This week’s Insight was written by Lorena Duguid, Intanify's Knowledge Engineer.
IT research firm Gartner’s most recent Impact Radar for Generative AI has highlighted Knowledge Graphs as a high-mass technology, reaching real use cases in less than a year.
Each of the 25 technologies and trends featured on the Impact Radar is categorised into one of four themes: Application-Related; Model-Related; Model Performance and AI Safety; and Build- and Data-Related Generative AI. The closer to the centre of the radar and the larger the mass of the technology, the higher and sooner the impact.
This is hardly surprising when you consider what knowledge graphs offer in this age of big data, black box models, and hallucinations.
Knowledge graphs are a form of artificial intelligence that was first developed at Stanford in the 1970s. They leverage the wisdom of human experts to construct a model of the logical rules that govern the real world, called a knowledge graph, and use that model to inform decision-making. By combining multiple experts’ knowledge in multiple knowledge graphs, you can cover a wider area of expertise in equal detail. Modern knowledge graphs are used in hundreds of applications, from determining therapies to assisting in loan decisions.
There aren’t many things better or more trusted for businesses than expert advice, but experts’ only have so many hours in their day.
Expert advice needs to have reasoning and explainability behind it, and knowledge graphs demonstrate that in a deterministic way. Making expert knowledge explicit allows greater access and preserves it for as long as you maintain your system, perhaps for many decades.
Key benefits of Knowledge Graphs
- Explainability - the capacity to follow each rule step by step allows for complete transparency and explainability, and the best systems offer full reports of every rule that led to a decision
- Accuracy - although ES are only as good as the knowledge contained in them, they offer controlled data input, mitigating the risk of the hallucinations that can be experienced by LLMs
- Industry-specific application knowledge graphs are built for one specific purpose, and have a highly focused use case, meaning that narrowing ES options for your business is not a mammoth task that requires deep domain knowledge
How we built – and continue to build – Intanify’s knowledge graphs:
In areas like intangible assets, where expertise is concentrated in the heads of world-class experts and a dearth of data suitable for training LLMs, knowledge graphs have massive potential. That’s why we chose to build our knowledge graphs in partnership with Dr Viktor Dörfler, a celebrated AI academic with more than 20 years of experience in building knowledge graphs and related expert systems. We tapped experts from world-class law and consulting firms and built a system that automates some of the most nebulous tasks in the industry – articulating the most important parts of businesses into an inventory and valuing them using industry standard methods. Our system also has the capability to evaluate the legal risks to those assets and guide users on mitigating those risks.
There are so many other inaccessible repositories of knowledge on this subject, and we’ve built an “assembly line” that means we’re poised for new builds in the coming months, guided by the deep needs we see in innovative businesses everywhere. What’s more, we’re going to combine the power of different AIs; using our unique data and reasoning to train generative AI and supercharge the process for our users.
Read further about Knowledge Graphs and related systems with these key resources:
Viktor Dörfler (2022) What Every CEO Should Know About AI, Cambridge University Press, Cambridge, UK. DOI: 10.1017/9781009037853 (particularly section 2.2)
Edward A. Feigenbaum (1992) A Personal View of Expert Systems: Looking Back and Looking Ahead (No. KSL 92-41), Department of Computer Science, Stanford University, Stanford, CA. Electronic version: https://purl.stanford.edu/gr891tb5766
Stuart Russell & Peter Norvig (2020) Artificial Intelligence: A Modern Approach (4th edition), Pearson Education, Harlow, UK. Electronic version: http://aima.cs.berkeley.edu/
Liao Shu-Hsien (2005) Expert system methodologies and applications—a decade review from 1995 to 2004, Expert Systems with Applications, 28(1): 93-103. DOI: 10.1016/j.eswa.2004.08.003