Spurious Correlations in Reference-Free Evaluation of Text Generation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1443 - 1454 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Spurious Correlations in Reference-Free Evaluation of Text Generation Esin Durmus1∗ Faisal Ladhak2∗ Tatsunori Hashimoto1 1Stanford University 2Columbia University esindurmus@cs.stanford.edu faisal@cs.columbia.edu thashim@stanford.edu Abstract Model-based, reference-free evaluation metrics have been proposed as a fast and cost-effective approach to evaluate Natural Language Genera- tion (NLG) systems. Despite promising recent results, we find evidence that reference-free evaluation metrics of summarization and di- alog generation may be relying on spurious correlations with measures such as word over- lap, perplexity, and length. We further observe that for text summarization, these metrics have high error rates when ranking current state-of- the-art abstractive summarization systems. We demonstrate that these errors can be mitigated by explicitly designing evaluation metrics to avoid spurious features in reference-free evalu- ation. 1 Introduction Building reliable automated evaluation metrics is a key factor for quick development of better NLG systems. Recent work has proposed reference-free evaluation metrics as a way to judge the quality of generated outputs without the need for human ref- erences (Celikyilmaz et al., 2020). Many of these reference-free evaluations achieve remarkably high correlations with human evaluations, raising hopes that they may soon become a viable alternative to expensive human evaluations (Kryscinski et al., 2020; Goyal and Durrett, 2020; Sinha et al., 2020; Phy et al., 2020; Gao et al., 2020). However, simply looking at correlation with hu- man scores may not be sufficient to determine the efficacy and robustness of an evaluation metric. In our work, we study recently proposed reference- free evaluation metrics of text summarization and dialog generation. We find that it is possible to achieve similar levels of correlation with human judgment, using simple spurious correlates such as word overlap, length, and perplexity. Further- more, we find that the learned metrics have a rela- ∗Equal contribution. tively high correlation with the spurious correlates as compared to human scores, which suggests that these metrics may rely heavily on spurious correla- tions. This may be a potential explanation for the robustness issues that are observed in recent work, despite the seemingly high reported correlations with human judgements (Gabriel et al., 2021; Yeh et al., 2021). We further analyze reference-free faithfulness evaluation metrics and show that the reliance on spurious correlations leads to errors in model se- lection and development. First, we show that word overlap, a spurious correlate for the task, does as well as recently proposed reference-free metrics at system-level ranking. Then, we look at rankings amongst systems that are relatively abstractive and faithful, i.e., the current state of the art, and find that these learned metrics perform significantly worse for these systems. This is because word-overlap is not a good measure for ranking these systems in terms of their faithfulness since all of these systems have similarly low word overlap. This suggests that we need metrics that are not overly reliant on word overlap in their faithfulness prediction. Finally, …

Content truncated to 500 words