A vectorial code for semantics in human hippocampus.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• The brain processes word meanings during speech comprehension.
• Large language models (LLMs) have shown success in representing meaning using vectors.
• The hippocampus is crucial for semantic coding.

Purpose of the Study:

• To investigate if the human brain uses vectorial coding principles for semantic representation, inspired by LLMs.
• To explore how neural activity in the hippocampus encodes contextual word meaning.

Main Methods:

• Recorded neural activity from hundreds of neurons in the human hippocampus during narrative speech.
• Analyzed neural population responses in relation to semantic distances derived from contextual (e.g., BERT) and non-contextual (e.g., Word2Vec) embedding models.
• Examined neural response variance in relation to lexical polysemy.

Main Results:

• Found that neural population responses encode contextual word meaning, with distances correlating with semantic distances in contextual models.
• Observed an inverse correlation between semantic and neural distances for highly similar words, potentially due to contrastive coding.
• Demonstrated that neural response variance increases with lexical polysemy, highlighting the role of context.

Conclusions:

• Neural coding of semantic meaning in the hippocampus aligns with vectorial principles, similar to contextual LLMs.
• Context plays a critical role in how the brain represents word meaning.
• The findings support a hypothesis of vector-based semantic coding in the human hippocampus.