Evaluating an Evaluation Method: The Pyramid Method Applied to 2003 Document Understanding Conference (DUC) Data

摘要

A pyramid evaluation dataset was created for DUC 2003 in order to compare results with DUC 2005, and to provide an independent test of the evaluation metric. The main differences between DUC 2003 and 2005 datasets pertain to the document length, cluster sizes, and model summary length. For five of the DUC 2003 document sets, two pyramids each were constructed by annotators working independently. Scores of the same peer using different pyramids were highly correlated. Sixteen systems were evaluated on eight document sets. Analysis of variance using Tukey's Honest Significant Difference method showed significant differences among all eight document sets, and more significant differences among the sixteen systems than for DUC 2005.