Estimating Association of Movie Segment by Co-occurrence Word from Japanese Subtitles of VOD Lecture

Abstract

Many universities are now using Video On Demand (VOD) lectures, which utilize the Internet to conduct lectures. However, because current systems are yet to implement search functionality for VOD, users must search manually by selecting movies based on their titles, playing a portion of the movies, and searching for the contents they are seeking. This study aims to create a search function that allows users to easily search important points and points to review. We propose a system in which the search function applies mixed beta distributions to the occurrence frequency of a search word based on the appearance frequency of the search word in subtitle data. Using the components of these distributions, the algorithm estimates the movie segments the users are searching for. The shapes of approximation distribution components differ in the mixed beta distribution. In this method, in addition to movie segment estimation for a search word, a frequency distribution is found for words appearing in the same text of the subtitles as the search word, called "co-occurrence words," and by applying mixed beta distribution in the same manner to the frequency distribution of the search word, the movie segments common to the original search word and the co-occurrence word are estimated. However, in cases where few words appear, approximation is difficult with the mixed beta distribution, movie segments of the slides where the words appear are provided. Furthermore, by creating a scatter graph related to the occurrence time and distance of the search word and co-occurrence words, the emergence of the words can be comprehended visually, and the movies with high association and low association among the co-occurrence words, as well as the scope of the slides with high association and low association can be confirmed.