Application Of Deep Learning In Food Safety Entity And Relation Extraction

With the rapid development of the catering industry,food safety issues have gradually received wide attention from all walks of life.Nowadays,ensuring food safety has become an important issue,affecting the daily lives of billions of people worldwide.Although many researchers in China have conducted research in this field,relative to some countries with earlier development in this area,research in this field in China is still in its infancy.Currently,there is no unified and publicly available food safety dataset in China for relevant researchers to use.The data used in related research mostly comes from the Internet,which is scattered and cannot be shared.In recent years,the application and development of knowledge graphs have caused a wave of research enthusiasm,and their excellent performance in various fields has been demonstrated.Constructing a knowledge graph for food safety can play a promoting role in China’s food safety research work,greatly reducing the difficulty of collecting and processing data for researchers from the Internet.And entity and relation extraction is the foundational work for constructing a knowledge graph.This paper focuses on the core technology of entity and relation extraction in constructing a food safety knowledge graph,and studies and explores the application of deep learning technology in entity and relation extraction in the field of food safety.The main work and contributions made by this paper are as follows:(1)Due to the lack of publicly available food safety datasets,this paper collected and organized two types of food safety data from the Internet: food inspection information data and food safety literature abstract data.After cleaning,this paper made them into standard entity and relation extraction datasets according to certain labeling rules,which were used for model training and performance verification.(2)The paper proposes a new entity and relation joint extraction model and designs a method of abstracting Chinese syntactic trees into text adjacency graphs.Based on this text adjacency graph,the paper extracts syntactic structure feature information through graph convolutional neural networks.Then,it inputs them together with BERT-encoded contextual information into a gate network for fusion to further enrich feature embedding.Before conducting relationship extraction,the paper uses a reinforcement learning trained relation selector to increase relation embedding information and improve the model’s accuracy in relation extraction tasks.Experiments on food safety datasets showed that the comprehensive performance of the proposed model in this paper surpassed several popular joint extraction methods currently available.(3)This paper collected five publicly available Chinese datasets from other fields to test the model’s versatility.The proposed model in this paper has better performance than some existing joint extraction models on these datasets,demonstrating its advantages.This paper constructed two food safety datasets,where the original text sources are from the inspection information published by the National Food Quality Supervision and Inspection Center and scientific literature abstracts related to food safety,ensuring the quality and safety of the datasets.On this basis,the paper also proposed a joint extraction model.Based on feature diversity learning and multi-feature fusion strategies,the model’s performance on food safety datasets surpassed many commonly used deep learning methods.This research not only enriches the data of the food safety field in China but also provides a new method for constructing a knowledge graph in this field.