One of the most crucial components of natural human-robot interaction is artificial intuition and its influence on dialog systems. The intuitive capability that humans have is undeniably extraordinary, and so remains one of the greatest challenges for natural communicative dialogue between humans and robots. In this paper, we introduce a novel probabilistic modeling framework of identifying, classifying and learning features of sarcastic text via training a neural network with human-informed sarcastic benchmarks. This is necessary for establishing a comprehensive sentiment analysis schema that is sensitive to the nuances of sarcasm-ridden text by being trained on linguistic cues. We show that our model provides a good fit for this type of real-world informed data, with potential to achieve as accurate, if not more, than alternatives. Though the implementation and benchmarking is an extensive task, it can be extended via the same method that we present to capture different forms of nuances in communication and making for much more natural and engaging dialogue systems.