Deep learning model for arabic question-answering chatbot

Abstract

In the era of innovation in information technology, one of the most outstanding and attractive applications to emerge in the field of mimicking human behavior is the chatbot. A chatbot is one of the most important examples of computer software based on Natural Language Processing and artificial intelligence. The chatbot is a conversational agent that attempts to mimic human behavior by conducting friendly conversations. Chatbots have been widely used to improve customer experiences, as well as to support learning, health care, and commercial social media interactions. By improving efficiency, making information more accessible, and personalizing the conversation. Chatbots are getting less difficult to train and deploy this is owing to the availability of development platforms and open-source code abundance. Most chatbot platforms, however, do not support conversation in Arabic, particularly in local dialects. Furthermore, there is a scarcity of Arabic datasets for training the system and designing applications to simplify discussions with Arabic users. The major goal of this research is to create deep learning models that use Recurrent Neural Networks (RNN), which can be extremely useful in learning natural languages, particularly Arabic conversational and local dialects. The proposed system achieved high-performance results, as experiments were carried out using different feature extraction techniques such as TF-IDF, BoW, N-grams, and a pre-trained, distributed word representation called AraVec. Moreover, many RNN models have been employed, such as standard RNN, LSTM, and GRU in order to obtain the best performance. The system was trained and evaluated over a dataset that was collected and built from the official Twitter governmental account (Tawakkalna) and official website. The system is able to respond to users' questions with an accuracy that reaches 99% in some models.