Automated searches to find scientific information require corpora (bodies) of annotated text, so words can be classified into appropriate categories. In biomedicine, such corpora are common, but in the food domain they are scarce. FoodBase is a text corpus (body) made of more than 22,000 recipes from the food-focused social network Allrecipes. Food items were classified into three categories (Food; Production of Food, Farming; and Acquisition of Animals for Food, Hunting) and, subsequently, into deeper levels within each, creating hierarchical standardised terminology (Hansard corpus). Using a model for food-specific text identification (FoodIE), five automated steps identified and categorised food items from the recipes. Annotation of 1000 recipes suggested numbers of words per recipe was 8% higher and food items 2% lower in manually-annotated recipes compared with those annotated using FoodIE. Otherwise, there were no significant differences between the two methods, meaning FoodBase a silver standard and the dataset a benchmark for further development of food- and nutrition-related corpora.
Excessive added sugar intakes, particularly of free sugars (all monosaccharides and disaccharides added to foods plus sugars naturally present in honey, syrups, and fruit juices), are associated with obesity, chronic diseases, and poor oral health. Non-alcoholic beverages are often high in free sugars, making reformulation with low and no-calorie sweeteners a common choice. This study analysed labelling information from more than 1000 non-alcoholic beverages (juices, nectars, energy drinks, sports drinks, and soft drinks) in Slovenian grocery stores, and found low and no-calorie sweeteners in 13.2% and 15.5% of non-alcoholic beverages in 2017 and 2019, respectively. The most commonly used low and no-calorie sweeteners were acesulfame K, aspartame, cyclamates, and sucralose. As expected, the use of low and no-calorie sweeteners in beverages was associated with lower calorie and sugar contents. Increased low and no-calorie sweeteners use in Slovenia was expected but long-term benefits are not clear. Low and no-calorie sweeteners are just one option for sugar reduction and producers can reformulate products in other ways, also encouraging consumers to adapt to less sweetened beverages.
Extraction of food information from published recipes can be done automatically. Different computer-based named-entity recognition (NER) methods can extract information about food, nutrients, quantities/units, and population groups, and relationships between factors determined. For example, from dietary recommendation ‘‘Babies need about 10g protein a day’’, NER can identify ‘‘babies’’ as a population group, ‘‘protein’’ as a nutrient entity, ‘‘10 g a day’’ as a quantity/unit entity, and ‘‘need’’ as the relationship between population group and nutrient. This automatic extraction of information can connect foods with other domains (i.e., health), allowing optimal development of, for example, public health goals and policies. In this article, you find an overview of four NER methods as well as comparisons of their precision and recall performance. The evaluation used 1000 recipes from allrecipes.com, and the recently proposed rule-based food NER system FoodIE found to be the best at extracting information from text.
According to analysis of almost 80 000 recipes on allrecipes.com people worldwide looked at different recipes during COVID-19 lockdowns. Recipes with beans, peas, and lentils were more visited more often during lockdown than before. We also looked for more soup recipes, as well as recipes for comfort foods like cupcakes, pancakes, and stews.
Studying foods and nutrition across cultures remains a challenge. In part, because data are recorded using different terms and standards and, therefore, not comparable (e.g., 250 ml porridge compared with a cup of oatmeal). This heterogeneity is addressed by FoodViz, a new tool that makes links between different food terms, standards, and resources. FoodViz helps users become more familiar with food annotation, particularly the semantics of food and ingredient names, ensuring that research or dietary menu preparation is easier and more accurate.
Dietary recommendations are based on population averages, but many would like to know what we should eat individually. Complex interactions among genetics and environment make individualised advice difficult at least for the moment, … read more
It is more and more common to calculate nutrients in meals or foods using mobile and web-based apps. In many cases, users need to enter exact quantities of ingredients, which often leads to bias due to recollection, error-prone manual input, and variability in recipes. P-NUT can be used predict macronutrient values of simple foods from short text descriptions in complex recipes. This new methodology combines representation learning, to cluster food in groups with similar characteristics, and machine learning to predict macronutrients (i.e., carbohydrates, fat, proteins) as well as water. Applying this workflow to 3 265 food items, including simple products and recipes with short descriptions, generated 86% accuracy in predicted macronutrient content. This innovative tool for nutrition professionals can facilitate dietary assessment, recommendations, and guidelines.
Estimation of nutrient content of foods made by following online recipes is valuable, especially when assessing diets of people living with non-communicable diseases (e.g., obesity, diabetes, cardiovascular disease). However, nutrient estimation from text descriptions is challenging, as most information is not standardised.
The P-NUT methodology, developed for predicting macronutrient values from short recipes, was examined for bias. Effect on clustering of foods by machine learning was compared using either FoodEx2 or the ‘traffic light’ systems. Five nutrients were predicted – salt, carbohydrates (sugar), fats including saturated fat, and protein. Both systems obtained 99% clustering accuracy, although FoodEx2 performed better in predicting sugar content.
Potato has a significant role in global food security, because this tuber can be easily grown and delivers yield high production. However, pesticides and other chemical contaminants have been detected in potatoes, and this is a potential health concern. However, early detection of pesticide residues can now be carried out mathematically rapidly and inexpensively using a dynamic mathematical model, based on diffusion of pesticide from soil. Potatoes are growing and changing entities but the current model treats them as a single compartment with homogeneous distribution of contaminants. Comparing classic and dynamic models for chlorpyrifos, a common crop pesticide, suggested a significant increase in chlorpyrifos concentrations during sprouting and/or growing. Despite little field data for comparison, the new model applies a more realistic heterogeneous distribution of contaminants along the radius of potato tubers and considers the changing dimensions of potatoes as they grows. This model could be useful for assessing health risk of pesticides applied at different stages in tuber growth, as well as supporting identification of optimal application times with respect to plant health.
Specific organoleptic characteristics of wine, beer or bread can come from the yeast communities used to produce them and, more precisely, environments where these yeasts grow. Scientists have used wasps to breed yeasts of biotechnological interest, like Saccharomyces cerevisiae (also known as brewer’s or baker’s yeast). By employing social insects to host yeasts in their gut, yeast survival and biodiversity can be enhanced naturally through formation of hybrid yeasts without the use of technology. In the future, the fermented beverages industry could benefit significantly from breeding insects and using them to produce new yeast varieties.
A seemingly simple food research question such as, ‘How does the sugar content of breakfast cereals differ among European countries?’ is not easy to answer. Analysing datasets from multiple sources that are increasingly diverse, heterogenous, fragmented, and/or have differences in syntax and semantics, make study of food systems difficult or impossible. Interoperable platforms that allow datasets to “communicate” and “work together” are needed to address questions that rely on diverse sources. Such platforms must use consistent file formats, terminology, and reporting systems, making data findable, accessible, interoperable, and reusable (i.e., FAIR). A step-by-step action plan for building FAIR data platforms is described and being tested with FNS-Cloud using typical food research questions. Platforms like this will allow (semi-)automated food nutrition security data integration needed to answer important research questions that involve multiple and diverse datasets: from food composition, authenticity, toxicity, and sustainability to food consumption, behaviour, and socioeconomic impacts and, finally, health biomarkers and disease outcomes.
Food Nutrition Security Cloud (FNS-Cloud) has received funding from the European Union’s Horizon 2020 Research and Innovation programme (H2020-EU.18.104.22.168. – A sustainable and competitive agri-food industry) under Grant Agreement No. 863059. Information and views set out across this website are those of the Consortium and do not necessarily reflect the official opinion or position of the European Union. Neither European Union institutions and bodies nor any person acting on their behalf may be held responsible for the use that may be made of the information contained herein.