Abstract
The majority of current researches in Machine Translation field are focalized essentially on spoken languages. The aim is to find a most likely translation for a given source sentence based on statistical learning techniques which are applied to very big parallel corpora. In this work, we focused on gesture languages especially on Sign Language in order to present a new methodological foundation for Sign Language Machine Translation. Our approach is based on Kernel Regression combined to Kernel Density Estimation method applied to Sign Language n-grams. The translation process is modelled as an n-gram to n-gram mapping with the consideration of the n-gram positions in the source and the target phrases. For doing so, we propose a new feature mapping process (Weighted Sub n-gram Feature Mapping) which is a modified version of the String Subsequence Kernel SSK feature mapping. The Weighted Sub n-gram aims to generate feature vectors mapping of both source and target n-gram. Afterwards, to learn the function that map source n-grams to target n-grams, we used and compared four learning techniques (Gaussian Process Regressor, K-Nearest Neighbors Regressor, Support Vector Regressor with Gaussian Kernel and Kernel Ridge Regression) for the purpose to choose the efficient one which minimizes the SSE (Sum of Squared Error). Even so, to find solution to the pre-image problem, we rely on the De-Bruijn Multi Graph search applied on n-grams target. For the purpose to obtain the best translation, we relied on the search of the most frequently observed bilingual n-gram alignment in term of the maximization of the translation probability. For unknown n-grams, we used kernel ridge regression for the purpose to predict the probability through learning the Density Estimation function of the bilingual n-grams alignments. We obtained encouraging experimental results on a small-scale reduced-domain corpus.