Iterative, MT-based sentence alignment of parallel texts

Abstract

Recent research has shown that MT-based sentence alignment is a robust approach for noisy parallel texts.
However, using Machine Translation for sentence alignment causes a chicken-and-egg problem: to train a corpus-based MT system, we need sentence-aligned data, and MT-based sentence alignment depends on an MT system.
We describe a bootstrapping approach to sentence alignment that resolves this circular dependency by computing an initial alignment with length-based methods.
Our evaluation shows that iterative MT-based sentence alignment significantly outperforms widespread alignment approaches on our evaluation set, without requiring any linguistic resources other than the to-be-aligned bitext.