Minding the “gApp” in the neural machine translation of discontinuous idioms (ES>EN/ZH)

Abstract

This study examines the effectiveness of gApp, an innovative text preprocessing system developed to automatically convert discontinuous idioms into their continuous forms in order to enhance current neural machine translation (NMT) systems. Through comprehensive testing with Google Translate and DeepL, we evaluated 500 instances (250 discontinuous and 250 continuous forms) of common Spanish Verb+ Prepositional Phrase (VPP) idioms like estar hasta los cojones, estar hasta los huevos, estar hasta las narices, estar hasta el gorro, and estar hasta la coronilla, comparing gApp’s automatic conversion with the manual conversion (the gold standard). To our knowledge, this is the first study evaluating gApp’s effectiveness in translating VPP idioms using NMT in the ES>EN and ES>ZH directions. In this context, the promising outcomes obtained for this idiom category offer valuable insights into potential improvements for idiom-aware NMT systems.