Artificial intelligence is able to help people with shopping, planning and writing, but it lags far behind in the field of cooking. Following step-by-step recipes in exact order proves to be difficult not only for humans, but also for AI. However, recent research from the Georgia Institute of Technology College of Computing promises to change that.

A team of researchers brings innovation to the culinary field by developing ChattyChef. This is a new project based on natural language processing models. They help users follow recipes using the GPT-J open language model.

The paper titled “Improved Instruction Ordering in Recipe-Grounded Conversation” was presented at the 61st Annual Meeting of the Association for Computational Linguistics and published in arXiv magazine.

Most attempts to use language models in cooking encounter problems due to the difficulty of keeping track of the current step of the recipe. It was these calls that became the starting point for the creation of ChattyChef. The project is developing in two key areas:

• Determining the user’s intentions allows you to find out what actions the chef plans to take from the proposed set of options. For example, clarify the following instructions or ask questions about the ingredients.
  
• Tracking the steps of a recipe helps to determine at what stage of the preparation of a dish the user is. This method works with an accuracy of up to 80%.
  

The combination of these two features provides the opportunity for ChattyChef’s third innovation: response generation. Understanding the user’s intentions and tracking the status of the instructions allows you to form the best answers to the user’s questions and suggest the most relevant recipe steps.

The data for ChattyChef comes from wikiHow positively rated recipes with no more than eight steps. The research team enlisted volunteers to model the use of ChattyChef and determine the optimal set of instructions to include in the database.

Beyond cooking, ChattyChef’s AI breakthrough could find applications in many other areas, including repair manuals or software documentation.