Ellipsis Translation for a Medical Speech to Speech Translation System

Ellipsis Translation for a Medical Speech to Speech Translation System Jonathan Mutal1, Johanna Gerlach1, Pierrette Bouillon1, and Herv´e Spechbach2 1FTI/TIM, University of Geneva, Switzerland 2Hˆopitaux Universitaires de Gen`eve (HUG), Switzerland {Jonathan.Mutal, Johanna.Gerlach, Pierrette.Bouillon}@unige.ch herve.spechbach@hcuge.ch Abstract In diagnostic interviews, elliptical utter- ances allow doctors to question patients in a more efficient and economical way. However, literal translation of such incom- plete utterances is rarely possible without affecting communication. Previous studies have focused on automatic ellipsis detec- tion and resolution, but only few specif- ically address the problem of automatic translation of ellipsis. In this work, we evaluate four different approaches to trans- late ellipsis in medical dialogues in the context of the speech to speech translation system BabelDr. We also investigate the impact of training data, using an under- sampling method and data with elliptical utterances in context. Results show that the best model is able to translate 88% of ellip- tical utterances correctly. 1 Introduction Ellipsis is one of the least studied discursive phe- nomena in automatic translation. Like anaphora, ellipsis require context to be understood, but con- trary to anaphora, there is no indicator that there is a missing part in the sentence1. The characteris- ing feature of ellipsis is that “elements of semantic content are obtained in the absence of any corre- sponding form. The syntax thus appears to be in- complete. More specifically, the implicit seman- tic context is recovered from elements of linguistic c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1Ellipsis is “a case of anaphora, where the anaphor is a null proform (zero-anaphora)” (Ginzburg and Miller, 2018) and extralinguistic context” (Ginzburg and Miller, 2018). In NLP, different studies have focused on auto- matic ellipsis detection and resolution either with rules (patterns or grammars) (for example, the pio- neer work from Hardt, 1992) or classification tech- niques (for example, Hardt and Rambow, 2001; Bos and Spenader, 2011; Liu et al., 2016; Kenyon- Dean et al., 2016; McShane and Babkin, 2016; Rønning et al., 2018). However, only few stud- ies specifically address this problem in machine translation (MT), despite the recent interest for context modelling in neural machine translation (see for example, Bawden et al., 2018). Very re- cently, some qualitative studies showed the nega- tive impact of ellipsis on generalist neural systems (DeepL, Google Translate, etc.) from a translation point of view in the English-French pair (for ex- ample, Hamza, 2019). In this paper, we focus on automatic translation of ellipsis in medical dialogues, in the particular context of BabelDr, a speech to speech transla- tion system for the medical domain (Spechbach et al., 2019)2. Elliptical utterances are very com- mon in dialogues, since they ensure the principle of economy and provide a way to avoid duplication (Hamza et al., 2019). In the medical dialogues we are interested in, ellipsis allows doctors to ques- tion patients in a more efficient way (Where is your pain? In the back? Is the pain severe? Moder- …

Content truncated to 500 words