Handle with Care: A Case Study in Comparable Corpora Exploitation for Neural Machine Translation

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3799–3807 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3799 Handle with Care: A Case Study in Comparable Corpora Exploitation for Neural Machine Translation Thierry Etchegoyhen and Harritxu Gete Ugarte Department of Speech and Natural Language Technologies, Vicomtech Mikeletegi Pasalekua, 57, Donostia, Gipuzkoa, Spain {tetchegoyhen, hgete}@vicomtech.org Abstract We present the results of a case study in the exploitation of comparable corpora for Neural Machine Translation. A large comparable corpus for Basque-Spanish was prepared, on the basis of independently-produced news by the Basque public broadcaster eitb, and we discuss the impact of various techniques to exploit the original data in order to determine optimal variants of the corpus. In particular, we show that filtering in terms of alignment thresholds and length-difference outliers has a significant impact on translation quality. The impact of tags identifying comparable data in the training datasets is also evaluated, with results indicating that this technique might be useful to help the models discriminate noisy information, in the form of informational imbalance between aligned sentences. The final corpus was prepared according to the experimental results and is made available to the scientific community for research purposes. Keywords: Comparable Corpora, Basque, Spanish, Neural Machine Translation 1. Introduction Comparable corpora are an important source of potential parallel data, suitable to train data-driven machine transla- tion systems (Munteanu and Marcu, 2005; Sharoff et al., 2016) or to create bilingual dictionaries (Rapp, 1995). The extraction of parallel sentences from this type of corpora faces a number of challenges, since potential parallel data are immersed in vast amounts of unrelated content which need to be efficiently mined. Over the years, several tech- niques have been designed to address comparable document alignment (Sharoff et al., 2015; Azpeitia and Etchegoyhen, 2019) and comparable sentence alignment (Munteanu and Marcu, 2002; Fung and Cheung, 2004; Smith et al., 2010; Etchegoyhen and Azpeitia, 2016b; Artetxe and Schwenk, 2019), leading to new parallel datasets that can support machine translation, in particular for under-resourced lan- guages (Etchegoyhen et al., 2016; Schwenk et al., 2019). Neural Machine Translation (nmt) is currently the domi- nant paradigm in research and development in the field of machine translation, having led to significant advances in recent years for most language pairs (Bahdanau et al., 2015; Bojar et al., 2016; Bojar et al., 2017; Vaswani et al., 2017). As a data-driven approach where model parameters are es- timated and optimised on large volumes of parallel data, nmt is particularly sensitive to the presence of noise in the training data (Belinkov and Bisk, 2018; Sperber et al., 2017; Cheng et al., 2018; Khayrallah and Koehn, 2018). Due to the nature of the task, parallel data extracted from compara- ble corpora are likely to introduce unwarranted noise in the training process and an evaluation of this potential issue is worth examining. In this article, we describe the preparation of a large comparable corpus for Basque-Spanish, composed of independently-produced news by …

Content truncated to 500 words