Parallel corpora segmentation using anchor words
Francisco Nevado, Francisco Casacuberta, Enrique Vidal · 2003
A new technique for monotone segmentation of parallel corpora is introduced. This segmentation is based on a set of anchor words which are defined manually. The parallel segments are computed using a dynamic programming algorithm. To assess this technique, finite-state transducers are inferred from both non-segmented and segmented corpora. Experiments have been carried out with Spanish-English and Italian-English translation tasks. This technique has proven useful in improving the results with respect to those obtained with unsegmented corpora.