Knowledge Graph Construction and Analysis Platform for Videos

Knowledge Analysis Platform for Videos


    06/2016 - 07/2017

Sponsored by

    Cooperation Research Program of Chinese Academy of Sciences and Huawei Inc. (one of the world's top 500 companies).


    Yantao Jia, Yuanzhuo Wang

Project Objectives:

    This project aims to construct a knowledge graph automaticly for videos, including movies, series, carton, etc. The knowledge graph includes 5 types of entities, i.e., videos, people, video companys, competitions, date, and also collect the abstract, comments and related news. Furthermore, the knowledge graph can be enriched automatically from different data source, and supports a plenty of applications such like relation extraction, tag inference, and so on.

My Work

  • Proposed and implemented ontology alignment method to align two knowledge graphs. It jointly embeded two knowledge graphs and overcame the dearth of alignment training data by co-training, which was accepted as a patent (No.CN201710230135.0.).
  • Designed and implemented relation extraction method for videos, which employed distant supervision using knowledge graph embedding, and which was regarded as one of the three main innovations by Huawei Inc.
  • Designed and implemented entity dereplication method, consisting of candidates generation by Elastic Search, similarity calculation and cluster process.
  • Participated in designing and implementing tag inference by constructing the tag-video graph and video knowledge graph, and learning the embeddings of two graphs jointly. Finally the tags are predicted by the semantic similarity of tag and video graphs. It was in a patent pending in China.
  • Participated in designing a parallel framework to learn knowledge graph embedding in parallel, and the paper was accepted by WI 2017 (Second Author).