Improving Maximum Likelihood Training for Text Generation with Density Ratio Estimation