A comparison of discriminative training criteria for continuous space translation models

摘要

This paper explores a new discriminative training procedure for continuous-space translation models (CTM s) which correlates better with translation quality than conventional training methods. The core of the method lays in the definition of a novel objective function which enables us to effectively integrate the CTM with the rest of the translation system through N-best rescoring. Using a fixed architecture, where we iteratively retrain the CTM parameters and the log-linear coefficients, we compare various ways to define and combine training criteria for each of these steps, drawing inspirations both from max-margin and learning-to-rank techniques. We experimentally show that a recently introduced loss function, which combines these two techniques, outperforms several objective functions from the literature. We also show that ensuring the consistency of the losses used to train these two sets of parameters is beneficial to the overall performance.