Getting More from Segmentation Evaluation

Martin Scaiano and Diana Inkpen
University of Ottawa


Abstract

The task of segmenting text is usually evaluated with WindowDiff but we have found WindowDiff has a number of limitations: near errors are masked, no separation of false positives or false negatives, no control over error tolerances. We present a new method of evaluation called WinPR, which resolves the previous issues and more. WinPR produces a confusion matrix from which precision, recall, and F-measure may be measured. We find that WinPR provides deeper insight into the functioning of segmentation algorithm, which can be important during selection or optimization of an algorithm.