NER may be sufficient, like you said it typically works best with full sentences but it is designed to handle entities within text (person, org). Aside from training/fine tuning a model to suit your data you may want to try out a BERT based model for zero-shot classification. I've found this to work fairly well for tasks like this.