Wmt Human All: Human Translation Quality Scores from WMT Campaigns

A continuously updated compilation of human translation quality assessments from past WMT (Workshop on Machine Translation) campaigns. The dataset merges scores from different annotation protocols onto a semi-unified 0 to 100 scale and includes human scores up to WMT 2025. It was created by the author 'zouharvi' and last updated on January 20, 2026.

Use Cases

• Benchmarking machine translation systems based on aggregated human assessment scores.
• Training quality estimation models using the unified 0-100 scale derived from multiple protocols.
• Analyzing trends in human evaluation across different WMT campaign years.
• Comparing the effectiveness of different annotation protocols (DA, MQM, ESA) for translation assessment.

Strengths

• Continuously updated, with the current version including scores up to WMT 2025.
• Merges scores from three distinct annotation protocols (DA, MQM, ESA) onto a common scale.
• Last updated on 2026-01-20, suggesting recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

Human assessments from WMT (Workshop on Machine Translation) campaigns.

Collection Method

Compilation and merging of scores from different annotation protocols using a custom script.

Time Range

Up to and including WMT 2025.

Freshness

Last updated 2026-01-20 08:45:45.