AgroNLP-SimAssist: a multilingual transformer-based decision support simulator for crop disease diagnosis, advisory generation, and market intelligence

There are three major shortcomings of agricultural NLP and need addressing, regardless of the recent developments in the field: (i) use of proprietary or small-scale datasets that affect reproducibility; (ii) one-task models that only focus on intent recognition but lack integration for crop disease diagnosis, advice, and market information; and (iii) the absence of simulators for evaluating the system’s performance during actual deployment. The application of precision agriculture is constrained by limitations in diverse farming environments. The study introduces AgroNLP-SimAssist, which represents a unified multilingual transformer architecture augmented by an easily deployable agricultural decision support simulator, as the proposed solution to overcome the challenges. The five major components constituting the proposed framework are: (1) domain-specific multilingual transformer embedding; (2) data collection and cleaning from publicly available agricultural datasets; (3) multiple prediction tasks including trend price estimation, crop disease detection, and advisory recommendation; (4) SHAP value explainability for expert validation; and (5) advisory simulation engine for evaluating latency and scalability. Multilingually available data sources, which include multilingual agricultural reports from the FAO organization, crop-disease descriptions from the PlantVillage knowledge base, and price-market narratives from the Agmarknet website, have been used for training and validation purposes. Among the 3,200 multilingual farmer queries dataset (in English, Hindi, and Kannada), there was an assignment of 70%, 15%, and 15% to the training, validation, and testing sets, respectively. In terms of accuracy compared to the existing classifiers, results from experiments yielded 96.1% disease classification, 94.3% advisory relevance, and 92.7% named entity recognition with F1-measure. The simulation system scaled up to 10,000 concurrent users with an inference latency of 0.35 s/query. The proposed data-driven AgroNLP-SimAssist framework facilitates extension services and precision agriculture through providing a scalable, reproducible, and interpretable solution for intelligent agricultural advisory systems.