Lost in Interpretation: Predicting Untranslated Terminology in Simultaneous Interpretation

Proceedings of NAACL-HLT 2019, pages 109–118 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 109 Lost in Interpretation: Predicting Untranslated Terminology in Simultaneous Interpretation Nikolai Vogler and Craig Stewart and Graham Neubig Language Technologies Institute Carnegie Mellon University {nikolaiv,cas1,gneubig}@cs.cmu.edu Abstract Simultaneous interpretation, the translation of speech from one language to another in real- time, is an inherently difficult and strenuous task. One of the greatest challenges faced by interpreters is the accurate translation of difficult terminology like proper names, num- bers, or other entities. Intelligent computer- assisted interpreting (CAI) tools that could an- alyze the spoken word and detect terms likely to be untranslated by an interpreter could re- duce translation error and improve interpreter performance. In this paper, we propose a task of predicting which terminology simultaneous interpreters will leave untranslated, and exam- ine methods that perform this task using super- vised sequence taggers. We describe a number of task-specific features explicitly designed to indicate when an interpreter may struggle with translating a word. Experimental results on a newly-annotated version of the NAIST Simul- taneous Translation Corpus (Shimizu et al., 2014) indicate the promise of our proposed method.1 1 Introduction Simultaneous interpretation (SI) is the act of trans- lating speech in real-time with minimal delay, and is crucial in facilitating international commerce, government meetings, or judicial settings involv- ing non-native language speakers (Bendazzoli and Sandrelli, 2005; Hewitt et al., 1998). However, SI is a cognitively demanding task that requires both active listening to the speaker and careful monitoring of the interpreter’s own output. Even accomplished interpreters with years of training can struggle with unfamiliar concepts, fast-paced 1Code is available at https://github.com/nvog/ lost-in-interpretation. Term annotations for the NAIST Simultaneous Translation Corpus will be provided upon request after confirmation that you have access to the corpus, available at https://ahcweb01.naist.jp/ resource/stc/. speakers, or memory constraints (Lambert and Moser-Mercer, 1994; Liu et al., 2004). Human short-term memory is particularly at odds with the simultaneous interpreter as he or she must con- sistently recall and translate specific terminology uttered by the speaker (Lederer, 1978; Dar`o and Fabbro, 1994). Despite psychological findings that rare words have long access times (Balota and Chumbley, 1985; Jescheniak and Levelt, 1994; Griffin and Bock, 1998), listeners expect inter- preters to quickly understand the source words and generate accurate translations. Therefore, pro- fessional simultaneous interpreters often work in pairs (Mill´an and Bartrina, 2012); while one inter- preter performs, the other notes certain challeng- ing items, such as dates, lists, names, or numbers (Jones, 2002). Computers are ideally suited to the task of re- calling items given their ability to store large amounts of information, which can be accessed al- most instantaneously. As a result, there has been recent interest in developing computer-assisted in- terpretation (CAI; Plancqueel and Werner; Fantin- uoli (2016, 2017b)) tools that have the ability to display glossary terms mentioned by a speaker, such as names, numbers, and entities, to an inter- preter in a real-time setting. Such systems have the potential to …

Content truncated to 500 words