Effect of Noise Sample Clauses

Effect of Noise. → Table 2 demonstrates the effect of noise on the de- velopment set. In row 1, “0+0” denotes there is no noise, which can be seen as an upper bound. Adding noise, either on the Chinese side or on the English side, deteriorates the F1 scores for all methods. Adding noise on the English side makes predicting phrase alignment in the C E direc- tion more challenging due to the enlarged search space. The situation is similar in the reverse di- rection. It is clear that agreement-based learning is more robust to noise: while independent train- ing suffers from a reduction of 40% in terms of F1 for the “20K + 20K” setting, agreement-based learning still achieves F1 scores over 70%. 0.50 0.45 0.40 0.35 F1 0.30 0.25 0.20 0.05 inner outer E−>C C−>E iteration adaptation for machine translation. The data set consists of two in-domain non- parallel corpora and an out-domain parallel cor- pus. The in-domain non-parallel corpora con- sists of 2.65M Chinese phrases and 3.67M English phrases extracted from LDC news articles. We use a small out-domain parallel corpus extracted from financial news of FTChina which contains 10K phrase pairs. The task is to extract a parallel corpus from in-domain non-parallel corpora start- ing from a small out-domain parallel corpus. We use the state-of-the-art translation system Xxxxx (Xxxxx et al., 2007) and evaluate the per- Figure 5: Comparison of F1 scores on the test set. Chinese jingji English economy Chinese jialebi English caribbean Chinese zhengzhi huanjing English political environment Chinese xxxxxxxxx xxxxxxxx xxxxx xxxxxx English exchange market price index Chinese qianding bianjing maoyi xieding English signed border trade agreements Table 3: Example learned parallel lexicons and phrases. New words that are not included in the seed lexicon are highlighted in italic.