Multilingual model for agro-information question answering

Abstract

This Dissertation introduces the Multilingual Model for Agro-information Question Answering (MAQA), a novel solution designed to bridge language gaps in agricultural knowledge dissemination. Leveraging cutting-edge Natural Language Processing (NLP) techniques, MAQA is adept at processing queries in multiple languages, notably in English and Luganda, addressing a critical need within the global agricultural sector. The model was rigorously developed, trained, and fine-tuned on a specialized dataset encompassing 2500 questionanswer pairs that span a broad spectrum of agricultural knowledge, from crop management to livestock care. The evaluation of MAQA’s performance was detailed, focusing on both English and Luganda datasets to ensure comprehensive linguistic coverage. The model’s e↵ectiveness is underscored by its ability to provide accurate, contextually relevant answers, to questions in the stated languages. Among the six models trained, significant outcomes were observed: Afriqa Afroxlmr was selected as the best model with a F1 of 85.25% and an EM of 81.61% while other models obtained scores as follows: BERT F1 score of 89.36% and an EM score of 76.26% , MobileBert with an F1 of 87.95% and an EM of 74.71%. The models were further evaluated using ROUGE score: MobileBert had 0.63 in English and 0.61 for DistilBert for Luganda dataset. These showed that the models ability to generate text that matches references. These results illustrate the models’ superior capacity for interpreting and responding to complex queries in both targeted languages. The contributions of MAQA are numerous, o↵ering a transformative tool for the agricultural sector by facilitating access to vital information across linguistic divides. This research not only propels forward the field of NLP but also lays the groundwork for future innovations in multilingual query answering models. Looking ahead, the Dissertation outlines potential pathways for scaling MAQA and tailoring it to a wider array of agricultural contexts, reinforcing the model’s utility and adaptability