Morpho-Syntax Based Statistical Methods for Automatic Sign Language Translation

Morpho-Syntax Based Statistical Methods for Automatic Sign Language Translation Daniel Stein, Jan Bungeroth and Hermann Ney Lehrstuhl f¨ur Informatik VI, Computer Science Department, RWTH Aachen University D-52056 Aachen, Germany {stein,bungeroth,ney}@i6.informatik.rwth-aachen.de Abstract We present a novel approach for the automatic translation of written text into sign language. A new corpus focussing on the weather report domain for the language pair German and German Sign Language is introduced. We apply phrase-based statistical machine translation, enhanced by pre- and post-processing steps based on the morpho-syntactical analysis of German. Detailed results are given based on automatic and manual evaluation. 1 Introduction The aim of this work is to employ an auto- matic translation system from written Ger- man to German Sign Language (DGS1), the primary means of communication for the deaf people in Germany. It may seem surprising at first to propose a translation of written text if the target group typically has no visual impairment. However, (Traxler, 2000) shows that the ma- jority of the deaf community possesses only poor to moderate reading skills. The lack of auditory feedback and the still common practice of oral teaching are two responsable factors. In this paper, we present our translation system from German sentences into DGS, a language independent from German. The results are visualized using an avatar devel- oped for the presentation of sign languages (Elliott, Glauert, Kennaway, & Marshall, 2000). We investigate in how far morpho- syntactic pre- and post-processing can en- hance the translation results. We also inves- tigate the specific demands for sign language translation. Finally, we present detailed re- sults based on both automatic and manual evaluation of the translation output. 1Deutsche Geb¨ardensprache 1.1 State-of-the-Art Several researchers deal with the challenges of automatic sign language translation. To the best of our knowledge, statistical meth- ods have not been used in translation from written text to sign language yet. Moreover, the other approaches did not present quan- titative results. Thus, performance compar- ison is not possible. (Morrissey & Way, 2005) investigate corpus-based methods for example-based sign language translation from English to the sign language of the Netherlands. With the small corpus and no available lexicon, the system is robust for sentences already encountered in the training set, but has problems with unseen combinations of cor- pus chunks as well as corpus parts that it is unable to align. (Huenerfauth, 2004) explores the chal- lenges of machine translation techniques from written text to sign languages. He pro- posed methods concerning the theoretical is- sues arising during translation, for example a notation for signs which use the 3D space around the signer to form complex expres- sions. (S´af´ar & Marshall, 2001) propose a de- compensation of the translation process into two steps: first they translate from written text into a semantic representation of the signs. Afterwards a second translation into a graphically oriented representation is done. Both steps use rule-based techniques for a specific domain in British Sign Language. However, no quantitative results were pub- lished. (Bauer, Nießen, & Hienz, 1999) propose the …

Content truncated to 500 words