Answer segmentation for question answering using latent dirichlet allocation and delta Bayesian information criterion

This study presents an approach to answer segmentation for the question answering system in a closed domain: a message board for emotional comfort. In this study, the unstructured articles were collected from the psychological consultation websites. In unstructured document processing, supervised latent Dirichlet allocation (SLDA) is employed for event detection and SLDA with delta Bayesian Information Criterion (delta-BIC) are employed for answer segmentation. In this study, the proposed method is applied to the negative emotion event detection of a user to provide appropriate answers segmented by the proposed method from the unstructured documents. K-Fold cross validation was employed to compare the performance of the proposed method. The evaluation result shows that the precision and recall of the proposed method achieved 96.62% and 84.2% when the distance between the detected segmentation points to the correct segmentation point is 2. The encouraging results confirm the usability of this proposed method for future applications.