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