BERT Pre-training for Cooking Time Prediction from Cooking Recipes

Recently, websites that allow ordinary users to share and search for recipes have become popular. Typically, each recipe contains various information such as a title, a list of ingredients, and descriptions of the cooking process through text and photos. The estimated cooking time is another valuable piece of information when selecting a recipe. However, it can be difficult to accurately describe cooking time because it depends on the cooking environment and conditions, such as heat level and quantity. Therefore, some recipes do not include cooking time information. In this study, we consider the prediction of cooking time in general cases from the list of ingredients and the textual description of the cooking process of each recipe by BERT, a natural language processing model. For this purpose, we propose an additional pre-training method that weights words related to cooking time from a cooking ontology. Our experimental results show that our methods outperform a fine-tuned BERT model with additional pre-training by a standard method.