Answering Arabic why-questions: Baseline vs. RST-based approach

Abstract

A Question Answering (QA) system is concerned with building a system that automatically answer questions posed by humans in a natural language. Compared to other languages, little effort was directed towards QA systems for Arabic. Due to the difficulty of handling why-questions, most Arabic QA systems tend to ignore it. In this article, we specifically address the why-question for Arabic using two different approaches and compare their performance and the quality of their answer. The first is the baseline approach, a generic method that is used to answer all types of questions, including factoid; and for the second approach, we use Rhetorical Structure Theory (RST). We evaluate both schemes using a corpus of 700 textual documents in different genres collected from Open Source Arabic Corpora (OSAC), and a set of 100 question-answer pairs. Overall, the performance measures of recall, precision, and c@1 was 68% (all three measures) for the baseline approach, and 71%, 78%, and 77.4%, respectively, for the RST-based approach. The recently introduced extension of the accuracy, the c@1 measure, rewards unanswered questions over those wrongly answered.