Standardizing food terms from product labels and menus into ontology concepts is a prerequisite for trustworthy dietary assessment and safety reporting. The dominant approach to Named Entity Linking (NEL) in the food and nutrition domains fine-tunes Large Language Models (LLMs) on task-specific corpora. Although effective, fine-tuning incurs substantial computational cost, ties models to a particular ontology snapshot (i.e., version), and degrades under ontology drift. This paper presents FoodOntoRAG, a model- and ontology-agnostic pipeline that performs few-shot NEL by retrieving candidate entities from domain ontologies and conditioning an LLM on structured evidence (food labels, synonyms, definitions, and relations). A hybrid lexical--semantic retriever enumerates candidates; a selector agent chooses a best match with rationale; a separate scorer agent calibrates confidence; and, when confidence falls below a threshold, a synonym generator agent proposes reformulations to re-enter the loop. The pipeline approaches state-of-the-art accuracy while revealing gaps and inconsistencies in existing annotations. The design avoids fine-tuning, improves robustness to ontology evolution, and yields interpretable decisions through grounded justifications.

This research focuses on developing advanced methods for assessing similarity between recipes by combining different sources of information and analytical approaches. We explore the semantic, lexical, and domain similarity of food recipes, evaluated through the analysis of ingredients, preparation methods, and nutritional attributes. A web-based interface was developed to allow domain experts to validate the combined similarity results. After evaluating 318 recipe pairs, experts agreed on 255 (80%). The evaluation of expert assessments enables the estimation of which similarity aspects--lexical, semantic, or nutritional--are most influential in expert decision-making. The application of these methods has broad implications in the food industry and supports the development of personalized diets, nutrition recommendations, and automated recipe generation systems.

In this article, we evaluate four Large Language Models (LLMs) and their effectiveness at retrieving data within a specialized Retrieval-Augmented Generation (RAG) system, using a comprehensive food composition database. Our method is focused on the LLMs ability to translate natural language queries into structured metadata filters, enabling efficient retrieval via a Chroma vector database. By achieving high accuracy in this critical retrieval step, we demonstrate that LLMs can serve as an accessible, high-performance tool, drastically reducing the manual effort and technical expertise previously required for domain experts, such as food compilers and nutritionists, to leverage complex food and nutrition data. However, despite the high performance on easy and moderately complex queries, our analysis of difficult questions reveals that reliable retrieval remains challenging when queries involve non-expressible constraints. These findings demonstrate that LLM-driven metadata filtering excels when constraints can be explicitly expressed, but struggles when queries exceed the representational scope of the metadata format.

Named Entity Recognition (NER) and Named Entity Linking (NEL) are key tasks in Information Extraction, addressing the identification and normalization of entity mentions from raw text. In the domain of food and nutrition, there have been several NER methods already developed, however, when applied to scientific text, they fail to generalize and produce large performance degradation. This introduces the need for new food NER and NEL models, developed specifically for extracting food entities from scientific text. In this paper, we present a scientific food NER and NEL model, SciFoodNER, obtained by fine-tuning transformer models on a corpus of scientific abstracts annotated with food entities. The models can identify mentions of food entites from raw text, and link the food entities to the Hansard Taxonomy, the FoodOn ontology and the Systematised Nomenclature of Medicine Clinical Terms (SNOMEDCT). Out of the evaluated models, the BioBERT model achieves the best results, reaching a median macro-averaged F1 score of 0.90 for the NER task, 0.66 for the NEL task linking to the Hansard Taxonomy, 0.43 for the NEL task linking to the FoodOn ontology and 0.58 for the NEL task linking to the SNOMEDCT ontology.

Although recipe data are very easy to come by nowadays, it is really hard to find a complete recipe dataset – with list of ingredients, nutrient values per ingredients, and per recipe, allergens, etc. Recipe datasets are usually collected from social media websites where users post and publish recipes. Usually written with little to no structure, using both standardized and non-standardized units of measurement. We collect six different recipe datasets, publicly available, in different formats and some including data in different languages. Bringing all of these datasets to the needed format for applying a machine learning (ML) pipeline for nutrient prediction [1], [2], includes data normalization using dictionary based named entity recognition (NER), rule based NER, as well as conversions using external domain specific resources. From the list of ingredients, domain-specific embeddings are created using the same embedding space for all recipes – one ingredient dataset is generated. The result from this normalization process are two corpora – one with predefined ingredient embeddings and one with predefined recipe embeddings. On all six recipe dataset the ML pipeline is evaluated. The results from this use case also confirm that the embeddings merged using the domain heuristic yield better results than the baselines.

In the modern era of big data, one of the biggest challenges is to find an efficient way of extracting information from unstructured data and structuring it in a form that can be interpreted and utilized by both humans and computers. In this paper, we focus on the domain of food and nutrition by introducing a Machine Learning (ML) based Named Entity Recognition (NER) method, which is a crucial step in extracting information from unstructured textual data. To the best of our knowledge, this is the first corpus-based food NER method that has been enabled by the recently published FoodBase corpus. The method is based on Bidirectional Long Short-Term Memory (BiLSTM) in conjunction with Conditional Random Fields (CRF) and Representation Learning (RL). Our experiments show that, despite the relatively small amount of annotated data, BuTTER is able to successfully identify food entities from raw text, with the best of the proposed models achieving an average macro F1 score of 0.946.

In the modern era of data, advanced approaches for extracting information and knowledge from data are required. Moreover, the extracted information needs to be formalised to be usable by information systems. Today, there exist several resources of semantics on food waste, which is a huge environmental problem that need to be fixed as soon as possible. Yet, the problem is that the existing semantic resources are not aligned and therefore needs to be linked. Only in this way, the complementary knowledge from different resources will become of real value. In the paper, an AutoMap algorithm for automated mapping of knowledge on food waste from different semantic resources is presented. By integrating such an algorithm in a web based tool, experts and the general public can get an insight into complex knowledge that is required for inventing new solutions for the food waste valorisation.

Human understanding and knowledge about food and nutrition is constantly evolving, and has significantly improved recently, one of the main contributor to this is data. The possibilities of gaining knowledge from food and nutrition-related data are yet to be explored. One of the most important information about food is nutrient content, which is very relevant for patients suffering from various diseases, professional athletes, and slowly part of everyday life of many for health or fitness goals. In this paper, we explore the effect of domain bias in a predictive study in the food and nutrition domain. Having a machine learning pipeline for predicting nutrient values with learned vector representations from short text description of recipes, we introduce domain knowledge before the prediction algorithms are applied. On a large corpus of recipe data containing short description and nutrient values we introduce word and paragraph embeddings, learn concept representations for the textual descriptions, introduce domain knowledge for clustering the data, and apply machine learning algorithms for predicting the nutrient content of the recipes. We explore the impact of the domain knowledge by introducing two different criteria of clustering the dataset - using graph embedding of the FoodEx2 codes, and using the traffic light labelling system from the Food Standards Agency; at the end we compare the two different criteria. The results from this study imply that inferring domain knowledge beforehand is crucial for the task of predicting nutrient content.

Food Frequency Questionnaires (FFQs) are inexpensive dietary assessment tools used in large epidemiological studies of diet and health. In the WellCo project, we developed the Extended Short Form Food Frequency Questionnaire (ESFFFQ) in order to monitor quality of users' nutrition. The developed questionnaire returns diet quality scores for eight targets - fruit intake, vegetable intake, fish intake, salt intake, sugar intake, fat intake, fibre intake and protein intake. In this paper we explored different feature selection methods in order to find the subset of the most important questions for five of the above-mentioned targets. This would help us monitor the most problematic areas of users' nutrition during a longer period of time without losing too much information and at the same time without overwhelming the user by asking the whole questionnaire. We were able to see that some of the feature selection methods found minimal subsets of questions (one or two questions), on which different machine-learning algorithms had better performance than on much bigger subsets.

Data normalization methodologies are extremely welcome to link extracted information from textual data to different semantic resources. These methodologies have been previously well researched especially in the biomedical domain, where health concepts were normalized and described using semantic tags. Recently, a methodology for normalizing food concepts has been proposed, based on Named-Entity Recognition methods resulting in the FoodOntoMap semantic resource. In this paper, we propose and evaluate a new architecture for linking phrases (i.e. textual name for foods) to concepts from semantic resources in the Food and Nutrition domain. We represent the food phrases (i.e. their textual name) in continuous vector space using state-of-the-art Natural Language Processing (NLP) embedding algorithms, and evaluate their proximity with respect to the annotated semantic food concepts. Additionally, indexing was incorporated to improve efficiency.The GloVe embedding with mean pooling provided best evaluation results, with maximum recall of 74% for the Snomed CT semantic dataset, which is promising result, but also opens a space for future improvement of the phrase representations, and their incorporation in this system.

Populations in countries undergoing rapid transition are experiencing food- and nutrition-related problems. To acquire high-quality nutrition information, we need beside adequate data about food consumption, also efficient methods for the extraction of information from the collected data. Our aim was to develop a methodology for analyzing and reasoning about dietary intake data collected by a food propensity questionnaire (FPQ) and dependent 24-hour recalls (24HRs). We analysed a subset of data (about 197 participants) in the SI.Menu survey carried out in 2016/17 in Slovenia. The participants completed FPQs and 24HRs. We were able to identify four clusters. Two clusters represented participants with more healthy habits, e.g., low intake of animal fats, high breakfast frequency, and high intake of fruits and vegetables. The other two clusters represented participants with less healthy habits, e.g., high intake of animal fats, low breakfast frequency and increased BMI. The four clusters can be well separated by only four variables. This interesting discovery could lead to simplified FFQ questionnaires, which could significantly decrease the participants' burden and could ensure participant compliance in similar studies. Having big national data set related to nutrition should ease the process of creating sustainable policies that will ultimately benefit agriculture, human health and the environment.

Although food and nutrition have been studied for centuries, modern nutritional science is surprisingly young. Human knowledge about food and nutrition has evolved drastically with time, and it has especially expanded the last few decades, with information and data being mass produced and available everywhere and in any form it is really easy to get overwhelmed and confused when it comes to what is right and what is wrong. Macronutrient assessment is a crucial task for individuals suffering from various diseases, and also very relevant for professional athletes, and nowadays it is becoming part of everyday life for many people, because of health or fitness reasons. Assessing the nutritional components of food is very challenging and requires reliant source of data. In this paper, we introduce an idea of finding similar recipes with regard to their macronutrient values based on learned recipe vector representation. On a scientifically proven dataset of recipe data containing description and macronutrient values we introduce word and paragraph embeddings, learn concept representations for the textual descriptions, proceed with calculating similarity between the embeddings, and then compare with the similarity between nutrient values. The results show a strong correlation between these two similarities. This study is a promising beginning for continuing in this direction-introducing multi-label classification or regression for predicting nutrient content of food.

This paper presents the recently finished EU-funded RICHFIELDS project aimed to design a new research infrastructure that would foster research in the areas of food and nutrition with a focus on consumers' behavior and lifestyle. In this project, an architecture of a new consumer data platform was designed and discussed from the researchers, business, management, ethical and legal points of view. Also new methodology for supporting big and open data standardization and interoperability was developed.

Recently, deep learning achieved the state of the art in the field of food image recognition. In this paper we describe our deep learning contributions to the field: NutriNet, a novel deep learning architecture, and a pixel-level classification solution for images of fake food. NutriNet was trained on a food image dataset of a larger size and containing more food classes than previous works, and was the first to recognize beverage images. Our work on fake-food image recognition includes the first automatic system for recognizing images of fake food, while the visual similarity of fake and real food makes it useful for fake-food experiments as well as real food recognition.

This position paper describes the METROFOODRI research infrastructure, which has been included in the ESFRI Roadmap 2018 in the domain Health and Food, with a focus on its electronic component. The Strategy Working Group and the Implementation Group in their evaluations clearly stated that the project fills a gap in the Health & Food domain and sufficiently mature to proceed with its implementation. One of the main tasks of the project is to integrate physical and electronic infrastructure based on FAIR data principles.

In this paper we explore the LanguaL hierarchy, which is a standard used to describe foods, in order to see if different food standards that are included in this hierarchy are linked together properly. For this reason, we have trained a vector representation (i.e. embedding) for each food product that is a part of the hierarchy and have found the most similar products for a subset of products. The results have shown that the LanguaL hierarchy consists of various food standards that are connected in one hierarchy, but without semantically linking the standards together. Such results indicate that further efforts should be made to link all these standards completely in order to provide a unified system for describing and standardizing food products.

Recently, as a part of an EU-funded project called REFRESH, a new web-based tool named FoodWasteEXplorer was developed. It provides an easy access to valuable data on unavoidable food waste that can be explored by researchers, industry, governmental agencies and the public to find ways of its valorization. The food waste data was manually collected and stored in a relational database. To enrich and make best use of it, we automatically transform the collected information into a new food waste ontology. The created Food Waste Ontology provides a formal description of knowledge from the food waste domain. Examples of its application are: (i) database querying based on natural language questions, and (ii) finding new or missing data from other datasets.

In the last decade, a great amount of work has been done in predictive modelling in healthcare. All this work is made possible by the existence of several available biomedical vocabularies and standards, which play a crucial role in understanding health information. Moreover, there are available systems, such as the Unified Medical Language System, that bring and link together all these biomedical vocabularies to enable interoperability between computer systems. However, in 2019, Lancet Planetary Health published that the year 2019 is going to be the year of nutrition, where the focus will be on the links between food systems, human health, and the environment. While there is a large number of available resources for the biomedical domain, only a limited number of resources can be utilized in the food domain. There is still no annotated corpus with food concepts, and there are only a few rule-based food named-entity recognition systems for food concepts extraction. There are also several food ontologies that exist, each developed for a specific application scenario. However there are no links between these ontologies. For this reason, we have created a FoodOntoMap resource that consists of food concepts extracted from recipes. For each food concept, semantic tags from four food ontologies are assigned. With this, we have created a resource that provides a link between different food ontologies that can be further reused to develop applications for understanding the relation between food systems, human health, and the environment.