ReciLex
A multilingual culinary lexicon containing approximately 15,000 translated entities across ten categories such as ingredients, utensils and chef actions, injected as external knowledge to enhance the cultural adaptation of recipe translation for low-resource languages.
Timeline (last 90 days)
The paper proposes ReciLex, a multilingual culinary lexicon containing about 15,000 translated entities spanning ten recipe entity categories.
The role of ReciLex is to serve as injected external knowledge rather than to replace model training.
The core question of the paper is whether lexicon augmentation alone, without large-scale collection of target-language recipe corpora and without fine-tuning, can improve the quality of culturally adapted translation.
On the selected models and languages, ReciLex augmentation raised average BLEU from 13.57 to 14.51.
On the selected models and languages, ReciLex augmentation raised chrF from 37.89 to 38.74.
On the selected models and languages, ReciLex augmentation lowered TER from 77.14 to 75.12.
The study focuses on three Nigerian languages and the specific domain of recipes, and whether the conclusions transfer to other low-resource languages or text types still requires subsequent verification.
All Facts (7)
The paper proposes ReciLex, a multilingual culinary lexicon containing about 15,000 translated entities spanning ten recipe entity categories.
50%UnverifiedThe role of ReciLex is to serve as injected external knowledge rather than to replace model training.
50%UnverifiedThe core question of the paper is whether lexicon augmentation alone, without large-scale collection of target-language recipe corpora and without fine-tuning, can improve the quality of culturally adapted translation.
50%UnverifiedOn the selected models and languages, ReciLex augmentation raised average BLEU from 13.57 to 14.51.
50%UnverifiedOn the selected models and languages, ReciLex augmentation raised chrF from 37.89 to 38.74.
50%UnverifiedOn the selected models and languages, ReciLex augmentation lowered TER from 77.14 to 75.12.
50%UnverifiedThe study focuses on three Nigerian languages and the specific domain of recipes, and whether the conclusions transfer to other low-resource languages or text types still requires subsequent verification.
50%