Parsing and Word Alignment Experiments. The models presented in Chapters 3 and 4 both use features derived from unsupervised word aligners and monolingual parsers. The unsupervised word aligner we used was the paired HMM included in the Berkeley Aligner (Xxxxx et al., 2006). The aligner was trained on all the available parallel data except for the MT tuning and test sets (i.e. the training set from Table 2.2 plus all the corpora in Table 2.3). To train the HMMs, we initialized with 5 EM iterations of joint Model 1 training, followed by 5 EM iterations of independent HMM training. The two directional models were symmetrized using the soft union heuristic with 1The tree transformation system described in Chapter 5 doesn’t actually require a bilingual treebank, but it is still useful to use one as a clean setting for qualitative assessment of the system. 2Many of the translations were of multiple English sentences translated from a single Chinese sentence, or more rarely, the other way around. competitive thresholding (XxXxxx and Xxxxx, 2007). When needed, a “one-best” alignment was generated by thresholding the symmetrized posteriors at 0.33. We used the Berkeley Parser (Xxxxxx and Xxxxx, 2007) for both our English and Chinese monolingual parsers. However, selecting the data to train the grammars for these parsers was a bit subtle. To maximize parser performance, we naturally wished to make use of all available annotated data. Thus, at test time, we used grammars trained on each language’s respective full monolingual training set plus the appropriate monolingual projection of the bilingual training set.3 However, at training time, it was important that the monolingual grammars not be trained on any of the bilingual training data – otherwise, the monolingual parsing features would be too powerful, and there would be no way to learn appropriate weights for the remaining features in the bilingual model. Thus, when training the bilingual models, we used monolingual grammars trained on less data. For Chinese, we simply omitted the entire bilingual corpus, using only the monolingual training data. For English, we had to be more careful, as the monolingual corpus (WSJ) was out-of-domain relative to the test data, and thus it was fairly important that some in-domain data be included when training the monolingual grammars. We solved this issue with a very simple jackknifing approach. We trained two English grammars: one that appended the first half of the English projection of the bilingual training cor...
