Training non-parametric features for statistical machine translation

Abstract

Modern statistical machine translation systems may be seen as using two components: feature extraction, that summarizes information about the translation, and a log-linear framework to combine features. In this paper, we propose to relax the linearity constraints on the combination, and hence relaxing constraints of monotonicity and independence of feature functions. We expand features into a non-parametric, non-linear, and high-dimensional space. We extend empirical Bayes reward training of model parameters to meta parameters of feature generation. In effect, this allows us to trade away some human expert feature design for data. Preliminary results on a standard task show an encouraging improvement.