Fine-tuned encoder models with data augmentation beat ChatGPT in agricultural named entity recognition and relation extraction

Agricultural research produces vast amounts of unstructured textual data, which remains largely underutilized due to the lack of robust tools for automated processing. If effectively processed, this underutilized data can provide critical insights to advance agricultural practices, decision-making, and sustainability. This work focuses on applying Named Entity Recognition (NER) and Relation Extraction (RE) to convert unstructured data into structured formats, addressing the challenges of domain-specific terminology and limited annotated datasets. This scarcity is primarily due to the domain-specific terminology, contextual complexity, and lack of annotated data in the agricultural domain. This study addresses these challenges by proposing sophisticated data augmentation techniques, validated using large language models and human reviewers, to enhance training data. We introduce AgNER-BERTa and AgRE-BERTa, two encoder-based models tailored for agricultural NER and RE tasks, and compare them with state-of-the-art (SOTA) baselines, including SciBERT, SpanBERT, and generative decoder models like ChatGPT. Our experiments demonstrate superior performance, achieving 98% accuracy for NER and 97% for RE outperforming SOTA models. The extracted entities and relations are used to construct the Agricultural Knowledge Graph (AgKG), providing structured, queryable insights to support precision agriculture, policy-making, and sustainable farming practices.