Assistive Recipe Editing through Critiquing

Abstract

Home cooks often have specific requirements regarding individual ingredients in a recipe (e.g., allergies). Substituting ingredients in a recipe can necessitate complex changes to instructions (e.g., replacing chicken with tofu in a stir fry requires pressing the tofu, marinating it for less time, and par-cooking)-which has thus far hampered efforts to automatically create satisfactory versions of recipes. We address these challenges with the RecipeCrit model that allows users to edit existing recipes by proposing individual ingredients to add or remove. Crucially, we develop an unsupervised critiquing module that allows our model to iteratively re-write recipe instructions to accommodate the complex changes needed for ingredient substitutions. Experiments on the Recipe1M dataset show that our model can more effectively edit recipes compared to strong language-modeling baselines, creating recipes that satisfy user constraints and humans deem more correct, serendipitous, coherent, and relevant.