Home » Posts tagged 'llm'
Tag Archives: llm
Semantic Data 2026 Themes and Trends
“Everywhere one seeks to produce meaning, to make the world signify, to render it visible. We are not, however, in danger of lacking meaning; quite the contrary, we are gorged with meaning and it is killing us.” – Jean Baudrillard, Simulacra and Simulation
I attended the annual Henry Stewart Semantic Data New York Conference for its third year on October 1st. Like last year, I’m using my monthly blog to perform a thematic roundup of the talks at the conference.
While of course the focused topic across much of the conference was AI, there has been a gradual shift from the uncertainty and lack of clarity brought about by rapid change and development to an acceptance that AI is here to stay. More practical use and evidence over the past several years has underscored the need for semantic practices paired with AI work of all kinds. Despite an ongoing lack of certainty as AI continues to rapidly evolve and find its place in the enterprise, there are more practical activities to create foundational semantic bedrock for AI endeavors.
Cross-System Alignment
One of the foundational principles of semantic data is to create a common language and domain model across all content. Many of the presenters emphasized how essential this is in the era of AI, particularly when it comes to aligning truth across systems in a semantic layer. Regardless of the system’s purpose, data, content type, or format, using taxonomies and ontologies to rationalize information across systems is essential for source of truth information for AI. A benefit of cross-system alignment is a cross-system and cross-source knowledge graph. The semantic layer of meaning created in taxonomies and ontologies paired with the things themselves containing data and knowledge allows humans and AI agents to find information grounded by semantic models.
A common theme across the conversations was how little has changed even as so much has changed. The rapid rise of AI is mind-blowing. Still, organizations suffer from poorly managed legacy data and systems requiring reconciliation through semantic practices to be AI-ready. Speakers covered multiple, disjointed taxonomies or metadata schemas, “islands” of differently governed information, and fragmented system landscapes that may perform in isolation but could not deliver as an organizational knowledge base for AI initiatives.
As always, cross-system alignment requires people working together across teams and functions to build common taxonomies and ontologies, integrate systems where possible, and automate processes that can be automated. While a semantic layer can provide a common truth, the passage of metadata and data across systems is still a common use case, so integrating systems and aligning to a common source of truth is essential.
Another common message that has not changed with AI is the use of semantic standards. Many of the speakers were using open standards like W3C’s RDF, SKOS, OWL, SPARQL, and SHACL, and other metadata and content representation standards. Interoperable standards is yet another foundational layer allowing organizations to move from AI proof of concepts to production-ready releases.
Deterministic Semantics
Semantic data practitioners agree using taxonomies and ontologies provides a foundational model for all types of data and content across the organization. Semantics are also key for preventing “hallucinations” in large language model (LLM) chatbots. When an LLM hallucinates, it’s not so much making up an answer out of thin air as providing an answer based on probability. At some point, if the LLM has very little contextual information, the most probable answer it provides is likely to be far off the mark. LLMs are great at text summarization, identifying and extracting new concepts, and automatically suggesting classification metadata, but they’re not very good at answering questions involving domains not covered or crossing between multiple documents or datasets to arrive at an answer for more complex questions.
In contrast, taxonomies and ontologies are deterministic. They are constructed at a slower pace and represent encoded truths in the domains they cover. If the data is not in the graph, then that information has no coverage. If the data is in the graph, then it can be contextually true if it is modeled correctly. Using a hybrid approach of deterministic semantic models and probabilistic AI results in much higher answer accuracy. Strictly deterministic use cases are often when the data or content structure is very well known, regular, and homogenous. In these cases, deterministic models and classifications are repeatable and can have high confidence in accuracy.
Another consideration was the nature of the work and industry. Speakers and attendees represented a broad range of industries, including retail, pharmaceuticals, libraries and museums, financial, and music and entertainment. For some industries, there can be general annoyance or reputational damage if chatbots answer information incorrectly. In other highly regulated industries, the wrong answers could lead to serious legal and financial repercussions or even severe problems for consumers or product users. While no organization wants to have their customers not trust them, the severity of the breach of trust matters. Therefore, accuracy in chatbot responses and recommendations is essential.
Knowledge and Change Management
Despite the dominance of AI in the conversation, the role of people is still ever-present and important to making organizations function. No matter how much knowledge LLMs have, it is not the experiential tacit knowledge of humans. Also, there are knowledge processes simply not captured at all in documentation: what decision was made, who made it, why was it made, and what uncaptured tacit knowledge helped to shape and form the decision? Meeting notes may capture a lot of information, especially with AI note takers in virtual meetings, but there is still unspoken knowledge feeding into decision-making. Human imperfection also leads to challenges in converting tacit knowledge to explicit, and structured, information. People, with or without the help of chatbots, create inconsistent and incomplete knowledge documents. Thus, there remains inconsistent word choice and descriptions still needing to be rationalized to fit into the taxonomy and ontology semantic representations.
Many of the speakers hit on change management as an essential, if often the most difficult, aspect of semantic practices across an organization. Adopting new systems, new ways of handling data in graph format, and agreeing on terms and definitions is still one of the most difficult tasks of anyone working in semantic data.
“Trust” was a word and a theme appearing across presentations and panel discussions. Users need to trust the data so they can trust the output of AI. Users trust what they know, however, so arriving at consensus across terminology, lines of work, and systems is feasibly still a greater challenge than the technical implementation of semantic structures and technologies.
Politics, Language, Power
“Clichés, stock phrases, adherence to conventional, standardized codes of expression and conduct have the socially recognized function of protecting us against reality, that is, against the claim on our thinking attention that all events and facts make by virtue of their existence.” – Hannah Arendt, The Life of the Mind
I’ve always been fascinated by language. I have a Bachelor of Arts and a Masters in English Language and Literature, including courses in American, European, and “World” literature, composition, linguistics, and philosophy. Paired with an undergraduate minor in History and a teaching degree, it’s no surprise that I often think about language and its connection to history and politics.
Like many taxonomists, I fell into this profession I didn’t even know existed. I never planned to be a business taxonomist, but I don’t think any other profession has quite captured my interest in language and keeping things organized quite the way taxonomy has. Taxonomists are typically methodical researchers, language savvy, and excellent at working with people. Many, but not all, come from a library science background.
I wonder, on nearly a daily basis—and especially at this moment in time—what our role as taxonomists is when building out semantic models. What is truth? Is there truth with a capital “t”? What are the ramifications of the words we choose as taxonomy concepts and the way we model ontologies? What power, if any, do we have in shaping thought and should we? Big questions, to be sure. But in those big questions, there are some very real consequences of our work. I believe we have the power to be conscientious at a minimum and thought-shapers with concerted effort.
You’ve Been Doing [Something] Wrong
I have never been a fan of clichés or hackneyed expressions no matter how accurate or timeless they are. Our pop culture world is full of them, from clickbait headlines to movie posters to commercials. Clichéd language is easy, familiar, and readily available. So easy, in fact, I wonder if we even really think about the meaning of these expressions anymore.
There’s a wonderful word I discovered a few years ago that describes a particular type of cliché, the snowclone:
“A snowclone is a phrasal template based on a cliché, substituting words to express a similar idea in a different context, often to humorous or sarcastic effect…The term snowclone was coined in 2004, derived from journalistic clichés that referred to the number of Inuit words for snow. The linguistic phenomenon of “a multi-use, customizable, instantly recognizable, time-worn, quoted or misquoted phrase or sentence that can be used in an entirely open array of different variants” was originally described by linguist Geoffrey K. Pullum in 2003.” (Wikipedia)
You already know exactly what a snowclone is because you’ve heard so many of them:
- The mother of all X,
- In space, no one can hear you X,
- X is the new Y,
- Have X, will travel, and
- X as a service.
In recent use, two of these snowclones—or to use a nuanced synonym, memechés—have been doing their best to drive me mad. The first is, “You’ve been doing X wrong.” I know it’s not deliberately directed at me, but I find it personally offensive in the insinuation that I am somehow so clueless that I’ve been doing so many things incorrectly my whole life. Since misery loves company (sigh), it only makes me feel slightly better that all of you have also been doing so many things wrong. Perhaps my favorite example was the email headline I received a few weeks ago claiming, “You’ve been sleeping wrong.”
The second, and this will speak to a larger point, is, “It’s not x, but y.” While negative parallelisms have been so long in use as to be cliché, there is also the overuse of such structures by LLMs:
“As a distinct element of style, however, it gained attention only in the 2020s when it became apparent that large language models (LLMs) such as ChatGPT use negative parallelism inordinately often (about three times more frequently than humans), to the point that such constructions came to be seen as a hallmark of machine-generated text.” (Wikipedia)
Memechés are the new clichés, but we’ve always had language that has become patterned and overused. Such use of language is to be, or not to be, but has never really been concerning enough to raise alarm bells. Until AI.
The Ouroboros and Model Collapse
Years ago I read a fantastic article about the death of journalism. Or, rather, journalism becoming a monoculture in which there were fewer and fewer media outlets, a polarization of the ones that still exist, and the reliance on single authorship and reuse from sources like the Associated Press (AP). It occurred to me even then that a lack of diversity of news sources would lead to a homogenization of viewpoints and facts. Perhaps with the ubiquity of self-publishing on social media and a distrust of traditional news outlets (actively fostered through misinformation campaigns), we actually have the opposite: so many varieties of truths and untruths as to be unable to determine fact from fiction, truth from falsehood.
The decline of investigative, fact-driven journalism has brought with it a polarized monoculture of thought, a dearth of interesting and novel ideas, and a replacement in the form of opinion and belief as fact. Without going into a deep dive on French Postmodern thought, the flattening of truth and the use of language as political power is not just philosophy (X), it is reality (Y).
I have been thinking if thoughtful content has been streamlined to surface homogenized ideas, and this homogenized content is food for LLMs, and LLMs in turn produce more homogenized output, then our bots are the electronic ouroboros, eating its own tail until it disappears into a point. In other words, garbage in, garbage out.
I felt absolutely vindicated in my thinking when I came across the article, “AI Is Becoming a Monoculture”, by Ayoub Nainia. I won’t summarize the whole article, because it is well worth the read, but I will focus on the notion of model collapse.
“Models trained on recursively generated data degrade, generation after generation, and the first thing to vanish is the tails of the distribution. The rare, the unusual, the outliers, the very diversity that gave the system its range, all of it erodes first, until the model converges toward a bland and increasingly wrong average.”
A monoculture of training data, and especially training data created by and then consumed by the models, leads farther and farther away from originality. The model may not be incorrect per se, but the idea of creating new and novel ideas may be harder and harder to achieve. Our LLMs are middling, averaging, and homogenizing their inputs.
Hierarchies as Power, Rhizomes as Truth
Ok, I said I wouldn’t say much about French Postmodern philosophy, but I will say this: Michel Foucault saw hierarchies as power structures dominated from the top; Gilles Deleuze and Félix Guattari (the absolute punk rockers of French Postmodernism) saw rhizomes as decentralized structures of connected points. Taxonomists should see the parallel right away: taxonomies are hierarchical and more or less rigid; graphs, and the ontologies and taxonomies built on them, are decentralized rhizomes. I’ve written and presented about this topic several times: Knowledge Graphs and Their Punk Rhizomatics, Why Graphs? Transcending the Limits of Hierarchical Thinking (with Bob Kasenchak), and Using Structure in Knowledge Organization.
If taxonomies are hierarchical power structures, are taxonomists fascists? We do, after all, advocate for strong central governance and control and rule over a fortified complex of taxonomies and ontologies that only a few may enter. Are we benign dictators? Well-meaning autocrats?
If language, politics, and power are hopelessly entwined and the role of the taxonomist is to model domain truth, then you have power. You have both hierarchical fascist power and punk rock rhizomatic power. With every concept label choice and modeling decision, you advocate for truth and facts. Perhaps not capital “t” truth, if such a thing exists, but researched, fact-based truth. Truth to power. In the world of “X as a service”, go do taxonomy as a service. Make it a fact-based, truth to power, rhizomatic service. Make your service in the service of balancing probabilistic statistics with LLMs fed on deterministic semantic models. Be a part of the neurosymbolism. Fight brainrot.
We might not write the words that make the whole world sing, but we build the semantic models that feed the hungry ouroboros.
Themes and Trends from Taxonomy Boot Camp London
“And history was the reason why she would never go to London. She saw it as dominated by the Bloody Tower, Fleet Street full of demon barbers, as well as dangerous escalators everywhere.” – Anthony Burgess, Inside Mr. Enderby
After a six year hiatus initiated by the Great Plague of 2020 and continued with a lost conference venue, Taxonomy Boot Camp London returned in 2026 at the America Square Conference Center. The event was co-located with the KMWorld Europe conference, allowing attendees to come together for keynote sessions while attending one of two tracks for each conference. Just blocks away from the Tower of London and literally encompassing part of the ancient walls of the City of Londinium, the steadfast and ancient hosted audience to the rapidly changing world of knowledge organization systems and artificial intelligence.
As with past conferences, I’m going to sum up some of the key themes and trends of Taxonomy Boot Camp as I heard them.
Working with AI
A prevalent theme across all sessions was working with AI rather than against it. While there is common concern that AI will replace jobs and remove the human from the work equation, most sessions focused on using AI as a tool to accomplish tasks that are repetitive, time-consuming, or inconsistent. In particular, using AI to identify and extract entities for taxonomy building or inclusion, summarizing large quantities of text, or automatically classifying content using taxonomy values.
Another key focus was on the probabilistic notion of machine learning models:
At a high level, probabilistic AI models uncertainty and provides outcomes based on likelihoods. This means that it doesn’t always offer one definitive answer but instead provides a range of possibilities with associated probabilities. Deterministic AI, on the other hand, is rule-based, designed to yield specific, predictable outcomes without room for variability once given a particular input. (Decision Point Advisors)
Machine learning models may generate different answers to the same question, often termed a “hallucination”. Grounding machine learning models in knowledge bases, including deterministic models like graphs in the form of taxonomies and ontologies, can create a neuro-symbolic AI approach providing more consistent answers.
Content and Training
Working with AI also means curating the training data made available to machine learning models. Publicly available large language models (LLMs) are trained on easily accessible large data sets…specifically, content available on the Internet. As we well know, content quality on the Internet is as varied as the people who create it and make it available. While LLMs then get the benefit of a variety of input, they also suffer from the biases inherent in that input. Using synthetic training data or retrieval-augmented generation (RAG) to supplement the pre-existing LLM training data can improve results. In particular, using organization-specific knowledge bases of training data can help provide more specific responses applicable to your domain with fewer erroneous answers.
Deciding what training data to use and how taxonomy and ontology structures become part of that training data is partially in the purview of taxonomists, so becoming familiar with which LLMs are being used and for which use cases will be important parts of a taxonomist’s changing role.
Language Consistency
While this is nothing new, many sessions focused on keeping the “controlled” in controlled vocabularies. Since nearly every session linked back to AI in one way or another, even the context within which we consider the basic tenant of control in a controlled vocabulary was emphasized as continually pertinent. With differences in language across one or more semantic models, machine learning outcomes are put at risk. As more areas of application for machine learning are being found, we are also venturing into areas involving more risk, like mental health or medical advice, legal enforcement, and financial decisions. Language consistency used in semantic models and as applied as metadata for training content is now more important than ever.
Among the use cases pertinent to this consistency is tagging documents at more granular levels, including inline tagging and tagging content chunks. Again, this is nothing new and has been a practice of DITA for 20 years. However, being able to consistently and accurately create training data to balance probabilistic large language models with deterministic knowledge can counter machine learning hallucinations and create more trustworthy AI agents.
“Bond”ing
And, although not a taxonomy or knowledge management theme, I did notice another commonality across at least two presentations: James Bond as an example. Perhaps it was the London venue that caused presenters to use our favorite secret agent as an example, but there he was, connected semantically to movies, sports cars, and identification numbers. I myself created a simple Bondtology for illustration purposes in past workshops and webinars. Interesting that a profession like spycraft met at the intersection of establishing deterministic truth through semantic models to avoid being deceived by artificial intelligence.
The AI-Taxonomy Disconnect
“Me, I disconnect from you.” – Gary Numan, Me! I Disconnect from You
I have seen a significant decline in the number of taxonomist positions available. Wondering if I was in a myopic bubble based on my location or choice of job boards, I’ve been asking colleagues if they see the same thing. The general consensus is, yes, there are fewer taxonomist jobs available even as the importance of foundational data structures are becoming more important with AI tools. I have seen an increase, or at least a steady availability of, ontologist jobs, many requiring more technical expertise (the ability to create Python data ingestion pipelines and retrieval augmented generation (RAG) systems) than I have seen in the past. Again, this may be a matter of where I live or where I am seeking jobs, but it seems jobs in the taxonomy and ontology field are becoming more technical and less focused on the business operations side of working with stakeholders to provide analysis and guidance on semantic frameworks.
I suspect this shift is driven by a wider adoption of AI tools and a contraction in the job market. Employers seem to be seeking a single resource to do both the business and technical sides of implementing semantic structures and the foundational components for building out applications. I wonder if another factor is also the belief that large language models (LLMs) are a replacement for semantic models. If true, there are several reasons I can see that may be driving these beliefs.
Speed to Business
As I have touched upon before in my blog (Friction and Complexity and The Taxonomy Tortoise and the ML Hare), the manual, slower, but deliberate curation of controlled vocabularies can be seen as a roadblock to business speed and agility. There are several valid points here.
One area of pushback I see frequently in organizations is the response time from initially requesting a new concept, taxonomy branch, or vocabulary until it is available in production. The ownership of the concepts and data is, to some degree, taken out of the hands of the business users and put under the control of taxonomists who incorporate these concepts into centralized semantic models (taxonomies, thesauri, ontologies). Even with governance models including service level agreements stating turnaround times and taxonomy availability, not every group in the organization is going to see this centralized service as a benefit. Rather than wait for enterprise-level support, stakeholders may develop workarounds in services and tools to support their own use cases. As we know, this decentralizing of schemas and tools creates a fragmented landscape of differing terminology, functional support in what tools have available to manage taxonomies, and processes in the way data and content is handled and tagged with metadata.
Similarly, enterprise taxonomists serving many areas of the organization may seem to be too domain agnostic to serve the variety of use cases served by controlled vocabularies. While taxonomists do not need to be a domain expert in the areas covered by semantic models, the perception may be that their time spent ramping up in a domain would be better served having the subject matter experts in those domains build their own models. There is some validity here if the domains are truly standalone and operate to serve only those domain use cases. However, tying together various domain areas to form enterprise-wide knowledge graphs seems to be the direction most organizations want to go. If that’s the case, then a centralized taxonomy team as a service to the entire enterprise makes a lot of sense.
Given these counterpoints to slowly developing semantic models, why not, one may ask, simply ask machine learning models to provide the schemes we need to organize and optimize information?
Words Are Words Are Worlds
Are LLMs seen as a replacement, or at least a viable alternative, to taxonomies (using the term as a broad umbrella for all controlled vocabularies?
A large language model (LLM) is a language model trained with self-supervised machine learning on a vast amount of text, designed for natural language processing tasks, especially language generation. The largest and most capable LLMs are generative pre-trained transformers (GPTs) and provide the core capabilities of modern chatbots. LLMs can be fine-tuned for specific tasks or guided by prompt engineering. These models acquire predictive power regarding syntax, semantics, and ontologiesinherent in human language corpora, but they also inherit inaccuracies and biases present in the data they are trained on (Wikipedia).
Certainly building your own language model reflecting all of the ways that natural language queries could be asked is non-starter when there are clearly highly performing, public LLMs available at our fingertips. The chatbots ChatGPT, Google Gemini, and Microsoft Copilot are some of the more familiar tools based on foundational LLMs. One of the primary, and arguably most successful, use cases for these tools is language generation. When prompted, a chatbot can generate text, format this text based on instructions or examples, and produce a slick product which, short of a quick review to ensure the content checks out, is ready to go. These LLMs are based on a “vast amount of text”; in short, a lot more text than you will be able to provide to train a whole model.
What LLMs are missing, however, is your context. There are many, fairly easy, methods for providing them context. You can allow them access to your documents, at work or at home, so the chatbot can “see” the content of your documents, the structures, and even the writing style. That is very much your context, allowing the chatbot to compose in a way that reflects your home or work artifacts and generate new text combining the existing LLM with your specific context. Using additional context moves an LLM from generic to specific, from a wider world of words to your world of words.
At enterprise scale, the same method applies. If your organization is large, has a lot of public exposure and presence, and has a history of public interaction on the Internet and in social media, LLMs are going to know a lot about you even without additional context. I think, however, for the same reason that the market for reusable, one-size-fits-all taxonomies has never taken off is that every organization—whether true or not—is unique and special. Outside of certain industries like healthcare, life sciences, pharmaceuticals, and finance which have well-defined, and often extremely complex, ontologies they can adopt, other industries or functions within a company do not. In my experience, marketing is a great example. Despite the common needs, I have never seen a marketing department adopt any public standard. They build from scratch even when a majority of the taxonomies are commonly available terms.
In these cases, building taxonomies and ontologies to add context specific to your organization provides LLMs with both words and structures modeling the world of your domain. In fact, it is becoming more common that the development of these taxonomies is being done with human-chatbot interaction. A taxonomist can provide glossaries, metadata schemas output in spreadsheets, and documents to provide chatbots with the raw materials to extract entities, cluster topics, compare values across documents, and other processes that once required text analytics tools and, sometimes, human intervention in the form of rule writing. The speed to taxonomy and ontology development is increasing. Like other iterative feedback processes, the taxonomist and LLM work together to create domain schemas in the form of taxonomies and ontologies that provide additional guidance to future “manual” and automated processes with LLMs in the mix.
Pure speculation on my part, but is the niche and sometimes still esoteric and obscure field of taxonomy and ontology design being replaced, for better or for worse, with LLM use? Or, more specifically, is it being viewed as a replacement for taxonomy and ontology building and the experts who do it? As I stated above, I have seen a shift from the business of taxonomies to the technology of ontologies.
I Disconnect from You
In my opinion, the roles of a taxonomist and of a technically skilled ontologist are still separate. While many people in the industry have the skills to do the work of both, the paths to the two roles have been different. Many taxonomists have library science degrees. They likely have technical skills, but are more focused on the information science aspect of taxonomy and ontology development, interfacing directly with the business and providing other services, such as research, business analysis, and support for use cases relying on semantic models. Ontologists are typically computer scientists who can code and develop the technical infrastructure for ontologies. Finding resources who can, or who like, to do both has not been common. This may be changing. Certainly the roles are asking for both, with, in my view, a leaning toward the technical.
Again, speculatively, is there a shift toward more technical resources in support of rising AI use in organizations? Is there a move to cut out the intermediary roles of taxonomists to let the business owners and technical implementers of taxonomies and ontologies interface more directly? If so, what do organizations lose in the process?
The Connect
If I were reading this blog, I would think the author was trying to sell you the value of taxonomists with more of the “soft” skills of research, business stakeholder interaction, and translation of business requirements into taxonomies and to the more technical resources who support their implementation and move to actionable production. After catching up on a few seasons of Landman and hearing in my head at some point in every episode, “Brought to you by [insert name of large oil company]”, maybe I’m a little sensitive to reading between the lines. You don’t have to. I am selling you on the value of taxonomists for all of the reasons I’ve listed above. If there is indeed a shift from the business skills of taxonomists to the technical skills of an ontologist instead of having the excellent skills of both, then your organization is missing out on a valuable resource who can work alongside AI technologies to bring the business requirements and practical domain building skills to bear.
In summary, believe there are two necessary components to bridge the seeming disconnect between AI and the foundational data quality governance needed to make AI operational:
- A taxonomist who can
- build taxonomies and ontologies to create domain-specific semantic models representing the business,
- provide business analysis and requirements for technical implementation, and
- be the human-in-the-loop working with AI tools to continue building out, expanding, and governing semantic models, and
- Technical engineers who can
- operationalize ontologies by building data pipelines to AI tools, and
- focus on the engineering aspects of sharing out and productionizing ontologies for use across the enterprise.
An enterprise with both of these roles isn’t creating unnecessary resource overhead or additional layers effectively slowing the path from automation to implementation; rather, the two roles work in harmony to clarify requirements and optimize the use of AI in a variety of applications meeting business use cases. Taxonomists and ontologists are bridges between the business and the technical implementation enabling business needs.
Semantic Data 2025 Themes and Trends
“Trust in me, just in me / Shut your eyes and trust in me” – The Jungle Book
I attended the Henry Steward Semantic Data Conference co-located with HS DAM in New York City a few weeks ago. As I’ve done with KMWorld in the past, I’m going to summarize some themes and trends I took away from both conferences, with an emphasis on Semantic Data.
Inevitable AI
The most common theme of both conferences, unsurprisingly, was artificial intelligence (AI) in all of its forms, applications, and impact. Broadly speaking, the key takeaway across all of the presentations and discussions was: this is happening. Whether it’s baked into digital asset management (DAM) systems (hint: it is), used wildly thrown at use cases until something sticks, or carefully governed with strict governance, guardrails to protect the organization, its people, and the people they serve, and measured to understand the effectiveness of different large language models (LLMs), AI is happening. So what do we, as digital asset and semantic data professionals, do about it? What is our role in the use of AI in the organization and in the public sphere? What are our responsibilities?
From the Semantic Data Conference, several themes emerged:
- Organizations are going to experiment with generative AI models to develop workable pipelines with humans in the loop;
- Context is key, and organizations can develop domain-specific and constrained semantic models to be used in conjunction with external LLMs;
- It’s incumbent upon all of us to develop valid, organizationally-specific and curated training data sets to provide machine learning models the context to output reasonable results.
Themes from the Digital Asset Management Conference included:
- AI can speed up the generation of assets and the automated application of metadata to those assets;
- Access to clean, curated metadata is critical, both from taxonomies and sources like data lakes;
- Metadata as a source of truth for embedded AI can lead to better analytics;
- Asset provenance is essential for usage and rights management, especially when AI is involved.
Metadata Is Critical
That’s it. That’s the story. Metadata is critical. It has been, and it will continue to be. But, maybe, organizations are more aware of the importance of metadata because of the lightning fast rise of AI. Metadata is critically important as applied to digital assets, and semantic metadata powers better asset connections, discovery, personalization, and analytics.
Core to the importance of metadata is the importance of trust. Metadata quality must be trusted. The data and content to which metadata is applied must be trusted. Quality, trusted data leads to quality, trusted content and training sets which can feed into AI pipelines. Similarly, legal and reputational risks can be mitigated by ensuring the quality of information and data, especially as applied as compliance and usage rights.
Since semantic models are a source of truth for quality metadata, developing taxonomies and ontologies over time can create more complexity as needed to support a variety of use cases. Complexity sounds like a negative, but the world is complex, and semantic models are meant to represent organizational domains, which are by necessity complex. Complex semantic models support a variety of use cases, even if they do take more conscientious planning, development, and governance. Within these complex models are fit-for-purpose structures addressing use cases.
As with AI processes, developing, managing, and governing metadata in all its forms involves humans in the loop. Even as the identification, extraction, and application of metadata improves with AI, humans need to be involved in the process to add, remove, and quality check automatically applied metadata. As pipeline processes improve, reaching a specified threshold of metadata accuracy may reduce the need for human intervention and review.
Context and Trust
If I had to boil the conference down to two keywords–or, maybe, if I could only apply two metadata tags to the conference–they would be context and trust. Data and content requires context and semantic models are one way to provide this context whether for use in machine learning pipelines or direct human interaction with content.
