AI can help people shop, plan and write, but not cook. Algorithms have a hard time following step-by-step recipes in the right order, but a new study conducted at the Georgia Institute of Technology College of Computing could make a difference.

The researchers created a dataset called ChattyChef, which uses natural language processing models that can help the user prepare a recipe. Using the large open source GPT-J language model, the ChattyChef dataset with cooking dialogs follows recipes along with the user.

The researchers presented their AI in the paper "Improved Instruction Ordering in Recipe-Grounded Conversation", presented at the 61st Annual Meeting of the Association for Computational Linguistics.

Although other researchers have theorized about the possibility of creating an AI chef, Georgia Tech's work is pushing this field forward. "We are one of the first research groups to analyze the problems of using large language models to create an AI cook," said Duong Le, a graduate student at the School of Interactive Computing.

Most attempts to use language models for cooking fail because GPT-J does not understand what the user wants to do next, or the user's intentions, and has difficulty tracking how far the user has progressed in the recipe. He also cannot answer clarifying questions, for example, about the amount of ingredients or cooking time.

For example, if someone is trying to cook draniki, AI tells him to pour oil into a frying pan and add potatoes. The user then asks about the next step. The algorithm may confuse the order of actions and say that it is possible to serve a dish, although it is not yet fully cooked. Or the user will ask a question about how long to cook the shingle, and the AI will not be accurate enough, instead it will indicate the total cooking time and will not specify the cooking time for each side.

With this in mind, the researchers made sure that their model had two key features:

determining the user's current intentions within a fixed set of possibilities, such as "Ask for the next instruction" or "Ask for details about ingredients";
recipe tracking to determine what stage the user is at, which works with 80% accuracy.
The combined information from these functions supports the third innovation of ChattyChef - response generation. The user's intention helps to generate the best answer to the user's question. The AI selects the most important parts of the recipe, rather than including the entire recipe, so as not to confuse the user and burden him with unnecessary actions during cooking.

The ChattyChef dataset is based on wikiHow recipes with positive ratings and less than eight steps. To determine which recipes are best included in the dataset, the researchers engaged people to role-play how they could use ChattyChef.

The researchers believe that ChattyChef's innovations can be used in many areas besides cooking, for example, in repair manuals or software documentation.