Fig. 2From: Customization scenarios for de-identification of clinical notesSystem performance as a function of number of labeled names. In each subfigure, an off-the-shelf system trained on dataset “A” (i2b2-2014) is partially customized using labeled examples from the target dataset “B”, then the system is evaluated on “B”. Training a system from scratch on “only B” is provided for comparisonBack to article page