A 2025 study using responses across TIMSS languages presented a validation and quality-control approach for multilingual automated scoring, including machine translation and neural scoring, and documented translation failures that required changing methods for Chinese, Turkish and Armenian responses.

What this does not establish

Technical validation on historical assessment data does not authorize unsupervised high-stakes scoring or prove equal validity for every language and subgroup.

Counterevidence & uncertainty

Translation/model versions and response distributions can change performance; language-specific error review remains necessary.

What would change the reading

Track operational audits, subgroup validity, drift and human adjudication rates.

Primary routes

1. EL-S015https://doi.org/10.1016/j.caeai.2025.100375Open source record →