Introduction
Adapting culinary traditions across cultural boundaries is a complex mapping problem that extends far beyond simple linguistic translation. When a recipe is moved from its source culture (e.g., Chinese) to a target culture (e.g., English), it often encounters significant semantic gaps and differences in ingredient availability. Traditional Large Language Models (LLMs) frequently fall into a "culinary uncanny valley," generating recipes that appear plausible but contain "hallucinated" steps or culturally inappropriate ingredient combinations.
To address these issues, researchers have proposed that Information Retrieval (IR) is a more robust solution than pure generation. By retrieving authentic, human-authored recipes from the target culture that capture the essence of the source dish, systems can ensure instructions remain cookable and culturally grounded.
The CARROT Framework
The authors introduce CARROT (Cultural-Aware Recipe Retrieval Tool), a modular framework designed to bridge the gap between disparate culinary traditions. CARROT moves away from literal translation, instead focusing on cultural adaptation through a three-stage pipeline.
- Query Processing & Rewriting: CARROT utilizes an LLM (such as Llama 3) to transform the source recipe title into a query optimized for the target culture. This involves masking the original title to generate a new one based on ingredients and steps, ensuring the query aligns with target naming conventions.
- Retrieval: Using a bi-encoder (MPNet), the system searches a corpus of over 2 million English recipes (RecipeNLG) to find the most relevant candidates based on the rewritten queries.
- Cultural-Aware Re-ranking: A final LLM-based stage evaluates the retrieved candidates. It prioritizes recipes that maintain the "soul" of the original dish while ensuring they are culturally appropriate for the target audience.
Data and Benchmarking
The study leverages two massive monolingual datasets: RecipeNLG (2M+ English recipes) and XiaChuFang (1.5M+ Chinese recipes). Recognizing the lack of standard evaluation metrics for this task, the authors introduced a new manually annotated benchmark. This dataset includes 98 queries and over 1,500 query-document pairs, labeled by bilingual experts on a 3-point scale of cultural equivalence: Exact Match, Partial Match, and No Match.
Results and Qualitative Analysis
CARROT-Llama3 demonstrates a significant performance improvement over baseline retrieval models, achieving an nDCG@10 of 0.346. Human evaluations further confirm that retrieval-based methods provide higher consistency (5.64/7) and cultural appropriateness (6.16/7) compared to direct generation by models like GPT-4.
As shown in the case studies, literal translations often lead to retrieval failures. For instance, translating Huiguo Rou as "Back to the pot" fails to find relevant matches, whereas CARROT's rewriting to "Stir-Fried Pork with Sichuan Pepper" successfully retrieves a culturally equivalent "Braised Pork with Pepper and Onion." This highlights the system's ability to navigate the naming divide where Chinese dishes are often named after legends or techniques, while English recipes focus on primary ingredients.
Conclusion
The CARROT framework demonstrates that retrieval is a superior strategy to generation for cross-cultural recipe adaptation. By grounding the output in real-world, human-written recipes, the system avoids the logical inconsistencies of LLMs while successfully navigating complex cultural naming conventions. Future work will look to expand this approach to more diverse language pairs and incorporate fine-tuned culinary "common sense" into the retrieval process.
Summary updated