In Search of the Best Method for Sentence Alignment in Parallel Texts
Alexandr Rosen · 2005
After a brief account of a parallel corpus project involving many diverse languages and a summary of two previous evaluations of sentential alignment tools, results are presented from tests of three automatic aligners on English-Czech and French-Czech literary and legal texts, clean and noisy. The results confirm that an alignment tool may perform well on one type of texts and fail on another type, and indicate that near-to-perfect alignment is possible when tools providing high precision are combined with manual checking, where the proofreader can focus only on those parts of the text that were either not aligned at all, or that were aligned less reliably. Further gains in precision are shown to be feasible when alignments proposed by multiple aligners are intersected.