Amharic Chabot for Diet Healthcare Consultation Using Deep Learning Approach

Conversational Chabot’s can serve by providing auto services such as advising, question answering, knowledge generation, and other services for people without the involvement of humans on the other side. These conversational agents are very versatile and can be applied to different areas of businesses including marketing and sales, psychology, news media, education, customer support and healthcare. So it is challenging to design a deep learning-based Chabot since different domains require different strategies and algorithms, depending on the types of the Chabot (task-oriented or non-taskoriented and retrieval or generative-based) and the nature of the dataset. This study focused on designing and implementing a retrieval-based Amharic Chabot model for diet-related healthcare consultancy by delivering domain-specific information through natural language conversation. To replay the response to each user's questions from the hand-crafted knowledge base, the syntactic and semantic structure of the dataset and the nature of the domain are considered. To track users’ conversations, dialog management of the model is included to save the current topic of conversation intent. Different consecutive Deep Learning (DL) and Natural Language Processing (NLP) model-building tasks were involved in designing the Chabot model. To carry out this study, the Deep Neural Network (DNN) and Gate Recurrent Unit (GRU) algorithms were used on the Keras sequential framework. The model was trained, validated, and tested using a dataset of 3121 pairs of questions and answers. During the model, training, and testing phases, overfitting and under fitting issues were addressed using regularization, dropout, and crossvalidation techniques. Different tests were carried out by adjusting the neural network's settings such as the number of training rounds, the layers inside the network, and how the data is split between training and testing to find the most effective model outcome. The proposed model is assessed using the current state-of-the-art deep learning Chabot evaluation metrics, including accuracy, precision, recall, F1-score, and user acceptance testing. The model achieved scores of 92.55% with DNN and 74.52% with GRU.