Reliability Crisis of Reference-free Metrics for Grammatical Error Correction

TL;DR

This paper reveals that popular reference-free grammatical error correction metrics are vulnerable to adversarial inputs that can artificially inflate scores. It presents targeted attack strategies against four metrics (SOME, Scribendi, IMPARA, and LLM-based metrics) and demonstrates, on the BEA-2019 development set, that adversarial systems can surpass current SOTA GEC systems. To mitigate this risk, the authors propose a naive metric ensemble that aggregates rankings across metrics, which dampens the impact of attacks and improves robustness. The work highlights the need for robustness-focused evaluation in GEC and discusses future directions for designing more reliable, multi-perspective evaluation frameworks.

Abstract

Reference-free evaluation metrics for grammatical error correction (GEC) have achieved high correlation with human judgments. However, these metrics are not designed to evaluate adversarial systems that aim to obtain unjustifiably high scores. The existence of such systems undermines the reliability of automatic evaluation, as it can mislead users in selecting appropriate GEC systems. In this study, we propose adversarial attack strategies for four reference-free metrics: SOME, Scribendi, IMPARA, and LLM-based metrics, and demonstrate that our adversarial systems outperform the current state-of-the-art. These findings highlight the need for more robust evaluation methods.

Figures (7)

• Figure 1: The situation we are concerned about. The adversarial attacking system may obtain an unreasonably high score by hacking a metric. This breaks the reliability of the automatic GEC evaluation.
• Figure 2: Illustrations of adversarial attack for SOME.
• Figure 3: Illustrations of adversarial attack for Scribendi.
• Figure 4: Illustrations of adversarial attack for IMPARA.
• Figure 5: A prompt example for LLM-S. Each corrected sentences are input as is.