The advance of technology, digital systems and communications has led to a continuous generation of massive amounts of data, mostly unstructured text, becoming infeasible to process and analyze them manually. The need for tools to automatically extract information from semistructured or uninstructed data gave rise to the field of Information Extraction (IE). IE has many sub-tasks, and one of them, among many others, is the Relation Extraction (RE). RE is the task of automatically extracting semantic relationships between named entities that were previously recognized in texts by means of the Named Entity Recognition (NER) task. The RE task is mainly used as a base for Question Answering (QA), Textual Entailment (TE) systems and Semantic Search. The solution proposed for the RE task is based on supervised learning and is composed by three steps. In the first step, embeddings are generated for each entity annotated in the data-set. In the second step, an auto-encoder is used to generate relation embeddings from each pair of entities that have a semantic relationship in the data-set. Finally, the embeddings generated in the first and second steps are then combined as features to build a classification model based on gradient boosting trees technique. The proposed solution was evaluated on two data-sets from different languages, Portuguese and English, and then compared to one of the benchmark solutions available in the literature, the OpenNRE. Compared to OpenNRE, the results showed that the proposal achieved an 8.1% higher F1 score for the Portuguese language and 16.7% higher F1 score for the English language.