Corpora for sentiment analysis of Arabic text in social media

Abstract

Different Natural Language Processing (NLP) applications such as text categorization, machine translation, etc., need annotated corpora to check quality and performance. Similarly, sentiment analysis requires annotated corpora to test the performance of classifiers. Manual annotation performed by native speakers is used as a benchmark test to measure how accurate a classifier is. In this paper we summarise currently available Arabic corpora and describe work in progress to build, annotate, and use Arabic corpora consisting of Facebook (FB) posts. The distinctive nature of these corpora is that they are based on posts written in Dialectal Arabic (DA) not following specific grammatical or spelling standards. The corpora are annotated with five labels (positive, negative, dual, neutral, and spam). In addition to building the corpora, the paper illustrates how manual tagging can be used to extract opinionated words and phrases to be used in a lexicon-based classifier.