anyantudre/moore-image-lexicon

Mooré image lexicon

This dataset has 4,108 photos taken from three Mooré dictionaries. Each photo comes with the Mooré word and its French and English glosses. For 162 test photos there are also three wrong options for multiple-choice questions and a label that says whether the thing is local to West Africa. A second set of 222 household photos from Dollar Street is in household_set/.

I built the dataset to test how small vision-language models handle West African objects in English, French and Mooré. The notebook that builds the table is 01_moore_dictionary_curation in the bucket anyantudre/moore-vision-blind-spot.

Splits

split	photos
train	2,870
validation	618
test	620

Columns

One row is one photo with its dictionary entry.

column	meaning
id	the dictionary (webonary or lexique), the page and the position on the page, as in webonary-p0522-04
image	the photo at its printed size (median 186 by 139 pixels)
moore, moore_variants	the headword and its spelling variants
tone_raw	the tone marking printed in brackets in the dictionary, when there is one
french, english	the glosses as printed. Several senses can be joined in one string
pos	part of speech
category, category_code	the category given by the dictionary, such as 1.2 Animals
scientific_name	the Latin name, when the entry has one
page, column, box	where the entry sits in its PDF
link_confidence	how the photo was linked to the word: 0.98 for a caption under the photo on the same page, 0.85 for a caption on the previous page or shared by two words, 0.9 for a Lexique photo linked by its position next to the entry
siglip_rank	the share of 99 random glosses that SigLIP scores below the true gloss for this photo (1.0 is a perfect match)
flags	notes on how the row was built, such as category_predicted
sources, edition	which dictionaries contain the word
split	train, validation or test
evaluation_item	true for the 162 test photos used in the evaluation
wrong_english, wrong_french, wrong_moore	three wrong options from the same category. Filled for the 162 evaluation photos and for 918 train photos
locality	local, global or unsure, for the 162 evaluation photos

The 162 evaluation photos

They are the test photos that have a specific common name in English and French (no "plant sp." or "kind of tree"), a category of concrete things and a link confidence of at least 0.85. I removed 33 entries by hand that cannot be drawn, such as latitude, swim or "in an irregular manner".

The locality label has three values. Local means found mainly in West Africa and the Sahel, global means found everywhere (cow, hen, fish), and unsure means found across Africa or a cultural object. 13 photos are local, 114 are global and 35 are unsure. For 14 plants and animals with a scientific name the label comes from the species range in GBIF. I judged the other 148 from the gloss.

A broad word such as bird, reptile or tool is never an answer or a wrong option. A wrong option may not share a word with the answer in English or French, may not be of the same genus, and may not share a Mooré form with the answer. A plum, a plum tree and a black plum are never offered together. Entries that share a Mooré word, a rare gloss, a scientific name or a near-identical photo are always in the same split.

Household set

household_set/items.csv and household_set/images/ have 222 photos from Dollar Street in 17 categories, such as cooking pot, cooking fire, broom, toilet, roof, door and footwear. 145 photos are from West African homes (Burkina Faso, Côte d'Ivoire and Togo, with Ghana, Nigeria and Liberia as fallback) and 77 are from Western homes.

Each photo has the name of the variant that is there (clay pot, three-stone fire, pit latrine) in English, French and Mooré, lists of accepted English and French names, and three wrong options. For West African photos there is also a list of the terms that would name the Western look-alike. I chose the photos and wrote the names. A Mooré name is missing for some variants where I did not find the word in the dictionaries. The photos are CC BY-SA 4.0, credit Gapminder and MLCommons.

How the table was built

Words and glosses come mostly from the Webonary print of the dictionary (2021), where the caption under a photo is the headword. The Lexique (2017) adds entries and photos, and its Mooré letters are repaired with a map learned from the entries that both dictionaries share. The body-parts list (2020) only tells which entries it also contains. The details and the code are in the notebook.

I ran these checks:

• Of the 1,091 Lexique photos whose word also has a Webonary photo, 1,036 (95.0%) are the same picture as the Webonary photo.
• For 88.1% of the photos, SigLIP scores the true gloss above 90% of the random glosses.
• 404 entries had no category. I predicted it from their nearest neighbours in SigLIP space and marked them category_predicted. The same rule gets 92.4% of the entries that already have a category right.
• I dropped 146 linked photos that are smaller than 60 pixels or very elongated.

Limits

• The photo-word link is automatic. Some links are wrong, mostly for homographs and for captions that sit on the next page.
• The photos come from dictionaries and are small. Many are drawings or cut-outs and not photos of daily life, and some are plates with printed text, including names in other languages.
• The locality labels and the household names are my own judgment.
• The categories with the most entries are plants, the body, food, diseases and animals.

Sources and rights

The three dictionaries are compiled by Urs Niggli: Dictionnaire Mooré - français - English avec images (Webonary, 2021, © 2016 SIL International), Mooré Lexique comprimé (2017) and Parties du corps en mooré (2020). They are public PDFs. None states a licence and the photo credits are not given. I share the words and photos for research, with attribution, and I will remove anything the authors ask me to remove. The PDFs are in the bucket for reference.