Artificial intelligence (AI) has revolutionized various aspects of our lives, from shopping to planning and even writing. However, when it comes to cooking, AI has struggled to follow step-by-step recipes in the correct order. Recognizing this challenge, researchers at the Georgia Institute of Technology’s College of Computing have made significant progress in this field with their new research.
The research team developed a dataset called ChattyChef, which utilizes natural language processing models to assist users in cooking recipes. By leveraging the power of the open-source large language model GPT-J, ChattyChef’s dataset consists of cooking dialogs that guide users through recipes.
In their paper “Improved Instruction Ordering in Recipe-Grounded Conversation,” the researchers delve into the complexities of using large language models to build an AI chef. Many previous attempts at employing language models for cooking have fallen short due to the model’s inability to understand user intent and track the progression of the recipe accurately, referred to as the “state of the conversation.” Additionally, these models struggle to provide precise answers to clarification questions regarding ingredient amounts or cooking times.
To address these challenges, the researchers incorporated two critical features into their model. The first feature is user intent detection, which helps determine the user’s intent within a predefined set of possibilities, such as requesting the next instruction or seeking details about ingredients. The second feature is instruction state tracking, which enables the model to identify the specific step the user is on, achieving an accuracy rate of 80%.
The combined utilization of user intent detection and instruction state tracking serves as the foundation for the third innovation of ChattyChef?response generation. By leveraging user intent, the model generates the most suitable response to answer a user’s question. At the same time, the instruction state enables the selection of the most relevant parts of the recipe. This approach aims to prevent confusion or overwhelming users with unnecessary steps during the cooking process.
The ChattyChef dataset is based on WikiHow recipes that have received positive ratings and contain fewer than eight steps. To create the dataset, the researchers employed crowdsourcing, where individuals role-played scenarios to determine the optimal instructions to include.
The potential applications of ChattyChef’s innovations extend beyond the realm of cooking. The researchers believe this approach can be utilized in various domains, such as repair manuals or software documentation.
In conclusion, the team has made significant strides in addressing the challenges of using large language models to build an AI chef. By incorporating user intent detection, instruction state tracking, and optimized response generation, their ChattyChef system demonstrates the promising potential for accurately assisting users in cooking recipes. This research opens doors to broader applications in other domains, enhancing user experiences and simplifying complex tasks through the power of AI.