Identification of sports game highlights using live comments data

In this paper, we propose a procedure for identifying sports game highlights using online live comments.Highlights in sports games can be defined as important events that capture the viewers' attention, often involving scoring.Internet portals and many other platforms may edit game videos and provide highlight scenes.Since reviewing each game and identifying highlights manually can be labor-intensive and time-consuming, statistical learning procedures are beneficial for the automatic identification of highlights.In this paper, we extract words from preprocessed live comments, and divide videos into one-minute intervals to create a document-term matrix summarizing the frequency of word occurrences at each timestamp.Next, employing the log-odds ratios as weights, the weighted sum of the term frequencies at each timestamp is calculated to generate an event score.Then, the nonparametric kernel regression model is applied to event scores to estimate the underlying trend of event scores.The residual for each timestamp is obtained as the difference between underlying trend and the actual event score.Finally, we predict events as those instances whose residuals are widely deviated from the estimated trend levels.Applying this approach to real soccer games, we could predict actual highlights more accurately compared to simply using comment frequencies.