According to ICW, corruption cases in Indonesia in the last 5 years have increased and the amount of losses suffered by the state from 2012-2022 reached Rp138.39 trillion. According to Transparency International, Indonesia's CPI ranking decreased in 2023 to 115 compared to 2022 at 110 out of 180 countries. These results show that the response to corruption is still slow and continues to deteriorate due to a lack of support from stakeholders. The purpose of this study is to test and compare the performance of deep learning model algorithms (RNN/LSTM/GRU/Bi-GRU/Bi-LSTM) on sentiment classification using word embedding, and obtain a model architecture that can determine the polarity of a sentence about public sentiment related to corruption in Indonesia, which can help governments, researchers, and practitioners in designing more effective anti-corruption strategies. The dataset used amounted to 1793 derived from crawling Twitter with 3 classes namely positive, negative and neutral. This research starts from data collection, preprocessing, word embedding, splitting the dataset which is divided into 80% training data and 20% test data, deep learning model testing, model evaluation and result representation. Word embedding uses word2vec with a dimension of 300. Based on the experimental results obtained, Bi-GRU has better performance than other architectural models with an accuracy value of 88%, precision 88.07%, recall 86.97% and f1-score 87.51%. The data used in this research is relatively small, it is recommended that future research can overcome it