Handling Ties Correctly and Efficiently in Viterbi Training Using the Viterbi Semiring

摘要

The handling of ties between equiprobable derivations during Viterbi training is often glossed over in research paper, whether they are broken randomly when they occur, or on an ad-hoc basis decided by the algorithm or implementation, or whether all equiprobable derivations are enumerated with the counts uniformly distributed among them, is left to the readers imagination. The first hurts rarely occurring rules, which run the risk of being randomly eliminated, the second suffers from algorithmic biases, and the last is correct but potentially very inefficient. We show that it is possible to Viterbi train correctly without enumerating all equiprobable best derivations. The method is analogous to expectation maximization, given that the automatic differentiation view is chosen over the reverse value/outside probability view, as the latter calculates the wrong quantity for reestimation under the Viterbi semiring. To get the automatic differentiation to work we devise an unbiased subderivative for the max function.