Relation extraction from unstructured text is essential for building knowledge bases or knowledge graphs, crucial in NLP tasks like question answering, information extraction, and intelligent applications (e.g., expert systems, search engines, chatbots). This paper introduces RExAS, a novel architecture leveraging adaptive self-attention for relation extraction. Our adaptive self-attention module learns a weighted adjacency representation from raw text, inputted into a Graph Convolution Network, bypassing dependency-parsed text and improving performance. Extensive experiments on NYT, Wikidata, GIDS, and SemEval datasets show RExAS outperforms or matches state-of-the-art methods across various evaluation setups.