Data mining and modelling for sign language

Abstract

Sign languages have received significantly less attention than spoken languages in the research areas of corpus analysis, machine translation, recognition, synthesis and social signal processing, amongst others. This is mainly due to signers being in a clear minority and there being a strong prior belief that sign languages are simply arbitrary gestures. To date, this manifests in the insufficiency of sign language resources available for computational modelling and analysis, with no agreed standards and relatively stagnated advancements compared to spoken language interaction research. Fortunately, the machine learning community has developed methods, such as transfer learning, for dealing with sparse resources, while data mining techniques, such as clustering can provide insights into the data. The work described here utilises such transfer learning techniques to apply neural language model to signed utterances and to compare sign language phonemes, which allows for clustering of similar signs, leading to automated annotation of sign language resources. This thesis promotes the idea that sign language research in computing should rely less on hand-annotated data thus opening up the prospect of using readily available online data (e.g. signed song videos) through the computational modelling and automated annotation techniques presented in this thesis.