Length Does Matter: Summary Length can Bias Summarization Metrics

AbstractEstablishing the characteristics of an effective summary is a complicated and often subjective endeavor. Consequently, the development of metrics for the summarization task has become a dynamic area of research within natural language processing. In this paper, we reveal that existing summarization metrics exhibit a bias toward the length of generated summaries. Our thorough experiments, conducted on a variety of datasets, metrics, and models, substantiate these findings. The results indicate that most metrics tend to favor longer summaries, even after accounting for other factors. To address this issue, we introduce a Bayesian normalization technique that effectively diminishes this bias. We demonstrate that our approach significantly improves the concordance between human annotators and the majority of metrics in terms of summary coherence.