Bridging the gap between sign language machine translation and sign language animation using sequence classification

To date, the non-manual components of signed utterances have rarely been considered in automatic sign language translation. However, these components are capable of carrying important linguistic information. This paper presents work that bridges the gap between the output of a sign language translation system and the input of a sign language animation system by incorporating non-manual information into the final output of the translation system. More precisely, the generation of non-manual information is scheduled after the machine translation step and treated as a sequence classification task. While sequence classification has been used to solve automatic spoken language processing tasks, we believe this to be the first work to apply it to the generation of non-manual information in sign languages. All of our experimental approaches outperformed lower baseline approaches, consisting of unigram or bigram models of non-manual features.