Realignment from Finer-grained Alignment to Coarser-grained Alignment to Enhance Mongolian-Chinese SMT

The conventional Mongolian-Chinese statistical machine translation (SMT) model uses Mongolian words and Chinese words to practice the system. However, data sparsity, complex Mongolian morphology and Chinese word segmentation (CWS) errors lead to alignment errors and ambiguities. Some other works use finer-grained Mongolian stems and Chinese characters, which suffer from information loss when inducting translation rules. To tackle this, we proposed a method of using finer-grained Mongolian stems and Chinese characters for word alignment, but coarser-grained Mongolian words and Chinese words for translation rule induction (TRI) and decoding. We presented a heuristic technique to transform Chinese character-based alignment to word-based alignment. Experimentally, our method outperformed the baselines: fully finergrained and fully coarser-grained, in terms of alignment quality and translation performance.