The question-answering task is a classical problem of natural language processing and is largely applied to a plethora of possible situations. There are different domains of application; one of these is providing an answer considering a text or a document, expecting to retrieve the user with the required information. For doing that, the state-of-the art approach is to represent each sentence of the document with a contextual embedding and select the closer sentence in terms of space (and consequential meaning) with respect to the question of the user. Indeed, the position of the sentences in the space can deeply influence the model and affect the correctness of the answer, and a crowded space will cause problems for the selection of the best answer. On the other side, a well-spaced space will provide better results. Starting from this idea, we will use weighted feature selection with the objective of obtaining more space among the points representing the document and consequently improving the performance.