Arabic Reading Comprehension Benchmarks Created Semiautomatically

摘要

Reading comprehension is the task of answering questions from paragraphs; it is also considered a subtask of question-answering systems. Although Arabic language is a language spoken by more than 330 million native speakers, it lacks the required resources, which are needed by the Arabic reading comprehension task to serve as a benchmark dataset. The goal of this work is to present the phases of creating Arabic reading comprehension benchmark dataset semiautomatically. The phases include; data collection, manual check, Google search, document retrieval, and paragraph retrieval. The paper also conducts a thorough evaluation for the created datasets.