From the outset of automated generation of summaries, the difficulty of evaluation has been widely discussed. Despite many promising attempts, we believe it remains an unsolved problem. Here we present a method for scoring the content of summaries of any length against a weighted inventory of content units, which we refer to as a pyramid. Our method is derived from empirical analysis of human-generated summaries, and provides an informative metric for human or machine-generated summaries.
展开▼
