Using subcategorization knowledge to improve case prediction for translation to German

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 593–603, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Using subcategorization knowledge to improve case prediction for translation to German Marion Weller1 Alexander Fraser2 Sabine Schulte im Walde1 1Institut f¨ur Maschinelle 2Centrum f¨ur Informations- Sprachverarbeitung und Sprachverarbeitung Universit¨at Stuttgart Ludwig-Maximilians-Universit¨at M¨unchen {wellermn|schulte}@ims.uni-stuttgart.de fraser@cis.uni-muenchen.de Abstract This paper demonstrates the need and im- pact of subcategorization information for SMT. We combine (i) features on source- side syntactic subcategorization and (ii) an external knowledge base with quantita- tive, dependency-based information about target-side subcategorization frames. A manual evaluation of an English-to- German translation task shows that the subcategorization information has a posi- tive impact on translation quality through better prediction of case. 1 Introduction When translating from a morphologically poor language to a morphologically rich language we are faced with two major problems: (i) the rich- ness of the target-language morphology causes data sparsity problems, and (ii) information about morphological features on the target side is not sufficiently contained in the source language mor- phology. We address these two problems using a two- step procedure. We first replace inflected forms by their stems or lemmas: building a translation system on a stemmed representation of the target side leads to a simpler translation task, and the morphological information contained in the source and target language parts of the translation model is more balanced. In the second step, the stemmed output of the translation is then inflected: the mor- phological features are predicted, and the inflected forms are generated using the stem and predicted morphological features. In this paper, we focus on improving case pre- diction for noun phrases (NPs) in German trans- lations. The NP feature case is extremely dif- ficult to predict in German: while the NP fea- tures gender and number are part of the stem or can be derived from the source-side input, respec- tively, the prediction of case requires information about the subcategorization of the entire clause. This is due to German being a less configurational language than English, which encodes grammati- cal relations (e.g. subject-hood, object-hood, etc.) through the position of constituents. German sen- tences exhibit a freer constituent order, and thus case is an important indicator of the grammatical functions of noun phrases. Correct case predic- tion is a crucial factor for the adequacy of SMT output, cf. the example in table 1 providing an erroneously inflected output (this is taken from a baseline “simple inflection prediction” system, cf. section 5.2). The translation of the English input sentence in terms of stems is perfectly acceptable; after the inflection step, however, the translation of NP4 ongoing military actions represents a geni- tive modifier of the subject NP2, instead of a direct object NP of the verb anordnen (to order). The meaning is thus why the government of the ongo- ing military actions ordered, which has only one NP and is completely wrong. The translation in table 1 needs verb subcatego- rization information. This is demonstrated …

Content truncated to 500 words