MMID: Massively Multilingual Image-Word Parallel Dataset

University of Pennsylvania researchers created a large-scale dataset of images paired with words in multiple languages. The dataset is doubly parallel, with words aligned to their representative images and to English translations with corresponding images. It is hosted on AWS Open Data.

Use Cases

• Train cross-lingual image retrieval models based on the parallel word-image alignments.
• Develop multilingual visual concept classifiers based on the image-word pairs.
• Benchmark machine translation or vision-language models based on the doubly parallel structure.
• Conduct research on visual semantics across languages based on the parallel translations.

Strengths

• Dataset is described as 'large-scale' and 'massively multilingual'.
• Features a doubly parallel structure aligning words, images, and English translations.

Limitations

• Row count, file size, and column-level documentation are unknown, limiting suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Penn NLP at the University of Pennsylvania.

Collection Method

Words and images were collected and paired, with translations into English.

Time Range

null

Freshness

Last updated date is unknown.

Geography

null