Word embeddings for Arabic sentiment analysis

Manual feature extraction is a challenging and time consuming task, especially in a Morphologically Rich Language (MRL) such as Arabic. In this paper, we rely on word embeddings as the main source of features for opinion mining in Arabic text such as tweets, consumer reviews, and news articles. First, we compile a large Arabic corpus from various sources to learn word representations. Second, we train and generate word vectors (embeddings) from the corpus. Third, we use the embeddings in our feature representation for training several binary classifiers to detect subjectivity and sentiment in both Standard Arabic and Dialectal Arabic. We compare our results with other methods in literature; our approach — with no hand-crafted features — achieves a slightly better accuracy than the top hand-crafted methods. To reproduce our results and for further work, we publish the data and code used in our experiments1.