<p>Political parties utilize social media platforms to transparently disseminate their policies, development plans, and strategies to the public. This approach encourages social media users and voters to actively engage by responding to and providing feedback on these political issues. As a result, political parties can consider the public feedback as an input for making different decisions according to the public comments and ideas with respect to their governmental policy and strategies. However, manually assessing those massive messages and feedbacks in order to make informed decisions is exceedingly tough. Because reading or reviewing such a big number of comments and reviews using standard methods is difficult and time consuming. As a result, sentiment classification, or the computational analysis of people's thoughts, attitudes, and opinions about a certain subject, has received a lot of interest. In this thesis, we used deep learning techniques such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Long Short-Term Memory (BiLSTM) to develop sentiment classification model. In addition, word embedding methods like Word2Vec and FastText were examined to improve classification performance. The goal of the thesis is to use deep learning approaches to classify Amharic social media political posts and comments into Positive, Negative, and Neutral, making the sentiment classification problem effective. We compared various deep learning algorithms and word embedding approaches in order to determine which is the best. The proposed method able to achieve increased accuracy, precision, recall, and F1-score with 0.84, 0.92, 0.86, and 0.89, respectively, utilizing BiLSTM with FastText embedding. This promising result can help political parties and voters in quickly analyzing various opinions by reducing the amount of time and effort required to read these posts and comments.</p>