What's for Dinner? AI Can Help
San Diego, Calif., Feb. 5, 2020 -- Research from computer scientists at the University of California San Diego could eventually lead to AI-generated recipes—customized to your personal taste. The study breaks new ground in natural language processing, which studies how AI understands and generates human (natural) language. The research was published on arXiv.org.
“You can imagine various applications like dialogue systems, question answering systems and recommender systems, which is a lot of what my lab builds,” said associate professor Julian McAuley from the UC San Diego Department of Computer Science and Engineering. “Our research is all about improving these models by personalizing them. Whenever you're dealing with text where there might be subjectivity or people having different writing styles or individuals liking or disliking some content, these natural language generation systems can potentially perform a lot better.”
In the study, the team pulled 180,000 recipes and 700,000 user reviews from Food.com to teach the neural network how to create recipes. The goal was to give the system enough data to produce personalized recipes based on individual likes and dislikes.
After that, the researchers would submit a recipe name, short ingredient list and calorie information, which the model extrapolated into a complete recipe. For example, the team asked the AI for a low-calorie pomberrytini recipe with pomegranate-blueberry juice, cranberry juice and vodka. The network produced several recipes, including:
Combine all ingredients except for the ice in a blender or food processor. Process to make a smooth paste and then add the remaining vodka and blend until smooth. Pour into a chilled glass and garnish with a little lemon and fresh mint.
The AI added lemon and mint based on the reviewer’s previous recipes – a nice touch.
McAuley and colleagues chose to study natural language processing with recipes because they were easy to access and would challenge the system.
“Recipes are interesting because they're a form of semi-structured texts where it's not just a bag of words the way some natural language models are,” said McAuley. “Potentially, there’s a lot of structured data – constraints on which ingredients are allowed or quantities or perhaps a nutritional profile. The model has to learn which combinations, and orders, of ingredients make sense together.”
The personalization serves two roles. In addition to providing recipes that are better suited to specific individuals, it helps teach the system.
“It should improve the quality of the text because we're helping the model along a bit,” said McAuley, “giving it some historical records of how an individual’s recipes typically look.”
While McAuley would love to generate an entire recipe without inputting any ingredients, the technology is not quite there. He looks forward to a time when this approach could help people navigate food preparation based on personal preferences, lifestyle, health issues and other factors.
“This is a prototype now, but maybe you could imagine building this into a system where you could say: I want more ingredients like this one,” said McAuley. “Or, I have an allergy, or I like this recipe, but can you give me a version that includes this ingredient or does not include this other ingredient? Or it's just a healthier version of the recipe.”
Generating Personalized Recipes from Historical User Preferences
Other authors included: UC San Diego Computer Science and Engineering PhD students Bodhisattwa Prasad Majumder, Shuyang Li and Jianmo Ni