Decomposition-Enhanced Training for Post-Hoc Attributions In Language Models