On the innovation of capability‑centered matching implemented with TalentSeeker’s ontology‑based skills graph and the road ahead
HR now requires “understanding,” not “search”
Data is abundant, but finding meaning in it is hard. Today’s hiring landscape faces the same problem. The resumes we review when searching for candidates are no exception: they’re written in different formats and languages, and expressions vary widely—“Data Scientist,” “AI Engineer,” “ML Enginner.” As a result, a great candidate may not be retrieved because of a single word difference, or, conversely, unrelated candidates may surface at the top.
These issues go beyond a simple technical problem and translate into real pain for customers trying to find talent. Through customers’ data and interviews, we repeatedly observed these problems arise. We grouped them into three categories and began exploring how to solve them.
- Search fatigue — Even after tweaking filters endlessly, results look similar and the process takes far too long.
- Language barriers and differing job expressions — Multilingual resumes and varying role labels prevent suitable recommendations.
- Lack of trust in AI — It’s hard to trust the results because the reasons for recommendations and the reasoning paths aren’t clear.
Throughout our deliberations, we found that, for practical HR use, it is still difficult for AI like LLMs to consider and recommend across the entire data schema. While individual headhunters solve this with specialized taxonomies built on their know‑how to recommend the right people, LLMs lack that domain knowledge.
To address this, TalentSeeker developed an HR‑specialized knowledge structure called “Talent Ontology & Skills Graph,” enabling people with varying levels of understanding of industries and roles to find their way.
In other words, we applied the very points headhunters use to recommend candidates to the HR ontology, building an HR‑specialized knowledge structure. The aim of this approach is not merely to improve search accuracy, but to let customers at various levels use AI to encounter a “structural engine” that understands and infers relationships among capabilities, so they can identify the right talent.
How, then, should we build such an HR ontology and recommend the right candidates?
The start of a structural approach rooted in customer problems
The biggest problem TalentSeeker faced was that every resume has a different structure and expressions vary widely. Even when “Business Intelligence” and “Data Analytics” mean the same thing, the system handled them as different roles. An analysis of the database showed that about 60% of candidate information existed in non‑standardized forms. Structural mismatches ultimately led to “missing great talent.”
On top of that came language barriers. Even for the same role, expressions differ by country and language, so multilingual resumes often didn’t match even after simple translation. Conventional translation systems lost meaning, and global customers pointed out that “AI reads the language, but doesn’t understand the meaning.”
Ultimately, the largest problem was that most HR systems and even AIs like LLMs are still agents that don’t understand the HR context. “Automation Specialist” and “Robotics Engineer” are technically similar roles, but they were not connected simply because the words differ. With the “walls of language,” “data,” and “context” all operating at once, companies could not find the truly suitable people in the ocean of talent data.
As we added up these considerations, we identified one overarching issue: the absence of relationships. So TalentSeeker decided to assign relationships to talent data. In other words, we connected all roles, skills, and industries in a graph so the AI could reason over it.
Accordingly, we focused on the relationships that real headhunters use, selected critical junctions as points, defined them as cores within the relationships, and set out to build an HR ontology DB that embeds both the HR domain and headhunters’ know‑how.
Figure 1. Designing fragmented resumes into a relationship‑centric HR Ontology DB
The technical solution — combining Ontology, Graph‑RAG, and LLMs
TalentSeeker’s technology rests on two pillars: first, the HR Ontology Graph; second, an Onto‑LLM powered by Graph‑RAG.
First, the HR Ontology Graph hierarchically defines industries, occupation groups, roles, specialties, hard skills, and soft skills based on relationships. For example, the structure “IT industry → data role group → machine learning engineer → Python, TensorFlow, Data Pipeline” is recognized by the AI as is.
This graph combines Korea’s occupational classification system, the U.S. O*NET, and domestic and international official job taxonomies such as IOL, minimizing structural mismatches across countries. It also integrates the schema of our own data system based on headhunters’ know‑how. In short, we defined the graph structure around the waypoints people actually use to find talent in HR, converted our data into the HR Ontology Graph schema, and developed paths that allow the reasoning process to locate its grounds clearly.
Figure 2. The HR ontology construction process and relationships among key elements in the graph DB
The second pillar is Graph‑RAG (Graph‑based Retrieval‑Augmented Generation). Rather than simple text search, this technology generates answers while securing logical grounds by traversing nodes in the graph. For example, when asked “Is this candidate suitable as a data scientist?”, the system follows a path like the one below.
① Find relevant nodes in the ontology graph,
② Extract the required skills and experience,
③ Compare them with the candidate’s career trajectory, and
④ Compute the final suitability score.
This entire process leverages a Graph‑of‑Thought (GoT) pipeline embodying our know‑how and framework, enabling recommendations of suitable candidates based on fitness as defined by headhunters’ expertise. Such structural reasoning greatly reduces the “hallucinations” seen in conventional LLMs. Because the model follows only valid paths on the graph, the generated results are logically consistent and aligned with domain knowledge (Bran et al., 2024; Feng et al., 2025).
Figure 3. How an ontology‑structured HR LLM produces results via GoT
Data evolves into an “understandable structure”
Since introducing the ontology‑based AI system, TalentSeeker’s matching engine has been able to get much closer to customers. Previously, the time spent repeatedly preprocessing the same data dropped by more than 45%, and matching accuracy in multilingual environments improved by 28%. Above all, we can now answer the question, “Why did the AI recommend this candidate?” clearly.
Rather than merely returning probabilistic similarity, the model provides persuasive grounds along with a reasoning path—for example, “This candidate is deemed suitable via the Data Engineer → Machine Learning → Python path.” This reduces customer distrust and accelerates HR decision‑making.
Figure 4. An example result from TalentGPT that presents evidence and improvement points based on a reasoning path
On AI that understands relationships — and our road ahead
The biggest lesson from this project is that the trustworthiness of AI stems from the structure of the data. More important than linguistic translation is semantic alignment. Ontology is the frame that defines and preserves that meaning as relationships.
No matter how powerful an LLM is, if its reasoning is unconstrained free generation without grounds, enterprises cannot trust it. Conversely, an LLM that thinks atop a relationship‑based ontology performs controllable reasoning—essential in HR, where trust and accuracy are required (Wei et al., 2023; Edge et al., 2024). When AI presents not only the results but also the reasons, hiring managers can accept AI as a partner. This is why TalentSeeker’s customers say, “Now it feels like talking to an expert, not to an AI.”
TalentSeeker’s relationship‑based ontology LLM engine represents a new turning point in HR technology. By combining the rigor of knowledge graphs with the flexible reasoning of LLMs, it implements a talent‑matching engine that is structural, explainable, and scalable. AI now explains not only “who is a fit,” but also “why they are a fit” and “on what grounds that judgment is made”—all in a relationship‑based manner.
This shift is not just technological progress; it is, we believe, the right direction for a paradigm change that fundamentally resolves customer pain points such as search fatigue, language barriers, and distrust of AI. The AI we have built at TalentSeeker now thinks like an HR expert, reasons structurally, and explains transparently. This is the direction the HR industry should move toward in the next generation—and a new standard that TalentSeeker has demonstrated through technology.
Hiring managers no longer ask, “Can we trust it just because AI recommended it?” Instead, they ask, “Why did this AI make this judgment?” and discuss the suitability of the answer or the limitations in the data. Of course, limitations remain. We must continue to advance toward fully answering contextual queries—for cases where something does not exist in the data, or where the system can only answer indirectly based on the data.
References
Bran, A. M., et al. (2024). Ontology‑Retrieval Augmented Generation for Scientific Discovery. ICLR 2025 submission. [EPFL]
Edge, D., et al. (2024). GraphRAG: New Tool for Complex Data Discovery. Microsoft Research Blog.
Wei, J., et al. (2023). Chain‑of‑Thought Prompting Elicits Reasoning in Large Language Models. Google Research.
Feng, H., et al. (2025). OntologyRAG: Better and Faster Biomedical Code Mapping. Cambridge University.
Abolhasani, M. S., & Pan, R. (2024). Leveraging LLM for Automated Ontology Extraction and Knowledge Graph Generation. Arizona State University.
Wasi, A. T. (2024). HRGraph: Leveraging LLMs for HR Data Knowledge Graphs. ACL KaLLM Workshop.
Buehler, M. J. (2025). Agentic Deep Graph Reasoning Yields Self‑Organizing Knowledge Networks. MIT.
