Multi-agent reinforcement learning(MARL) is receiving increasing attention due to its significance in real-world applications. To improve the efficiency of learning, most of the recent MARL works use homogeneous system. This implies that all agents share the same policy or value network. However, fully-shared parameters may have detrimental effects, such as causing all agents to behave similarly, ignoring information from different actions, and so on. To alleviate these issues, we introduce a novel framework known as Role-Based Heterogeneous Action Representation(RBHAR). Specifically, we first learn a vector representation for each role, with the same roles specialising in the certain subtasks. To reasonably assign agents to different roles, we encode empirical trajectories as implicit information that represents their current identity. All agents determine their individual roles by matching information with role vector representations. We further integrate information from different actions into the role policies to learn heterogeneous action representations based on roles, which is beneficial for the agent’s role to know what other agents it should pay attention to. Empirical results show that RBHAR is capable of learning reasonable role assignment and distinct action representations. These capabilities significantly improve the learning performance of our RBHAR on the StarCraft II micromanagement benchmark.