Gappy Phrasal Alignment by Agreement

TitleGappy Phrasal Alignment by Agreement
Publication TypeConference Paper
Year of Publication2011
AuthorsBansal, M., Quirk C., & Moore R. C.
Page(s)1308-1317
Other Numbers3269
Abstract

We propose a principled and efficient phraseto-phrase alignment model, useful in machinetranslation as well as other related natural languageprocessing problems. In a hidden semi-Markov model, word-to-phrase and phraseto-word translations are modeled directly bythe system. Agreement between two directionalmodels encourages the selection of parsimoniousphrasal alignments, avoiding theoverfitting commonly encountered in unsupervisedtraining with multi-word units. Expandingthe state space to include “gappyphrases” (such as French ne pas) makes thealignment space more symmetric; thus, it allowsagreement between discontinuous alignments.The resulting system shows substantialimprovements in both alignment quality andtranslation quality over word-based HiddenMarkov Models, while maintaining asymptoticallyequivalent runtime.

Acknowledgment

This work was partially supported by funding provided to ICSI by a gift from Microsoft Research.

URLhttp://www.icsi.berkeley.edu/pubs/speech/gappyphrasalalignment11.pdf
Bibliographic Notes

Proceedings of the 49th annual Meeting of the Association for Computational Linguistics (ACL HLT 2011), pp. 1308-1317 Portland, Oregon

Abbreviated Authors

M. Bansal, C. Quirk, and R. C. Moore

ICSI Research Group

Speech

ICSI Publication Type

Article in conference proceedings