More information
Description
The TAISTI project is designed to answer specific questions aimed at solving practical problems of detecting ingredients in a recipe that should be replaced concerning a special diet, dish or other constraints and recommending their valid substitutes. The project will focus on providing practical solutions in the domain of information engineering researching various designs and experimentally evaluating them with a purpose to propose a new technology.
TAISTI will increase the share of female researchers in technical sciences: four female researchers in technical sciences, including three in information engineering, and one in food and beverages, will participate in TAISTI, one in the roles of the PI aand WP leader, and two in the roles of WP leaders. Moreover, PI will establish a new research collaboration by going abroad for research to visit Norwegian University of Science and Technology (NTNU).
The specific objectives of the project are to provide: 1) integrated knowledge and data resources on culinary recipes and their ingredients to fuel artificial intelligence algorithms, 2) novel data-driven (machine learning-based) methods to recommend candidate ingredient substitutes and predict their characteristics, 3) novel knowledge-driven (logic reasoning-based) methods to select and explain target ingredients and their valid substitutes, and 4) a proof-of-concept system for recommending ingredient substitutes to integrate and demonstrate the developed technologies. The project will result in conference and journal publications as well as in a patent application. The result of the project will be at TRL level 6.
Summary of project results
The TAISTI project was designed to solve the problems of detecting ingredients in a recipe that should be substituted due to a special diet, dish or other restriction and recommending their correct replacements. The project focused on providing practical solutions in the field of information engineering, investigating different designs and evaluating them experimentally in order to propose a new technology.
The ambition of the project was to develop technology for healthy, intelligent substitution of ingredients in recipes.
An analysis of available resources with the goal to integrate adequate knowledge was carried out. A central food ontology for the project and a main reference for food products was FoodOn. A knowledge graph based on the FoodOn recipe model was built.
A training datasets by annotating entities in a recipe corpus with a tagset previously identified was prepared. As well as TASTEset – a novel dataset of recipes and prepared several baselines for fine-grained analysis and extraction of information from recipes.
The RecipeNLG corpus was prepared in a form enabling the creation of embeddings. The project team developed a procedure for automatic creation of representations for cooking recipes, suitable for machine-learning algorithms.
Development of a machine learning-based models for recommending substitutable ingredients was mainly devoted to development of the methods to prepare data for machine learning: (1) the BERT language model fine-tuned on prepared recipe collections, (2a/b) PPMI measures determining the co-occurrence of recipes and ingredients, (3) the FastText method based on lists of ingredients in recipes, (4) frequent itemset mining, and (5) the Cleora method using hypergraph representations of ingredients.
The technology, which is a neural-symbolic learning and reasoning solution, was developed. The technology avails of the developed machine learning models to recommend candidate ingredients and an ASP, logic-based solution, for pruning and recommending relevant ingredients from the candidate ones.
An interesting set of datasets, models, methods and tools were performed to tackle the project main goal. A total of six scientific publications were produced, some are of high quality, such as the ACM CIKM conference paper.
The project results could be potentially used by sectors of the economy related with food, namely restaurants, food and distribution industry. As for a sociological impact, the project results could be potentially valuable for humans that have food ingredient restrictions, such as allergies or diabetes condition. There is some potential for the project results to be used by more than one entity (e.g., restaurant and hospitals, both have food serving needs). After the funding project, there is some potential for a product commercialization if the TAISTI project is moved towards a more industry-oriented research. Also, the performed research could be further improved by making use of more recent Artificial Intelligence technology (e.g., multi-modal LLMs and diffusion models).