Smart Computer-Aided Translation Environment (SCATE): Highlights

Smart Computer-Aided Translation Environment (SCATE): Highlights Vincent Vandeghinste Tom Vanallemeersch Bram Bulté Liesbeth Augustinus Frank Van Eynde Joris Pelemans Lyan Verwimp Patrick Wambacq Geert Heyman Marie-Francine Moens Iulianna van der Lek-Ciudin Frieda Steurs KU Leuven first.lastname@kuleuven.be Ayla Rigouts Terryn Els Lefever Arda Tezcan Lieve Macken Ghent University first.lastname@ugent.be Sven Coppers Jens Brulmans Jan Van den Bergh Kris Luyten Karin Coninx UHasselt – tUL – EDM first.lastname@uhasselt.be Abstract We present the highlights of the now fin- ished 4-year SCATE project. It was com- pleted in February 2018 and funded by the Flemish Government IWT-SBO, project No. 130041.1 We present key results of SCATE (Smart Com- puter-Aided Translation Environment). The pro- ject investigated algorithms, user interfaces and methods that can contribute to the development of more efficient tools for translation work. Improved fuzzy matching: Levenshtein dis- tance is not the best predictor for post-editing ef- fort. Linguistic metrics and different metrics (such as TER) combined show better results. Integration of Translation Memory (TM) and Machine Translation (MT): Combining TM matches, fuzzy match repair and SMT shows improvements over a baseline SMT. Informed Quality Estimation: Accuracy and fluency error detection systems form the basis of the sentence-level Quality Estimation system, which results in better correlations with temporal post-editing effort compared to the Quest++ base- line. Detected errors can additionally be high- lighted in the MT output. Identifying bilingual terms in comparable texts: We found improvements when combining word embeddings with character-based models  © 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attrib- ution, CC-BY-ND. using a neural classifier trained on a seed lexicon. This includes short multi-word term phrases. Post-Editing via Automated Speech Recog- nition (ASR): ASR for post-editing can benefit from additional information sources, such as the source language, the MT translation model and the activation of domain-specific terminology, for which we boosted ASR language model probabil- ities. The ASR language model is also enriched with character-level information, making it possi- ble to model out-of-vocabulary words, which are very common in new domains. Intelligible Translator Interfaces: We itera- tively developed a functional prototype that inte- grates several of the aforementioned translation aids. In contrast with other approaches, our sys- tem applies the design concept of intelligibility to support translators’ decision-making process when they interact with their translation environ- ment. The evaluation showed that the prototype allows translators to better evaluate translation suggestions from MT, TM and term base but it had no major impact on their performance in terms of speed and quality. Furthermore, a small- scale lab experiment revealed no significant dif- ference in efficiency between translating with the prototype and with a commercial tool, which shows less suggestions by default. Integration: We created an interactive demo so that translators can experience and evaluate our research results: http://scate.edm.uhasselt.be/. 1 http://www.ccl.kuleuven.be/scate P´erez-Ortiz, S´anchez-Mart´ınez, Espl`a-Gomis, Popovi´c, Rico, Martins, Van den Bogaert, Forcada (eds.) Proceedings of the 21st Annual Conference of the European Association for Machine Translation, p. 367 Alacant, Spain, May 2018.