Performance values for models that use combinations of text, positional and linkage features.

Values calculated across the whole-corpus as well as per-article averages are shown. Each row gives the performance measures for one model. The T, P and L columns indicate with a those models that use text, positional, and linkage features, respectively. The sentences-in-states CRF approach, CRF (SIS), has the top performance on all measures. Asterisks denote -values (*denotes and **denotes ) from paired -tests comparing the per-article performance of our methods to those of Baseline+DM, the current state-of-the-art.