Building an Extremely Low Resource Language to High Resource Language Machine Translation System from Scratch

Building an Extremely Low Resource Language to High Resource Language Machine Translation System from Scratch Flammie A Pirinen UiT Norgga ´arktalaˇs universitehta Tromsø, Norway firstname.lastname@uit.no Linda Wiechetek UiT Norgga ´arktalaˇs universitehta Tromsø, Norway firstname.lastname@uit.no Abstract Building a machine translation system for an extremely low-resource language is a prob- lem in contemporary computational linguistics. In this article, we show how to use existing morpho-syntactic analysers and a modern rule- based machine translation system to rapidly build a baseline system for a language pair where a neural model approach is not feasi- ble due to the total lack of high-quality paral- lel corpora. Our experiment produces a freely available open-source North S´ami to German machine translator, which provides us useful insights into rule-based machine translation of unrelated languages with varying levels of mor- phological complexity. As German is a lan- guage taught in Scandinavian schools this MT system would be of immediate relevance for S´ami school children learning German. In ad- dition, there is a strong Finno-Ugric tradition in the German linguistics space that has in the past produced important publications on the S´ami language, so the system is immediately useful for researchers and enthusiasts as well as language users. 1 Introduction 1.1 Motivation Machine translation is an important tool for lan- guage users. The most common contemporary method for implementing machine translation is to curate professionally translated texts and use machine learning methodology to learn the transla- tions. This presupposes the availability of perhaps several millions of professionally translated sen- tences, which is unfeasible for under-resourced marginalised languages, where very little paral- lel corpora or even monolingual corpora are avail- able. To put the low-resourcedness of North S´ami in context, the largest available monolingual cor- pus (SIKOR, 2018) is only 38 million tokens, and for the bilingual corpora at most 10,000s of aligned phrases, most of which are from Linux program GUI translations1. Given the circumstances, we do not find it reasonable to try to train a neural net- work for this task. The sensible solution is to use linguistic knowledge to build a rule-based machine translation system. What we are presenting in this article is a machine translation from North S´ami to German, a language pair that to our knowledge has not brought forth any system before, and that does not have enough resources for a neural machine translation system. Furthermore, our contribution consists in exploring a newly created module in a rule-based machine translation system, and we are looking at workflows for the rapid development of a baseline machine translator. The rule-based system took us only some 100 hours to write and is the work of one program- mer/linguist/advanced learner of German and na- tive speaker of Finnish, an expert on Apertium - and one computational linguist, native speaker of Ger- man with high proficiency in North S´ami (but not a native speaker of it). The system described here is a work-in-progress, yet it is a proof-of-concept that rapid building of a machine translation system is plausible without big …

Content truncated to 500 words