Creative Culinary Recipe Generation Based on Statistical Language Models

Many works have been done in an effort to create systems for automatic generation of creative culinary recipes. Although most of them are related to the recipe ingredient lists, few works have been done to evaluate and generate the preparation steps of culinary recipes. This work proposes the use of statistical Language Models, as well as the perplexity metric, for the generation of culinary recipes. In this work, we also developed a system for automatic generation of creative culinary recipes using two approaches: one based on a genetic programming algorithm guided by the proposed language model; and the other based on a decomposition of existing recipes and recomposition of new recipes through a genetic algorithm guided by the proposed language model. This second approach achieved the best results. For this approach, a total of 6 recipes were generated to evaluate, through an online survey, the influence of the Language Model in the generation of recipes with better use of secondary ingredients, oils and seasonings, throughout the preparation steps. In the comparison between these two groups of recipes, the respondents considered the recipes generated using the language model as having the best quality, presenting an average evaluation of 63.6% of the scale (i.e. between medium and good use of oils and seasonings compared to recipes from the other group). In addition, a recipe from this approach was cooked and tasted for taste assessment, obtaining an average evaluation of 93% of the scale.