Assessing Three Representation Methods for Sign Language Machine Translation and Evaluation
Sara Morrissey · 2011
Having no standard written format, sign languages must be transcribed in some way in order to be processable for machine translation (MT). Previous research into MT for sign languages (SLs) has shown little consistency or agreement on the appropriate transcription methodology for the SLs. In this paper, we take a corpus of 200 SL utterances and explore the effects of three different representations on the MT process using a randomly and a specially selected testset. We use the DCU MA-TREX MT system and show that using an XML-based markup achieves the best results over other formats in terms of BLEU scores. We discuss the meaning of these results in the context of evaluating a representation of a language as opposed to the final form. 1