We present a workflow that combines the data-driven aspects of traditional machine learning algorithms with the formal representation of domain knowledge. It is composed of a toolset that may have a significant impact on the seismic interpretation process, speeding up data annotation for applications with a massive demand for labeled data such as deep learning. This workflow has been adopted by an oil company to leverage the traditional machine learning tools such as image retrieval, in combination with a geological knowledge base, to help geoscientists quickly expand their knowledge over a region or transfer it to a new area. We demonstrate the applicability of our workflow in a real-case scenario using the Netherlands F3 dataset by expanding four annotations performed by an expert to more than one thousand of automatically labeled images. We validate the proposed workflow by running a classification task.