Deep Ensemble Methods for Food Ingredient Entity Recognition in Natural Language Processing

Abstract

In recent years, recipe-sharing websites are becoming popular among those who wish to learn how to cook or plan their menu. Individuals can choose ingredients that suit their lifestyle and health condition using online food recipes. The information from online recipes can be used to build various food, nutrition, and healthcare applications. However, the information collected from online food recipes lacks structured information. To extract such information into well-structured data, we can use a technique in natural language processing called Named Entity Recognition or NER. NER is a technique of recognizing key information or entities in a text and categorizing them into a predetermined category. However, three major issues arise when developing named-entity recognition in the food domain: (1) The availability of datasets for the food domain is still quite limited; (2) How to design a machine learning model that is effective and efficient in recognizing food entities; and (3) Existing NER models relied solely on a single model, and just a few studies employ ensemble learning for NER, particularly none for the food domain. This study aims to solve these problems via an ensemble learning technique, combining various learning algorithms to obtain a collective performance beyond existing models' performance. Drawing upon the ensemble technique, we propose a solution to the challenges mentioned above in two stages: first, we built an iterative self-training approach called SMPT (Semi-supervised Multi-model Prediction Technique). SMPT is a deep ensemble learning model that employs the concept of self-training and builds on multiple pre-trained language models in the iterative data labeling process, with a voting mechanism used as the final decision to determine the entity's label. Utilizing the SMPT, we have created a new annotated dataset of ingredient entities named the FINER dataset; and second, we proposed a food ingredient NER model called the Recurrent Network-based Ensemble model or RNE. RNE is a novel model for extracting food-related entities by incorporating deep ensemble learning with recurrent network models, including RNN, GRU, and LSTM. The experimental findings demonstrate that the proposed RNE model could extract information from food recipes more effectively than a single model. In future development, such information can support numerous food-related information systems.