Achieving an robust and universal semantic representation for action description remains a key challenge in natural language understanding. Current approaches often struggle to capture the nuance of human actions, leading to imprecise representations. To address this challenge, we propose a novel framework that leverages hybrid learning techniques