Joint Translation and Unit Conversion for End-to-end Localization

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 265–271 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 265 Joint translation and unit conversion for end-to-end localization Georgiana Dinu Prashant Mathur Marcello Federico Stanislas Lauly Yaser Al-Onaizan Amazon AWS AI {gddinu,pramathu,marcfede,laulysl,onaizan}@amazon.com Abstract A variety of natural language tasks require pro- cessing of textual data which contains a mix of natural language and formal languages such as mathematical expressions. In this paper, we take unit conversions as an example and propose a data augmentation technique which lead to models learning both translation and conversion tasks as well as how to adequately switch between them for end-to-end localiza- tion. 1 Introduction Neural networks trained on large amounts of data have been shown to achieve state-of-the art solu- tions on most NLP tasks such as textual entail- ment, question answering, translation, etc. In par- ticular, these solutions show that one can success- fully model the ambiguity of language by making very few assumptions about its structure and by avoiding any formalization of language. However, unambiguous, formal languages such as numbers, mathematical expressions or even programming languages (e.g. markup) are abundant in text and require the ability to model the symbolic, “proce- dural” behaviour governing them. (Ravichander et al., 2019; Dua et al., 2019). An example of an application where such ex- amples are frequent is the extension of machine translation to localization. Localization is the task of combining translation with “culture adapta- tion”, which involves, for instance, adapting dates (12/21/2004 to 21.12.2004), calendar conversions (March 30, 2019 to Rajab 23, 1441 in Hijri Cal- endar) or conversions of currencies or of units of measure (10 kgs to 22 pounds). Current approaches in machine translation han- dle the processing of such sub-languages in one of two ways: The sub-language does not receive any special treatment but it may be learned jointly with the main task if it is represented enough in the data. Alternatively, the sub-language is decoupled from the natural text through pre/post processing techniques: e.g. a miles expression is converted into kilometers in a separate step after translation. Arguably the first approach can successfully deal with some of these phenomena: e.g. a neural net- work may learn to invoke a simple conversion rule for dates, if enough examples are seen training. However, at the other end of the spectrum, cor- rectly converting distance units, which itself is a simple algorithm, requires knowledge of numbers, basic arithmetic and the specific conversion func- tion to apply. It is unrealistic to assume a model could learn such conversions from limited amounts of parallel running text alone. Furthermore, this is an unrealistic task even for distributional, unsuper- vised pre-training (Turney and Pantel, 2010; Baroni and Lenci, 2010; Peters et al., 2018), despite the success of such methods in capturing other non- linguistic phenomena such as world knowledge or cultural biases (Bolukbasi et al., 2016; Vanmassen- hove et al., 2018).1 While the second approach is currently the pre- ferred one in translation technology, …

Content truncated to 500 words