An Empirical Study of Translation Rule Extraction with Multiple Parsers

Translation rule extraction is an important issue in syntax-based Statistical Machine Translation (SMT). Recent studies show that rule coverage is one of the key factors affecting the success of syntaxbased systems. In this paper, we first present a simple and effective method to improve rule coverage by using multiple parsers in translation rule extraction, and then empirically investigate the effectiveness of our method on ChineseEnglish translation tasks. Experimental results show that extracting translation rules using multiple parsers improves a string-to-tree system by over 0.9 BLEU points on both NIST 2004 and 2005 test corpora.