Results of Machine Translation with FLExTrans in Production-Mode

The linguistically-based machine translation tool FLExTrans has been around for a few years. This paper shows the results of successfully using FLExTrans in “production-mode” for doing ongoing translation in a real language project. We have seen significant speed-up of translation over human translation.
FLExTrans is conceptually similar to the legacy CARLA system. It is based on lexical mappings between source and target languages and transfer rules that transform the source words and phrases into target ones. It leverages lexicons residing in SIL’s Fieldworks Language Explorer (FLEx) tool. This paper outlines the first attempt to put FLExTrans into production-mode in a project with a source lexicon and target lexicon and with experts to evaluate and polish the resulting translations. The data presented come after translating the book of Joshua using FLExTrans. I present statistics on how long it took to do different parts of the machine translation task for each chapter. This includes preparing the text, parsing the source text, drafting the target text and polishing the target text. I present lessons learned in the process and suggestions for implementation of FLExTrans in other Bible Translation projects.