Revisiting sentence alignment algorithms for alignment visualization and evaluation
Résumé
In this paper, we revisit the well-known problem of sentence alignment, in a context where the entire bitext has to be aligned and where alignment confidence measures have to be computed. Following much recent work, we study here a multi-pass approach: we first compute sure alignments that are used to train a discriminative model; then we use this model to fill in the gaps between sure links.
Experimental results on several corpus show the effectiveness of this method as compared to alternative, state-of-the-art, proposals.