It has been a year of advances in AI with new tools that create art, write essays, and have conversations that are often (but not always) startlingly eloquent. It is that space of "not always" that keeps these tools from working out of the box to solve actual business problems without human intervention.
In 2022, we saw organizations showing interest in knowledge graph technology but not just for the shiny new tools of a few years ago. Knowledge graphs are now getting attention from the business community in the context of their specific challenges and pain points. While many organizations have conducted “science experiments” with knowledge graph technologies (“we just need to get started with knowledge graphs”) business is paying more attention to how they support many practical functions.
These include improving ecommerce through personalization and better use of first party customer data management, which is a key ingredient to personalization. In addition, knowledge graphs can provide improved employee and customer experience using semantic search and search based applications, making information easier to find and contextualize. Other knowledge graph applications can reduce the cognitive load on humans in a wide range of work tasks.
Knowledge graphs are also foundational to integrating structured and unstructured data and content, which is an enabler of process efficiencies, reduced IT spend, and improved usability. Artificial intelligence applications need reference data to make sense of data features, so organizations have looked at how knowledge graphs provide a holistic framework for AI algorithms and the data that powers them. Finally, our ever-expanding range of virtual assistants and bot platforms benefit from knowledge that is componentized with a standard domain content model and tagged from a knowledge graph.
We expect continued interest in knowledge graphs going forward into 2023 as more businesses come to understand knowledge graphs and how to implement this technology.
Many of the customer service and employee help desk automation vendors area building ontologies on a massive scale, with millions of concepts designed to correctly identify intent (the action or thing people want to accomplish) based on utterances (the endless ways in which people can make requests). These ontologies map related concepts in a semantic vector space. This space represents the way in which similar concepts are derived from human judgment as well as analysis of large amounts of data. The understanding of intent data is growing rapidly. Increasingly, preprocessing of the utterance can extract entities that act as additional signals for retrieval.
Greater recognition is emerging of the need to curate knowledge for chatbots and virtual assistants. Automated approaches for componentizing (for example, breaking large PDFs into chunks for question answering systems) are adding scale to the painstaking approach of populating bots with training data. Auto-tagging of components further improves efficiency. Many vendors still claim their systems simply ingest the content and data as it generates questions and answers. The quality of that result will always depend on the quality of the content. The more signals that are provided through metadata, the better, in terms of speed and retrieval accuracy.
Ontologies are essential in understanding and representing core information in organizations, and the relationships among different groups of information within them. Therefore we expect continued focus on organizing information and establishing robust ontologies to reflect enterprise knowledge.
ChatGPT is producing a lot of interest and excitement. ChatGPT is an open-source conversational bot that carries out a dialog that is flexible, allowing it to respond to questions beyond the initial one asked by the user. It can also explain why it gave a particular response and evaluate both the inquiries and its own responses for accuracy.
This software has an out of the box demo response about quantum computing that is good of course. Was it curated? Not according to the tool. Here is my question about that, and the response:
Was that example curated? How was it generated?
"The example I provided was not curated. It was generated by me based on my understanding of quantum computing. As a large language model trained by OpenAI, I have been trained on a vast amount of text data, which allows me to generate responses to questions based on the information I have learned from that data. In this case, I used my understanding of quantum computing to provide a simple explanation of the concept."
I then asked about the role of AI in knowledge management and received this well-written response:
"The role of AI in knowledge management is to help organizations efficiently store, organize, and access their knowledge and expertise. AI technologies, such as natural language processing and machine learning, can be used to automate the process of capturing, organizing, and disseminating knowledge within an organization. This can help improve the efficiency and effectiveness of knowledge management processes, allowing organizations to make better use of their collective knowledge and expertise."
"For example, AI-powered tools can be used to automatically categorize and tag knowledge assets, such as documents and articles, making them easier to search and access. AI can also be used to analyze large volumes of data and extract key insights and information, which can be used to improve decision-making and problem-solving. Additionally, AI can be used to personalize knowledge experiences for individual users, providing them with tailored recommendations and suggestions based on their interests and needs. Overall, the use of AI in knowledge management can help organizations better leverage their knowledge and expertise to improve their performance and competitiveness."
ChatGPT is certainly a tremendous step forward, but like any emerging technology it will require continued development. Some of its answers reportedly do not make sense, and in some cases, just a slight modification in the phrasing of a question makes the difference between getting an answer and causing the system to say it does not know the answer.
It’s important to remember that very specialized knowledge domains will require specialized language models. As the industry matures, there will be more domain specific knowledge graphs and ontologies available, but a layer of differentiated knowledge will always be needed. Right now, most organizations are still just trying to get their foundational house in order. This is an important Just as there are simple answers to questions no matter what the mechanism, some complex questions and problems are unique and often require human creativity. The low hanging fruit for AI is still out there. As the common tasks and typical use cases are further exploited and automated, though, the gap will be closing between AI and the unique human judgment and problem solving.