A New Hybrid Machine Translation Approach Using Cross-Language Information Retrieval and Only Target Text Corpora

Parallel corpora play a vital role in Statistical Machine Translation. Nonavailability of these corpora is a major barrier for adding new languages pairs. In this paper, we propose a new hybrid approach for English-French machine translation combining a cross-language search engine and a statistical language model trained from a monolingual corpus. The cross-language search engine returns the translation candidates ordered by their relevance and the language model of the target language is used to disambiguate the translation. This approach has been evaluated and compared to Moses. We used 100000 French sentences of the Europarl corpus to train the language model, 1103 English-French sentences of the Arcade-II corpus as the translation reference and the BLEU score. The obtained scores are 21.33% for our approach and 21.45% for Moses. The experimental results also showed that our approach provides better translation performance in terms of grammatical coherence.