Hi, On Thu, Apr 4, 2013 at 9:04 AM, Per Tunedal <[email protected]> wrote: > Hi, > obviously word alignment is very important for extraction of the most > useful phrases. What about other ways for phrase extraction? > > In your text book 'Statistical Machine Translation' you do mention an > other approach: > > "we may also use the expectation maximization algorithm to directly find > phrase alignments for sentence pairs." (p. 8) > > Any work on that? Any success? Pros and cons?
Some old work on this is: Marcu, Daniel and Wong, Daniel (2002): A Phrase-Based, Joint Probability Model for Statistical Machine Translation, Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) Some more recent promising work is, for instance: Wuebker, Joern and Mauser, Arne and Ney, Hermann (2010): Training Phrase Translation Models with Leaving-One-Out, Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics > Further in the paper "Philipp Koehn, Franz Josef Och, Daniel Marcu: > Statistical Phrase-Based Translation (2003)" I've found the surprising > conclusion: > "straight-forward syntax-based mappings do not lead to better > translations > than unmotivated phrase mappings." > > Has anyone done any more research on this? How do this compare to > tree-based models and factored translations models, respectively? Hierarchical models are not different from phrase-based models, but syntax-based models have very strong restrictions to only allow rules for spans that match constituents on whatever side(s) syntax is used. -phi _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
