الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 78 |  |  |
| | | from | 78 | | |
Abstract
Throughout the era of artificial intelligence (AI), AI has become an interdisciplinary industry in all domains. These days AI is used in Technology, Education, Hiring, Business, Agriculture, Manufacturing, Economics, Weather, Medicine, Sports, and other domains. Computer vision CV and Natural language processing NLP are two sub-fields in AI. These two sub-fields mimic human behavior in terms of vision and understanding human language. NLP is one of the most intriguing tasks in AI which is responsible for understanding the outer world during the human natural language. It has a variety of tasks: question answering (QA), text summarization, text generation, machine translation, topic modeling, entity linking, and text-to-speech (TTS). The QA can be used in many applications like Feedback systems, recommender systems, chatbot systems, etc.…
This study focuses on the (QA) task. It imitates the ability of the human to answer any given question. The QA task is either Open-Domain or Closed-Domain. It focuses on Open-Domain Question Answering (ODQA). These questions may be limited to a domain or more than one domain. In this research, (QA) is shown in religious field, although it emphasizes the long-form, open-domain question-answering (ODQA) task. The long-form means that the QA system supports the long answers (justification answers).
The QA task here provides a model that can be used in developing an IslamBot QA system. This model is a free-form question-and-answer chatbot that can answer questions about Islam. The ODQA model can be built using several architectures. The retriever-reader and the retriever-generator are the most common architectures in the QA task. The ODQA model here was built using deep learning-based retrieval-reader models because it depends on reading the answers without any text generation. This occurs because the religion domain is sensitive so, the answers are required to be shown as it without any changes. This model makes use of scraped data, and the data is an English Islamic Articles Dataset (EIAD). Each article has its content and some metadata like title, date of publication, number of views, and rate of the article.
These articles were collected from the most trusted Islamic websites on the internet. This dataset is a crowd-sourced, labeled (ODQA) dataset which are questions and answers for these articles. The EIAD dataset is about 10k articles; with actually 7.5k crowdsourced, and about 10k question-answer pairs. Each article has at least one question-answer pair. Moreover, each question has at least one answer. As any deep learning model requires the input data to be in a specific format, the QA models have their formats too. There is a replica of (EIAD) dataset in a SQUAD format. SQUAD refers to Stanford QUestion Answering Dataset (SQUAD) that is one of the most popular annotation formats for QA tasks. The retriever model also requires the data in some other format like Dense Passage Retriever (DPR) format. Therefore, there was another copy from (EIAD) dataset in the retriever format.
This study shows a suggested system architecture for the ODQA system. This system was SQL database-based. It achieved an end-to-end ODQA system with a new baseline model and benchmark using (EIAD) dataset. It also attains a new state-of-the-art result with some of the recent Dense Passage Retriever models and QA models. It results in 78% Recall@100 and 78.8% accuracy at Top-100 articles. The ODQA model achieves new results too as it reached a 71.5% Exact Match (EM) and a 75.8% F1 score. Furthermore, the model’s input was only the question without context, and the use of the long-form open-domain type is a difficult point during the length of the answer (justification answer)