That’s what I’m wondering about. I’m thinking split your test (unseen job descriptions) set into noun chunks and run the inference on each new job description separately. This way the model will classify each noun chunk