Advances in English language representation enabled a more sample-efficient\npre-training task by Efficiently Learning an Encoder that Classifies Token\nReplacements Accurately (ELECTRA). Which, instead of training a model to\nrecover masked tokens, it trains a discriminator model to distinguish true\ninput tokens from corrupted tokens that were replaced by a generator network.\nOn the other hand, current Arabic language representation approaches rely only\non pretraining via masked language modeling. In this paper, we develop an\nArabic language representation model, which we name AraELECTRA. Our model is\npretrained using the replaced token detection objective on large Arabic text\ncorpora. We evaluate our model on multiple Arabic NLP tasks, including reading\ncomprehension, sentiment analysis, and named-entity recognition and we show\nthat AraELECTRA outperforms current state-of-the-art Arabic language\nrepresentation models, given the same pretraining data and with even a smaller\nmodel size.\n