Extraction of novelty concepts from TV broadcasts with longitudinal user experiments

Recent popularity of online catch-up TV services has facilitated time-shifted TV viewing. However, contemporary services do not utilize the rich information available in broadcast TV content streams. Richer program descriptions and summaries help on-demand viewers to employ new information seeking behaviour to find interesting content. Techniques for content-based analysis of broadcast TV streams aim to improve access to relevant archived TV content and assist in efficient on-demand viewing. In this paper we introduce a methodology that extracts novelty concept words from Finnish broadcast TV stream. The methodology is employed in an online content analysis system, which executes near real-time analysis and indexing of seven free-to-air DVB TV channels. The methodology uses machine learning and statistical data mining techniques to extract descriptive novelty concepts automatically from TV program subtitles. Extracted concepts are further used to summarize and access program content in end-user services that facilitate search and browsing of archived TV content. We show results from user logs of nearly 3 000 sessions to demonstrate how novelty words have been used in our prototype services.