Minimal sentence pairs are frequently used to analyze the behavior of language models. It is often assumed that model behavior on contrastive pairs is predictive of model behavior at large. We argue that two conditions are necessary for this assumption to hold: First, a tested hypothesis should be well-motivated, since experiments show that contrastive evaluation can lead to false positives. Secondly, test data should be chosen such as to minimize distributional discrepancy between evaluation time and deployment time. For a good approximation of deployment-time decoding, we recommend that minimal pairs are created based on machine-generated text, as opposed to human-written references. We present a contrastive evaluation suite for English–German MT that implements this recommendation.

Lexical disambiguation is a major challenge for machine translation systems, especially if some senses of a word are trained less often than others. Identifying patterns of overgeneralization requires evaluation methods that are both reliable and scalable. We propose contrastive conditioning as a reference-free black-box method for detecting disambiguation errors. Specifically, we score the quality of a translation by conditioning on variants of the source that provide contrastive disambiguation cues. After validating our method, we apply it in a case study to perform a targeted evaluation of sequence-level knowledge distillation. By probing word sense disambiguation and translation of gendered occupation names, we show that distillation-trained models tend to overgeneralize more than other models with a comparable BLEU score. Contrastive conditioning thus highlights a side effect of distillation that is not fully captured by standard evaluation metrics. Code and data to reproduce our findings are publicly available.