Traces of Meaning Itself: Encoding Distributional Word Vectors in Brain Activity

Area of Science:

• Cognitive Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• The storage of semantic information in the brain is debated, with vectorial representations proposed.
• Distributional word vector models show success in explaining context-based language processing in EEG data.

Purpose of the Study:

• To investigate if distributional vector representations of word meaning can model brain activity for isolated words.
• To assess the viability of prediction-based vectorial representations for human semantic knowledge.

Main Methods:

• Utilized electroencephalography (EEG) data (event-related brain potentials) from English and German experiments.
• Encoded and decoded word vectors from prediction-based Word2vec algorithms using isolated word stimuli.
• Compared distributional models against a human-created taxonomic baseline (WordNet).

Main Results:

• Word position in vector space predicted neural activity patterns between 300-500 ms post-word onset.
• Distributional models outperformed the WordNet baseline across multiple vector types.
• Multiple latent semantic dimensions were successfully decoded from brain activity.

Conclusions:

• Empiricist, prediction-based vectorial representations are a strong candidate for modeling human semantic knowledge.
• These findings support the role of distributional semantics in understanding brain-based semantic representation.