In this work, we investigate whether unsupervised learning can be useful for extracting well-separated latent semantic representations from EEG data. To train and evaluate the model, we used the University of Bath's publicly available dataset of semantic concepts for imagination and perception tasks, which contains data acquired from 12 subjects under various stimuli (auditory, pictorial, and orthographic) in two different cognitive conditions (imagination and perception) for three different semantic concepts (guitar, flower, penguin). EEG data were preprocessed, segmented into epochs and then convolutional autoencoders, integrated with temporal sequence modeling blocks (LSTMs and Transformers), were used to extract latent representations. Clustering metrics (Adjusted Rand Index, Normalized Mutual Information, Silhouette Score) and visualizations (t-distributed Stochastic Neighbor Embedding and Uniform Manifold Approximation and Projection) reveal that these latent representations are subject-specific and, for several subjects, can be distinguished based on cognitive state.