An Axiomatic Approach to Regularizing Neural Ranking Models.

摘要: Axiomatic information retrieval (IR) seeks a set of principle properties desirable in IR models. These when formally expressed provide guidance the search for better relevance estimation functions. Neural ranking models typically contain large number parameters. The training these involve appropriate parameter values based on quantities labeled examples. Intuitively, axioms that can guide traditional should also help machine learning rankers. This work explores use to augment direct supervision from data neural We modify documents our dataset along lines well-known during and add regularization loss agreement between model which version document---the original or perturbed---should be preferred. Our experiments show achieves faster convergence generalization with axiomatic regularization.