Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets

Abstract Recent machine translation (MT) metrics calibrate their effectiveness by correlating with human judgment. However, these results are often obtained by averaging predictions across large test sets without any insights into the strengths and weaknesses of these metrics across different error types. Challenge sets are used to probe specific dimensions of metric behavior but there are very few such datasets and they either focus on a limited number of phenomena or a limited number of language pairs. We introduce ACES, a contrastive challenge set spanning 146 language pairs, aimed at discovering whether metrics can identify 68 translation accuracy errors. These phenomena range from basic alterations at the word/character level to more intricate errors based on discourse and real-world knowledge. We conducted a large-scale study by benchmarking ACES on 47 metrics submitted to the WMT 2022 and WMT 2023 metrics shared tasks. We also measure their sensitivity to a range of linguistic phenomena. We further investigate claims that large language models (LLMs) are effective as MT evaluators, addressing the limitations of previous studies by using a dataset that covers a range of linguistic phenomena and language pairs and includes both low- and medium-resource languages. Our results demonstrate that different metric families struggle with different phenomena and that LLM-based methods are unreliable. We expose a number of major flaws with existing methods: Most metrics ignore the source sentence; metrics tend to prefer surface level overlap; and over-reliance on language-agnostic representations leads to confusion when the target language is similar to the source language. To further encourage detailed evaluation beyond singular scores, we expand ACES to include error span annotations, denoted as SPAN-ACES, and we use this dataset to evaluate span-based error metrics, showing that these metrics also need considerable improvement. Based on our observations, we provide a set of recommendations for building better MT metrics, including focusing on error labels instead of scores, ensembling, designing metrics to explicitly focus on the source sentence, focusing on semantic content rather than relying on the lexical overlap, and choosing the right pre-trained model for obtaining representations.