Learning Food Composition Knowledge from scientific literature

Abstract

Food Composition Tables (FCT) or Food Composition Databases (FCD) contains the food we eat and what it contains. It is built by using chemical analysis to determine the different composition and structure of foods. However, the chemical analysis of food requires significant financial resources and skilled laboratory investigators. These resources are not always available. Thus, in many cases, to build Food Composition Tables, many people rely on existing resources such as scientific papers. Scientific papers contain key-insights organized in text, tables figures, etc. that are used to understand the scientific contribution of its author. Many FCT are stored in scientific papers related to food, nutrition, food chemistry, etc. in the form of tables. Acquiring these tables manually as it is currently done by domain experts is costly, not scalable and cumbersome work because one has to open the paper, copy the elements one by one and save in a file such as CSV files. This paper proposes to learn Food Composition Knowledge (FCK) stores in tables of scientific papers. It consists of using Deep Learning techniques for the automatic detection of tables, text recognition from these tables, text extraction and table reconstruction. This approach was used to extract over 10,000 tables from around 5000 scientific papers. To validate the knowledge extracted, we presented 100 tables selected manually to a Professor in Food Science and Nutrition. On the other hand, the validation by Ontology Recommender of Bioportal showed that the knowledge extracted are relevant to the biomedical domain in general and can be used to enrich food ontologies.