A Neural Machine Translation System for the Low-Resource Sixian Hakka Language

A neural machine translation system using CNN with the Attention mechanism is constructed to translate Chinese sentences into the scarce Sixian accent of the Hakka language. We analyzed the dif-ferences in idioms and phrases between the collected corpora and defined a common word usage lexicon for different accents of Sixian Hakka. The proposed algorithm adopts forced segmentation of Hakka idioms and phrases, along with common Chinese word substi-tution, to tackle the unseen word problem frequently encountered during real-world translation. The final translation results are more comprehensible following this procedure. The proposed method shows promising results for machine translation using a low-resource corpus and could be useful for subsequent language processing tasks such as language learning and speech generation.