Prompsit’s Submission to the IWSLT 2018 Low Resource Machine Translation Task

Prompsit’s Submission to the IWSLT 2018 Low Resource Machine Translation Task V´ıctor M. S´anchez-Cartagena Prompsit Language Engineering Av. Universitat s/n. Edifici Quorum III E-03202 Elx, Spain vmsanchez@prompsit.com Abstract This paper presents Prompsit Language Engineering’s sub- mission to the IWSLT 2018 Low Resource Machine Trans- lation task. Our submission is based on cross-lingual learn- ing: a multilingual neural machine translation system was created with the sole purpose of improving translation qual- ity on the Basque-to-English language pair. The multilin- gual system was trained on a combination of in-domain data, pseudo in-domain data obtained via cross-entropy data selec- tion and backtranslated data. We morphologically segmented Basque text with a novel approach that only requires a dic- tionary such as those used by spell checkers and proved that this segmentation approach outperforms the widespread byte pair encoding strategy for this task. 1. Introduction This paper presents Prompsit Language Engineering’s sub- mission to the IWSLT 2018 Low Resource Machine Trans- lation task. The objective of this task is building an MT sys- tem for translating TED talks from Basque to English from a very limited amount of in-domain Basque–English parallel data. We relied on cross-lingual learning via a multilingual approach [1] to neural machine translation (NMT), extraction and cleaning of pseudo in-domain parallel text from out-of- domain data, and backtranslation of Spanish text into Basque for building our submission. Moreover, we applied morphological segmentation to the Basque text. We took advantage of an existing spell check- ing dictionary and its inflection paradigms and used an au- tomatic morphology inference model to decide between am- biguous segmentations. We proved that this method, that re- quires shallower linguistic information1 than other segmen- tation approaches based on full morphological analysis and disambiguation [2, 3], outperforms the widespread byte pair encoding (BPE) segmentation strategy [4] in terms of trans- lation quality for Basque-to-English NMT. 1Neither part of speech/morphological information in the dictionary nor a part of speech tagger/parser are needed. In principle, this approach could be applied to any language for which a Hunspell-based (http: //hunspell.github.io/) spell checker exists. Table 1: Size of in-domain data. Processed segments are those that remain after removing talks included in the devel- opment and test sets. Language pair # raw segments # processed segments eu-en 5 687 5 687 eu-es 6 742 5 610 eu-fr 7 021 5 878 es-en 280 947 279 737 fr-en 290 961 289 722 The remainder of the paper explains the steps followed to build the submitted NMT system. Next section explains how the in-domain and out-of-domain parallel corpora were processed and filtered, while Section 3 focuses on describing and assessing the impact of the morphological segmentation approach followed. Section 4 describes the NMT architec- ture and training process. Section 5 depicts the process fol- lowed to obtain the data set used to train our submission. Finally, the most relevant related approaches are reviewed in Section 6 and the paper ends with some concluding remarks. 2. Data acquisition and cleaning Our submission was trained on a combination …

Content truncated to 500 words