Understanding the content of self-reported meals and online-published recipes is a basic requirement for further linking food and dietary concepts to heterogeneous health networks. Despite the huge amount of work that is done in the biomedical domain, the food and nutrition domains are relatively low-resourced. DietHub represents a step forward in food science & technology that requires knowledge from a broad spectrum of areas.
Scope and approach:
DietHub is an AI workflow methodology that annotates online-published recipes or self-reported meals with the food concepts that are mentioned in them. The food semantic labels that are used are hierarchical food semantic tags from the Hansard taxonomy. DietHub overviews and exploits several state-of-the-art methods from two areas of AI: representation learning and predictive modelling. We evaluated DietHub by applying it on a corpus of online-published recipes of different styles, such as health, cooking and region. Once the selected recipes were annotated, we compared them considering their styles. The results show justifiable comparison of Mediterranean diet recipes with recipes from other diets.
Key Findings and Conclusions: The experimental evaluation reveals that DietHub has high predictive power and correctly annotate the recipes with semantic tags. The analysis of the annotations shows that there is no statistically significant difference between Mediterranean diet and each of the diets: diabetic, weight loss, heart healthy recipes, low fat, low calorie, high fiber, and dairy free. All in all, the presented work shows that DietHub can be successfully used to analyze corpora of food-related textual documents and provide a deeper insight into the human dietary behaviour.