Rule-based Machine translation from English to Finnish

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 323–329 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics Rule-based Machine Translation from English to Finnish Arvi Hurskainen University of Helsinki arvi.hurskainen@helsinki.fi Jörg Tiedemann University of Helsinki jörg.tiedemann@helsinki.fi Abstract The paper describes a rule-based machine translation system adapted to English to Finnish translation. Although the translation system participates in the shared task of news transla- tion in WMT 2017, the paper describes the strengths and weaknesses of the approach in general. 1 Credits We are grateful to Pasi Tapanainen from Connexor OY for allowing us to use the en-fdg analyser of English as well as the Constraint Grammar (CG-3)1 environment for a number of translation phases. 2 Introduction The translation system described here is in stark contrast to the majority of systems partici- pating in this conference. There are a number of reasons why we are interested in developing rule-based translation systems. One is the obser- vation that, if we use statistical or neural transla- tion systems, we will exclude 99.8 percent of languages out of development. Digitalization is supposed to break barriers between language groups, but in fact it currently increases them. The current hype on neural methods still acceler- ates the break between the small group of domi- nant languages and the less-resourced ones. If we want to avoid the break, we do not see any other way out than to put efforts in developing such systems that are affordable for less-resourced languages. At the same time, efforts for finding ways to overcome the problems of statistical sys- tems are needed (Tiedemann et al 2016). The approach described here deals with the English to Finnish translation system. However, the basic components of the system were devel- oped with the language pair Swahili and English, for which Hurskainen developed rule-based translation systems to both directions (Hurskai- nen 1992, 1996, 2004, 2006, 2007, 2012). 1 CG-3 is also termed as FDG-3, because within this envi- ronment it is possible to write also functional dependency rules. A number of approaches for constructing rule- based systems have been studied. These include OpenLogos (Scott and Barreiro 2009; Barreiro et al 2011), Apertium (Forcada 2006), Grammatical Framework (Ranta 2011), and Nooj (Silbertztein 2015). Common to these approaches is the use of grammatical knowledge and lexicon of lan- guages in translation. Although the approach that we have used has much in common with those, we did not implement any of them directly. The main reason is that we find it useful to have full control of all phases of the translation process, so that corrections can be made instantly at the cor- rect point of the process. For the same reason we did not adapt such resources as Omorfi (Pirinen 2015). Instead we developed our own system for generating Finnish word forms. The system described here deliberately avoids any statistical elements in translation process. The basic assumption is that running text can al- ways be decomposed into structured units, …

Content truncated to 500 words