In this work, we propose a novel approach for generating Knowledge- Graph (KG) embeddings that addresses potential limitations associated with existing methods relying on random walks. Our method focuses on capturing the faithfulness of entity embeddings to the real-world texts in which the corresponding entities appear. Our hypothesis is that the quality of embeddings can be improved for applications that require such faithfulness by applying a non-random procedure. Specifically, we introduce a customized strategy that captures real-world semantic properties of KG entities as an alternative to traditional graph walks, grounded in the observation that written language reflects how humans organize and convey knowledge. We support our approach with a theoretical background in cognitive science and linguistics, including details about the sequential perception of stimuli in humans and their interplay with attention and memory, which are fundamental elements in language production. We present a case study in the domain of music biographies and detail each step of the process, from collecting and processing raw data to performing Entity Linking and augmenting the resulting sequences of entities with additional KG-specific information in the form of semantic triples. To verify the effectiveness of our approach, we conduct machinelearning experiments based on models using the LSTM-RNN architecture. We train seven models on different features and feature combinations. We utilize cross-entropy loss as a proxy for perplexity to evaluate the quality of the models during training. The results indicate that the proposed approach offers a promising direction for enhancing the quality of KG embeddings for future applications aiming to maximize faithfulness to real-world texts.