A scaling law of contextual persistence in human language

Short summary

This paper discovers a scaling law governing how word arrangement in human language persists contextually over distance. Using LLMs as probabilistic probes across ten corpora in six language families, the authors show that contextual persistence P(d) decays approximately as 1/d (mean exponent α=1.04, median r²=0.96). The effect vanishes in scrambled controls and does not appear in genomic or protein sequences, establishing it as a property unique to human language structure.

• •Contextual persistence in human language follows a power law P(d) ∝ d^-1.04 across ten corpora and six language families
• •Effect vanishes in scrambled text and is absent in genomic/protein sequences, confirming it reflects word arrangement
• •Exponent near 1 distributes contextual influence approximately uniformly across logarithmic timescales

Generated with AI, which can make mistakes.