Annotation of LSF subtitled videos without a pre-existing dictionary
Résumé
This paper proposes a method for the automatic annotation of lexical units in LSF videos, using a subtitled corpus without annotation. This method, based on machine learning and involving linguists for added precision and reliability, comprises several stages. The first consists of building a bilingual lexicon (including potential variants of a given lexical unit) in a weakly supervised manner. The resulting lexicon is then refined and cleaned by LSF experts. This data serves next to train a supervised classifier for automatic annotation of lexical units on the Mediapi-RGB corpus. Our Pytorch implementation is publicly available.
