We show how to address the problem of bilingual data scarcity in machine translation. We propose a method that generates aligned sentences which may be not perfect translations. It consists in 'hallucinating' new sentences which contain small but well-attested variations extracted from unaligned unrelated monolingual data. We conducted various experiments in statistical machine translation between Chinese and Japanese to determine when adding such quasi-parallel data to a basic training corpus leads to increases in translation accuracy as measured by BLEU.
@InProceedings{LEPAGE18.11, author = {Yves Lepage}, title = {Quasi-Parallel Corpora: Hallucinating Translations for the Chinese-Japanese Language Pair}, booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {may}, date = {7-12}, location = {Miyazaki, Japan}, editor = {Reinhard Rapp and Pierre Zweigenbaum and Serge Sharoff}, publisher = {European Language Resources Association (ELRA)}, address = {Paris, France}, isbn = {979-10-95546-07-8}, language = {english} }
Summary of the paper