Active Learning for Massively Parallel Translation of Constrained Text into Low Resource Languages

We translate a closed text that is known in advance and available in many\nlanguages into a new and severely low resource language. Most human translation\nefforts adopt a portion-based approach to translate consecutive pages/chapters\nin order, which may not suit machine translation. We compare the portion-based\napproach that optimizes coherence of the text locally with the random sampling\napproach that increases coverage of the text globally. Our results show that\nthe random sampling approach performs better. When training on a seed corpus of\n~1,000 lines from the Bible and testing on the rest of the Bible (~30,000\nlines), random sampling gives a performance gain of +11.0 BLEU using English as\na simulated low resource language, and +4.9 BLEU using Eastern Pokomchi, a\nMayan language. Furthermore, we compare three ways of updating machine\ntranslation models with increasing amount of human post-edited data through\niterations. We find that adding newly post-edited data to training after\nvocabulary update without self-supervision performs the best. We propose an\nalgorithm for human and machine to work together seamlessly to translate a\nclosed text into a severely low resource language.\n