BERTScore is Unfair: On Social Bias in Language Model-Based Metrics for Text Generation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3726 - 3739 December 7-11, 2022 ©2022 Association for Computational Linguistics BERTScore is Unfair: On Social Bias in Language Model-Based Metrics for Text Generation Tianxiang Sun♢♡∗ Junliang He♢♡∗ Xipeng Qiu♢♡† Xuanjing Huang♢♡ ♢School of Computer Science, Fudan University ♡Shanghai Key Laboratory of Intelligent Information Processing, Fudan University {txsun19,xpqiu,xjhuang}@fudan.edu.cn jlhe22@m.fudan.edu.cn Abstract WARNING: This paper contains examples that are offensive in nature. Automatic evaluation metrics are crucial to the development of generative systems. In recent years, pre-trained language model (PLM) based metrics, such as BERTScore (Zhang et al., 2020), have been commonly adopted in various generation tasks. However, it has been demon- strated that PLMs encode a range of stereotypi- cal societal biases, leading to a concern on the fairness of PLMs as metrics. To that end, this work presents the first systematic study on the social bias in PLM-based metrics. We demon- strate that popular PLM-based metrics exhibit significantly higher social bias than traditional metrics on 6 sensitive attributes, namely race, gender, religion, physical appearance, age, and socioeconomic status. In-depth analysis sug- gests that choosing paradigms (matching, re- gression, or generation) of the metric has a greater impact on fairness than choosing PLMs. In addition, we develop debiasing adapters that are injected into PLM layers, mitigating bias in PLM-based metrics while retaining high per- formance for evaluating text generation. 1 Introduction In text generation tasks, for example machine trans- lation, text summarization, and caption generation, automatic evaluation metrics are widely adopted for model selection. Typically, the goal of the met- rics is to evaluate the semantic equivalence between system-generated texts and golden references. Tra- ditional metrics such as BLEU (Papineni et al., 2002) and ROUGE (Lin, 2004) are usually based on n-gram matching, regardless of the semantic similarity. In recent years, pre-trained language models (PLMs) (Devlin et al., 2019; Lan et al., 2020; Yang et al., 2019; Raffel et al., 2020; Qiu ∗ Equal contribution. † Corresponding author. Data Generative Model Evaluation Metric Training Evaluation Model Selection Generation Human Bias Model Bias Metric Bias Pre-Trained Language Model Pre-Training As Initialize Intrinsic Bias Extrinsic Bias Figure 1: Impact of the social bias in PLM-based met- rics. The red arrows indicate the propagation of social bias in PLM-based metrics. et al., 2020) have been exploited for evaluating text generation. In contrast to traditional metrics that merely consider surface-form similarity, PLM-based met- rics such as BERTScore (Zhang et al., 2020) and BARTScore (Yuan et al., 2021) can well capture the semantic similarity between system outputs and ref- erences, and therefore achieve higher correlation with human judgements. Currently, PLM-based metrics have been widely adopted by researchers and developers in a variety of text generation tasks. Although these PLM-based metrics have been well studied from many perspectives such as robust- ness (Hanna and Bojar, 2021) and efficiency (Pu et al., 2021; Eddine et al., 2021), the fairness of these metrics has not yet been investigated. The fairness of the text generation metrics has …

Content truncated to 500 words