that uses the original NER model; and

storing the retrained NER model.