TargetSide Generation of Prepositions for SMT

Target-side Generation of Prepositions for SMT Marion Weller1,2, Alexander Fraser2, Sabine Schulte im Walde1 1 IMS, University of Stuttgart – (wellermn|schulte)@ims.uni-stuttgart.de 2 CIS, Ludwig-Maximilian University of Munich – fraser@cis.uni-muenchen.de Abstract We present a translation system that mod- els the selection of prepositions in a target- side generation component. This novel ap- proach allows the modeling of all subcate- gorized elements of a verb as either NPs or PPs according to target-side requirements relying on source and target side features. The BLEU scores are encouraging, but fail to surpass the baseline. We addi- tionally evaluate the preposition accuracy for a carefully selected subset and discuss how typical problems of translating prepo- sitions can be modeled with our method. 1 Introduction The translation of prepositions is a difficult task for machine translation; a preposition must convey the source-side meaning while also meeting target- side constraints. This requires information that is not always directly accessible in an SMT sys- tem. Prepositions are typically determined by gov- ernors, such as verbs (to believe in sth.) or nouns (interest in sth.). Functional prepositions tend to convey little meaning and mostly depend on target- side restrictions, whereas content-bearing preposi- tions are largely determined by the source-side, but may also be subject to target-side requirements, as in the following example: go to the cinema/to the beach → ins Kino/an den Strand gehen. In this paper, we treat prepositions as a target- side generation problem and move the selection of prepositions out of the translation system into a post-processing component. During translation, c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. we use an abstract representation of prepositions as a place-holder that serves as a basis for the gen- eration of prepositions in the post-processing step. In this step, all subcategorized elements of a verb are considered and allotted to their respective func- tions – as PPs with an overt preposition, but also as NPs with an “empty” preposition, e.g. to call for sth. → ∅ etw. erfordern. In a standard SMT system, subcategorization is difficult to capture in the lan- guage model or by the translation rules if the verb and its subcategorized elements are not adjacent. In the following, we outline a method to handle prepositions with a target-side generation model in an English-German morphology-aware SMT sys- tem. We study two aspects: (i) features for a mean- ingful abstract representation of prepositions and (ii) how to predict prepositions in the translation output using a combination of source and target- side information. In addition, we compare prepo- sitions in the machine translation output with those in the reference translation for a selected subset. Finally, we discuss examples illustrating typical problems of translating prepositions. 2 Related Work Most research on translating prepositions has been reported for rule-based systems. Naskar and Bandyopadhyay (2006) outline a method to han- dle prepositions in an English-Bengali MT system using WordNet and an example base for idiomatic PPs. Gustavii (2005) uses bilingual …

Content truncated to 500 words