Sign Language Interpreting - Assessment of the Efficiency of the Translation Model

The article assessed the effectiveness of a model translating text written in Polish into Polish Sign Language using animation identifiers (animation identifiers) based on verbal data. Collaboration with sign language experts aims to identify and address performance issues, emphasizing irregularities in generated sentences. The study focuses on offering recommendations for improving translation quality by analyzing problems in the training data and incorporating expert insights. Initially considering dissimilarity algorithms proved unsuitable due to the model’s non-autoregressive architecture. Instead, a graphical user interface web application facilitated expert input, fostering an open dialogue on model effectiveness. Results revealed incorrect model operation, illogical outputs, and issues with translations beyond the training set. Low data volume and quality, along with nonrepresentative examples, contributed to subpar performance. This research offers valuable insights for future Polish Sign Language translation model enhancements within a concise framework.