Measuring Syntactic Priming in Dialogue Corpora

The tendency to reuse syntactic constructions is often attributed to syntactic priming. We devise a simple, distribution-based measure of priming between linguistic constructions (or syntactic rules/categories), and find priming in treebanks of dialogue corpora, both for context-free production rules and for Combinatory Categorial Grammar categories. It is stronger for task-oriented dialogues, and stronger in lexical categories than in syntactic categories. As priming cannot be measured directly in language corpora, we use the decay of rule repetition probability as a proxy. A limitation of the method presented here is that it conflates self-priming and other-priming. However, we consider it a great advantage that our method takes into account all rules or categories occurring in a given corpus, not just a few carefully selected constructions.