VIREO @ TRECVID 2015: Video Hyperlinking (LNK)
Lei Pang, Chong‐Wah Ngo · 2015
This paper presents an overview and comparative analysis of our system designed for TRECVID 2015 [1] video hyperlinking (LNK) task. The application scenario for video hyperlinking is to satisfy the needs of users to find further information on the content of interest contained within an anchor. The task here is given an anchor to generate a ranked list of video segments relevant to the name entities extracted from the anchor. Our four runs are summarized below: - tf-idf : The subtitles of video segments are indexed with Lucene [2] and more like this (MLT) query is adopted for tf-idf based retrieval. - word2vec: Each video segment is represented by uniformly summing the vector representations of words in subtitles using Word2Vec [3, 4] and the relevant segments are ranked based on cosine similarity.