Visual Context and AI Translation Accuracy in Chinese Museum Object Descriptions: An Exploratory Paired Comparison of Text Only and Text Plus Image Conditions

A photograph can make an object feature easier to inspect without resolving how that feature should be translated. This study examines that problem in 12 Chinese museum descriptions through a paired comparison of 24 retained English outputs from the National Museum of China. An adapted Multidimensional Quality Metrics scheme separates confirmed accuracy and terminology errors from unresolved judgements, recorded as reservations. The text-only condition contained 11 confirmed errors and 7 reservations; the text-plus-image condition contained 9 errors and 13 reservations. Accuracy errors numbered 9 in each condition, and terminology errors 2 and 0. Five samples had fewer confirmed errors with an image, five had the same count, and two had more. Case analysis shows contrasting changes: some outputs retain quantities and local forms more fully, while others omit details or substitute an animal category. Technical identification still depends on terminology and curatorial evidence. Checking was assisted by artificial intelligence (AI), with condition labels visible and no second independent evaluator. Because the comparison uses single outputs under uncontrolled generation conditions and leaves some judgements unresolved, it cannot establish an overall quality ranking or attribute the differences to image input. Its contribution is an account of specific translation decisions that require review.