ADKG-LLM: Construction of Agricultural Disease Knowledge Graph with Large Language Models

The transformation of information storage methods has prompted the construction of targeted knowledge bases to become crucial. Traditional knowledge management suffers from low efficiency and poor scalability. With the development of agricultural intelligence, a disease detection and prevention Q&A system based on knowledge graphs has emerged. Knowledge graphs can structurally store heterogeneous data and construct semantic connections of diseases. The construction methods of knowledge graphs have become a hotspot in natural language processing, but they require a large amount of manually annotated data. The emergence of LLMs(Large Language Models) provides new ideas for the construction of knowledge graphs, reducing the dependence on manual annotation. This study proposes the ADKG-LLM(Agricultural Disease Knowledge Graph) method, which utilizes LLMs to analyze agricultural disease texts, automatically extract entities and relationships to construct knowledge graphs, and perfects them through Lora fine-tuning. In the experiments, data from professional literature were collected using web crawlers and character recognition, cleaned, and then used for few-shot prompting and fine-tuning of LLMs. The results show that GPT-4 performs best in the named entity recognition task, followed by GPT-3.5, and Llama3-8B outperforms Llama2-7B, indicating that GPT series models outperform the LLAMA series in the construction of Chinese agricultural pest and disease knowledge graphs. Compared with traditional models, LLMs perform better in Named Entity Recognition (NER) and Relation Extraction (RE) tasks. Moreover, this method is not only applicable to the field of agricultural pests and diseases, but can also be extended to the construction of knowledge graphs in other fields, saving human effort and time.