Lexicons or phrase tables? An investigation in sampling-based multilingual alignment

Adrien Lardilleux (1) , Jonathan Chevelu (1, 2) , Yves Lepage (1) , Ghislain Putois (2) , Julien Gosme (1)

1 Equipe Hultech - Laboratoire GREYC - UMR6072
2 Orange Labs [Lannion]

Sampling-based multilingual alignment is an example-based approach to sub-sentential alignment that has proven to be able to outperform ubiquitous statistical models on some tasks. As for machine translation however, it still typically does not provide as good results as was first expected. In this paper, we propose two experiments to determine the nature of alignments produced by this method, and what they would still lack of. We then deduce what possible improvements will make the method perform better on machine translation tasks.

Domaines: Informatique et langage [cs.CL]; Traitement du texte et du document

Adrien Lardilleux, Jonathan Chevelu, Yves Lepage, Ghislain Putois, Julien Gosme. Lexicons or phrase tables? An investigation in sampling-based multilingual alignment. 3rd Workshop on Example-­Based Machine Translation, Nov 2009, Dublin, Ireland. pp.45-52. ⟨hal-00439806⟩