A Chatbot for Recipe Recommendation and Preference Modeling

This paper describes the main steps and challenges in building a chatbot for a nutritional recommendation system addressed to the elderly population. We identified 4 main components: Natural Language Understanding (NLU), Dialogue Management, Preference Modeling and Ingredient Matching and Extraction. To address the specific challenges of a chatbot for this domain we have tested transformer-based models both in the development of the NLU component and the Dialogue Management component. Moreover, we explored word embeddings and nutritional knowledge bases combined with sentiment analysis for user preferences modeling. The sentiment analysis algorithms used to model food preferences showed to correctly match the real feeling of the users. Each one of these components were evaluated individually using appropriate metrics. Moreover, the developed chatbot was successfully tested by users and its opinions were recorded by means of usability and user experience questionnaires. The results of usability tests show that the components were well integrated. The scores obtained were higher than the benchmark values for both the System Usability and the User Experience Questionnaires.