Mean-reverting self-excitation drives evolution: phylogenetic analysis of a literary genre, waka, with a neural language model

Abstract To elucidate the evolutionary dynamics of culture, we must address fundamental questions such as whether we can interpolate and extrapolate cultural evolution, whether the time series of cultural evolution is distinguishable from its reverse, what factors determine the direction of change, and how the cultural influence of a creative work from the viewpoint of an instant is correlated with that from the viewpoint of a later instant. To answer these questions, the evolution of classical Japanese poetry, waka , specifically tanka , was investigated. Phylogenetic networks were constructed on the basis of the vector representation obtained using a neural language model. The parent–child relationship in the phylogenetic networks exhibited significant agreement with a previously established honkadori (allusive variation) phrase-borrowing relationship. The real phylogenetic networks were distinguishable from the time-reversed and shuffled ones. Two anthologies could be interpolated but not extrapolated. The number of children of a poem in the phylogenetic networks, the proxy variable of its cultural influence, evaluated at an instant, was positively correlated with that evaluated later. A poem selected for an authoritative anthology tended to have 10–50% more children than a similar but nonselected poem, implying the existence of the Matthew effect. A model with mean-reverting self-excitation replicated these results.