EEG Decoding Achieves 78% Accuracy on Semantic Categories from Picture Naming

A new pre-print from researchers Wei Hu and Binbin Xu demonstrates that semantic-category information can be reliably decoded from high-density EEG during overt picture naming. The study, posted on arXiv on June 12, 2026, involved 16 native French-speaking participants naming line drawings while their EEG was recorded. The researchers embedded picture labels using a multilingual text-embedding model, grouping them into nine interpretable semantic categories to form a data-driven target space. EEG activity was processed channel-wise using a pre-trained single-channel EEG encoder across three temporal windows: an early post-stimulus window, a later naming-related window, and their combination. Nine-class decoding consistently exceeded chance across all representations, with balanced accuracy rising from 0.562 in the early window to 0.610 in the naming window, and reaching 0.781 when both windows were combined. The maximum Macro-F1 score was 0.784, with class-level F1 scores showing consistent gains across categories. Sensor-level decoding maps further revealed spatially distributed category information.

The findings suggest that semantic-category structure is clearly reflected in EEG activity and that early perceptual and later naming-related windows provide complementary information. This supports the use of modern neural decoding methods (e.g., EEG encoders combined with embedding models) as tools for investigating lexical-semantic processing in spoken language production. The work opens new avenues for non-invasive brain-computer interfaces and cognitive neuroscience research, demonstrating that covert semantic content can be recovered from scalp EEG with practical accuracy.