A study on the use of sequence-to-sequence neural networks for automatic translation of brazilian portuguese to LIBRAS

Abstract

The World Health Organization estimates that approximately 466 million people have some level of hearing loss. This significant portion of the world population faces several challenges in accessing information. The main problem is that the languages in which the deaf community can perceive and produce in a natural way are sign languages (SL). An alternative to dealing with this would be the translation of the content from an oral language to SL. However, when it comes to accessing online content, it is necessary to consider translating SL not only for audio or video content, but also for more complex text on websites. This is already a difficult task by itself for the volume involved, and it also addresses some additional challenges, related to the high cost of human interpreter service and the great dynamism of Internet content. In this context, one of the most promising approaches to such scenarios is the use of machine translation applications from oral to sign language. This work evaluates the use of neural network models usually used in natural language processing for the production of LIBRAS glosses from texts in Portuguese. Using a 2k factorial experiment design, we evaluated the impact of several aspects such as database size, types of models and training parameters in the quality of automatic translation obtained. The results of the experiments were very promising and point to an initial superiority of the LightConv model in most of the evaluated scenarios.