Learning to rank has shown great value in web search, question answering and recommender systems. This paper focuses on supervised listwise document ranking, where documents in a query response are assigned scores with a ranking function, whose dissimilarity to ground truth relevance levels is quantified with a loss function and then minimized. We propose to use \(\alpha \) -divergence to replace Kullback-Leibler divergence or cross-entropy as the loss function. Specifically, \(\alpha \) -divergence is a generalization of Kullback-Leibler divergence; by tuning the hyper parameter \(\alpha \) , trained models can have better balance between assigning high scores to relevant documents (inclusiveness) and assigning low scores to irrelevant documents (exclusiveness). Since ground truth relevance levels have much less granularity than score values, we further propose to use score entropy regularization to encourage score diversity among documents at the same relevance level. To validate the effectiveness of the proposals, we conduct a series of experiments on two benchmark datasets. Specifically, the improvements of \(\alpha \) -divergence on nDCG@1, ERR@1 are up to 3.37%, 4.45%, and entropy regularization improves up to 3.69%, 5.46%, respectively. The combination of the two methods can further improve metrics under optimization. Some theoretical properties of the \(\alpha \) -divergence loss function are provided. We further prove that the loss function has all its stationary points as its global minimum if the Jacobian of the transformation function has left null space as \(Span({\textbf {1}})\) .