Desenvolvimento de um Componente de Sumarização de Sentenças em Português para Uso em Acessibilidade Digital

Abstract

Summarizing texts means making them smaller while keeping all the important information coherent. Some experiments described in scientific literature show that a person speaks an average of four words per second, while an interpreter is capable of signaling only two signals in the same time interval. In the context of translation/interpretation into sign language in videos, therefore, there is a need to extract from speech only the most important words to be interpreted so that the times between speech and interpretation in sign language are similar. Considering this type of problem, this work presents an analysis and solution for automatic extractive summarization of texts in Portuguese, so that it is possible to apply it, in the future, in automatic translations into Libras. For this, Deep Learning Networks and a dataset in English were used. Using the ROUGE-1 metric, the model presented an average of 66% accuracy in generating summaries of a set of tests. Tests were carried out with random users to evaluate the understanding and naturalness of the summaries in Portuguese obtained with the automatic summarizer in conjunction with an automatic translator.