Semantic decoding aims to identify, directly from brain activity, which concept a person is perceiving or imagining, without any motor or behavioural response. Although electroencephalography (EEG) is the most widely adopted recording modality in brain–computer interfacing, its use for semantic decoding remains an emerging line of research because of its practical advantages: non-invasiveness, low cost, portability, and high temporal resolution. Realising this potential, however, requires overcoming two limitations that affect most decoding work: poor cross-subject generalisation, since within-subject and pooled evaluation reward subject-specific idiosyncrasies, and limited interpretability, since the field is increasingly dominated by opaque deep-learning models. This work addresses these limitations on the multimodal semantic EEG dataset of Wilson et al., applying a strict Leave-One-Subject-Out (LOSO) protocol to all three classification tasks: perception versus imagination, sensory modality, and semantic concept.