Different computational relations in language are captured by distinct brain systems

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• Understanding how language shapes semantic representations in the brain is crucial.
• Previous research has not fully elucidated the specific statistical properties of language that drive neural semantic encoding.

Purpose of the Study:

• To investigate whether statistical properties of language, specifically graph-topological relations, explain neural semantic representations.
• To determine if these findings are specific to language or generalize to other information types.

Main Methods:

• Utilized two functional magnetic resonance imaging (fMRI) experiments with shared semantic processes.
• Compared the explanatory power of three computational principles: co-occurrence, graph-topological relations, and vector embeddings.
• Analyzed neural patterns associated with semantic representations of words and visual objects.

Main Results:

• Graph-topological word relations uniquely explained neural patterns in specific brain regions, including the anterior temporal lobe (graph-common-neighbors) and inferior frontal gyrus (graph-shortest-path).
• Simple co-occurrence and vector embedding relations did not show similar explanatory power.
• These language-specific topological effects were distinct from sensory-motor similarities and visual object processing.

Conclusions:

• Neural semantic representations are specifically tied to the graph-topological properties of language.
• Different brain regions are specialized for processing distinct information types based on topological computations.
• The brain encodes information in a manner specific to both the information type and its statistical properties.