Abstract
There has been an enormous increase in the volume of Arabic posts on many social networks, providing a rich source of opinions on a variety of topics. Opinion mining has attracted much research interest in the last decade, but has mostly been conducted using the English language. Opinion mining from Arabic social media is lagging behind, primarily due to challenges with processing the morphologically complex Arabic natural language and the lack of available tools and resources for extracting Arabic sentiments from text. This complex task is further exacerbated when dealing with Saudi dialects that do not abide by the formal grammatical structure of Modern Standard Arabic (MSA). Targeting the use-case of unemployment in Saudi Arabia, this paper investigates the major challenges facing sentiment analysis researchers of the informal Arabic language and documents how these challenges can be initially addressed by means of linguistic pre-processing of raw text and supervised machine learning (ML) to identify sentiment polarity. The initial experimental evaluation yielded satisfactory sentiment classification results while also demonstrating the benefit of using ML to identify the dominant features in deciding sentiment polarity, which can in turn be used in constructing a problem-domain knowledge base that can further improve the accuracy of sentiment classification. (C) 2017 The Authors. Published by Elsevier B.V.