Public administrations around the world produce large volumes of documents in many areas, including Human Resource Management (HRM). These documents include valuable data about people, positions, events, processes, and locations. However, the unstructured format of the documents makes it dificult to extract necessary information and limits their efective use by government agencies. The aim of the paper is to develop a semanticdriven HRM data model for education to enable a model-driven zero-shot Large Language Model (LLM) prompting approach for information extraction. Using various LLMs (e.g., Gemini, Llama, Claude4, ChatGPT), the data model is applied to administrative documents issued by the Greek Ministry of Education and published in DIAVGEIA.gov.gr, the national open government portal of Greece. Although at an exploratory stage, the study presents promising results regarding the capability of modeldriven LLMs in knowledge engineering, especially in text-intensive domains such as public administration.
Public administrations around the world generate vast amounts of documents (e.g., laws, administrative decisions) across various functional areas (e.g., HRM). These documents include valuable data about people (e.g., education personnel), things (e.g., job positions), events (e.g., position assignments), processes (e.g., employment applications), and locations (e.g., school facilities). However, because this data is frequently provided unstructured, extracting the necessary information can be challenging. Therefore, public administrations are unable to properly utilize the potential of this important data.
The literature has introduced Generative AI, particularly LLMs, as efective tools for information
extraction tasks [
Although ontologies and KGs are frequently employed in education for pedagogical reasons (such as
modeling learning domains), little is known about how they may be applied to education administration
tasks, especially when managing educational staf [
Given the limited number of modeling initiatives in the HR domain of education, this study aims to develop a semantic-driven data model that captures the key elements of “Employment” and “Position Assignment” of education personnel suited to the Greek education system. This model will form the basis for applying a model-driven zero-shot LLM prompting approach to administrative documents issued by the Greek Ministry of Education agencies and published in DIAVGEIA.gov.gr, the national open government portal of Greece. This study is guided by the following research questions: 1. What are the fundamental building blocks of data that describe the employment and the position assignments of education personnel? 2. Can these building blocks be aligned with generic HRM data models, making the resulting HRM data model applicable to educational systems worldwide? 3. How efectively can Greek-capable LLMs (e.g., ChatGPT) structure unstructured data from Greek language administrative documents using the proposed HRM data model and a model-driven zero-shot LLM prompting approach?
The rest of the paper is structured as follows: Section 2 develops the background and related work of the study. Section 3 analyses the scope, goal, requirements, and procedure of developing the data model. Section 4 presents the experimental evaluation results and analysis. Finally, section 5 presents conclusions, study limitations, and recommendations for future work.
The literature has proposed data models, that represent the key entities and relationships in the HRM
domain. In particular, Hay [
M. Jarrar et al. [
HRM-specific data models and ontologies have also been developed in education. The European
Commission has created controlled vocabularies (e.g., taxonomies, thesauri) for education-related
concepts, including teaching staf [
According to the literature, KGs can integrate large volumes of data in a meaningful manner. To
solve various issues in hiring, training, payroll, and HR systems, researchers in HRM have used KG
technologies. In particular, HR-focused KGs have been created to assist with tasks like job market
analysis [
According to the literature, there is a strong interplay between LLMs and models/ontologies/KGs [24],
especially in text-intensive domains such as the public sector. For example in [
In 2010, the Greek government established the national open government portal, "DIAVGEIA" (https://diavgeia.gov.gr/) (meaning "clarity"), to promote government accountability. Since then, all government institutions have been required to upload their acts and decisions (e.g., budgets, appointment decisions) in the platform to be accessible to the public. Each uploaded document is assigned a unique ID number to certify its authenticity and legality. As of now, 68.4 million documents have been uploaded to DIAVGEIA, with an ongoing rate of 16,000 decisions being published each working day. Specifically, in 2024, agencies from the Greek Ministry of Education uploaded 54,305 administrative decisions including HRM-related acts.
The study scopes the administrative acts that are published in DIAVGEIA and pertain to the employment of education personnel and their assignment to primary and secondary public education positions. It encompasses all professionals directly involved in the educational process (e.g., teachers, psychologists, social workers) through all types of employment (e.g., regular, substitute). Positions like school headmasters, which include administrative and instructional duties, are also included. The study aims to create a semantic-driven data model that identifies the fundamental building blocks of data (such as actors, events, and locations) related to educational staf’s employment and position assignments. The classes and their properties will give LLMs the metadata they need to organize the textual data from DIAVGEIA documents meaningfully. This way, the model will help with information extraction using a model-driven zero-shot LLM prompting strategy, resulting in better model-driven prompts and more organized results.
To ensure alignment with the study’s overall goals, we established a set of modeling requirements. The four fundamental principles outlined by [28] were used to engineer the modeling requirements systematically. First, the model should be accurate for the purpose at hand, thus guiding LLMs to automatically extract and structure textual information from DIAVGEIA (Req. 1 – Validity). Second, it should be aligned with generic HRM data models and, at the same time, represent the specifics of the Greek public education system. In particular, in Greece, teachers are classified according to their qualifications into specific branches (e.g., mathematics), levels (e.g., primary/secondary education), and education types (e.g., general/special education). The Ministry of Education centrally hires them and assigns them to particular educational regions within the country. Depending on local educational needs, the regional agencies (e.g., Directorates of Primary Education) assign them to one or more positions (Req. 2 – Credibility). Third, it should enable the transformation of the unstructured textual data into structured, queryable datasets (Req. 3 – Utility). Fourth, it should be implementable within time, resources, and data constraints, relying on the lightweight and scalable use of pre-trained language models, without domain-specific finetuning (Req. 4 – Feasibility).
Our modeling procedure is grounded in modeling hierarchy theory [29], which defines a four-layer meta-model architecture. The raw data at Level 0 serves as the foundation for the Level 1 model. At Level 2, the model is abstracted into a meta-model that organizes and classifies the Level 1 components. At the top, Level 3 comprises the meta²-model, which defines the foundational concepts used to develop the lower levels. Moving from Level 0 to Level 3 involves a process of classification, while moving downward follows a process of instantiation.
The modeling hierarchy schema provides a solid framework and the necessary mechanisms to align our modeling procedure with the previously outlined requirements. To this purpose, we adopted a combination of bottom-up and top-down steps (see Fig. 1). Rather than following a linear sequence, these two approaches are applied in a complementary manner. While the top-down approach helps us focus on high-level system needs, the bottom-up approach allows us to construct models from lower-level components.
More precisely, we created the model according to the recommendations made by [30]. After establishing the domain and scope, we chose generic HRM data models from the literature examined in Section 2. We also collected documents from DIAVGEIA and their underlying legal texts related to employment and assignments of education personnel. Our review, supported by our experience as domain experts, enabled a thorough understanding of the domain. After collecting key terms as the basis of our ontology, we selected terms with independent existence and organized them into a class hierarchy. We then analyzed the remaining terms to define class properties (e.g., value types, allowed values, cardinality). Throughout this process, we aimed to align our ontological building blocks with generic HRM data models. Finally, we created individual instances of each class.
Using a bottom-up approach to describe data from DIAVGEIA documents was considered suitable, as it can enrich LLMs with detailed metadata for organizing unstructured text (Req. 1). This also supports prompt optimization through the application of ontology-based prompts, which help LLMs extract key information and convert it into triples (Req. 3). This process can lead to a queryable KG supporting advanced analytics (Req. 3). Moreover, drawing from generic HR data models and refining them into domain-specific models minimizes bias and enhances overall model reliability (Req. 2). 3.4. Design and development
Using our methodology, we initially searched DIAVGEIA for documents issued by agencies of the Greek Ministry of Education, scoping the employment and assignment of educational personnel. Due to the inconsistent terminology used across government agencies [31], we also reviewed the overarching legal texts. Afterward, we created an initial list of terms such as "teacher", "special education", and "position". We then classified them according to the six Zachman interrogatives [32]. This process resulted in the development of an initial controlled vocabulary that encompasses key business elements (e.g., position), processes (e.g., applying for an appointment), locations (e.g., region of appointment), educational organizations (e.g., Directorate of Secondary Education), events (e.g., position assignment), and goals (e.g., meeting teaching needs).
We then identified terms representing entities with independent existence (e.g., "secondary education teacher") and terms that describe these entities (e.g., "substitute"). The former were classified as classes within the ontology, while the latter were considered as their attributes. Next, we organized the classes hierarchically. For instance, "Position type" was defined as a super-class of "Secondary education teacher”. In contrast, more specific concepts (e.g., "Secondary education mathematics teacher"), were classified as sub-classes of broader categories and positioned at the bottom. Finally, we attached slots to the most general classes and specified their characteristics. This process produced the core building blocks for an HRM data model representing the employment and position assignments of educational personnel (see Fig. 2).
Special care was taken to ensure that the developed building blocks are aligned with established
generic HRM data models proposed in the literature, such as those referenced in [
A person employed by an organization is assigned, for a specific period of time, to a particular
position (e.g., mathematics teacher). In the Greek education system, teaching positions are not
linked to the hiring organization (e.g., Ministry of Education) but usually to specific schools (e.g.,
1st High School of Athens). Therefore, the “Position” entity is modeled as distinct from
“Employment” and is directly linked to the “Organization” through a “defined by” relationship [33],
[
Since many positions share common characteristics (e.g., job title), the entity “Position Type” is
introduced [
While the concept of "space" is not traditionally regarded as a building block of HR models in the literature, it plays a significant role in national-level education staffing. The Ministry of Education centrally manages recruitment, but teachers are assigned to regional organizations (e.g., Directorates of Secondary Education), which are in charge of specific management areas and assign teachers to positions. The Ministry of Education establishes the boundaries of these areas, which may not align with the nation's recognized geopolitical boundaries. Schools are also classified according to their geographic location. As a result, factors such as a teacher’s years of service in remote geographic areas (e.g., islands) receive special consideration in administrative processes, including transfer eligibility.
Because of the significant role of geography in education personnel management, we introduced
the concept of “geographic location” into the HR model. Based on [
This section presents the results of experimental evaluations on using LLMs to extract structured information from DIAVGEIA documents through a model-driven approach. The experiment aimed to assess how effectively Greek-capable LLMs can structure unstructured government data. The experiment was conducted in seven phases. First, we developed an initial prompt template to be used for information extraction from documents describing teachers’ position assignments. The prompt is model-driven, meaning that the HRM data model structure is embedded in the prompt and requests results based on this specific format, to enforce semantic consistency in LLM outputs.
Second, we defined the test set of documents to apply the model-based prompt by searching DIAVGEIA for various types of position assignments (e.g., secondment) and employment relationships (e.g., regular, substitute) of teachers. Given that teachers can hold multiple positions simultaneously, we included a range of documents, including those detailing the assignment of several teachers to various roles. Given the variability in wording and terminology across public organizations, we selected documents from 14 public entities, comprising a total of 36 documents— 18 involved assigning each teacher to a single position, and 18 involved assigning each teacher to multiple positions. All documents were public records, containing only basic information such as names and fathers' names, and adhered to GDPR requirements, excluding sensitive personal data.
Some interesting remarks about the documents (see Fig. 3) include: i) most of the documents refer to multiple (some times more than 50) position assignments, ii) the documents usually begins with a preamble text that refers to all assignments, including info about things like the start and the end date, iii) after the preamble text follows a table with specific information for each assignments such as information about the person, management area, working hours, iv) in some cases the table lists more than one assignment for the same person (e.g., she/he may be assigned at two schools on a part-time basis). These structural characteristics make information extraction particularly challenging.
Third, we manually annotated the 36 documents based on the HRM data model. This manually generated information was regarded as the ground truth. Fourth, we applied the prompt format to LLMs that understand Greek, and we had access. Seven LLMs were employed: Claude 3.5, Claude 3.7, Claude 4, Llama (lama 3.3 30b parameters), Deepseek- R1, ChatGPT 4, and Gemini 2.5 Pro. The documents were first converted from PDF to text, and then the extracted table was converted to a dataframe to have more structured text. The LLMs produced output in triples (e.g., Person_1 has_first_name Chris), which were evaluated against the ground truth.
Fifth, we improved the prompt after identifying several extraction errors during the initial phase of the experiment across all LLMs. These errors included failing to identify the correct start and end dates and omitting multiple position assignments. The errors were due to the documents' complex structure, which contained multiple dates (such as the document issue date and the council opinion date), and to the domain-specific semantic complexity. To improve the accuracy of the returned information, we decided to enhance the prompt by adding a section with domainspecific logic and rules. This led to the development of the final prompt template used for evaluation (see Fig. 4).
Sixth, we applied the final version of the prompt to the LLMs and, seventh, we compared the returned AI-generated triples against the ground truth. The evaluation was performed semiautomatically due to the large number of triples per document. An automatic script first normalized the values (e.g., by removing symbols and accents) and then identified identical triples. The script calculated a similarity score for triples with value differences. If the score was above 85%, the triple was considered correct. After the script returned the correct triples, we doublechecked them manually to correct any remaining mistakes. Ground truth and AI-generated triplets that were the same were considered correct. Minor differences, such as special symbols (e.g., <>, ‘, ", etc.) or variations in wording/case of values (e.g., 'Western', 'western'), were not taken into account and were also considered correct.
Through the evaluation process, we identified: i) True Positives: Triples correctly identified by AI, ii) False Positives: Incorrectly added AI-triples, and iii) False Negatives: Missed triples by AI. Finally, we calculated Precision, Recall, and F1-score (check Table 1 and Fig. 5) using the following formulas: • • •
Precision = True Positives / (True Positives + False Positives) Recall = True Positives / (True Positives + False Negatives)
F-1 = 2*(Precision*Recall)/ (Precision + Recall)
According to the results, Gemini achieved the highest Precision (0.913), with Claude's versions hovering around 0.841-0.856. Deepseek and ChatGPT also perform well with scores of 0.814 and 0.774, respectively. In contrast, with 0.657, Llama has the lowest Precision. Regarding Recall, Gemini has the highest score (0.889), while Claude4, with 0.768, shows a remarkable improvement over previous versions, around 0.673-0.687. Deepseek's recall (0.699) is similar to ChatGPT's (0.694) and better than versions 3.7 and 3.5 of Claude. With 0.532, Llama has the lowest recall, which means it misses a significant number of cases. Overall, among the LLMs tested, Gemini 2.5 Pro and Claude4 performed the best, with Gemini 2.5 Pro achieving the top F1-score (0.898). Table 2 and 3 present separate scores for documents describing “one-to-one” and “one-to-many” person-position assignments respectively. In general, the “one-to-many” documents are more complex than the “one-to-one” and this is reflected in the scores. Except for Gemini 2.5 Pro, all LLMs struggled to extract information from “one-to-many” documents, achieving lower scores.
Regarding the information extraction errors, serious issues were identified in cases where a teacher holds two or more positions. While Gemini and Claude4 perform well, other LLMs often omit position assignment information, which is reflected in the Recall values which are lower than Precision. Except for Gemini, LLMs also face challenges in accurately identifying position assignments' dates. The semantic complexity (e.g., simultaneous presence of multiple dates) and the conceptual ambiguity (e.g., lack of explicit reference to the start and end dates) of the documents often led LLMs to incorrect results. Finally, issues such as ambiguities in table parsing (e.g., unclear cell boundaries and structure, and header misinterpretation) made it even more difficult for LLMs to identify information correctly. Among these factors, table parsing ambiguity (e.g., simultaneous presence of multiple values in the table header) often leads to challenges in accurately identifying the correct position assignment type, the kind of position, and the working hours per week.
The primary objective of this study was to structure unstructured government data in the field of Human Resource Management (HRM) using a model-driven zero-shot LLM prompting approach. To this end, the study developed a semantic-driven HRM data model tailored to the specifics of public education. This model captures the fundamental building blocks of data associated with the concepts of “Employment” and “Position Assignment”. Beyond key HR building blocks—such as employment, position, position type, and position assignment—the model systematically integrated the spatial dimension, often overlooked in HRM ontologies, to address the complexities of nationallevel education staffing. Although the model was specifically designed for the context of public education in Greece, it is closely aligned with generic HRM data models, allowing for adaptability in various educational systems worldwide. Furthermore, the model was embedded in a prompt template that ensures semantic consistency in the outputs generated by the LLM.
The study used real-world, high-volume government data from thirty-six (36) documents issued by fourteen (14) regional government agencies of the Greek Ministry of Education and published on DIAVGEIA.gov.gr. These documents scope the assignment of sixty-four (64) teachers to various positions. More specifically, they contain valuable information about teachers’ personal information (e.g., first and last name), position assignments (e.g., start and end date), position types (e.g., job title, employment relationship), schools that define the positions, education management areas, and competent authorities having jurisdiction over them. To assess how effectively Greekcapable LLMs could extract and structure this information, the study utilized seven competitive Greek-capable LLMs for evaluation. The assessment results are promising since most LLMs achieved good scores. However, in more complex cases, there is still room for improvement.
A limitation of the study is that it focuses on the employment and the position assignment of teachers, which are the most frequently categories in the HR-centric documents of the Greek Ministry of Education published on DIAVGEIA. As future work, we plan to expand the model by incorporating additional, less frequently published concepts. These concepts cover various HRMrelated processes, such as benefit payments and employee terminations. We also plan to use and evaluate other prompting approaches (few-shot, chain of thought, etc.) and prompt designs (e.g., provide the structure of the ontology in the prompt as RDF or JSON, Pydantic code). Lastly, we aim to leverage European open-access LLMs, such as the Mistral model, Aleph Alpha Luminous, and other models emerging from EU research initiatives.
Although still in its early stages, the research shows encouraging results regarding the interaction between LLMs, domain-specific models, and KGs. In particular, it demonstrates the effectiveness of a model-driven zero-shot LLM prompting approach in extracting structured information from a large amount of unstructured government data. This method can help public administration worldwide to unlock the full value of its dispersed, document-based data. Moreover, the demonstrated capability of model-driven LLMs in knowledge engineering suggests strong potential for broader impact, especially in text-intensive domains such as public administration.
During the preparation of this work, the authors used Grammarly and ChatGPT in order to: Paraphrase and reword, Improve writing style and Grammar and spelling check. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.