A vectorial code for semantics in human hippocampus

Abstract

As we listen to speech, our brains track the meanings of the words we hear. Recent successes of large language models suggest that distributed population geometry can capture rich semantic relationships between words. Motivated by this idea, we hypothesized that semantic information in the brain may likewise be expressed in distributed patterns of activity across neurons, rather than in the activity of neurons narrowly tuned to a specific word. We recorded responses of hundreds of neurons in the human hippocampus while participants listened to narrative speech. We find encoding of contextual word meaning in the simultaneous activity of neurons whose individual selectivities span multiple unrelated semantic categories. Decoding and population geometry analyses revealed distinct neural coding principles for low-versus high-frequency words, likely reflecting the greater polysemy of common words. Similar to embedding vectors in semantic language models, distance between neural population responses correlates with semantic distance; however, this effect was only observed in contextual embedding models (GPT-2 and BERT), suggesting that the semantic distance effect depends critically on contextualization. Consistent with this, we find that neural population activity supports a multidimensional semantic subspace that aligns most closely with the contextual structure captured by GPT-2. Moreover, for semantically similar words, even contextual embedders showed an inverse correlation between semantic and neural distances; we attribute this pattern to the noise-mitigating benefits of contrastive coding. Ultimately, these results provide a neurocomputational account for understanding how neural populations track word meaning.