InDeep × NMT: Empowering Human Translators via Interpretable Neural Machine Translation

InDeep × NMT: Empowering Human Translators via Interpretable Neural Machine Translation Gabriele Sarti and Arianna Bisazza Center for Language and Cognition (CLCG) University of Groningen, The Netherlands {g.sarti, a.bisazza}@rug.nl Abstract The NWO-funded InDeep project aims to empower users of deep-learning models of text, speech, and music by improving their ability to interact with such models and in- terpret their behaviors. In the translation domain, we aim at developing new tools and methodologies to improve prediction attribution, error analysis, and controllable generation for neural machine translation systems. These advances will be evalu- ated through field studies involving profes- sional translators to assess gains in post- editing efficiency and enjoyability. 1 Introduction In recent years, the widespread adoption of deep learning systems in neural machine translation (NMT) led to substantial performance gains across most language pairs. Consequently, the focus of human professionals gradually shifted towards the post-editing of machine-generated content. De- spite the indisputable quality of NMT, the ques- tion of why and how these systems can effectively encode and exploit linguistic information stands unanswered. Indeed, NMT systems are intrin- sically opaque due to their multi-layered nonlin- ear architecture. This fact significantly hinders our ability to interpret their behavior (Samek et al., 2019), an essential prerequisite to their appli- cation in real-world scenarios requiring account- ability and transparency. For this reason, the in- terpretability of neural models has grown into a prolific field of research, developing multiple ap- © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. proaches aimed at analyzing models’ predictions and learned representations (Belinkov et al., 2020). While most explainable NMT studies focus on analyzing model learning and predictive behav- iors to gain theoretical insights, interpretability ap- proaches have seldom been applied from a user- centric perspective. This criticality was high- lighted by exponents of the interpretability field, among which the necessity of grounding future re- search in practical applications found broad con- sensus (Doshi-Velez and Kim, 2017). In light of this, the development of methods that are self- contained, generalizable, and scalable would en- able the identification of widespread issues char- acterizing NMT predictions such as hallucina- tions (Raunak et al., 2021), under- and over- translation, and inadequate terminology (Vamvas and Sennrich, 2021; Vamvas and Sennrich, 2022). 2 Project Description As part of the broader consortium ‘InDeep: Inter- preting Deep Learning Models for Text and Sound’ funded by the Dutch Research Council (NWO)1, we aim to build upon the latest advances in inter- pretability studies to empower end-users of NMT via the application of interpretability techniques for neural machine translation. The InDeep project will run from 2021 to 2026, involving a number of academic and industrial partners such as the uni- versities of Groningen and Amsterdam, KPN, De- loitte and Hugging Face. Central to this project is improving the subjective post-editing experience for human professionals, promoting a shift from a passive proofreading routine to an active role in the translation process by employing interac- tive and intelligible …

Content truncated to 500 words