ИСТИНА

The talk addresses the issue of speech segmentation. One Russian spoken narrative was segmented into elementary discourse units by trained experts who followed an explicit instruction; afterwards, the very same narrative was annotated by “naïve” annotators not familiar with the instruction. The inter-annotator agreement for “naïve” annotators reached 0.65 (using Fleiss’ kappa); compared to the model annotations, the median value for “naïve” annotations was 0.76. Both values indicate a substantial agreement. To account for the decisions taken by “naïve” annotators, a multi-factored model was be proposed that includes pauses, accent placements, pitch movements and syntactic structure.