Building a knowledge graph for dietary services targeting specific groups of people

Currently, most research focuses on users' behavioral preferences while ignoring the impact of food properties on human health. This article extracts a large amount of knowledge about the relationship between food properties and human health from multiple heterogeneous data sources. Based on this, a knowledge graph in the field of food is constructed for special populations to help them plan their diet more reasonably and reduce the risk of common diseases. Using background data sources such as Baidu Baike and Wikipedia, the BERT-BiLSTM-MHA-CRF method is proposed to extract food-related attributes from more than 14,731 descriptions of food properties. Combined with the differentiated features of special populations, a knowledge graph in the field of food is constructed. The knowledge graph mainly includes six entity types: food nutrition, efficacy and function, food name, population, dish name, and seasoning, with a total of 11,218 entities and 96,186 relationships. The experiment shows that compared with traditional static word vector models, BERT can generate dynamic word vectors based on context in large-scale corpus, making semantic encoding more accurate. The multi-head self-attention mechanism weights various entities in the food domain to reduce the interference of invalid information, making the model more accurate in capturing entity features. The BERT-BiLSTM-MHA-CRF method proposed in this article achieves P, R, and F1 greater than 90%.