Hacking Neural Evaluation Metrics with Single Hub Text
Hiroyuki Deguchi, Katsuki Chousa, Yusuke Sakai
TL;DR
Embedding-based neural evaluation metrics correlate with human judgments but are vulnerable to hubness in discrete text space. The authors present a three-step pipeline—hub training in continuous space, hub decoding via an inversion model, and local search over the vocabulary—to produce hub texts that consistently receive high COMET scores across inputs. In WMT'23/24 En--Ja and En--De, a single hub text outperformed M2M100 translations on COMET scores and generalized to cross-language pairs, highlighting reliability and safety concerns for single-metric evaluation. The work argues for sanity checks and multi-metric evaluation to mitigate exploitation of metric vulnerabilities and provides NP-hardness and practical complexity analyses.
Abstract
Strongly human-correlated evaluation metrics serve as an essential compass for the development and improvement of generation models and must be highly reliable and robust. Recent embedding-based neural text evaluation metrics, such as COMET for translation tasks, are widely used in both research and development fields. However, there is no guarantee that they yield reliable evaluation results due to the black-box nature of neural networks. To raise concerns about the reliability and safety of such metrics, we propose a method for finding a single adversarial text in the discrete space that is consistently evaluated as high-quality, regardless of the test cases, to identify the vulnerabilities in evaluation metrics. The single hub text found with our method achieved 79.1 COMET% and 67.8 COMET% in the WMT'24 English-to-Japanese (En--Ja) and English-to-German (En--De) translation tasks, respectively, outperforming translations generated individually for each source sentence by using M2M100, a general translation model. Furthermore, we also confirmed that the hub text found with our method generalizes across multiple language pairs such as Ja--En and De--En.