Dynamic Meta-Metrics: Source-Sentence Conditioned Weighting for MT Evaluation
Published in Association for Computational Linguistics (ACL) Student Research Workshop 2026, 2026
Dynamic Meta-Metrics learns source-sentence conditioned combinations of existing machine translation metrics. Instead of using one static ensemble of metric weights, the framework adapts metric weighting based on source-segment properties and evaluates these combinations across WMT Metrics Shared Task data.
Recommended citation: Luke Zhang, Justin Vasselli, Aditya Khan, York Hay Ng, and En-Shiun Annie Lee. 2026. Dynamic Meta-Metrics: Source-Sentence Conditioned Weighting for MT Evaluation. In Proceedings of the Association for Computational Linguistics (ACL) Student Research Workshop 2026.
Download Paper