Exploiting Abduction for Learning from Incomplete Interpretations