Detecting Non-literal Translations by Fine-tuning Cross-lingual Pre-trained Language Models

Proceedings of the 28th International Conference on Computational Linguistics, pages 5944–5956 Barcelona, Spain (Online), December 8-13, 2020 5944 Detecting Non-literal Translations by Fine-tuning Cross-lingual Pre-trained Language Models Yuming Zhai BFSU Artificial Intelligence and Human Languages Lab Beijing Foreign Studies University 100089, Beijing, China zhaiyuming@bfsu.edu.cn Gabriel Illouz Universit´e Paris-Saclay CNRS, LIMSI 91400, Orsay, France illouz@limsi.fr Anne Vilnat Universit´e Paris-Saclay CNRS, LIMSI 91400, Orsay, France vilnat@limsi.fr Abstract Human-generated non-literal translations reflect the richness of human languages and are some- times indispensable to ensure adequacy and fluency. Non-literal translations are difficult to pro- duce even for human translators, especially for foreign language learners, and machine transla- tions are still on the way to simulate human ones on this aspect. In order to foster the study on appropriate and creative non-literal translations, automatically detecting them in parallel corpora is an important step, which can benefit downstream NLP tasks or help to construct materials to teach translation. This article demonstrates that generic sentence representations produced by a pre-trained cross-lingual language model could be fine-tuned to solve this task. We show that there exists a moderate positive correlation between the prediction probability of being human translation and the non-literal translations’ proportion in a sentence. The fine-tuning experiments show an accuracy of 80.16% when predicting the presence of non-literal translations in a sen- tence and an accuracy of 85.20% when distinguishing literal and non-literal translations at phrase level. We further conduct a linguistic error analysis and propose directions for future work. 1 Introduction Translation is a cross-lingual and intercultural process helping to communicate across language barri- ers. Human-generated parallel corpora have been largely exploited to develop machine translation (MT) systems, which rapidly evolve and are widely used in production nowadays (Koehn et al., 2003; Och and Ney, 2003; Wu et al., 2016; Gehring et al., 2017; Vaswani et al., 2017; Hassan et al., 2018). Multi- ple human-generated references could be used to evaluate each MT hypothesis, since the same meaning could be expressed in different ways. When we take a closer look at human translations, apart from literal translations, fixed corresponding expressions or the most commonly used renderings, various translation techniques could be employed to produce non-literal translations. The resulting variations reflect the richness of human languages. Fur- thermore, because of the existing differences between languages and cultures, it is sometimes inevitable to translate non-literally to ensure adequacy and fluency. Table 1 presents two English-French transla- tions found in the subtitles of TED Talks1, produced by volunteer translators. We mark the non-literal translations in bold and present the literal translation of the French rendering in brackets. As shown below, non-literal translations can occur at word, phrase and even sentence level. Different techniques of producing non-literal translations are systematically studied by linguists and translation scholars, such as generalization, particularization, modulation, transposition, etc. (Vinay and Darbelnet, 1958; Chuquet and Paillard, 1989; Molina and Hurtado Albir, 2002). Based on these studies, Zhai (2019) investigated the possibility of annotating and automatically recognizing different translation techniques at sub-sentential level. For natural language …

Content truncated to 500 words