Discovering Human-Object Interaction Concepts via Self-Compositional Learning