The emergence of verse templates through iterated learning

Every language produces some type of verse in the form of songs, poems or nursery rhymes, which can be analysed as a layer of words set to a template (e.g. a tune, a poetic metre). Verse templates typically consist of hierarchically organised sections: songs are made up of stanzas, divided into lines, containing bars, etc. We hypothesise that this kind of patterns may emerge in the process of cultural transmission; unstructured sound sequences impose a challenge to short-termmemory, but chunking the input makes it easier to parse and reproduce the sequences accurately. In order to test this hypothesis, we have run an iterated learning experiment where random sequences of syllables are evolved across four transmission chains with ten generations of subjects each (all native Dutch speakers). The initial random sequences are generated by concatenating twelve tokens of the set {ban, bi, ta, tin}, as a way to materialise the abstract verse templates without using content-words. More precisely, the experiment aims to model the sequences of nonsense syllables used in many traditions to communicate the rhythmic patterns underlying songs (e.g. bols in Hindustani music, lalay patterns in Berber verse). Participants listened to the sequences of syllables, and tried to reproduce them using four computer keys, eachmapped to one of the four syllables used in the input sequences. The relative timing of the participants’ responses were normalised so that the input always consisted of completely isochronous sequences. Overall, the results show that sequences become shorter, easier to recall and more structured in the transmission process. Some regularities can be related to a global tendency to chunk the input and increase the popularity of a handful of ngrams. Besides, sequences increasingly tend to be opened by a heavy syllable (e.g. ban) and closed by a light syllable (e.g. ta), which can derive from a Dutch-specific bias.