Dynamic Meta-Metrics: Source-Sentence Conditioned Weighting for MT Evaluation

ACL ARR 2026 January Submission7663 Authors

Keywords: Machine Translation Evaluation

Abstract: We propose Dynamic Meta-Metrics (DMM), a framework for machine translation evaluation that learns source-sentence conditioned combinations of existing metrics. Unlike static meta-metrics that use a single set of weights per language pair or globally, DMM conditions the combiner on clusters of source segments. We study hard conditioning, which fits an interpretable combiner per cluster, and a soft-conditioned extension based on a gated mixture of linear experts. We evaluate DMM on WMT21--24 Metrics Shared Task data across multiple English-to-X language pairs using pairwise agreement measures at system and segment level. Across both average system-level SPA and average segment-level agreement, our DMM variants generally outperform competitive single-metric baselines, and often static meta-metrics.

Paper Type: Short

Research Area: Machine Translation

Research Area Keywords: Machine Translation Evaluation

Languages Studied: English, Czech, Chinese, Japanese

Submission Number: 7663