O uso da análise de emoções como auxílio na tradução automática de português brasileiro para libras

Abstract

Across the world, the spread of information is predominantly directed towards oral languages, which harms a portion of the population, the deaf, whose main language is sign language. In this sense, Information and Communication Technologies have appeared to minimize this barrier, however there is still some criticism from this community regarding

the use of these technological alternatives. Therefore, this study appears with the objec- tive of helping automatic translators of content in oral language to sign language, through

the use of a strategy of emotion analysis in texts, using machine learning, filling in a pro- totype of an automatic translator that brings more acceptability by that community. To

this end, a neural network based on the MultiFit method for the classification of emotions in Brazilian Portuguese texts is built, trained from an existing database, which will be

expanded through experiments involving data augmentation. In addition, the Su ́ıte VLi- bras automatic translator has been adapted to that the platform’s 3D avatar acts with

emotions in the translation through facial expressions. Although initial, to approach to do the emotion analysis presented an accuracy of 94.29%, much higher than the work that originated the database. Thus, the results are useful because they show that the adaptation of the automatic translator offers a starting point for the adaptation of other existing automatic translators. In addition to that, the study brings promising results, which lead to believe that the use of the prototype developed here may come to minimize the usual criticism by the deaf community when using automatic translators. In addition, the results found here also collaborate to fill the gap existing in the literature regarding research focused on low-resource languages, as is the case of Libras.