New Chef dataset brings AI to cooking

Artificial intelligence (AI) can help people shop, plan and write – but not cook. It turns out humans aren’t the only ones who have difficulty following step-by-step recipes in the correct order, but new research from Georgia Institute of Technology’s College of Computing may change that.

The researchers created a dataset called ChattyChef, which uses natural language processing models that can help a user cook a recipe. Using the open-source large language model GPT-J, ChatyChef’s dataset of cooking dialogues follows recipes with the user.

The researchers presented their AI in a paper titled “Improved Instruction Ordering in Recipe-Grounded Conversations.” 61st Annual Meeting of the Association for Computational Linguistics, and the study has also been published on arXiv preprint server

Although other researchers have theorized about the possibility of an AI chef, Georgia Tech’s work continues to advance the field. “We are one of the first research teams to analyze the challenges of using large language models to build AI chefs,” said Duong Le, PhD. student at the Interactive Computing School.

Most attempts to use language models for cooking fail because GPT-J doesn’t understand what the user wants to do next, or what the user’s intent is, and it is difficult to track how much the user adds to the recipe. away—what the researchers call a “situational interaction.” It also cannot easily answer clarifying questions such as quantity of ingredients or cooking times.

For example, someone might be trying to cook hashbrowns. AI asks them to melt butter in a pan and add potatoes. After this the user asks about the next step. A rogue bot could mess up orders and ask them to serve hashbrowns even though they haven’t finished cooking them yet. Or a user asks a follow-up question about how long to cook hashbrowns, and the AI ​​won’t be precise enough, instead giving a general time and not specifying the cooking time for each side.

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

• User intent detection to determine the user’s current intent within a given set of possibilities, such as “ask for further directions” or “ask for details about materials.”
• Instruction position tracking to identify which recipe step the user is at, which works with 80% accuracy.

The combined knowledge of these features supports the third innovation of ChattyChef – response generation. User intent helps generate the best response to answer a user’s question. Direction positioning selects the most relevant parts of a recipe rather than including the entire recipe, so that the user is not confused or burdened with additional steps when cooking.

The ChattyChef dataset is built on WikiHow recipes with positive ratings and fewer than eight steps. The researchers crowdsourced people to role-play how they might use ChattyChef to determine which instructions would be best to include in the dataset.

The researchers believe ChatyChef’s innovations could be used in many areas besides cooking, such as repair manuals or software documentation.