A Fully Unsupervised Approach for Mining Parallel Data from Comparable Corpora

This paper presents an unsupervised method for extracting parallel sentence pairs from a comparable corpus. A translation system is used to mine the comparable corpus and to detect parallel sentence pairs. An iterative process is implemented not only to increase the number of extracted parallel sentence pairs but also to improve the overall quality of the translation system. A comparison between this unsupervised method and a semi-supervised method is also presented. The unsupervised method was tested in a hard condition: no available parallel corpus to bootstrap the process and the comparable corpus contained up to 50% of non parallel data. The experiments conducted show that the unsupervised method can be really applied in the case of lacking parallel data. While preliminary experiments are conducted on French-English translation, this unsupervised method is also applied successfully to a low e-resourced language pair (French-Vietnamese).