Analyzing the Influence of Parsing Errors on Pre-reordering Performance for SMT

Word alignment for long distance language pairs is problematic in state-of-the-art phrasebased statistical machine translation. Linguistically motivated reordering models have been widely studied to conquer this challenge. One of the most popular and effective methods is called pre-reordering, where words in sentences from the source language are re-arranged with the objective to resemble the word order of the target language. There are mainly two ways to formulate re-arranging rules. One is to learn automatically from the data (Xia and McCord, 2004; Genzel, 2010); while another one is to handcraft reordering rules based on linguistic studies (Isozaki et al., 2010; Han et al., 2012; Han et al., 2013). In both methods, syntactic information are obtained by using automatic parsers. However, although these parsers produce parsing errors, current reordering methods do not include any parsing error identification or correction mechanism. In order to improve the robustness of pre-reordering method, it is useful to explore the relationship between parsing and reordering. In this work, we use both empirical and descriptive approaches to analyze the effects of parsing errors on pre-reordering performance for Chinese-to-Japanese statistical machine translation. We examine the impact of parsing errors that are produced by a dependency parser called Corbit1 (Hatori et al., 2011), on a pre-reordering framework called unlabeled dependency parsing