The development of summarization systems requires reliable similarity (evaluation) measures that compare system outputs with human references. A reliable measure should have correspondence with human judgements. However, the reliability of measures depends on the test collection in which the measure is meta-evaluated; for this reason, it has not yet been possible to reliably establish which are the best evaluation measures for automatic summarization. In this paper, we propose an unsupervised method called Heterogeneity-Based Ranking (HBR) that combines summarization evaluation measures without requiring human assessments. Our empirical results indicate that HBR achieves a similar correspondence with human assessments than the best single measure for every observed corpus. In addition, HBR results are more robust across topics than single measures.
展开▼
