Tree-Based Target Language Modeling

Proceedings of the 13th Annual Conference of the EAMT, pages 152–159, Barcelona, May 2009 Tree-based Target Language Modeling Vincent Vandeghinste Centre for Computational Linguistics - KULeuven Leuven, Belgium vincent@ccl.kuleuven.be Abstract In this paper we describe an approach to target language modeling which is based on a large treebank. We assume a bag of bags as input for the target language gener- ation component, leaving it up to this com- ponent to decide upon word and phrase or- der. An experiment with Dutch as target language shows that this approach to can- didate translation reranking outperforms standard n-gram modeling, when measur- ing output quality with BLEU, NIST, and TER metrics. 1 Acknowledgements The development of this system and research is made possible by the STEVIN-programme of the Dutch Language Union, Project Nr. STE-07007, which is sponsored by the Flemish and Dutch Governments, and by the SBO-programme of the Flemish IWT, Project Nr. 060051. 2 Introduction In this paper we describe an approach to target lan- guage modeling using large treebanks. This intro- duction starts with a description of the MT system for which this target language modeling compo- nent is intented and continues with a short descrip- tion of related research. In section 3 we describe the details of the target language modeling component and in section 4 we describe an evaluation experiment for this compo- nent. Section 5 draws conclusions and sketches future work. 2.1 System description We are developing a data-driven hybrid approach towards machine translation, reusing as much as c � 2009 European Association for Machine Translation. possible already existing tools and resources to set up an MT architecture much like a classic rule- based transfer system. Instead of manually design- ing the rules, we intend to derive them from large parallel and monolingual (uncorrected) treebanks. The system requires a source language parser and a parallel treebank, aligned from the sen- tence level up to the word level (Och and Ney, 2003), including sub-sentential alignment (Tiede- mann, 2003; Tinsley et al., 2007, Macken and Daelemans, 2008). To get a parallel treebank we parse both the source and target language com- ponents of parallel corpora `a la Europarl (Koehn, 2005). Each tree pair, sub-tree pair or word pair presents an example translation pair, and becomes a dictionary entry. This way we are removing the conceptual distinction between a dictionary and a parallel corpus, like Vandeghinste (2007). In a similar fashion, but making abstraction of the concrete words, we derive a set of transfer rules from the available alignments. A translation model is built by counting the frequencies of occurrence of all these alignments. The source language sentence is syntactically parsed, and the parse tree (and its sub-trees) is matched with the source language side parse trees of the dictionary/parallel treebank. The retrieved target fragments are then restructured according to the information in the transfer rules resulting in a target language bag of bags, which is structured like a parse tree, but without implying any surface order in the daughters …

Content truncated to 500 words