Sensing social representations (e.g. relationships and common interests) of a conversation and recognizing affect and metaphor from context are challenging but essential for the building of an intelligent agent with emotion and social intelligence. In this paper, we report contextual affect detection based on emotion modeling of personal and social improvisational context. We also discuss cooking and sensory metaphor (including temperature, light, taste, and smell metaphors) interpretation. The overall affect sensing model has been embedded in an AI agent interacting with human users. The evaluation results indicated that the new development made the AI agent perform generally better in improvisational interaction.