Enhancing Medical Named Entity Recognition with Features Derived from Unsupervised Methods

Creating the annotated corpus for training a named entity recognition model is expensive, particularly in specialised domains, such as medicine, which require expert annotators. Moreover, a model trained on text from one medical sub-domain often shows a drop in performance when applied on texts from another sub-domain, and annotated text from this other sub-domain might be required. When incorporating features from unsupervised methods, to what extent is it possible to: • Reduce the amount of annotated data needed to achieve a fixed level of performance? • Reduce the amount of additional annotated data needed for adapting a model to a new sub-domain?