Phrase alignment for integration of SMT and RBMT resources
Akira Ushioda · 2007
A novel approach is presented for extracting syntactically motivated phrase alignments. In this method we can incorporate conventional resources such as dictionaries and grammar rules into a statistical optimization framework for phrase alignment. The method extracts bilingual phrases by incrementally merging adjacent words or phrases on both source and target language sides in accordance with a global statistical metric. Phrase alignments are extracted from parallel patent documents using this method. The extracted phrases used as training corpus for a phrase-based SMT showed better cross-domain portability over conventional SMT framework.