AMHARIC QUESTION GENERATION FROM ETHIOPIAN HISTORICAL TEXT DOCUMENTS BY USING DEEP LEARNING APPROACH

Advisor: Kemal Mohammed (Assistant Professor) In our daily lives, the most common way to learn new things is by asking questions. It improves thinking and rehearsal capacity, as well as the learning environment, due to its emphasis on a clear idea. However, manually generating questions is time-consuming, labor-intensive, and requires the use of experts. As a result, establishing automatic question generation can reduce construction time and the demand for human labor. Question generation systems can be applied on different areas, including Chabot, automatic teaching systems, and question-answering models. There are various approaches to develop a question generation models like rule based, machine learning based and deep learning based. This study focused on developing an Amharic question generation model from Ethiopian historical text documents by using deep learning approach. By using Python as programming language, different consecutive deep learning and NLP model building tasks like tokenization, normalization, stop word removal and feature extraction were involved for model development. To implement this study, DNN and GRU algorithms were employed on Keras sequential architecture by using 5144 pairs of question and answer datasets for train, validate, and test the developed model. During model, training and testing, overfitting and underfitting problems were handled by regularization, dropout, and cross-validation techniques. Different experiments have been done by changing neural network hyperparameters like the number of epochs, the number of hidden layers, and type of optimizers and train-test data distributions to select the best model result. Finally, the developed model is evaluated by the current state-of-the-art of deep learning model evaluation metrics which are accuracy, precision, recall, f1-score and user acceptance testing. The developed model scores performance of DNN, GRU, and user acceptance testing 73.76%, 74.52%, and 79.25 % respectively. Key words: Amharic, Deep Learning, DNN, GRU, Question Generator.