Mining Parallel Data from Comparable Corpora via Triangulation

Abstract

This paper improves an unsupervised method for extracting parallel sentence pairs from a comparable corpus by using the triangulation through a third language. Before, an unsupervised method for extracting parallel sentence pairs from a comparable corpus has been proposed. This method is based on technique of cross-language information retrieval with iterative process and requires no more additional parallel data. The method has been validated on the Vietnamese-French and Vietnamese-English bilingual data. In this paper, we address the problem of using triangulation through a third language to improve the parallel data mining processes: English is used in the Vietnamese-French parallel data mining process, and French is used in the Vietnamese-English parallel data mining process. The experiments conducted show that using triangulation can improve the quality of the extracted data and the quality of the translation system as well.