- Natural language processing models are difficult to personalize
- Online food data gives scientists plentiful information on how to personalize recipes
- AI-generated recipes may be the first step to something bigger
Cooking is a decidedly human affair. It’s laughable to think of a turtle spending hours simmering a soup to perfection or a chimpanzee slaving over a stove to make a full Thanksgiving dinner. However, we may soon be adding machines to the list of cuisine-capable creatures—at least in terms of recipe generation.
Researchers at the University of California San Diego have created a natural language processing (NLP) model that can personalize a recipe to a user’s taste. Julian McAuley and his colleagues scraped 700,000 reviews from 180,000 recipes on Food.com in order to train a neural network to create original recipes.
NLP is the technology that allows people and machines to communicate via naturally-developed human languages. This study relies on recipes, but the overall goal was to improve NLP models to help break down barriers between users and machines.
Digging for data
When they began this research, McAuley and his colleagues Bodhisattwa Majumder, Shuyang Li, and Jianmo Ni weren’t just looking for something new to put in their lunchboxes. Instead, they’re pursuing state-of the-art techniques for NLP, focusing on individual needs.
“One thing the lab is really interested in is personalization of machine learning,” says McAuley. “We thought recipes were a great example because there is a lot of personal variation in the kind of recipes that individuals consume.”
The first step was training the neural network to recognize diners’ preferences from the Food.com data. Recipes have unique qualities that make them an interesting form of semi-structured text. There are constraints on which ingredients blend well with others or what are appropriate quantities. The NLP model must use the text to learn which combinations make sense.
“We were looking at the activity trace of the user,” says McAuley. “We observed the sequence of recipes the person has consumed. We might have additional information as to whether you liked or disliked the recipe or whether it was successful or not, but that was less important.”
Once the model was trained, the researchers could submit a recipe name and an ingredient list with caloric requirements. The model then takes this data and extrapolates into a complete personalized recipe, such as one for a low-cal pomberrytini.
“If the word lime shows up, the individual user is probably more likely to consume recipes that use lime,” says McAuley. “But it could be other things, like techniques or cuisine types or even the complexity of the recipe. These are more subtle things that the model might be able to learn from reparsing the natural language recipe data.”
Sadly, McAuley reports that no one in the research team has yet attempted to create one of the recipes generated from these user reviews. But after all, it wasn’t their own preferences that the model had learned.
Cooking up something personal
Personalized recipes aren’t the endgame for this research. Instead, McAuley says the work he and the team have accomplished via recipes has broader implications.
“There's a lot of cases where people are currently studying a system’s natural language generation,” says McAuley. “This could mean things like dialogue systems or Siri or anything you interact with like that. It could include automated question answering systems or automated tech support systems.”
Personalized NLP models could even make a splash in industries like healthcare, but getting access to private health information for training and testing is difficult. Online recipes, on the other hand, present an enormous amount of easily accessible data that’s open to the public.
“A really big direction in the natural language processing community is, ‘How can we build better models that will generate language for their output?’” says McAuley. “And I think a lot of what's missing there is how can those models be personalized.”
Personalized models work with the user in order to create a more seamless experience. As McAuley points out, getting this technology right could literally save lives in the future.
“We can think of other things like personalized help if you're talking about modeling things like clinical documents,” says McAuley. “We can also think of a question-answering system for patient triage in a hospital—but this is very, very far-fetched stuff and for the next decade.”
While we’re clearly far from handing over medical decisions to a neural network, every major breakthrough has countless steps leading up to it. AI-generated recipes are a fun way to experiment with a promising technology.
If we could make the machine-user interaction as seamless as two humans speaking, there’s almost no limit to what we could accomplish.