This paper presents the results of the DogMa Matcher in the OAEI 2025 competition. DogMa is a Large Language Model-based system for the Knowledge Graph Entity Matching (KGEM) task in the OAEI 2025 Knowledge Graph Track. The system formulates entity matching as a Retriever-LLM selection problem, where entities are represented as concise, graph-aware textual summaries. Each entity in the source Knowledge Graph is summarized using an instruction-tuned LLM that integrates both textual descriptions and local relational context. Candidate entities are then retrieved from the target graph via dense vector similarity, and a second LLM determines whether any candidate represents the same real-world entity. This combination of graph-informed summarization and selective LLM reasoning yields high precision and recall across heterogeneous Knowledge Graphs, outperforming embedding-based methods in our internal evaluations. DogMa demonstrates that context-engineered LLMs can efectively exploit semi-structured graph information for accurate and interpretable entity alignment.
in the initial candidate selection process, thereby enabling a more robust and comprehensive alignment pipeline.
Our proposed system follows a Retriever–Augmented Generation (RAG)-like design with three main stages: 1. Entity Representation (Dogtag) Generation: Each node in the input KGs is converted into a compact textual description called a Dogtag. These are produced by an LLM-based summarizer that compresses node-level textual content (labels, abstracts, types, infoboxes) and incorporates salient information from local graph neighbors like wiki mention links or other relations. 2. Candidate Retrieval: For each entity in the smaller KG (anchor entity), we retrieve the top- most similar entities from the larger KG using an SBert [4] embedder model (BAAI/bge-large-en-v1.5). 3. LLM Selection: The retrieved candidates are then passed to an instruction-tuned LLM (Llama-3.1-70B-Instruct [5]), which decides whether any candidate refers to the same real-world entity. The LLM can also abstain if no suitable match exists.
All operations are performed in inference mode only, with deterministic decoding (temperature = 0, sampling disabled). The system enforces one-to-one alignments through post-hoc deduplication.
The prompt used for the LLM-based selection step is shown below:
For the OAEI 2025 Knowledge Graph Track, the following adaptations were applied: • Input preprocessing: OWL files were parsed into node-level representations using RDFLib.
Entity attributes, labels, and neighbor relations were serialized into a textual context for summarization. • Summarization model: Dogtags were generated by prompting Llama-3.1-70B-Instruct with a zero-shot summarization template designed to include both textual and relational content. • Retriever configuration: Top- retrieval was set to = 10 using cosine similarity of bge-large-en-v1.5 embeddings. • Merging alignments: To improve recall while preserving precision, the final alignment was obtained by merging the high-confidence Exact match pairs (entity pairs with corresponding labels) with additional matches from the LLM-based method. A dynamic similarity threshold — set as the median similarity score of the exact matches — filtered secondary pairs, which were only added if they exceeded this threshold and did not violate 1:1 consistency.
DogMa can be downloaded from https://github.com/kiscsonti/DogMa-Ontology-Matcher alongside the alignment files.
Our system was evaluated on the Knowledge Graph Track KG pairs. We present our results in Table 1.
mcu-marvel Method P R F1 ExactMatch 88.39 66.22 75.72 DogMa 52.89 45.05 48.66 DogMa Merged 82.61 78.70 80.61 memoryalpha-memorybeta stexpanded-memoryalpha
P R F1 P R F1 94.49 88.23 91.25 96.59 90.89 93.66 95.46 86.90 90.98 95.22 91.74 93.44 92.93 93.64 93.28 94.83 94.94 94.89
swg-starwars
P R F1 93.67 62.00 74.61 93.44 80.11 86.26 92.06 85.82 88.83 swtor-starwars
P R F1 94.20 88.66 91.35 90.76 86.56 88.61 93.67 93.21 93.44
Strengths: DogMa demonstrates that high-quality textual summarization of graph nodes enables robust retrieval and matching. The LLM Selector efectively reasons over candidate sets and leverages implicit world knowledge to resolve dificult equivalences.
Weaknesses: Our approach is computationally expensive: each node in the smaller KG triggers one LLM call. This limits scalability for very large graphs.
Future improvements include: • Reducing cost via lightweight selectors or small LLMs distilled from the current system. • Incorporating global structural information (e.g., communities, influence paths) beyond local neighbors.
• Exploring prompt-order-invariant selection mechanisms to mitigate positional bias.
The OAEI framework provides a solid and reproducible evaluation setup. However, due to partial gold standards, true precision and recall are underestimated. We recommend augmenting gold alignments to include additional verified correspondences.
We introduced DogMa, an LLM-based system for Knowledge Graph Entity Matching in the OAEI 2025 track. DogMa formulates alignment as a Retriever–LLM selection task, where entities are represented through graph-aware textual summaries and matched via dense retrieval followed by LLM-based verification.
Results indicate that integrating local relational context into entity representations substantially improves matching quality, yielding high precision and recall across heterogeneous graphs. Despite its computational overhead, DogMa shows that context-engineered LLMs can efectively reason over semi-structured data, bridging symbolic and neural alignment methods.
This research has been supported by the European Union project RRF-2.3.1-21-2022-00004 within the framework of the Artificial Intelligence National Laboratory. Miklós Krész was supported by the Slovenian Research and Innovation Agency (ARIS) through research program P1-0404 and by the research program Cognicom (0013103) at the University of Primorska. He is also grateful for the support of ARIS grant J7-60122.
During the preparation of this work, the authors used ChatGPT for rephrasing, grammar and spell checking. After using these tools, the generated text was reviewed and edited and the authors take full responsibility for the publication’s content. Heidelberg, 2023, p. 561–578. URL: https://doi.org/10.1007/978-3-031-47240-4_30. doi:10.1007/ 978-3-031-47240-4_30. [4] N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Association for Computational Linguistics, Hong Kong, China, 2019, pp. 3982–3992. doi:10.18653/ v1/D19-1410. [5] A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024). doi:10.48550/arXiv.2407.21783. arXiv:2407.21783.