On the Blind Spots of Model-Based Evaluation Metrics for Text Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12067–12097 July 9-14, 2023 ©2023 Association for Computational Linguistics On the Blind Spots of Model-Based Evaluation Metrics for Text Generation Tianxing He∗ Univ. of Washington goosehe@cs.w*.edu Jingyu Zhang∗ Johns Hopkins Univ. jzhan237@jhu.edu Tianle Wang Shanghai Jiao Tong Univ. wtl666wtl@sjtu.edu.cn Sachin Kumar Carnegie Mellon Univ. sachink@cs.cmu.edu Kyunghyun Cho New York Univ. kyunghyun.cho@nyu.edu James Glass Mass. Institute of Technology glass@mit.edu Yulia Tsvetkov Univ. of Washington yuliats@cs.washington.edu Abstract In this work, we explore a useful but often ne- glected methodology for robustness analysis of text generation evaluation metrics: stress tests with synthetic data. Basically, we design and synthesize a wide range of potential errors and check whether they result in a commen- surate drop in the metric scores. We examine a range of recently proposed evaluation met- rics based on pretrained language models, for the tasks of open-ended generation, translation, and summarization. Our experiments reveal interesting insensitivities, biases, or even loop- holes in existing metrics. For example, we find that BERTScore is confused by truncation er- rors in summarization, and MAUVE (built on top of GPT-2) is insensitive to errors at the be- ginning or middle of generations. Further, we investigate the reasons behind these blind spots and suggest practical workarounds for a more reliable evaluation of text generation. We have released our code and data at https://github. com/cloudygoose/blindspot_nlg. 1 Introduction Automatic evaluation of machine-generated text (Celikyilmaz et al., 2020) has been a core research challenge in the field of natural language gener- ation (NLG), as difficult as language generation itself. Encouraged by the phenomenal success of large-scale pretraining (Devlin et al., 2019), a re- cent series of work proposed to base evaluation met- rics on pretrained language models (PLMs) (Zhang et al., 2020; Yuan et al., 2021; Pillutla et al., 2021). For example, BERTScore (Zhang et al., 2020) com- putes a similarity score between the contextualized embeddings of the hypothesis and the reference text. PLM-based metrics have been shown to have higher correlations with human annotations for var- ious tasks (Yuan et al., 2021), and are becoming increasingly popular in practice. ∗ Equal contribution. Both are corresponding authors. w* in the email refers to washington. Figure 1: Motivation: The flaws of the underlying PLMs or certain design choices in the metrics could lead to potential blind spots in the evaluation. However, PLMs have flaws. They could assign a high likelihood to degenerate, repetitive text (Holtz- man et al., 2020) and could be insensitive to pertur- bations such as word order shuffling (Pham et al., 2021), negation (Ettinger, 2020), etc. These flaws, in combination with certain design choices, may lead to the metrics based on such PLMs being brit- tle and open to manipulation (Figure 1). In this work, we develop a suite of stress tests with synthetic data for the robustness analysis of NLG metrics. In essence, we induce a variety of potential errors in clean text and examine the re- sulting drop in the …

Content truncated to 500 words