Semantic relation classification is the task of identifying or recognizing the relation between entities. It is mainly used to understand and analyses the meaning of the sentence. Semantic relation classification helps to easily classify the relation which exists between entities. Relation classification is a crucial component for other application areas of natural language processing. In the previous study to classifying semantic relation between entities, traditional approaches such as (feature-based, kernel based), and deep learning approaches are used by different researchers for other languages such as English, Arabic, French, and Chines language. Currently to classify semantic relation between entities deep learning approach attract great attention. Amharic language has its own grammatical rules, syntax and it uses geez characters (Fidel). For this morphologically rich and under-recourse language, there are some works in different application areas such as named entity recognition, information extraction, and text summarization, but there is no work to classifying semantic relation between Amharic entities. Semantic relation classification used as supportive tools for other studies including information extraction, question answering, and natural language inference. In this study, we designed BiLSTM network model to classify the semantic relation between Amharic entities. In our study, we used spatial drop out to reject less significant feature vectors and to reduce the complexity of feature selection. We used weighted attention mechanisms to select the relevant feature dynamically which helps to improve the performance. To handle the problem of overfitting we used drop-out to reject less weighted network nods outputs. To handle the sequence and to recognize the pattern of the sentence we used BiLST, and softmax used as classifiers. In our model, we used Keras to build the model and TensorFlow as backend. We collected 4171 Amharic sentences from Fana broadcasting corporation, Walta information center, and Ethiopian news agency. We carried out experiments using LSTM, BiLSTM, and BiLSTM network models with attention mechanisms. The proposed model is evaluated by using different VIII performance evaluation metrics. Finally, the proposed semantic relation classification model is evaluated through testing dataset and achieves 77.34% accuracy