MetricX-25 and GemSpanEval: Google Translate Submissions to the WMT25 Evaluation Shared Task

TL;DR

This paper introduces MetricX-25 and GemSpanEval, advanced models for translation quality assessment and error span detection, leveraging fine-tuned multilingual models to improve accuracy and interpretability in WMT25 shared task evaluations.

Contribution

The paper presents novel models MetricX-25 and GemSpanEval, enhancing translation evaluation by improving quality score prediction and error span detection using state-of-the-art multilingual models.

Findings

MetricX-25 outperforms its predecessor in quality score prediction.

GemSpanEval achieves competitive error span detection accuracy.

Error span detection as a generative task improves unambiguous identification.

Abstract

In this paper, we present our submissions to the unified WMT25 Translation Evaluation Shared Task. For the Quality Score Prediction subtask, we create a new generation of MetricX with improvements in the input format and the training protocol, while for the Error Span Detection subtask we develop a new model, GemSpanEval, trained to predict error spans along with their severities and categories. Both systems are based on the state-of-the-art multilingual open-weights model Gemma 3, fine-tuned on publicly available WMT data. We demonstrate that MetricX-25, adapting Gemma 3 to an encoder-only architecture with a regression head on top, can be trained to effectively predict both MQM and ESA quality scores, and significantly outperforms its predecessor. Our decoder-only GemSpanEval model, on the other hand, we show to be competitive in error span detection with xCOMET, a strong encoder-only sequence-tagging baseline. With error span detection formulated as a generative task, we instruct the model to also output the context for each predicted error span, thus ensuring that error spans are identified unambiguously.