Publication at Faculty of Mathematics and Physics |
2012
Abstract
The paper describes the Charles University setup used in the Search and Hyperlinking task of the MediaEval 2012 Multimedia Benchmark. We applied the Terrier retrieval system to the automatic transcriptions of the video recordings segmented into shorter parts and searched for those relevant to given queries.
Two strategies were applied for segmentation of the recordings: one based on regular segmentation according to time and the second based on semantic segmentation by the TextTiling algorithm. The best results were achieved by the Hiemstra and TF-IDF models on the LIMSI transcripts and various segmentation.