Automatic Amharic Question Generation System Using Recurrent Neural Network

Abstract

A question is a linguistic term that is used to make a request for information and is essentially posed in order to meet informational needs. These days electronic documents written in different languages are available over the internet or any other storage media. These documents however do not contain enough questions. This is because manually preparing questions is very time-consuming and tedious task. The solution for such problem is an Automatic Question Generation System, which is a very challenging task in NLP that is designed to automatically create questions from sentences. In this study an attempt is made to design an Automatic Question Generation System from Amharic sentence using Recurrent Neural Network. To train the model, 60,023 sentence-question pair and/or sentence question-answer triple dataset collected from the internet, with a data collection system that has been specifically designed for this task, is used. To make the system generate more than one question from a single sentence a beam search decoder is used. The study achieved an accuracy of 88.36% and 82.54% for the model trained with the sentence-question pair dataset and sentence-question-answer triple dataset respectively. The former model generated Amharic questions for the given Amharic sentence while the later generated both questions and answers for the given sentence.