Character-Level System Combination: An Empirical Study for English-to-Chinese Spoken Language Translation

This paper proposes a character-level system combination strategy for English -- Chinese spoken language translation. For languages like Chinese that the word boundaries are not orthographically marked, word segmentation which segments a Chinese sentence into a sequence of words, is often required for many Natural Language Processing tasks. In this paper we evaluate the impact of segmentation (spoken data) on the performance of system combination, and show that using inappropriate segmentation in system combination can result in inferior performance compared to single systems. We further demonstrate that using characters as basic translation unit in system combination on IWSLT ASR translation task leads to significant gains in translation quality in terms of BLEU and NIST scores.