The increasing availability of large corpora – a set of text corpus – creates the expectation of synthesizing, organizing and inferring an unprecedented amount of information. Automatic Knowledge Extraction (AKE), an area that is at the heart of Natural Language Processing (NLP) and Artificial Intelligence (AI), aims at using structured knowledge acquisition techniques from unstructured data such as text documents. In essence, AKE is comprised of two tasks: Named Entity Recognition and Classification (NERC) and Relation Extraction (RE). Recently, the solutions proposed for these tasks are mostly deep learning (DL) approaches. Today, models can accurately extract relation information between entities in different sentences. This work deals with the creation of DL mechanisms that allow increasing the extracted information storage capacity in order to recognize more complex patterns. This work aims to develop, implement and evaluate DL techniques for NLP applications such as NERC and RE from a raw text. Therefore, models based on Transformer structures, that optimize sequence processing, are explored together with mechanisms based on memory networks, in order to increase the reasoning capacity of these models. The training dataset is based on a set of labeled texts – annotations – indicating the presence of entity types or relationships between them. The model receives text as its input and must learn to recognize the entities and relationships contained therein. Results show the effectiveness of the proposed models when compared to those based on existing neural networks approaches.