Global Explainability of BERT-Based Evaluation Metrics by Disentangling along Linguistic Factors

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8912–8925 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8912 Global Explainability of BERT-Based Evaluation Metrics by Disentangling along Linguistic Factors Marvin Kaster, Wei Zhao, Steffen Eger Natural Language Learning Group (NLLG) Technische Universit¨at Darmstadt, Germany marvin.kaster@stud.tu-darmstadt.de {zhao,eger}@aiphes.tu-darmstadt.de Abstract Evaluation metrics are a key ingredient for progress of text generation systems. In recent years, several BERT-based evaluation metrics have been proposed (including BERTScore, MoverScore, BLEURT, etc.) which correlate much better with human assessment of text generation quality than BLEU or ROUGE, in- vented two decades ago. However, little is known what these metrics, which are based on black-box language model representations, actually capture (it is typically assumed they model semantic similarity). In this work, we use a simple regression based global explain- ability technique to disentangle metric scores along linguistic factors, including semantics, syntax, morphology, and lexical overlap. We show that the different metrics capture all as- pects to some degree, but that they are all sub- stantially sensitive to lexical overlap, just like BLEU and ROUGE. This exposes limitations of these novelly proposed metrics, which we also highlight in an adversarial test scenario. 1 Introduction Evaluation metrics are a key ingredient in assess- ing the quality of text generation systems, be it machine translation, summarization, or conversa- tional AI models. Traditional evaluation metrics in machine translation and summarization, BLEU and ROUGE (Papineni et al., 2002a; Lin, 2004), have measured lexical n-gram overlap between system prediction and a human reference. While simple and easy to understand, early on, limitations of such lexical overlap metrics have been recognized (Callison-Burch et al., 2006), e.g., in that they can only measure surface level similarity, and they are especially inadequate when it comes to assessing current high-quality text generation systems (Rei et al., 2020; Mathur et al., 2020; Marie et al., 2021). Recently, a class of novel evaluation metrics based on BERT and its variants has been explored that correlates much better with human assessments of translation quality. For example, BERTScore (Zhang et al., 2020), MoverScore (Zhao et al., 2019), BLEURT (Sellam et al., 2020), XMover- Score (Zhao et al., 2020), and COMET (Rei et al., 2020) all use large-scale pretrained language mod- els, but differ in whether they compare hypotheses to references, to source texts, to both, on the one hand, and whether they use human scores for super- vision or not, on the other. Since these models all leverage large-scale language models which have pushed the state-of-the-art in many areas of NLP, their success comes with little surprise. To better understand these novel metrics based on black-box language representations is a prereq- uisite for identifying their limitations, e.g., to adver- sarial inputs. For example, if an evaluation metric is sensitive to lexical overlap, it can be fooled by using the same words but in different order. In this work, we fill the existing ‘explainability gap’ and introspect linguistic properties encoded in BERT-based evaluation metrics. Although there is already considerable …

Content truncated to 500 words