Figure 12.

The error in the tagger-corpora pairs have been categorized according to their relevance. The error types and counts from Figures 8 and 10 have been grouped into the categories “Core” (for a-PG and PGT), “Margin” (for 2C, ea-PG, BMT) and “Artefact” (for 1C, xa-PG, GE) representing different levels of relevance for the PGN tagging task. The FN results are on the left, the FP results are to the right, the upper part contains the absolute numbers and the lower figures have been normalized according to the total number of annotations in the corpora. Only the results for the cos98 evaluation are shown.