Measuring syntactic complexity in spontaneous spoken Swedish

Hesitation disfluencies after phonetically prominent stranded function words are thought to reflect the cognitive coding of complex structures. Speech fragments following the Swedish function word att 'that' were analyzed syntactically, and divided into two groups: one with att in disfluent contexts, and the other with att in fluent contexts. Complexity was calculated in terms of a number of measures related to syntactic tree structures produced by the analysis tool GRAMMAL. Results showed that disfluent att is in general associated with significantly higher mean complexity values than fluent att. This information can be used to predict whether the function word at the beginning of a fragment is likely to be disfluent or not. Two kinds of statistical classification algorithms (Bayesian and neural networks) were used to test this hypothesis. The best result was 71% correctly classified cases, which is significantly better than a system that is based on selecting the data's majority class.