Afaan Oromo Interactive Virtual Assistant for Coffee Production Enhancement

Abstract

Coffee is Ethiopia's top export in terms of revenue, accounting for more than 60% of all foreign exchange earned by the nation. In order to produce quality coffee crops, follow-up is necessary particularly during production processes and collection. Traditionally, agricultural professionals help farmers with the best coffee-production’s agricultural practices. The development of text-based chatbots for agriculture has made significant strides in assisting farmers with valuable information and solutions. These systems have leveraged various deep learning techniques and data preprocessing methods to achieve high accuracy rates. However, there is room for improvement, particularly in handling new queries, incorporating more diverse datasets, and enhancing the chatbot's natural language understanding capabilities. Further research and development in this field hold great potential for improving agricultural productivity and supporting farmers in their endeavors. This study proposed a chatbot system that addresses this issue by allowing machines to provide virtual assistance for users via messaging platforms in Afaan Oromo Language. The dataset for this study was collected from the agriculture bureau, a book titled “Teeknoloojii oomisha, qulqullinaa fi walitti hidhamiinsa gabaa”, and farmers. The study proposed four models: BiLSTM, BiGRU, CNN-BiLSTM and CNN-BiGRU-AM and two feature extraction techniques: word2vec and fasttext were used. Hyperparameter tuning was conducted using grid search to optimize model performance, considering factors like learning rate, batch size, and network architecture. CNN-BiGRU-AM outperformed other models on our dataset achieving accuracy of 91.01% with word2vec and 89.47% with fasttext. Human evaluation played a vital role in understanding user-centric aspects of chatbot performance, including response time, user friendliness, usability, completeness, relevance, and overall user experience. The proposed model achieved an average of 86.88% acceptance. Flask framework and telegram bot API were used to create a prototype of the model.