Representation Learning for Large Knowledge Graph Completion

点击翻译按钮获取中文摘要

Knowledge graphs are data structures that represent knowledge in machine readable format and are being used extensively in several applications including search engines, recommender systems and digital assistants. Such Knowledge Graphs are usually incomplete because they do not contain all the existing knowledge of a domain and the knowledge they do contain may become outdated. As such, Knowledge Graph completion is the process of adding missing knowledge in a Knowledge Graph. One of the most prominent approaches for identifying this missing information is using Knowledge Graph embedding methods to learn representations. A large body of research has focused on learning such representations based on the structure of the Knowledge Graph and only recently researchers have started exploring the use of additional modalities such as text, images, and audio to produce better representations. Moreover, most of the available embedding methods are not inductive, which means that they only work on existing elements of the graph and cannot accommodate new elements being added as the knowledge of a domain evolves. Finally, an important consideration for embedding methods is their scalability to Knowledge Graphs such as Wikidata, which has more than 100 million nodes. Based on the above, this research project will focus on the following questions:i) How can the structure (e.g. patterns in paths surrounding nodes) of a large and evolving Knowledge Graph be used to predict missing links?iii) To what extent do Knowledge Graph embeddings work in real-world Knowledge Graphs?ii) Which learning strategies are best suited to foundational Knowledge Graph embedding models?