The ParTrans Project

The Parallel Grammar Project (Pargram) is interesting for what it tells us about the robustness of our grammatical formalisms and techniques and about the extent to which these support solutions to problems that generalize across languages. It also provides a natural framework in which to develop new approaches to a problem that has always been a primary source of motivation for computational linguists, namely machine translation.

XLE gives us powerful technology for both analyzing and generating sentences, and Pargram gives us grammars that enable us to apply it to similar phenomena in a number of languages. Under the title of Partrans, we also have worked on an experimental program for effecting what is usually called "transfer" between sentence structures in a pair of languages, thus providing us with the core of an experimental machine translation system. The program applies to the output of a parser for the source language to provide input to the generator for the target language.

Our transfer system lets you write hand-crafted transfer rules between a pair of languages. As an experiment Anette Frank spent about a year writing the transfer rules needed to translate part of a tractor operations manual from French into English. Also, Stefan Riezler and John Maxwell wrote a program for automatically extracting transfer rules from a corpus of pairs of f-structures that were in German and English. We are interested in applying the same technique to pairs of f-structures that are in Chinese and English.

An important innovation of our transfer system is that it uses the same compact representation of alternative parts of a linguistic structure that make these structures so compact inside the XLE system and which are one of the main sources of its speed. Generally speaking, when a rule is applied in our transfer system, contributions are made to several different eventual translations.

The use of these techniques in the central component of a translation system has led us to explore other new avenues in translation. In particular, it has led us to a notion that we refer to as "triangulation". The key idea is this. Much of the ambiguity of a text that makes it hard to translate into another language is resolved if a translation into some third language is available. The trick is to find general techniques for bringing the information in the other translation to bear.

Our way of representing alternative structures suggests an approach. Each small alternative in the representation reflects a choice that the program faced but was unable to resolve. When the same text is translated into other languages, the program will therefore often have the information available to make the proper choice when it arises.

Selected papers

  • Stefan Riezler and John Maxwell (2006). "Grammatical Machine Translation" In HLT-NAACL'06 New York, New York.

  • Anette Frank (1999). "From Parallel Grammar Development towards Machine Translation" In Proceedings of MT Summit VII. "MT in the Great Translation Era", September 13-17, Kent Ridge Digital Labs, Singapore, 134-142, 1999.

  • Anette Frank (1999). "From Parallel Grammar Development to Machine Translation" (shortened version) Section 4 of Miriam Butt, Stefanie Dipper, Anette Frank, Tracy Holloway King (1999): ``Writing Large-Scale Parallel Grammars for English, French, and German'' (ps), (html) in Miriam Butt and Tracy Holloway King (eds): Proceedings of the LFG99 Conference, University of Manchester, CSLI Online Publications .

  • Michael Dorna, Anette Frank, Josef van Genabith and Martin Emele (1998). "Syntactic and Semantic Transfer with F-Structures" in Proceedings of the 17th International Conference on Computational Linguistics (COLING '98) and the 36th Annual Meeting of the Association for Computational Linguistics (ACL'98) (COLING-ACL'98), Montréal, Québec, Canada. August, 1998.

    Last updated: .