This project examines, five different approaches for the Turkish sentiment analysis task. Besides, the classical fine-tuning process of the BERT models, different hybrid BERT-CNN models that operate activations from different layers are proposed and the results are compared. To supervise the training and the fine-tuning processes a manually labeled Turkish tweet dataset that includes random tweets from various fields is created. The fine-tuned Turkish BERT model demonstrates the best performance among the five approaches. Moreover, this model achieves 0.82 accuracy on an open-source dataset that includes product reviews and texts from Wikipedia.