I investigate strategies for semantic decoding from minimal-channel, consumer-grade EEG systems. Using only four electrodes and 50–100 word stimuli, I evaluate convolutional architectures on two semantic tasks including emotional valence and part-of-speech discrimination (specifically, noun/verb classification). To address limited data, I introduce (1) a data amplification method based on short-time FFT snapshots and (2) an embedding-constrained EEG architecture that leverages clustered word embeddings to design specialized processing branches without requiring embeddings at inference. Spectral (FFT) data improved accuracy by~8% over time-series models, while the embedding-constrained architecture reached ~93.5%, outperforming baselines and multi-head models.