German light verb construction in the course of the development of Machine Translation

The German light verb construction (lvc) is commonly used despite its relative complexity. Different writing guidelines recommend avoiding lvcs and replacing them with the base verb constructions (bvcs). However, since not every lvc has an equivalent bvc, replacement is not always possible. The present study addresses two aspects: first, how the machine translation (mt) of lvc has evolved in light of recent progress in mt and the increasing dominance of neural machine translation (nmt), and second, whether the use of bvcs improves mt output compared to lvcs. The analysis of the mt output of both scenarios, lvc and bvc, is performed for different mt approaches in terms of number and types of mt errors, style and content quality ratings, and scores from two automatic evaluation metrics (aems). For this, a mixed-methods triangulation approach that includes error annotation, human evaluation, and automatic evaluation was applied and five mt systems were examined: a rule-based system (rbmt), a statistical system (smt), two differently constructed hybrid systems (hmt), and a neural system (nmt). The study is conducted for the language pair German-to-English in the technical domain. The results show that systems that employ earlier mt approaches (rbmt, smt, hmt) benefited from replacing the lvc with the corresponding bvc as their output was improved (i.e., mt errors were reduced; quality and aems scores were increased). On the contrary, the nmt system was able to produce mt with minimal number of errors both for lvcs and bvcs and recorded the highest quality levels in both scenarios among the analyzed mt systems