First-order regret bounds for combinatorial semi-bandits