Study On Construction Of Knowledge Graph For Pest Control Of Storage-Oriented Grain

The 20th National Congress of the Communist Party of China pointed out that a strong agricultural country is not only the foundation of a modern and powerful socialist country,but also the cornerstone of national security.Currently,ensuring food security is the key to building an agricultural powerhouse,and grain storage pests are one of the important factors that restrict food security.It is conservatively estimated that 30% of the total annual food loss is caused by insect pests.The main reason is that farmers and grain storage personnel have low entry barriers to their profession,often relying on existing grain storage experience,lacking scientific knowledge of grain storage pest control,and unable to make scientific grain storage decisions in a timely manner based on the grain situation.Researchers propose to use knowledge graph technology,which contains a large amount of prior knowledge and is easy to query,to provide intelligent grain storage pest control and query services for farmers,reducing food losses.Knowledge graphs integrate and utilize multiple sources of heterogeneous data to establish a new way of organizing and presenting knowledge,solving problems such as the fragmentation and isolation of knowledge about grain storage pest control,and facilitating the sharing and dissemination of this knowledge.Grain storage pest control data has characteristics such as heterogeneity and non-readability,making it difficult to directly construct a knowledge graph.To address this problem,this paper proposes a named entity recognition and entity relationship joint extraction method to extract knowledge triplets and construct a high-quality knowledge graph of grain storage pest control,providing technical support for farmers and grain storage personnel to store grain scientifically.In summary,the research in this paper is mainly as follows:(1)The data sources of this paper mainly include relevant texts collected from books,literature,and web pages.First,the text is preprocessed and other operations are performed,and ontology concept relationships are constructed under the guidance of experts.Under the guidance of experts,ten types of entities and nine types of relationship types are identified.Then,two improved data augmentation methods are used to enrich the diversity of data and fill in the missing annotated data.Finally,a comparative experiment is designed to establish a pre-training model for subsequent research.(2)To address the issue of recognizing specialized terms and out-of-vocabulary words in the field of storage pest control,a named entity recognition model based on new word discovery and fusion of multi-feature attention mechanism is proposed.In the corpus of storage pest control,the N-grams new word discovery algorithm is used to construct a dictionary of storage pest control,which is matched with the text to obtain a Lattice structure,and the relationship between the lattice structures is learned by the Transformer encoder.Experimental results show that the F1 value of the model reaches 91.29%,and the model can effectively utilize new word information to recognize out-of-vocabulary words in the text of storage pest control,thereby improving the entity recognition performance.(3)A joint extraction algorithm based on BERT attention threshold neural network is proposed to address the problem of the large number of texts in storage pest control data containing only one head entity and multiple tail entities.Through a decomposition strategy,all candidate head entities are first extracted from the text and passed to the subsequent tail entity and relationship extraction,which reduces the problem of redundant entity pairs.At the same time,the attention and threshold mechanism is designed to promote the interaction between the head entity and global information,making the model pay more attention to information related to the head entity.The performance of the model is verified on the publicly available dataset DUIE and the self-built dataset GraSitution.Experimental results show that the model outperforms other baseline models and can effectively extract triples in the text of storage pest control,with an F1 score of 83.20%,providing data support for constructing a knowledge graph of storage pest control.