INVESTIGATION AND DEVELOPMENT OF BASIC AFAAN OROMOO GRAMMAR LEARNING CHATBOT USING DEEP LEARNING APPROACH

Abstract

Artificial intelligence has several uses, one of which is chatbot technology. It is a computer program that allows human like conversation between users and computers or other communication devices using natural language. Chatbot technology has advanced to a new level thanks to the application of machine learning algorithms and natural language processing methods. Chatbots can be used in different areas, i.e., business, health, and education. Language learning is one application area for chatbots. One of the applications of chatbots is language learning. Using chatbots for language learning has several benefits. For example, chatbots provide continuous practice and engagement for the learners; the application is available for the learners 24/7; and chatbots provide a personalized learning experience for the learners. Deep learning and NLP are the two emerging technologies to develop effective and efficient language-learning chatbots. For the benefits of applying chatbot technologies to language learning and the adoption of deep learning and NLP techniques, we have investigated and developed the basic Afaan Oromoo grammar learning chatbot. We started by collecting data from sources that we believed were important for the purpose. Then, after the collection of the data, we converted it to a semi structured format, which is the JSON data format. To remove inconsistency, incompleteness, and normalization processes, we have applied NLP techniques to the dataset we prepared. After preprocessing our data, the task that follows is text feature generation, i.e., our deep learning model will understand what we input for training purposes. We have used the Bag of Words (BOW) technique for feature extraction. The model we have used is the RNN (LSTM), and we created our model and performed the prediction task. Finally, we have used different metrics to evaluate the performance of our model. One of the metrics we have used is training accuracy, and the result showed about 88% of it for the final five epochs. The other metric we have used is training loss, and the result showed about 35% of it for the final five epochs.