Artificial inelligent (AI) chatbot for scientific knowledge delivery in black pepper cultivation

Abstract

The research work entitled “An artificial intelligent (AI) chatbot for scientific knowledge delivery in black pepper cultivation” was conducted at the College of Agriculture, Vellayani, Thiruvananthapuram, Kerala, from 2022 to 2024. The objectives of the study were to develop an intelligent chatbot specifically tailored to the needs of farmers and extension workers in black pepper cultivation and to assess the suitability of an embedding model and a fine-tuned model of OpenAI for developing the chatbot. A detailed survey was conducted in Idukki and Wayanad, selecting 60 farmers with the assistance of the Department of Agriculture Development and Farmers’ Welfare, Government of Kerala. The survey focused on gathering critical data about the key concerns of black pepper farmers enabling us to build a comprehensive chatbot database. The survey outcomes highlighted major challenges faced by the farmers are disease management, pest management, lack of awareness about government schemes, insufficient extension services, nutrient management and climate change-related issues. These critical findings laid the foundation for the subsequent development of a comprehensive chatbot database tailored to address these farmers identified issues effectively. The database was then enriched by incorporating black pepper cultivation practices recommended by Kerala Agricultural University (KAU) alongside crop management strategies from the Indian Institute of Spices Research (IISR) as standard references. This approach combined direct farmer perspectives with scientifically backed methods creating a robust database that the chatbot could use to deliver accurate responses to farmer queries. The process of data preparation included key preprocessing steps to improve model performance. Techniques such as text cleaning, lowercasing, stop word removal and stemming were done sequentially. These steps were crucial for reducing data noise, simplifying complexity, minimizing overfitting and ensuring that the chatbot could generalize across varied questions. The development of the chatbot started with generating the necessary token keys for Telegram and OpenAI. The embedded model chatbot is designed to answer questions related to black pepper cultivation by leveraging the text-embedding-ada-002 model. The embedding model converts the query of the user into embeddings and matches the query to the database to retrieve relevant answers. A custom fine-tuned chatbot was developed using OpenAI’s GPT-3.5 turbo-0125 model. Both models were integrated into Telegram as Telegram is free, cross-compatible and has unlimited user limits. The study employed a quantitative research methodology to assess chatbot acceptance and compare models using a survey based on the Use and Gratification Model (UGM). The survey included 150 participants, comprising farmers, scientists, agriculture officers and research scholars. The performance of each model was assessed across three dimensions: technology, hedonics and risk. Technology was evaluated for authenticity and convenience, hedonic for enjoyment and entertainment and risk for privacy concern and immature technology. Behavioural intention was the dependent variable. The demographic analysis showed that most respondents were aged between 28 and 35. The reliability of the survey scale was confirmed through Cronbach's alpha. The Mann-Whitney U test and regression analysis, provided insights into the comparative performance of the two models. The Mann-Whitney U test highlighted significant differences in three areas: authenticity, immature technology and behavioural intention. The embedding model had a higher mean score for authenticity and behavioural intention, while the fine-tuned model exhibited higher concerns with immature technology. The multiple linear regression analysis was done for both models keeping behaviour intention as the dependent variable. Under both regression models, authenticity of conversation, privacy concern and immature technology were found to be significant. Authenticity of conversation has a slightly higher impact in the embedding model compared to the fine-tuned model. Privacy concern has a negative relationship in both models. Immature technology shows a stronger negative effect in the fine-tuned model compared to the embedding model. Based on the statistical analysis the embedding model was found to be better than fine-tuned model. The embedding-based model's adaptability and broad language representation made it ideal for diverse queries, effectively addressing farmer concerns and enhancing user satisfaction. The model’s strength in similarity based search provided a robust solution proving more versatile than the fine-tuned model for real-time, varied interactions in agriculture.