Cross-cultural Adaptation of Recipes for Low-Resource Languages: A Benchmark and Multilingual Culinary Lexicon for Hausa, Yorùbá, and Igbo
Organizations: Department of Computer Science, The University of Manchester, United Kingdom
Abstract
Recent multilingual large language models (LLMs) have improved machine translation, yet culturally adaptive translation for low-resource languages (LRLs) remains underexplored. This is particularly challenging for recipes, where ingredients, cooking tools, measurements, and procedures may require adaptation to the target cultural context. In this work, we introduce ReciLoReL, a benchmark of 600 expert-translated and culturally adapted recipes across Yorùbá, Hausa, and Igbo (three of the most prominent languages in Nigeria), and ReciLex, a multilingual culinary lexicon containing approximately 15,000 translated entities across ten recipe entity categories, including ingredients, tools, chef actions, state of tools, and durations. We evaluate state-of-the-art multilingual LLMs for English-to-LRL recipe translation under zero-shot and few-shot settings and investigate whether augmenting LLMs with ReciLex improves culturally adaptive recipe translation without requiring large target-language recipe corpora for training or fine-tuning. Beyond conventional reference-based translation evaluation using BLEU, chrF, and TER, we introduce five criteria for evaluating cultural adaptation, adapted from established cross-cultural adaptation methods. Across the selected LLMs and languages, ReciLex improves average BLEU from 13.57 to 14.51 and chrF from 37.89 to 38.74, while reducing TER from 77.14 to 75.12. Human evaluation further shows improvements ranging from 0.11 to 0.36 points across the five cultural adaptation evaluation criteria on a five-point scale. Overall, our work provides new resources, evaluation criteria, and empirical evaluation for studying culturally adaptive recipe translation in low-resource settings.
Figures & tables
| Element | Adapt when | Preserve |
|---|---|---|
| Ingredients | Unavailable/unfamiliar | Culinary function |
| Measurements | Convention differs | Quantity/proportion |
| Tools/equipment | Unavailable/uncommon | Intended function |
| Methods/procedures | Impractical/unfamiliar | Preparation logic |
| Culinary expressions | Unnatural when translated | Intended meaning |
| Language | Without ReciLex | ReciLex | ||||
|---|---|---|---|---|---|---|
| BLEU | chrF | TER | BLEU | chrF | TER | |
| Yorùbá | 8.14 | 29.48 | 83.93 | 2.36 | 17.23 | 96.63 |
| Hausa | 16.29 | 43.55 | 71.03 | 21.28 | 52.76 | 58.81 |
| Igbo | 16.28 | 40.64 | 76.43 | 19.90 | 46.24 | 69.93 |
| Language | Method | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|---|
| Yorùbá | Without ReciLex | 3.29 | 3.17 | 3.23 | 2.48 | 2.98 |
| ReciLex | 4.13 | 3.55 | 3.69 | 3.03 | 3.34 | |
| Hausa | Without ReciLex | 2.94 | 2.88 | 3.07 | 2.97 | 2.68 |
| ReciLex | 3.07 | 2.72 | 2.93 | 3.05 | 2.66 | |
| Igbo | Without ReciLex | 2.81 | 3.03 | 2.61 | 2.33 | 2.58 |
| ReciLex | 2.93 | 3.15 | 2.73 | 2.41 | 2.67 |
| Setting | BLEU | chrF | TER |
|---|---|---|---|
| Without ReciLex | 13.57 | 37.89 | 77.14 |
| ReciLex | 14.51 | 38.74 | 75.12 |
| Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|
| Without ReciLex | 3.01 | 3.03 | 2.97 | 2.59 | 2.75 |
| ReciLex | 3.37 | 3.14 | 3.12 | 2.83 | 2.89 |
| Setting | BLEU | chrF | TER |
|---|---|---|---|
| Zero-shot + ReciLex | 10.70 | 34.96 | 78.78 |
| Few-shot + ReciLex | 18.32 | 42.53 | 71.47 |
| Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|
| Zero-shot + ReciLex | 3.30 | 3.11 | 3.07 | 2.78 | 2.85 |
| Few-shot + ReciLex | 3.45 | 3.17 | 3.16 | 2.88 | 2.92 |
| Lang. | Model | Shot | Without ReciLex | ReciLex |
|---|---|---|---|---|
| Yorùbá | GPT-5.1 | Zero-Shot | 8.84 | 79.42 |
| Few-Shot | 12.17 | 75.24 | ||
| Llama-70B | Zero-Shot | 1.90 | 80.74 | |
| Few-Shot | 7.97 | 71.15 | ||
| Gemma-27B | Zero-Shot | 3.43 | 76.62 | |
| Few-Shot | 11.82 | 72.73 |
| Language | Prompt | NFC changes | Marks/100 | ||
|---|---|---|---|---|---|
| Without ReciLex | ReciLex | Without ReciLex | ReciLex | ||
| Hausa | Zero-Shot | 0/80 | 0/80 | 0.00 | 0.00 |
| Hausa | Few-Shot | 0/80 | 0/80 | 0.00 | 0.00 |
| Igbo | Zero-Shot | 3/80 | 2/80 | 0.03 | 0.00 |
| Igbo | Few-Shot | 0/80 | 0/80 | 0.00 | 0.00 |
| Yorùbá | Zero-Shot | 17/80 | 80/80 | 1.96 | 11.30 |
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | Setting | BLEU | chrF | TER |
|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 11.65 | 34.65 | 83.70 |
| ReciLex | 4.05 | 19.70 | 98.40 | |
| Llama-70B | Without ReciLex | 5.15 | 24.85 | 86.05 |
| ReciLex | 1.95 | 17.50 | 94.35 | |
| Gemma-27B | Without ReciLex | 7.60 | 28.95 | 82.05 |
| ReciLex | 1.05 | 14.50 | 97.15 |
| Model | Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 3.68 | 3.60 | 3.87 | 3.04 | 3.44 |
| ReciLex | 4.28 | 3.85 | 4.10 | 3.34 | 3.63 | |
| Llama-70B | Without ReciLex | 3.04 | 2.88 | 2.85 | 2.12 | 2.72 |
| ReciLex | 4.10 | 3.39 | 3.50 | 2.94 | 3.20 | |
| Gemma-27B | Without ReciLex | 3.16 | 3.03 | 2.98 | 2.27 | 2.78 |
| ReciLex | 4.00 | 3.42 | 3.47 | 2.81 | 3.18 |
| Model | Setting | BLEU | chrF | TER |
|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 13.30 | 43.75 | 76.80 |
| ReciLex | 19.45 | 52.30 | 60.95 | |
| Llama-70B | Without ReciLex | 16.75 | 41.40 | 69.40 |
| ReciLex | 22.90 | 53.25 | 58.15 | |
| Gemma-27B | Without ReciLex | 18.80 | 45.50 | 66.90 |
| ReciLex | 21.50 | 52.75 | 57.30 |
| Model | Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 3.43 | 3.51 | 3.68 | 3.33 | 3.21 |
| ReciLex | 3.03 | 2.72 | 2.92 | 3.07 | 2.72 | |
| Llama-70B | Without ReciLex | 2.50 | 2.23 | 2.56 | 2.60 | 2.06 |
| ReciLex | 3.04 | 2.69 | 2.90 | 3.01 | 2.58 | |
| Gemma-27B | Without ReciLex | 2.88 | 2.91 | 2.98 | 2.98 | 2.77 |
| ReciLex | 3.13 | 2.74 | 2.99 | 3.06 | 2.67 |
| Model | Setting | BLEU | chrF | TER |
|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 16.45 | 43.10 | 81.25 |
| ReciLex | 20.25 | 47.10 | 71.90 | |
| Llama-70B | Without ReciLex | 14.40 | 37.55 | 75.90 |
| ReciLex | 18.90 | 45.00 | 69.70 | |
| Gemma-27B | Without ReciLex | 18.00 | 41.30 | 72.20 |
| ReciLex | 20.55 | 46.60 | 68.20 |
| Model | Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|---|
| GPT-5.1 | Without ReciLex | 3.06 | 3.20 | 2.79 | 2.44 | 2.73 |
| ReciLex | 3.16 | 3.37 | 2.91 | 2.60 | 2.88 | |
| Llama-70B | Without ReciLex | 2.54 | 2.82 | 2.40 | 2.19 | 2.39 |
| ReciLex | 2.70 | 3.04 | 2.58 | 2.29 | 2.58 | |
| Gemma-27B | Without ReciLex | 2.83 | 3.09 | 2.64 | 2.35 | 2.62 |
| ReciLex | 2.94 | 3.05 | 2.70 | 2.35 | 2.56 |
| Model | Evaluation | BLEU | chrF | TER |
|---|---|---|---|---|
| GPT-5.1 | Original | 4.06 | 19.70 | 98.41 |
| NFC | 6.11 | 29.11 | 93.39 | |
| Diacritic-insensitive | 13.47 | 42.64 | 81.46 | |
| Llama-70B | Original | 1.95 | 17.50 | 94.32 |
| NFC | 2.99 | 25.55 | 90.79 | |
| Diacritic-insensitive | 7.50 | 36.82 | 82.77 |
| System | BLEU | chrF | TER |
|---|---|---|---|
| LLMs (Zero-shot) | 6.36 | 29.80 | 84.93 |
| LLMs (Zero-shot + ReciLex) | 10.70 | 34.96 | 78.78 |
| M2M100-1.2B | 0.90 | 13.57 | 96.20 |
| NLLB-200-600M | 6.83 | 26.43 | 89.47 |
| System | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|
| LLMs (Zero-shot) | 2.76 | 2.84 | 2.84 | 2.42 | 2.49 |
| LLMs (Zero-shot + ReciLex) | 3.30 | 3.11 | 3.07 | 2.78 | 2.85 |
| M2M100-1.2B | 1.66 | 1.58 | 1.43 | 1.33 | 1.19 |
| NLLB-200-600M | 2.66 | 2.57 | 2.50 | 2.22 | 2.19 |
| Setting | BLEU | chrF | TER |
|---|---|---|---|
| Zero-shot | 8.53 | 32.38 | 81.86 |
| Few-shot | 19.55 | 44.26 | 70.41 |
| Setting | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|
| Zero-shot | 3.03 | 2.98 | 2.96 | 2.60 | 2.67 |
| Few-shot | 3.36 | 3.19 | 3.13 | 2.82 | 2.96 |
| Prompt Language | BLEU | chrF | TER |
|---|---|---|---|
| Target LRL | 14.04 | 38.32 | 76.13 |
| English | 15.02 | 39.61 | 75.74 |
| Prompt | INGR | PROC | SEM | CULT | LING |
|---|---|---|---|---|---|
| Target LRL | 3.18 | 3.11 | 3.06 | 2.77 | 2.83 |
| English | 3.22 | 3.15 | 3.06 | 2.76 | 2.92 |
| Language | Shots | BLEU | chrF | TER |
|---|---|---|---|---|
| Yorùbá | 3-shot | 13.39 | 38.26 | 82.65 |
| 5-shot | 16.51 | 41.13 | 78.64 | |
| 10-shot | 16.28 | 39.31 | 79.52 | |
| Hausa | 3-shot | 20.09 | 50.91 | 72.39 |
| 5-shot | 19.83 | 49.65 | 69.87 | |
| 10-shot | 20.33 | 51.66 | 71.34 |
| Category | Configuration | Total |
| Overall annotation coverage | ||
| Source recipes | Unique recipes | 80 |
| Translated recipes | All configurations | 3,900 |
| Translation system | ||
| System | LLM translations | 3,420 |
| MT translations | 480 | |
| Evaluation Dimension | Yorùbá | Hausa | Igbo |
|---|---|---|---|
| Ingredient localisation | 0.631 | 0.747 | 0.831 |
| Procedural localisation | 0.644 | 0.697 | 0.816 |
| Semantic preservation | 0.556 | 0.694 | 0.844 |
| Cultural authenticity | 0.650 | 0.766 | 0.878 |
| Linguistic naturalness | 0.669 | 0.772 | 0.881 |
| Average | 0.630 | 0.735 | 0.850 |
| Recipe Details | English Source | Translation | CA Criterion | Error Analysis |
| Chicken Casserole (Hausa, Few-Shot) | Mix soup and milk together; stir until smooth. Fold in sour cream and stir until smooth. | ReciLex: Ka cakua Miya da Madara tare; ka motsa har sai ta zama lallausa. Sai ka lankafa Madara mai tsami ka motsa har sai ta zama lallausa. Without ReciLex: Ka haa miya (soup) da madara a wuri guda; ka juya har su yi laushi su zama aya. Ka ara sour cream ka juya har ya yi laushi sosai. | INGR | ReciLex replaces the untranslated sour cream with the Hausa description Madara mai tsami . This improves lexical localisation, although the expression is less specific than the original ingredient. |
| Fish A La Creme (Igbo, Zero-Shot) | Marinate fish in lemon juice for 10 minutes. Mix butter, soup and sour cream. Pour over fish. | ReciLex: Tee ngwa nri azụ n’ime mmanụ lemuni (lemon juice) maka nkeji 10. Gwakọ bọta, supu na sour cream. Gbanye ya n’elu azụ. Without ReciLex: Tinye anụ ọhịa n’ime mmanya ọsịsa lemon ka ọ mara na ya nkeji 10. Kpọọ ọnụ ma kụọ butter, soup na sour cream. | SEM | The baseline incorrectly translates fish as anụ ọhịa (wild animal). ReciLex preserves the intended ingredient using azụ (fish), although its rendering of lemon juice remains semantically inaccurate. |
| Hotel Roanoke Spoon Bread (Igbo, Zero-Shot) | Combine water, butter, sugar and salt in a saucepan; bring to a boil. Stir in cornmeal and cook until thickened. | ReciLex: Gwakọta mmiri, bọta, shuga na nnu n’ite sọọsụ nta; ruo ọsụsụ. Gbanye cornmeal ma gbọọ agbọ Nkeji otu. Without ReciLex: Tinye mmiri, butter, sugar na salt n’ime obere saucepan; kụdaa ya ruo mgbe ọ ga-abụ boil. Tinye cornmeal ma kụdaa ya. | LING | The baseline retains several English terms, including butter , sugar , salt , and boil . ReciLex provides more localised lexical forms, reducing code-mixing, although cornmeal remains untranslated. |
| Shrimp Fried Rice (Yorùbá, Few-Shot) | Add green onions, garlic, shrimp and peas. Stir-fry for 2 minutes. | ReciLex: Fi àlùbọ́sà eléwé, àlùbósà ááyù, edé àti ẹ̀pà kun un. Yan lásán fún ìṣẹ́jú méjì. Without ReciLex: Fi alubosa àtàrẹ aláwọ̀ ewé, allaró (garlic), shrimpu àti èpà (peas) sí i. Stir-fry fún ìṣẹ́jú méjì. | CULT | ReciLex replaces the English-derived shrimpu with edé , a more recognisable Yorùbá culinary term. However, its rendering of stir-fry is less procedurally precise. |
| Fish A La Creme (Hausa, Few-Shot) | Marinate fish in lemon juice for 10 minutes. Mix butter, soup and sour cream. Pour over fish. Broil for 10 minutes. | ReciLex: Ka tsoma a ruwa Kifi a cikin Kunun lemun tsami na tsawon daia goma. Ka cakua Man shanu, Miya da Madara mai tsami. Ka zuba a kan Kifi. Ka tafasa na daia goma. Without ReciLex: Ka jika kifin da ruwan lemun tsami na tsawon minti goma. Ka haa bota, miya (soup) da madarar sour cream. Ka zuba a kan kifin. Ka soya a arashin broiler na tsawon minti goma. | PROC | ReciLex changes both ten-minute durations to ten seconds and replaces broiling with boiling . Despite localising some ingredient terms, it alters the cooking method and duration, making the instructions unreliable. |
| Statistic | English | Yorùbá | Hausa | Igbo |
| Recipes | 200 | 200 | 200 | 200 |
| Avg. title words | 3.10 | 3.80 | 3.80 | 3.90 |
| Avg. title characters | 20.90 | 23.10 | 21.90 | 22.90 |
| Unique title words | 295 | 391 | 254 | 280 |
| Avg. recipe words | 68.20 | 108.60 | 103.20 | 104.00 |
| Avg. recipe characters | 383.90 | 503.90 | 524.50 | 525.60 |
| Statistic | English | Yorùbá | Hausa | Igbo |
| Lexicon entries | 5000 | 5000 | 5000 | 5000 |
| Entity categories | 10 | 10 | 10 | 10 |
| Avg. words | 1.60 | 2.90 | 2.30 | 2.40 |
| Avg. characters | 7.90 | 17.60 | 11.80 | 12.20 |
| Unique words | 1,789 | 2,294 | 1,324 | 1,839 |
| Lexical diversity | 0.22 | 0.16 | 0.11 | 0.18 |
| Tag | Name | Definition |
|---|---|---|
| F | Food | Edible items, including both raw ingredients and intermediate products |
| T | Tool | Cooking tools such as knives, bowls, pans |
| D | Duration | Time durations in cooking ( 20 minutes ) |
| Q | Quantity | Quantities associated with ingredients |
| Ac | Action by chef | Verbs for deliberate cook actions ( bring in “Bring the mixture to a boil”) |
| Ac2 | Discontinuous Ac | Non-contiguous parts of compound chef actions ( to a boil ) |
| Model | Params | Type | Inference Backend | Quant. | Max. Input Tokens | Access |
| GPT-5.1 | Unknown | Proprietary | OpenAI API | – | 2048 | API |
| Llama-3.3-70B | 70B | Open-weight | Unsloth | 4-bit | 2048 | Local/HF |
| Gemma-27B | 27B | Open-weight | Unsloth | 4-bit | 2048 | Local/HF |
| NLLB-200 | 600M | Multilingual MT | HuggingFace Transformers | – | 2048 | Local/HF |
| M2M-100-1.2B | 1.2B | Multilingual MT | HuggingFace Transformers | – | 2048 | Local/HF |
| Evaluation Criteria | Evaluation Focus |
|---|---|
| Ingredient Localisation | Are the ingredients familiar, culturally appropriate, and reasonably available in the target context? |
| Procedural Localisation | Are the cooking methods, equipment, and measurements appropriate for the target context? |
| Semantic Preservation | Does the adapted recipe preserve the intended culinary meaning and outcome of the source recipe? |
| Cultural Authenticity | Does the adapted recipe represent a culturally plausible recipe within the target context? |
| Linguistic Naturalness | Does the recipe read naturally using culinary language and conventions of the target language? |
Machine Translation
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.
See all 287 papers