Text Segmentation Methods for Annotation on eHealth Consultation with Interview Function Labels: A Comparative Study

There have been several existing text segmentation methods. Nevertheless, no study has provided the experimental result on the performance of those methods in segmenting text of health consultation data based on sentence context and provided automatic annotation on the segmented results in the form of interview function labels. This study compares four methods with different text segmentation approaches and analyzes their performances based on their reliability concerning human expert judgment. The methods are Content Vector Segmentation (CVS), GraphSeg, K-Means, and Latent Dirichlet Allocation (LDA). This study used a greedy similarity approach to perform automatic annotation by selecting candidate labels based on the maximum value. The annotation results were evaluated using Gwet's AC1 method to assess the integrity among evaluators in clinical research. The evaluation results indicate that the CVS method outperforms other methods and has a substantial level of agreement (0.67). It is also relatively stable, with a standard deviation of 0.12.