Adaptable transformer models via key term replacement

Abstract

Techniques are disclosed relating to storing, by a computer system, a map file that includes a new term that is mapped to a particular one of a plurality of category names that a natural-language processing (NLP) model is trained to identify, wherein the NLP model is not trained to identify the new term. The technique may further include receiving, by the computer system, user-generated content (UGC) that includes the new term. Additionally, the technique may include generating, by the computer system using the map file, a modified version of the UGC by substituting the particular category name for text in the UGC that is associated with the new term, and using, by the computer system, the NLP model on the modified text to classify the modified version of the UGC without retraining the NLP model to identify the new term.