Machine Translation Experiments Sample Clauses

Machine Translation Experiments. ‌ All our machine translation experiments were carried out from Chinese to English, using the syntactic tree-to-string pipeline included in the latest release of Moses (Xxxxx et al., 2007). This pipeline requires target-side parse trees, so we automatically annotated 3In Chinese, this corresponded to the full standard training set from the Chinese Treebank v5.0. In English, this augments the standard WSJ training set with an additional 2200 sentences.‌ 4Chapter 3 also contains development set results that were obtained with a slightly different experimental setup. The setup was similar, but the baseline parsers used during development were weaker, so the dev set results aren’t directly comparable to the final test results. the training data with English parses and with word alignments. The baseline setup uses parse trees from a monolingual English parser and word alignments from the same HMM word aligner described in Section 2.2. The only differences between our various MT condi- tions were in the annotations. Typically, only the English parse trees changed, although we did also swap in different word alignments for experiments that involved the joint parsing and word alignment model. We trained, tuned, and tested the MT system using the data in Table 2.2. The evaluation metric we used for both tuning and testing was multi-reference BLEU (Xxxxxxxx et al., 2001).5 We tuned parameters with MERT (Och, 2003). However, since XXXX is known to be unstable (Xxxxx et al., 2011), we ran MERT 10 times for each data condition; the scores reported here are averages taken across these 10 runs.