Enforcing Consistent Translation of German Compound Coreferences

Coreferences to a German compound (e.g. Nordwand) can be made using its last constituent (e.g. Wand). Intuitively, both coreferences and the last constituent of the compound should share the same translation. However, since Statistical Machine Translation (SMT) systems translate at sentence level, they both may be translated inconsistently across the document. Several studies focus on document level consistency, but mostly in general terms. This paper presents a method to enforce consistency in this particular case. Using two in-domain phrase-based SMT systems, we analyse the effects of compound coreference translation consistency on translation quality and readability of documents. Experimental results show that our method improves correctness and consistency of those coreferences as well as document readability.