As a side effect of increasingly popular social media, aggression has emerged as a serious problem afflicting individuals or groups. However, aggression classification plays an important role in the user’s feedback evaluation and decrease the challenge. At present, deep learning techniques such as long short-term memory (LSTM), gated recurrent unit (GRU), bidirectional long short-term memory (BiLSTM), or convolutional neural network (CNN) are prevalently preferred in aggressive comment classification. Moreover, word embedding such as Word2Vec and FastText is closely examined in text for mapping closely related to the vectors of real numbers. In this thesis, we present BiLSTM model with FastText word embedding to classify Amharic Facebook and Twitter posts and comments into non-aggressive, covertly aggressive, and overtly aggressive comments to reduce the challenges of Amharic aggression using deep learning approach. The proposed model extracts features from FastText word embedding and classifies texts in terms of aggression levels. The FastText word embedding capable to handle out-of-vocabulary problems, spelling error for words and another challenging issue lack of efficient in representation of rare words. By comparing, the performance of the proposed word embedding with word2vec, the proposed word embedding better for aggression detection classification performance with 1.23% accuracy difference score. We developed BiLSTM model using FastText word embedding to classify Amharic social media posts and comments into neutral, covertly, and overtly aggressive. We compared with BiGRU and BiLSTM-CNN based on word2vec and FastText word embedding, the proposed model based on FastText embedding performed best with 0.7631, 0.71, 0.71, and 0.71 accuracy, F1-score, Precision and Recall respectively.