Aligning Words in English-Hindi Parallel Corpora

In this paper, we describe a word alignment algorithm for English-Hindi parallel data. The system was developed to participate in the shared task on word alignment for languages with scarce resources at the ACL 2005 workshop, on "Building and using parallel texts: data driven machine translation and beyond". Our word alignment algorithm is based on a hybrid method which performs local word grouping on Hindi sentences and uses other methods such as dictionary lookup, transliteration similarity, expected English words and nearest aligned neighbours. We trained our system on the training data provided to obtain a list of named entities and cognates and to collect rules for local word grouping in Hindi sentences. The system scored 77.03% precision and 60.68% recall on the shared task unseen test data.