Abstract:

[Objective] The rural revitalization strategy presents novel requisites for the extension of agricultural technology. However, the conventional method encounters the issue of a contradiction between supply and demand. Therefore, there is a need for further innovation in the supply form of agricultural knowledge. Recent advancements in artificial intelligence technologies, such as deep learning and large-scale neural networks, particularly the advent of large language models (LLMs), render anthropomorphic and intelligent agricultural technology extension feasible. With the agricultural technology knowledge service of fruit and vegetable as the demand orientation, the intelligent agricultural technology question answering system was built in this research based on LLM, providing agricultural technology extension services, including guidance on new agricultural knowledge and question-and-answer sessions. This facilitates farmers in accessing high-quality agricultural knowledge at their convenience. [Methods] Through an analysis of the demands of strawberry farmers, the agricultural technology knowledge related to strawberry cultivation was categorized into six themes: basic production knowledge, variety screening, interplanting knowledge, pest diagnosis and control, disease diagnosis and control, and drug damage diagnosis and control. Considering the current situation of agricultural technology, two primary tasks were formulated: named entity recognition and question answering related to agricultural knowledge. A training corpus comprising entity type annotations and question-answer pairs was constructed using a combination of automatic machine annotation and manual annotation, ensuring a small yet high-quality sample. After comparing four existing Large Language Models (Baichuan2-13B-Chat, ChatGLM2-6B, Llama 2-13B-Chat, and ChatGPT), the model exhibiting the best performance was chosen as the base LLM to develop the intelligent question-answering system for agricultural technology knowledge. Utilizing a high-quality corpus, pre-training of a Large Language Model and the fine-tuning method, a deep neural network with semantic analysis, context association, and content generation capabilities was trained. This model served as a Large Language Model for named entity recognition and question answering of agricultural knowledge, adaptable to various downstream tasks. For the task of named entity recognition, the fine-tuning method of Lora was employed, fine-tuning only essential parameters to expedite model training and enhance performance. Regarding the question-answering task, the Prompt-tuning method was used to fine-tune the Large Language Model, where adjustments were made based on the generated content of the model, achieving iterative optimization. Model performance optimization was conducted from two perspectives: data and model design. In terms of data, redundant or unclear data was manually removed from the labeled corpus. In terms of the model, a strategy based on retrieval enhancement generation technology was employed to deepen the understanding of agricultural knowledge in the Large Language Model and maintain real-time synchronization of knowledge, alleviating the problem of LLM hallucination. Drawing upon the constructed Large Language Model, an intelligent question-answering system was developed for agricultural technology knowledge. This system demonstrates the capability to generate high-precision and unambiguous answers, while also supporting the functionalities of multi-round question answering and retrieval of information sources. [Results and Discussions] Accuracy rate and recall rate served as indicators to evaluate the named entity recognition task performance of the Large Language Models. The results indicated that the performance of Large Language Models was closely related to factors such as model structure, the scale of the labeled corpus, and the number of entity types. After fine-tuning, the ChatGLM Large Language Model demonstrated the highest accuracy and recall rate. With the same number of entity types, a higher number of annotated corpora resulted in a higher accuracy rate. Fine-tuning had different effects on different models, and overall, it improved the average accuracy of all models under different knowledge topics, with ChatGLM, Llama, and Baichuan values all surpassing 85%. The average recall rate saw limited increase, and in some cases, it was even lower than the values before fine-tuning. Assessing the question-answering task of Large Language Models using hallucination rate and semantic similarity as indicators, data optimization and retrieval enhancement generation techniques effectively reduced the hallucination rate by 10% to 40% and improved semantic similarity by more than 15%. These optimizations significantly enhanced the generated content of the models in terms of correctness, logic, and comprehensiveness. [Conclusion] The pre-trained Large Language Model of ChatGLM exhibited superior performance in named entity recognition and question answering tasks in the agricultural field. Fine-tuning pre-trained Large Language Models for downstream tasks and optimizing based on retrieval enhancement generation technology mitigated the problem of language hallucination, markedly improving model performance. Large Language Model technology has the potential to innovate agricultural technology knowledge service modes and optimize agricultural knowledge extension. This can effectively reduce the time cost for farmers to obtain high-quality and effective knowledge, guiding more farmers towards agricultural technology innovation and transformation. However, due to challenges such as unstable performance, further research is needed to explore optimization methods for Large Language Models and their application in specific scenarios.

Key words: LLM, generative pre-trained transformer, agricultural technology knowledge, intelligent question answering, name entity identity