Proceedings of the Annual Meeting of the Cognitive Science Society

Across languages, word order within phrases shows robust statistical regularities: dependents are typically ordered con- sistently by type, word order tends to be harmonic with the head near an edge, and more strongly associated dependents tend to occur closer to the head. Existing accounts often derive these properties from separate mechanisms or stipulations. I propose a unified efficiency-based explanation grounded in in- formation theory: the idea is that languages are organized so that it is possible to make good predictions using only small amounts of memory. Using simulations of nouns modified by multiple adjectives differing in association strength to the noun, I compare linguistically motivated ordering policies that vary in consistency, harmony, and locality. Across sampled source distributions, I find that memory requirements for prediction are minimal in orders that are simultaneously harmonic, consistent, and local. I analyze why these effects arise: consistency makes upcoming category structure predictable from small amounts of context, and harmony makes phrase boundaries easy to antici- pate. The results suggest that diverse crosslinguistic patterns of phrasal order may reflect a general pressure to make incremental prediction simple.