Large Language Models (LLMs) are increasingly used to automate knowledge engineering tasks such as generating RDF mappings. While promising, their outputs often lack semantic precision, syntactic correctness, and contextual metadata. This paper investigates whether structured metadata aligned with the mapping lifecycle can improve the quality and reusability of LLM-generated mappings. We present a metadata model that covers key phases of the mapping process and integrate it into the MetaSEMAP tool to support context-aware prompting. Using real-world uplift scenarios, we compare RML outputs generated from unguided prompts with those informed by lifecycle metadata. Our initial findings show that guided prompts consistently produce syntactically valid, semantically rich, and FAIR-aligned mappings. These results highlight the potential of structured metadata to guide LLMs toward generating higher-quality and reusable semantic artifacts in knowledge graph construction.
Declarative mappings like RML1 are essential for transforming structured data into RDF for use in
knowledge graphs. However, creating such mappings is a technically demanding task, often requiring
expertise in both syntax and domain-specific ontologies [
This paper explores whether incorporating structured metadata into LLM prompts can improve the
quality of generated mappings. This approach aligns with recent work in context engineering, which
focuses on enriching prompts with structured and task-specific context to enhance the reliability and
relevance of LLM outputs [
The remainder of this paper is structured as follows: we review related work on declarative mappings, metadata standards, and the role of LLMs in semantic data generation. We then introduce our lifecyclebased metadata model and its integration into the MetaSEMAP prompting interface. Next, we present the experiment design and results comparing guided and unguided prompt outputs. Finally, we conclude the paper and outline future research directions.
Declarative mapping languages such as R2RML and its extension RML enable structured transformation
of heterogeneous data sources into RDF for use in knowledge graphs [
To enhance transparency and reusability, several metadata standards have been proposed.
Generalpurpose vocabularies like Dublin Core and DCAT provide terms for describing datasets and publishing
context. More specialized eforts, such as SSSOM [
Large Language Models (LLMs), such as GPT-3 and GPT-4, have demonstrated the ability to generate
RDF triples, SPARQL queries, and declarative mappings such as RML from natural language inputs [
To address these challenges, we propose a structured metadata-driven approach that guides LLMs using context aligned with the mapping lifecycle. By injecting metadata such as the mapping’s purpose, data source characteristics, design decisions, and provenance, we provide the model with domainspecific constraints and intentions that help shape the generated outputs. This approach improves the accuracy, interpretability, and reusability of the resulting mappings by grounding generation in explicit and verifiable context.
This aligns with emerging practices in context engineering [
To support high-quality and reusable semantic mappings, we proposed a lifecycle-based metadata model
that organizes metadata across five key phases: analysis, design, development, testing, and maintenance.
This model was first introduced in our earlier work [
The model includes fields for describing the mapping purpose, input sources, stakeholders, design
decisions, tooling, validation outcomes, and publishing context. Table 1 presents a summary of the key
metadata fields grouped by lifecycle phase. The full specification is publicly available on GitHub [
In this paper, we extend the model’s use beyond documentation by incorporating selected fields into LLM prompts. This approach evaluates whether structured metadata can guide LLMs to generate more accurate, standards-compliant, and reusable RML mappings.
To evaluate whether lifecycle-based metadata improves the quality of LLM-generated RML mappings, we designed an experiment comparing two prompting strategies: (1) unguided prompts with only a task description, and (2) guided prompts augmented with structured metadata. We used gpt-3.5-turbo via the OpenAI Python SDK v1.0 (chat.completions.create) to generate RML outputs. All prompts were submitted through a controlled interface (Figure 1) to ensure consistent model behavior and prompt formatting across both guided and unguided conditions.
We evaluate our approach using three real-world scenarios derived from publicly available datasets on Ireland’s open data portal2. Full dataset links and mapping files are available in our GitHub repository 3. In Scenario 1 (S1), we use the Counties, National Statutory Boundaries, 2019 dataset to convert a CSV file containing administrative boundary data for Irish counties into RDF. Although the dataset includes multiple attributes such as object IDs, area size, Irish names, and shape geometry, only a relevant subset was selected for the purpose of the experiment.
In Scenario 2 (S2), we use the G0421, Population per NUTS 3 Region dataset to convert a JSON file containing population data for Irish regions into RDF. Although the dataset includes various statistical ifelds such as region codes, units of measurement, and percentage values, only a relevant subset was selected for the purpose of the experiment. The mapping focused on region identifiers, population counts, and temporal attributes to construct RDF triples that reflect regional population distributions in a given year.
In Scenario 3 (S3), we use the CSO Electoral Divisions, National Statistical Boundaries, 2022 dataset to convert a CSV file containing administrative division data for electoral districts in Ireland into RDF. Although the dataset includes various attributes such as object IDs, Irish and English names, shape geometry, and codes from multiple authorities, only a relevant subset was selected for the purpose of the experiment. The mapping focused on modeling spatial hierarchy and containment relationships. Each electoral division was assigned a hierarchical URI based on its ED_ID and linked to its parent county and province using a custom property ex:containedIn.
To compare the efects of metadata guidance, each mapping scenario was submitted to gpt-3.5-turbo
using two types of prompts: unguided and guided. The unguided prompt included only a task description,
while the guided prompt was enriched with structured metadata drawn from our lifecycle-based model,
which defines 37 fields across five phases. We selected 17 fields most relevant to prompt-based mapping
generation, focusing on descriptors from the Analysis and Design phases (e.g., Purpose, Mapping Type,
Mapping Domain, Input Description, Final Design Decisions, Justification, and Quality Metrics). We also
included contextual publishing and output fields from the Maintenance phase such as Publisher Source,
Version Number, Version Date, Output Format, and Output Syntax, which clarify expected structure
and encourage standards compliance. These fields were chosen based on their impact on design-time
decisions and feedback from previous user studies [
Fields from the Development and Testing phases (e.g., Tools, Mapping Algorithm, Testing Result) were excluded as they relate to post-generation steps. Similarly, fields with limited influence on generative quality (e.g., Risks, Stakeholder Background) were omitted. The selected metadata was embedded in natural language form to simulate structured context engineering and guide the LLM in producing syntactically valid, semantically rich, and reusable RML mappings. Table 2 shows an example of both prompt types used in Scenario 1 (S1). The full prompt sets for Scenarios 2 and 3 are available in our GitHub repository3.
Each mapping was assessed based on three key dimensions. Correctness refers to whether the output is syntactically valid RML and can be parsed without errors. Structure awareness evaluates how well the mapping captures the structure of the input data, including the correct use of rml:logicalSource, rml:referenceFormulation, and iterators. Semantic quality considers whether the mapping uses meaningful classes, properties, and URIs aligned with domain semantics and whether it includes relevant metadata to support reuse. Across all three scenarios, guided prompts produced mappings that 3Project repository: https://github.com/sarah-alzahrani/LLM. were consistently more correct, structurally aware, and semantically richer than those from unguided prompts. In particular, all guided prompts correctly used rml:logicalSource, specifying input type and reference formulation (e.g., ql:CSV or ql:JSONPath), and included iterator when handling JSON inputs. In contrast, unguided prompts often defaulted to rr:logicalTable, which is valid R2RML syntax but insuficient for non-tabular or nested data. This distinction is important because logicalSource is the required mechanism in RML for handling diverse data sources, and its absence leads to incorrect or incomplete mappings, especially for formats like JSON. For example: • Scenario 1 (Counties – CSV): The unguided mapping used rr:logicalTable, ignoring CSVspecific reference formulation. The guided version used rml:logicalSource with ql:CSV, along with appropriate classes and coordinate properties. • Scenario 2 (Population – JSON): The unguided prompt omitted both the iterator and JSON path references, resulting in an unusable mapping. The guided prompt correctly used ql:JSONPath, specified the iterator, and aligned the schema with population data standards. • Scenario 3 (Electoral Divisions – CSV): The unguided output lacked hierarchical URI structure and containment logic. The guided version constructed meaningful URIs using region and division identifiers, and modeled geographic relationships with custom and standard vocabularies. These results highlight how metadata-enriched prompting helps LLMs handle diferent mapping complexities by injecting contextual knowledge into generation. Guided prompts also produced reusable outputs that included metadata blocks aligned with FAIR and provenance principles, which were entirely missing from unguided versions. Overall, our initial evaluation shows that context-aware prompting grounded in lifecycle metadata improves both the technical correctness and semantic value of LLM-generated mappings, supporting better reuse, validation, and documentation. While this early evaluation ofers qualitative insights, future work will expand the analysis with quantitative scoring, completeness metrics, and broader use case coverage to more systematically validate the approach.
This study investigated the efect of lifecycle-based metadata guidance on the quality of RML mappings generated by LLM. By comparing guided and unguided prompts, we observed that metadata-enriched prompting significantly improves the syntactic accuracy, semantic richness, and standards compliance of LLM-generated mappings. Guided prompts led to the consistent use of appropriate vocabularies, correct handling of input formats, and the inclusion of metadata blocks that enhance provenance and reuse. These initial findings support a context-aware approach that combines the generative flexibility of LLMs with structured metadata to improve the reliability, interpretability, and reusability of semantic outputs. In future work, we plan to extend this evaluation framework to additional mapping types, compare performance across diferent LLMs, as well as investigate whether prompting LLMs to generate declarative mappings (like RML) results in more reusable outputs than directly generating RDF triples. Our initial findings highlight the value of structured metadata not just for documentation but as a guide in prompt-based knowledge graph construction. As LLMs become embedded in semantic workflows, the community must consider not only what outputs they generate but also how those outputs are guided, contextualized, and made reusable for others.
The first author acknowledges financial support from Al Imam Mohammad Ibn Saud Islamic University and the Saudi government — represented by the Royal Embassy of Saudi Arabia — Cultural Bureau in Dublin. The second author is partially supported by the SFI ADAPT Research Centre (grant number 13/RC/2106_P2).
During the preparation of this work, the author used ChatGPT (GPT-4) and Grammarly for the purposes of grammar improvements. All AI-generated content was thoroughly reviewed, edited, and validated by the author, who takes full responsibility for the final manuscript and all its content.