Microsoft Research Treelet Translation System: NIST MT Evaluation 06
Xiaodong He, Arul Menezes, Chris Quirk, Anthony Aue, Simon H. Corston-Oliver, Jianfeng Gao, Patrick Nguyen · 2006
The Microsoft Research Treelet Translation System is a syntax-based SMT system employing a source language dependency parser. The system translates by composing dependency treelet translation pairs, ranked by a log-linear combination of models including a global reordering model informed by a source language dependency parser. The source language analysis is produced by a Chinese Treebank parser. The log-linear combination also includes standard channel models, language models, and common feature functions. These weights are optimized by maximizing BLEU on a development set. 1.