COMET: A Neural Framework for MT Evaluation

TL;DR

COMET is a neural framework that leverages cross-lingual pretrained models to evaluate machine translation quality more accurately, achieving state-of-the-art correlation with human judgments across multiple datasets.

Contribution

The paper introduces COMET, a novel neural evaluation framework that utilizes multilingual pretrained models and source-reference information for improved MT quality assessment.

Findings

Achieves new state-of-the-art correlation with human judgments.

Demonstrates robustness across different types of human evaluation data.

Performs well on the WMT 2019 Metrics shared task.

Abstract

We present COMET, a neural framework for training multilingual machine translation evaluation models which obtains new state-of-the-art levels of correlation with human judgements. Our framework leverages recent breakthroughs in cross-lingual pretrained language modeling resulting in highly multilingual and adaptable MT evaluation models that exploit information from both the source input and a target-language reference translation in order to more accurately predict MT quality. To showcase our framework, we train three models with different types of human judgements: Direct Assessments, Human-mediated Translation Edit Rate and Multidimensional Quality Metrics. Our models achieve new state-of-the-art performance on the WMT 2019 Metrics shared task and demonstrate robustness to high-performing systems.