Learning semantic templates for example-based machine translation (EBMT)


The aim of my research is to investigate dynamic automatic learning of semantic templates for example based translation. This implies two important problems in EBMT: involvement of semantic information in the translation templates and dynamic learning of templates.

I focus on two highly inflected languages - Romanian and German, which is itself an interesting problem, as most part of the example-based translation systems involved English, Japanese or French. Moreover this kind of approach will prove the utility of example-based translation systems for languages for which electronic linguistic resources are still quite poor.

In my presentation I will show you how I tend to develop a language neutral approach for extracting semantic templates, where the templates are stored as strings, contain semantic information and are generalizations of examples. Many papers report problems of memory usage and time of processing during template extraction. In order to overcome this problem, I came up with a solution, not found in any of the consulted literature, namely extraction of a similarity matrix, which limits the search space and overall time used for extracting semantic templates.

I will also give examples of how generalization by syntactic and morphologic categories can produce wrong templates, due to the fact that a word with the same syntactic function and the same morphologic characteristics can have two translation equivalents in a target language, which is why semantic information in the template is useful, at least for two common problems: learning false templates and translation of homonyms.

-- GavrilaMonica -- 19 Jun 2006
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Foswiki? Send feedback