Knowledge Graphs as a Test-Bed for Current Generation AI Algorithms

Knowledge graphs have a two way relationship with the current generation AI algorithms. On one hand, knowledge graphs enable many of the current AI applications, and on the other, many of the current AI algorithms are used in creating the knowledge graphs. We will consider this symbiotic synergy in both directions.

Personal assistants, recommender systems, and search engines are applications that exhibit intelligent behavior and have billions of users. It is now widely accepted that these applications behave better if they can leverage knowledge graphs. A personal assistant using a knowledge graph can get more things done. A recommender system with a knowledge graph can make better recommendations. Similarly, a search engine can return better results when it has access to a knowledge graph. Thus, these applications provide a compelling context and a set of requirements for knowledge graphs to have an impact on immediate product offerings.

To create a knowledge graph, we must absorb knowledge from multiple information sources, align that information, distill key pieces of knowledge from the sea of information, and mine that knowledge to extract the wisdom that would influence the intelligent behavior. The AI techniques play an important role at each step of knowledge graph creation and exploitation. For extracting information from sources, we considered entity and relation extraction techniques. For aligning information across multiple sources, we used techniques such as schema mapping and entity linking. To distill the extracted information, we can use techniques such as data cleaning and anamoly detection. Finally, to extract the wisdom from the graph we used inference algorithms, natural language question answering, etc.

Hence, knowledge graphs enable the current generation of AI systems, which provide motivation and set of requirements for them. Current AI techniques are also fueling our ability to create the knowledge graph economically and at scale.

You may learn more about this perspective in this excellent lecture by Luna Dong.