As one of the most widely used data storage and exchange formats, tabular data can be challenging to be integrated, interpreted, and reused when they lacks accurate semantic annotations, particularly when data come from heterogeneous sources. However, the annotation process is often time-consuming and requires a deep understanding of the internal structure of the target ontology. Therefore, developing eficient and accurate semi-automatic or fully automatic annotation tools is very important. Most existing approaches often rely on textual similarity to match column headers to ontology terms, and fail to efectively leverage the rich relational semantics representation within the ontology. To address this issue, we propose a reranking approach that combines semantic similarity with ontology structure. Specifically, we first generate a set of candidate ontology terms based on semantic similarity. For each source table header and its candidate ontology terms, we construct subgraphs and train a lightweight Graph Neural Network (GNN) model on these graphs to learn structure-aware representations. These representations are then used to improve the ranking of candidate ontology terms. To validate our approach, we performe experiments on the OAEI dataset. The results demonstrate that our approach improves Hit@1 by 4% compared to a baseline model that only relies on lexical similarity. This result shows that learning on local subgraphs is a promising direction for ontology alignment and schema matching.
Interoperability and knowledge integration between heterogeneous data sources have always been key
challenges in the semantic web domain. A large amount of tabular data are often generated and stored in
separate databases across diferent infrastructures. The semantics of such data are always ambiguous and
non-standardized, which impedes the implementation of the FAIR principles [
Inspired by recent research in the application of Graph Neural Networks (GNNs) to knowledge graph completion and reranking tasks [11, 12, 13], we propose a lightweight reranking approach that integrates ontology structure into the matching process. We construct a subgraph for each source table header and its candidate ontology terms, and train a GNN model on these graphs. By passing and aggregating messages among nodes, the model evaluates both the semantic and structural similarity and generates the final ranking of candidate terms. The proposed approach significantly reduces computational cost and improves annotation accuracy. The main contributions of this paper are as follows: • We propose an approach for dynamically constructing context graphs for semantic annotation and reranking tasks, which improves the matching accuracy and enhances computational eficiency; • We perform our approach on several real-world datasets, the approach achieves significant performance gains over the baseline model and is able to generate higher quality semantic annotations.
The remainder of this paper is structured as follows: Section 2 shows related work on graphbased reranking techniques. Section 3 introduces the proposed methodology. Section 4 describes the experimental setup and results. Section 5 concludes the paper with future work.
Previous work [12] shows that graph-learning methods can be efectively used for reranking and Retrieval Augmented Generation (RAG) tasks. In graph-based reranking approaches, candidate documents are modeled as nodes, and candidate-candidate edges are constructed from semantic similarity and external knowledge. Then, the message passing or aggregation will be used for structured reasoning within the candidate set to generate more reliable candidates. The training methods of graph-based reranking models can be categorized into three types [14], they are point-wise [15, 16], pair-wise [17, 18], and list-wise [19]. Motivated by these works, we adapt this idea to Column Type Annotation (CTA) tasks. We construct subgraphs for each table header, where nodes of the subgraph are the top-K candidate ontology terms, and edges are derived from semantic similarity between candidates and structural relations in the target ontology (such as subClassOf, part_Of, has_quality). A graph-based reranking model then scores the nodes on this subgraph to get the final ranking.
In this section, we provide a brief description of our approach and its implementation details. As shown in Figure 1, the proposed approach can be divided into two stages. In the first stage, we use an SBERT to retrieve the top-K candidate ontology terms based on semantic similarity. In the second stage, we construct a local subgraph for each table header and its candidate ontology terms. These subgraphs are then used as input to a GNN, which learns structure-aware representations to rerank the candidate ontology terms. In the following, we first define the problem formally and then describe in detail how the subgraphs are constructed.
Our task can be defined as: Given a target ontology that contains a set of terms = {1, 2, . . . , } and an input table header ℎ, we first apply an SBERT bi-encoder (·) to obtain embeddings ℎ = (ℎ) and () = (), compute cosine similarities (ℎ, ) = ⟨ℎ, ()⟩/(‖ℎ‖ ‖()‖), and return a top- candidate list cand = [︀ (1 , 1 ), . . . , ( , )︀] sorted by similarity score . Then for each table header ℎ we construct a header-specific candidate subgraph ℎ = (, , ): the node set = {ℎ, 1 , . . . , } contains the candidates for that header ℎ, edges and weights (, ) capture pairwise relatedness, for example, semantic similarity or ontology relations. Each node has features = [︀ ( ); ︀] and embeddings ℎ. Then we define a reranking function rerank based on a pre-trained GNN model. This function takes the table header ℎ and the candidate list =
{1 , 2 , ..., } as input, and outputs final results as = {1 , 2 , ..., }. ontology terms with SBERT. Stage 2 uses GNN to rerank candidate ontology terms.
To enable the reranking process, we construct a subgraph for each source header ℎ and its candidate ontology terms = {1 , 2 , ..., }, which are shown in Algorithm 1. We represent ℎ and = {1 , 2 , ..., } as nodes in a graph. To connect the nodes, we add edges between the ℎ and each candidate . The edge weights are the semantic similarity score calculated from the first-stage retrieval. In order to include structural information of the target ontology , we search for the relations between candidate terms = {1 , 2 , ..., } in the target ontology (such as subClassOf, part_Of, has_quality). If the relation exists and (, , ) is true, we add an edge between and . Each edge contains two features: a similarity score and the binary value of whether it is a structural edge of the ontology. In addition, we add self-loop edges to all nodes with a fixed weight of 1. These self-loops help preserve their own node features during message propagation.
Algorithm 1: Graph Construction
Output: Graph = (, ) with node features and edge weights // Step 1: Graph Nodes
ontology terms. ← {ℎ
, 1, . . . , }, where ℎ is the source table header and {1, ..., } are the candidate // Step 2: Add Edges to Graph
return = (, ) with node features and edge weights
To learn structural representation for re-ranking candidate terms, we train a lightweight Graph Attention Model (GAT) based on GATv2 [20]. The GAT model consists of two GATv2 convolutional layers, followed by a linear classifier. In the training process, we use the RankNet loss [21]: = log 1 + −( − )︁) ︁( (1) For each graph, we sample all positive and negative candidate pairs and compute the average pairwise ranking loss. The goal is to rank the correct term as high as possible in the final reranking list . Baseline (SBERT only)
Rerank with MMR
Rerank with CE Rerank with GCN Rerank with GAT
We conduct experiments on the Bio-ML track 1 of the OAEI (Ontology Alignment Evaluation Initiative) benchmark in 2024, which focuses on ontology alignment tasks in the biomedical domain. The dataset2 used in our experiments consists of three parts: • Source Header: Each class label in the source ontology is treated as a source header to be annotated. We use the NCIT ontology as the source ontology in this experiment. • Target Ontology: The complete target ontology is used for candidate retrieval. We select DOID as the target ontology. • Ground Truth Dataset: The oficial reference alignment file with a unique correct match in the target ontology for each source header.
We use two standard ranking metrics to evaluate the performance of the model: Hit@K to evaluate topK accuracy and Mean Reciprocal Rank (MRR) to evaluate the overall quality of the reranked results [22]. We evaluate all methods on the same candidate set generated by the first-stage SBERT bi-encoder. The systems compared are as follows: SBERT-only, using the first-stage similarity score as the final score; A non-graph reranker based on Maximal Marginal Relevance (MMR) that post-processes the SBERT list to balance relevance and diversity; A lightweight Cross-Encoder (CE) that concatenates the table header with candidate terms and inputs them into a single transformer and rescoring relevance to generate the ifnal score; And two graph-based reranking models, Graph Convolutional Neural Network (GCN) and GAT that operate on the candidate subgraph.
Preliminary experimental results are shown in Table 1. The proposed GAT model achieves the best overall performance. It reaches a Hit@1 of 0.824, with an accuracy improvement of 4% over the SBERTonly baseline model (Hit@1 of 0.782), and GCN model (Hit@1 of 0.629). In addition, the GAT model also achieves the highest MRR score of 0.863. The results demonstrate the efectiveness of incorporating ontology structure into the reranking process and highlight the significant potential for enhancing schema matching tasks.
In this paper, we propose a graph-based reranking approach, which improves the performance of semantic annotation tasks. By constructing a local subgraph for each table header and its candidate ontology terms, our method efectively integrates lexical semantic similarities with structural knowledge. Experiments on the OAEI Bio-ML track dataset show that our approach results in a Hit@1 of 0.824 1https://krr-oxford.github.io/OAEI-Bio-ML/ 2https://zenodo.org/records/13119437 and a 4% improvement compared to the baseline model. These results provide a new perspective on performing eficient annotation solutions with reduced computational and cost demands.
For future work, we plan to enrich the representation of the constructed graphs by adding additional node and edge features beyond simple relations. Furthermore, we aim to extend the model to support multiple ontologies, enabling it to better support annotation tasks in multi-domain scenarios.
The authors would like to thank the German Federal Government, the German State Governments, and the Joint Science Conference (GWK) for their funding and support as part of the NFDI4Energy consortium. The work was funded by the German Research Foundation (DFG) – 501865131 within the German National Research Data Infrastructure (NFDI, www.nfdi.de).
This work is supported by the Helmholtz Association Initiative and Networking Fund on the HAICORE@KIT partition and the Helmholtz Metadata Collaboration (HMC).
During the preparation of this work, the author(s) used GPT-4o and Grammarly for: Grammar and spelling checks. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [11] J. Dong, B. Fatemi, B. Perozzi, L. F. Yang, A. Tsitsulin, Don’t forget to connect! improving rag with graph-based reranking, arXiv preprint arXiv:2405.18414 (2024). [12] M. S. Zaoad, N. Zawad, P. Ranade, R. Krogman, L. Khan, J. Holt, Graph-based re-ranking: Emerging techniques, limitations, and opportunities, arXiv preprint arXiv:2503.14802 (2025). [13] H. Zhu, D. Xu, Y. Huang, Z. Jin, W. Ding, J. Tong, G. Chong, Graph structure enhanced pre-training language model for knowledge graph completion, IEEE Transactions on Emerging Topics in Computational Intelligence 8 (2024) 2697–2708. [14] Z. Cao, T. Qin, T.-Y. Liu, M.-F. Tsai, H. Li, Learning to rank: From pairwise approach to listwise approach, volume 227, 2007, pp. 129–136. doi:10.1145/1273496.1273513. [15] K. Reed, H. {Tayyar Madabushi}, Faster bert-based re-ranking through candidate passage extraction, in: The Twenty-Ninth Text REtrieval Conference (TREC 2020), 2020, pp. 1–5. [16] A. G. D. Francesco, C. Giannetti, N. Tonellotto, F. Silvestri, Graph neural re-ranking via corpus graph, 2024. URL: https://arxiv.org/abs/2406.11720. arXiv:2406.11720. [17] J. Luo, X. Chen, B. He, L. Sun, Prp-graph: Pairwise ranking prompting to llms with graph aggregation for efective text re-ranking, 2024, pp. 5766–5776. doi: 10.18653/v1/2024.acl-long.313. [18] L. Gienapp, M. Fröbe, M. Hagen, M. Potthast, Sparse pairwise re-ranking with pre-trained transformers, in: Proceedings of the 2022 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR ’22, ACM, 2022, p. 72–80. URL: http://dx.doi.org/10.1145/3539813. 3545140. doi:10.1145/3539813.3545140. [19] M. Rathee, S. MacAvaney, A. Anand, Guiding retrieval using llm-based listwise rankers, 2025. URL: https://arxiv.org/abs/2501.09186. arXiv:2501.09186. [20] S. Brody, U. Alon, E. Yahav, How attentive are graph attention networks?, arXiv preprint arXiv:2105.14491 (2021). [21] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender, Learning to rank using gradient descent, in: Proceedings of the 22nd international conference on Machine learning, 2005, pp. 89–96. [22] Y.-M. Tamm, R. Damdinov, A. Vasilev, Quality metrics in recommender systems: Do we calculate metrics consistently?, in: Proceedings of the 15th ACM conference on recommender systems, 2021, pp. 708–713.