SBIR Phase I: A Deep-learning-based Chatbot and Personalized Recommendations: Application to Nutrition

Mealmate Inc., Hermosa Beach CA

Investigators

Abstract

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to advance the health and welfare of the American public. Obesity among American adults has risen from 12% in 1990 to over 40% today, leading to an estimated medical cost of $260 billion in 2016, according to the Center for Disease Control (CDC). According to the National Institute for Health (NIH), 70% of American adults were overweight or obese in 2014. In 2013, American adults were spending $60 billion annually on weight loss, according to US News & World Report. A 2008 American Journal of Preventive Medicine study showed that those who kept daily food journals lost twice as much weight as those who did not. However, existing diet tracking methods are often too time-consuming for maintaining long-term weight loss. A personalized artificial intelligence (AI) chatbot could make food logging fun and easy, benefitting millions of Americans who are trying to lose weight and furthering knowledge on spoken dialogue systems. This Small Business Innovation Research (SBIR) Phase I project will advance knowledge in the field of spoken dialogue systems in several ways. First, the project establishes a new research area by noting that AI and spoken dialogue systems have yet to be applied to nutrition. Typically, conversational agents focus on factual question answering or tasks such as flight booking, but there is an opportunity to leverage big data for learning relationships between diet and health. Second, this project will develop a neural generative chatbot model with memory, demonstrating the benefit of personalized conversational interactions with intelligent agents that remember the history of conversations and personal details about the user. While manually writing chatbot responses ensures more control over the output, the drawback is that the responses are less interesting, diverse, and flexible. This work proposes generative Transformers in order to generate more realistic, human-like responses and knowledge graphs as a novel method for remembering the conversation and diet tracking history of each user for personalized feedback. Finally, this project proposes the application of causal inference, often used for medical diagnosis, to the new, challenging task of predicting which foods lead to outcomes such as gut symptoms, weight loss, or muscle building. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.