The utilization of Large Language Models (LLMs) has revitalized the pursuit of human-centered Artificial General Intelligence (AGI), offering a promising path toward machines capable of nuanced social interaction. Among the multifaceted traits of human intelligence, empathy represents a fundamental component, crucial for fostering meaningful and trust-driven communication. While prior studies have predominantly sought to augment cognitive empathy via external knowledge integration, comparatively little attention has been given to the intrinsic conversational qualities of sensibility and rationality—two dimensions essential to authentic empathetic exchange. In practice, conversational rationality is often constrained by limited contextual cues, and naive knowledge augmentation methods risk semantic conflicts or a restricted single-role perspective. The paper introduces an encoder grounded in the sociological theory of self-presentation, explicitly disentangling and modeling sensible and rational utterances within dialogues. We further employ an LLM as a “rational reasoning engine” to infer deep-seated logical structures embedded in conversational history, enabling the system to assess and balance sensibility with rationality in generating empathetic responses. Experimental evaluations, across both automated metrics and human judgment, demonstrate that our approach delivers responses that are not only more emotionally attuned but also more contextually coherent than state-of-the-art baselines.