Cross-Cultural Recipe Transformation via Neural Network and Encoder-Based Models

Vibhuti Dhar Khanduri, Mansi Goel, Lakshay, Harsh Kumar Agarwal, Ganesh Bagler

Keywords: recipe transformation, ingredient embeddings, ingredient replacement, cuisine embeddings, Word2Vec, BERT, SBERT, Doc2Vec, computational gastronomy

TL;DR: A comparison of neural network and encoder-based models for cross-cultural recipe transformations using clustering-based approach.

Abstract: Every cuisine has a culinary fingerprint characterized by its idiosyncratic ingredient composition. Transforming the culinary signature of a recipe is a creative endeavor. Traditionally, such fusion recipes have arisen from creative human interventions as a product of trial and error. Herein, we present a framework to transform the culinary signature of a recipe from one regional cuisine to another. A clustering-based computational strategy was developed, which replaces the ingredients of a recipe, one at a time, to achieve the transformation of the cuisine. We used a neural network-based Word2Vec-Doc2Vec model and three encoder-based BERT models to capture the context of an ingredient within the culinary landscape. The performance of recipe transformation strategies was evaluated by scoring their success at ‘Recipe Transformation’ and manually assessing the most frequent ingredient replacements for every fusion experiment. We observe that the encoder-based models perform better at transforming recipes with fewer ingredient replacements needed, suggesting that BERT-based models are better at providing more meaningful ingredient replacements to transform the culinary signature of recipes. The percentage of successful recipe transformations in the case of Word2Vec-Doc2Vec, BERT-Mean Pooling, BERT-CLS Pooling, and BERT-SBERT model are 99.95%, 43.1%, 41.65%, and 41.45% respectively, indicating that the neural network-based model can better cluster the cuisine-wise ingredient embeddings. On the other hand, for a successful recipe transformation, the average percentage of ingredients replaced for Word2Vec-Doc2Vec, BERT-Mean Pooling, BERT-CLS Pooling, and BERT-SBERT model are 77%, 52.3%, 51.6% and 51.5%, respectively. Our study shows a way forward for implementing cross-cultural fusion of recipes.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10259