Recalling sequences from memory can explain the distribution of recursive structures in natural languages

Abstract

Language operates within the cognitive machinery of its users. Hence language structure is likely to evolve under the pressure of cognitive constraints (Christiansen & Chater, 2008). The challenge remains, however, in determining precisely how this may have occurred. Hierarchical recursive structures are especially difficult to relate to finite cognitive features. Here, we propose a new cognitive account explaining why Center Embedded recursive structures of relative clauses (as in The boy A₁ the dog A₂ chases B₂ falls B₁) (A₁A₂B₂B₁) are ubiquitous among thousands of languages, whereas Crossed-Dependent (CD) structures (A₁A₂B₁B₂) hardly ever occur. The preponderance of CE grammars is surprising considering they can produce dependent elements at longer distances than CD. We propose that this can be explained by memory retrieval mechanisms combined with linguistic word binding operations (role assignment). Processing CE requires the sequential retrieval of referent words in a backward direction, and CD in a forward direction. We first specify two Retrieval-and-Binding (R&B) functions, from which we derive mathematically that R&B performance under backwards recall (CE) exceeds performance under forward recall (CD). Next, we reanalyze an existing dataset that investigated strategies of recall and review the literature on sequential recall strategies under conditions that mimic sentence processing. The reanalysis verified the predictions of our account and showed that a backwards recall (CE) strategy is superior under conditions relevant to language processing. We suggest that the productive power of recursive embeddings is best conserved in a CE instantiation because memory mechanisms optimally support the processing of this structure, which might explain why CE has prevailed during language evolution.