Statistical Machine Translation without Parallel Data

We examine approaches of statistical machine translation without parallel data (SMT). SMT has achieved impressive performance by leveraging large amounts of parallel data in the source and target languages. But such data is available only for a few language pairs and domains. Using human annotation to create new parallel corpora sufficient for building a good translation system is too expensive. On the other hand, there are many resources of text for many languages. This has raised a new research challenge in SMT: How can we train a statistical language translation system without parallel data? There has been a long line of research on learning translation from monolingual data, beginning with Rapp (1995). Many of these works have focused on extracting a translation lexicon by mining monolingual resources of data. More recent works have extended research from building translation lexicons to translation systems. We categorize methods of SMT without parallel data based on end product (translation lexicon or translation system) and type of available resources which can be used (limited parallel information or just monolingual data). We characterize the methods in terms of quality of translation and scalability.