New Chef Dataset Introduces AI into Cooking

2023-07-06

Artificial intelligence can help people with shopping, planning, and writing, but it struggles with cooking. However, a new study from the Georgia Institute of Technology's School of Computing may change that. The researchers have created a dataset called ChattyChef, which uses natural language processing models to assist users in cooking based on recipes. Using the open-source language model GPT-J, ChattyChef's cooking dialogue dataset can follow a user's recipe. In their paper "Improved Recipe Dialogue Instruction Ordering," the researchers describe their AI, which was presented at the 61st Annual Meeting of the Association for Computational Linguistics and published on the arXiv preprint server. While other researchers have theorized about the possibilities of AI chefs, the work from the Georgia Institute of Technology advances the field. Duong Le, a doctoral student from the Interactive Computing School, said, "We are one of the first research teams to analyze the challenges of building AI chefs using large language models." Most attempts to use language models for cooking have failed because GPT-J does not understand user intent or what the user wants to do next, making it difficult to track the user's progress in the recipe, known as "dialogue state." It also struggles to answer specific questions about ingredient quantities or cooking times. For example, someone might want to make potato pancakes. The AI tells them to melt butter in a pan and add the potatoes. Then the user asks what to do next. A poor robot might get confused and tell them to serve the potato pancakes, even though they are not cooked yet. Or if the user asks about the cooking time for the pancakes, the AI might provide a general time instead of specifying the cooking time for each side. To address this, the researchers ensured that their model has two key functionalities: 1. User intent detection, which determines the user's current intent from a set of predetermined possibilities, such as "asking for the next instruction" or "inquiring about ingredient details." 2. Instruction state tracking, which determines the recipe step the user is on with an accuracy of 80%. The combined information from these functionalities supports ChattyChef's third innovation—response generation. User intent helps generate the best response to answer the user's questions, while instruction state selects the most relevant part of the recipe instead of including the entire recipe to avoid confusion or additional burden. The ChattyChef dataset is built on positively rated WikiHow recipes with fewer than eight steps. The researchers had crowdsourced individuals play the role of ChattyChef users to determine which instructions to include in the dataset. The researchers believe that ChattyChef's innovations can be applied in many fields beyond cooking, such as repair manuals or software documentation. Source: https://techxplore.com/news/2023-07-chef-dataset-ai-cooking.html

RECENT AI NEWS

RECENT AI TOOLS