Ratatouille: A tool for Novel Recipe Generation

TL;DR

This work tackles generating novel recipes from a given set of ingredients, a semi-structured text generation problem. It compares LSTM-based and transformer-based approaches, training on the RecipeDB dataset to produce complete recipes with titles, ingredients, and instructions. GPT-2-based generation significantly outperforms LSTMs in preserving context and structure, enabling a functional web application (Ratatouille) for user-driven recipe synthesis. The study demonstrates a practical pipeline for recipe generation and highlights evaluation and hardware challenges, suggesting GPT-Neo as a future direction to scale performance. Overall, it advances ingredient-conditioned culinary text generation with a deployable interface for broader use.

Abstract

Due to availability of a large amount of cooking recipes online, there is a growing interest in using this as data to create novel recipes. Novel Recipe Generation is a problem in the field of Natural Language Processing in which our main interest is to generate realistic, novel cooking recipes. To come up with such novel recipes, we trained various Deep Learning models such as LSTMs and GPT-2 with a large amount of recipe data. We present Ratatouille (https://cosylab.iiitd.edu.in/ratatouille2/), a web based application to generate novel recipes.

Figures (5)

• Figure 1: Dataset before preprocessing
• Figure 2: Dataset after preprocessing
• Figure 3: Flow diagram of recipe generation
• Figure 4: Website interface to choose ingredients and generate recipe
• Figure 5: Recipe Generated using GPT2 model