Providing neural models with a persona profile has demonstrated that conditioning such models on specific characteristics produces more coherent and engaging conversations. However, current representations remain shallow, typically relying on static biographical facts. In reality, human communication is shaped by both profound internal traits – such as goals, opinions, values, and beliefs – and dynamic situational factors, including emotional states, conversational context, and the specific relationship between speakers. The goal of this PhD Thesis is to bridge these dimensions, moving beyond rigid profiles to develop adaptive, context-sensitive persona models. By integrating deep profile characteristics with real-time social dynamics, the research aims to advance the state of language generation for conversational agents across multiple domains and languages.