Taming the beast ... or ... How to make sense of knowledge graphs

June 20, 2022 by Tassilo Pellegrini
Umutcan Simsek - Poster & Demo co-chair of Semantics 2022

In this interview Umutcan Simsek, Poster & Demo co-chair of SEMANTiCS 2022, gives you a head start on knowledge graphs and what role they might play in an AI-powerred future.

You have been working intensely on knowledge graphs over the previous years. How does the concepts of knowledge graphs relate to other forms of knowledge representations such as ontologies or taxonomies?

Umutcan Simsek: Knowledge graphs and ontologies typically live together symbiotically. An ontology can be used as a schema for the facts in a knowledge graph. This is of course not in the sense of relational databases but more with the purpose of describing the facts. Thanks to the vocabularies like SKOS, taxonomies can also be connected with the entities in a knowledge graph rather easily. That being said, we see in our empirical analysis that knowledge graphs contain much more factual knowledge comparing to terminological knowledge and ontologies used are less complex comparing to the traditional knowledge bases. 

Knowledge graphs need to be constantly maintained. What are the major challenges in this respect technologically and – maybe even more important – organizationally?

Umutcan Simsek: From our perspective, building a knowledge graph is a not a one off endeavor but a lifecycle that includes the creation, hosting and maintenance of knowledge graphs in the form of curation. There are of course many technological challenges here. For starters, each step in the knowledge graph lifecycle such as creating semantically annotated data from heterogeneous sources, hosting them in a way that supports modularization and provenance tracking, assessing the quality and improving the quality through cleaning and enrichment of the knowledge graph is a different beast. Each of these tasks requires efficient methods considering the heterogeneity and the vast size of knowledge graphs. From an organizational point of view, these tasks may require different expertise and may have to be done in a distributed manner. Proper pipelines to support these aspects must be developed. 

 What will be the role of knowledge graphs in future AI systems?

Umutcan Simsek: There are obviously many use cases for knowledge graphs in the context of AI, for example as part of the Explainable AI work. However to me the foremost role of knowledge graphs will be as a knowledge source, which is particularly critical for any useful intelligent application. ML works great for speech recognition which brought us conversational assistants, however they are only useful if they have domain and task knowledge in real life. For this purpose knowledge graphs are ideal as they can integrate large heterogeneous sources rather flexibly. A self-driving car would be making less mistakes due to misrecognized objects, if it had access to a knowledge graph of city roads and let's say the traffic signs on them.