In this year's IWSLT evaluation campaign (TALK task), we applied three adaptation techniques: (1) training data selection based on information retrieval approach, (2) subsentence segmentation, and (3) language model adaptation using source-side of the test set. We also applied a sequential labeling method based on conditional random fields for restoring punctuation markers in the ASR input condition. We present and discuss these techniques in this paper, based on the automatic evaluation results.
Cite as: Sudoh, K., Duh, K., Tsukada, H. (2010) NTT statistical machine translation system for IWSLT 2010. Proc. International Workshop on Spoken Language Translation (IWSLT 2010), 147-152
@inproceedings{sudoh10_iwslt,
author={Katsuhito Sudoh and Kevin Duh and Hajime Tsukada},
title={{NTT statistical machine translation system for IWSLT 2010}},
year=2010,
booktitle={Proc. International Workshop on Spoken Language Translation (IWSLT 2010)},
pages={147--152}
}