An Action Is Worth Multiple Words: Handling Ambiguity in Action Recognition