藉由同步文字評論以鑑別影片中的主題和人物

Abstract

我們提出了一個新的方法，藉由分析同步的討論推文去鑑別影片中出現了哪些主題和哪些人物，而此結果可以視為被偵測出的影片事件。我們採用了交互強化的概念建構我們的系統，在判斷同步討論文章中出現的各個人名是否代表同個角色的同時，這個結果也會影響判斷各句文章是否在談論同個主題的機率，而最終這個結果又會回頭影響角色的判斷，如此不斷相互強化影響，將能讓雙方達到較好的成果。分析的結果除了可以讓後續的使用者根據出場人物和主題選擇觀看他個人有興趣的討論內容，也可以在未來依據主題去協助調整影片分段的結果，進而擷取出更好的影片摘要。

We propose a novel method to discriminate which topics and which characters are participated in the video by using synchronous comments of the video, and the results can be regarded as events found in the video. We adopt the conception of mutual reinforcement to construct our system, when determining whether those names indicate the same character, in the meanwhile, the result also influences the probability of whether those sentences refer to the same topic, after that, the new result influences the discrimination of names again. Thus we can make both ends better performances with mutual reinforcements continuously. The results not only can let follow-up user choose the content he interested in according to topics and characters participated in, but also help adjusting video segmentation with topics, then extracting better parts in video to make better video summarization.