Pipeline Signed Japanese Translation Using PBSMT and Transformer in a Low-Resource Setting

We propose a novel pipeline method for translating signed Japanese sentences into written Japanese. Sign languages often suppress functional words such as particles, and most are not morphologically inflected they in spoken languages. Our explicitly compares contrasts the two divides translation process tasks: first, it translates glosses lemmatized or phrases, followed by complementing particles conjugating predicates verbs, auxiliary adjectives. is especially effective when size of parallel corpus very limited costly to obtain, but there plenty monolingual corpora target. Specifically, our first uses phrase-based statistical machine (PBSMT) map sign corresponding then employs transformer-based neural (NMT) model trained with refine output translation. Experimental results show that outperforms direct PBSMT competitive NMT models data augmentation, including back-translation transfer learning low-resource setting on order 104 words.