Measuring the Similarity between TV Programs using Semantic Relations

Proceedings of COLING 2012: Technical Papers, pages 2945–2960, COLING 2012, Mumbai, December 2012. Measuring�the�Similarity�between�TV�Programs� ����������������� Using�Semantic�Relations� Ichiro YAMADA Masaru MIYAZAKI Hideki SUMIYOSHI Atsushi MATSUI Hironori FURUMIYA Hideki TANAKA Japan Broadcasting Corporation, 1-10-11 Kinuta, Setagaya-ku, Tokyo, JAPAN yamada.i-hy@nhk.or.jp, miyazaki.m-fk@nhk.or.jp, sumiyoshi.h-di@nhk.or.jp, matsui.a-hk@nhk.or.jp, furumiya.h-ke@nhk.or.jp,tanaka.h-ja@nhk.or.jp ABSTRACT This paper presents a novel method of measuring the similarity between TV programs by using summaries of the Electronic Program Guide (EPG). Most previous methods use statistics such as the TFIDF based cosine measure of word vectors, whose elements are words appearing in the summaries. However, these approaches are not effective because TV program summaries, especially short ones, do not necessarily share many words even when they have similar meanings. The proposed method generates a graph structure whose nodes are TV programs and nouns. These nouns are connected by semantic relations that are extracted from the Web automatically. The similarity between two TV programs is measured in terms of the relativeness of two TV program’s nodes in the graph structure by using a random walk algorithm. Experiments showed that our method is better at measuring similarities between two TV programs compared with baseline methods. KEYWORDS : Measuring similarity, Semantic relation, Recommendation system, Graph structure, Random walk 2945 1 Introduction Japanese broadcasting stations have started services to provide viewers with previously broadcast TV programs on demand. Since these services are becoming popular, it is important for them to have an efficient way to find programs that a viewer wants to watch amidst huge program archives. Many on-demand services have the ability to present TV programs related to a selected program and rank them. This function makes it possible for viewers to find programs they would be interested in but would not have known about in advance. The TV program that the viewer selects from such a list is highly dependent on its presentation rank given by the on-demand service. FIGURE 1 shows how the number of selected programs depends on the presentation rank in a major Japanese on-demand service, NHK on demand. The higher ranked TV programs are selected more frequently than the lower ranked ones. For this reason, it is important to find out how to select related programs in huge program archives and how to rank them for better on- demand services. In order to select the related programs and rank them, NHK on-demand system measures the similarity between TV program summaries in Electronic Program Guides (EPGs) in advance. The system ranks TV programs according to their similarities to the selected program by a viewer. Since the viewer might interested in watching TV programs which are relevant, but not exactly similar in content, the technique for measuring the similarity between programs can exclude exactly similar content and select only related programs and rank them1. NHK on-demand adopts a method of Goto et al. (2010) for measuring the similarity between TV program summaries. This method is based on word co-occurrences. Each summary is represented using words or n-grams of words as a vector, and the similarity between two …

Content truncated to 500 words