CosyLab : named entity recognition on recipe dataset

Abstract

Traditional cooking recipes follow a structured format that can be effectively modelled by analysing and accurately representing the rules and semantics of different sections within the recipe text. In this paper, we propose a structured framework for representing cooking recipes and a pipeline to infer the most suitable representation in a uniform format. The Ingredients section, which typically lists ingredients along with attributes such as quantity, temperature, and processing state, is modelled by defining these attributes and their corresponding values. We categorise the physical entities involved in recipes into utensils, ingredients, and their combinations, all of which are related through various cooking techniques. The Instructions section outlines a sequence of events where cooking techniques or processes are applied to these ingredients and utensils, which we model as relational tuples. We apply this method to the RecipeDB dataset [1], demonstrating the effectiveness of our approach. This model has various applications, including recipe translation, recipe similarity determination, novel recipe generation, and nutritional profile estimation.