A Dictionary Lookup Strategy for Translating of Discontinuous Phrases

A Dictionary Lookup Strategy for Translating Discontinuous Phrases Michael Carl and Ecaterina Rascu Institut f¨ur Angewandte Informationsforschung, Martin-Luther-Str. 14, D-66111 Saarbr¨ucken, Germany {carl | kati}@iai.uni-sb.de Abstract Translation of discontinuous phrases is a major challenge in Machine Trans- lation. Within METIS-II we developed a dictionary lookup strategy by mapping the items of a dictionary entry on non-adjacent words in an input text. Mapping is controlled through so-called contextual rejection, i.e. in- appropriate mappings are discarded if they fail to satisfy a predefined set of constraints. We present various dictionary preprocessing steps to transform the entries into a suitable and more effective format for lookup. Then we describe the dictionary matching of discontinuous phrases. We illustrate this process on German verbs with detachable prefixes and support verb constructions. 1 Introduction This paper describes a dictionary lookup strategy for the German to English compo- nent of the METIS-II MT system. METIS- II1 (Dologlou et al., 2003) is a research prototype of a hybrid statistical machine translation system, financed under the EU STREP programme. Hand-crafted bilin- gual dictionaries are used to transfer lex- ical tokens (words, multi-word units and phrases) into the target language. A large target language corpus is consulted to ad- just the transferred token according to the target language syntax and/or to rank the generated translation candidates. Within METIS-II, a number of modules have been proposed and developed for re-ordering and ranking the translation candidates (Badia, Boleda, Melero, & Oliver, 2005; Carl, Rascu, & Schmidt, 2005; Dirix, Schuurman, & Vandeghinste, 2005; Markantonatou, Sofi- anopoulos, Spilioti, & Tambouratzis, 2005). The general layout of the architecture is shown in figure 1. Translation of discontinuous phrases is a major challenge when translating from Ger- man into another language. In German 1http://www.ilsp.gr/metis2/ there are not only detachable (verbal) pre- fixes that produce ‘non-monotonic’ transla- tions (Turcato & Popowich, 2003) but also a large number of light and semantically weak verbs that participate in support verb con- structions (SVCs). In such constructions, (e.g. in Gefahr bringen = in danger bring = endanger) the verb is semantically empty while the noun carries the information. Homonymy and changing word order may further complicate the translation process. Detached prefixes and prepositions have the same form — but can be distinguished on the basis of their different syntactic position within the sentence. For instance the verb bringen can participate in a support verb construction together with the noun Gefahr or it can occur as a ‘full’ verb on its own with a different meaning (i.e. bring) even if the listed noun (i.e. Gefahr) co-occurs in the same sentence. In addition, German has productive in- flection, derivation and composition pat- terns that make the enumeration of all word forms or even of all different lemmas in a dictionary impossible. There are thus at least two problems when matching a German sentence on a dictio- nary: i) find the ‘right’ translation entry Figure 1: General System Flow and Resources Used in METIS-II even if the words are differently inflected and and have …

Content truncated to 500 words