Multimodal Corpora Annotation: Validation Methodsto Assess Coding Scheme Reliability

摘要

Many multimodal corpora have been collected and annotated in the last years. Unfortunately, in many cases most of the multimodal coding schemes have been shown not to be reliable. This poor reliability may be caused either by the nature of multimodal data or by the nature of statistic methods to assess reliability. In this paper we will review the statistical measures currently used to assess agreement on multimodal corpora annotation. We will also propose alternative statistical methods to the well known kappa statistics.