Alignment Inference and Bayesian Adaptation for Machine Translation
Kevin Duh, Katsuhito Sudoh, Tomoharu Iwata, Hajime Tsukada · 2011
We propose a flexible and efficient domain adaptation method that yields consistent im-provements in machine translation (for 11 lan-guage pairs). The idea is to decompose the word alignment process into two steps, model training and alignment inference, and perform Bayesian adaptation on the latter. This modu-larity allows one to incorporate out-of-domain data without the need to modify existing train-ing algorithms. We show how ideas in sequen-tial Bayesian methods can be naturally applied to the word alignment problem and demon-strate various positive results on EMEA and NIST datasets. 1