6th International Conference on Pervasive Computing Technologies for Healthcare

Research Article

Ingredient Matching to Determine the Nutritional Properties of Internet-Sourced Recipes

Download338 downloads
  • @INPROCEEDINGS{10.4108/icst.pervasivehealth.2012.248681,
        author={Manuel M\'{y}ller and Morgan Harvey and David Elsweiler and Stefanie Mika},
        title={Ingredient Matching to Determine the Nutritional Properties of Internet-Sourced Recipes},
        proceedings={6th International Conference on Pervasive Computing Technologies for Healthcare},
        publisher={IEEE},
        proceedings_a={PERVASIVEHEALTH},
        year={2012},
        month={7},
        keywords={lifestyle health prevention recommender systems},
        doi={10.4108/icst.pervasivehealth.2012.248681}
    }
    
  • Manuel Müller
    Morgan Harvey
    David Elsweiler
    Stefanie Mika
    Year: 2012
    Ingredient Matching to Determine the Nutritional Properties of Internet-Sourced Recipes
    PERVASIVEHEALTH
    ICST
    DOI: 10.4108/icst.pervasivehealth.2012.248681
Manuel Müller1, Morgan Harvey1, David Elsweiler2, Stefanie Mika1,*
  • 1: University of Erlangen-Nuremberg
  • 2: University of Regensburg
*Contact email: stefanie.mika@cs.fau.de

Abstract

To utilise the vast recipe databases on the Internet in intelligent nutritional assistance or recommender systems, it is important to have accurate nutritional data for recipes. Unfortunately, most online recipes have no such data available or have data of suspect quality. In this paper we present a system that automatically calculates the nutritional value of recipes sourced from the Internet. This is a challenging problem for several reasons, including lack of formulaic structure in ingredient descriptions, ingredient synonymy, brand names, and unspecific quantities being assigned. We present a system that exploits linguistic properties of ingredient descriptions and nutritional knowledge modelled as rules to estimate the nutritional content of recipes. We evaluate the system on a large Internet sourced recipe database (23.5k recipes) and examine performance in terms of ability to recognise ingredients and error in nutritional values against values established by human experts. Our results show that our system can match all of the ingredients for 91% of recipes in the collection and generate nutritional values within a 10% error bound from human assessors for calorie, protein and carbohydrate values. We show that the error is less than that between multiple human assessors and also less than the error reported for different standard measures of estimating nutritional intake.