Embedding representation of words in sign language

Word Embedding is currently the standard in machine learning methods for natural language processing. It is a matrix that represents the interdependence between words in a given linguistic corpus. This matrix is N x dimension, where N is the number of words in a given linguistic corpus, and the dimension is most often 100 or 300. The embedding matrix mathematically represents the semantic distance between individual words. Various methods exist for generating such a matrix for natural language, such as Word2Vec or GloVe.

In this work, we want to focus on creating an embedding matrix for Polish Sign Language (PSL). Sign language has different characteristics than the spoken language; it is the so-called spatial language, encompassing not only gestures but also facial expressions and body language. As a result, it has no official written form, though signs can be represented using glosses. With a dataset of sentences annotated with glosses, we attempted the generation of an embedding matrix that could be used in further researches on translation between Polish and PSL. For this purpose, the abovementioned Word2Vec and GloVe methods, with addition of fastText, ELMo and BERT algorithms, will be employed.