A Simple, Fast, and Effective Reparameterization of IBM Model 2

TLDR

The paper proposes a simple log-linear reparameterization of IBM Model 2 to address Model 1’s strong assumptions and Model 2’s overparameterization. The authors provide efficient inference, likelihood evaluation, and parameter estimation algorithms for the reparameterized model. The reparameterized model trains ten times faster than Model 4 and outperforms it on three large-scale translation tasks, with an open-source implementation available at http://github.com/clab/fast‑align.

Abstract

We present a simple log-linear reparameterization of IBM Model 2 that overcomes problems arising from Model 1’s strong assumptions and Model 2’s overparameterization. Efficient inference, likelihood evaluation, and parameter estimation algorithms are provided. Training the model is consistently ten times faster than Model 4. On three large-scale translation tasks, systems built using our alignment model outperform IBM Model 4. An open-source implementation of the alignment model described in this paper is available from http://github.com/clab/fast align .