Figure 5

Decision surfaces of the VTT¹ (top) and VTT⁵ (bottom) classifiers with SP (left) and bigram (right) textual features, for the documents in one of the 4 cross-validation subsamples using the training data. The decision surfaces are plotted with the parameters in Table 3, and x(d) and y(d) are computed according to Eq. (7) for every document d. The plots for VTT¹ surfaces display many documents d with the same values of y(d), plotted in horizontal rows, while VTT⁵ displays a smoother ranking of documents. This happens because VTT¹ uses information from a single NER tool (ABNER protein mentions), while VTT⁵ uses information from five such tools; thus, while in the VTT¹ plot many documents have the same value of ABNER protein mentions, in the VTT⁵ plot the various NER measurements lead to a finer distinction between documents.