Cybercrime is the bane of technology and computer networks, instilling anxiety in everyone who uses computers and the internet. While traditional crimes are on the decline, cybercrime is on the rise, with activities such as hacking, cyberstalking, denial of service, and phishing. Phishing assaults are a kind of social engineering, and one of the most common types of cybercrime to which most people are exposed to. Blacklists, which are lists of phishing URLs, have traditionally been used to detect phishing URLs. However, the blacklist method has a drawback, as new phishing URLs absent in the blacklist cannot be avoided. As a result, deep learning algorithms and Natural Language Processing techniques are employed to create and develop a model for real-time anti-phishing systems that understands the core semantic pattern differences between phishing and legitimate URLs. Hence, a deep learning architecture is designed over the concatenated features, where the concatenated features are the hand-crafted features and the convoluted character level embeddings of the URL. Furthermore, the model was experimented in three different ways with respect to activation functions, which included the model with only ReLU activation functions, the model with only Swish activation functions and the model with both activation functions. After a comparative analysis using various metrics, it is found that the model with ReLU activation functions performed better and yielded the highest accuracy of 96.83%.