Purpose: Existing relaxation works either focus on ontology-based or entity-based elements, leaving a gap in handling queries that combine both. The purpose of this paper is to fill this gap and relax real-world SPARQL queries combining ontology-based and entity-based techniques to retrieve top-k relevant results. Methodology: We propose a hybrid query relaxation solution with a ranking system based on information content, integrating ontology-based and entity-based strategies by extending similarity measures beyond classes to entities and literals. Findings: The model retrieves relevant top-k results and better aligns with user expectations across diverse query types. It effectively ranks relaxed queries involving ontology-based and instance-based relaxations. Value: This work introduces a unified relaxation framework and an evaluation benchmark grounded in human-judgment-based relevance. It advances the literature by providing a relaxation system that aligns closely with user-validated relevance, especially for queries involving annotations.
Purpose: The purpose of this paper is to determine whether a domain-based bias exists within entity linking techniques and how to potentially mitigate it. Methodology: To study how annotation results of entity linkers are domain-dependent, we annotate 6 widely used benchmarks from varying domains with 12 entity linkers and identify 37 distinct topics. We adopt a predictive approach that leverages topic modelling to assess system performance across domains determine whether domain information by itself is a sufficient signal to determine an optimal system. Findings: By analysing the interaction between domains and annotation behaviour, we find that no single system consistently outperforms across all domains. Instead, performance varies significantly depending on domain, suggesting that optimal results require domain-specific system selection. Value: To detect and mitigate domain effects, we provide our approach, tools, and benchmarks for enhanced domain-balanced (DOMiNO) and domain-aware (DOMeX) evaluation. Our findings underline the critical role of domain awareness in the development and deployment of text-processing systems, providing a pathway for more adaptable and robust methodologies. We release and open source all generated data, code and findings for future research endeavours on our repository at https://github.com/kmdn/linking_bias/.
Purpose: Ontology Matching (OM) has been studied for decades, yet fully automatic solutions remain elusive because ontologies differ in structure, granularity and vocabulary. Nevertheless, the abundant textual content attached to ontology entities suggests that the task could benefit from modern language-representation models. We therefore present the Semantically-Informed Similarity Matching Algorithm (SISMA), a novel system that matches concepts by leveraging the similarity of SBERT embeddings computed over pseudo-sentences extracted from the ontologies. Methodology: We focus on the task of class and property equivalence. We represent each ontology concept as a set of SBERT embeddings associated with each predicate. For every pair, a similarity matrix is computed and reduced to a score via linear operations with two learnable matrices. These are trained on a dedicated dataset. We evaluated our system on the OAEI benchmark alignments, training on the Conference track and testing on the Circular Economy (CE) and Material Sciences and Engineering (MSE) tracks. Findings: Our experiments reveal that the SISMA method achieves performance comparable to the state of the art. On the CE track our system achieves a higher F1-score than the participating systems, while on the MSE track it performs slightly lower. We also compared our results with a baseline across the parameter space, confirming that the training step is key to overall performance. Value: We have designed, implemented, and evaluated a novel system for ontology matching that achieves performance comparable to state-of-the-art methods. Our approach is readily extensible-primarily by training and testing on additional datasets-and the underlying idea can be realized in alternative ways, for example by replacing the current linear-operator scoring and threshold-filtering approach with a classifier that operates directly on the similarity matrix space.
Purpose: Existing approaches for generating Knowledge Graphs (KGs) from file-based, non-RDF data using declarative mappings are either limited by language-specific engines or lack optimization with language-independent relational algebra backends, resulting in suboptimal performance. This research proposes an integrated framework that tightly couples logical and physical plan optimizations, enabling high-performance, language-agnostic RDF graph construction. Methodology: We formalize the KG construction process using relational algebra with a dedicated RDF term generation function within the projection operator, resulting in one of two canonicalized logical plans, one with a join and one without. We then introduce tightly coupled physical operators used to define concrete execution pipelines. We propose and evaluate two optimizations, logical-level constant-folding to reduce redundant computations and a physical-level heuristic scheduling strategy to optimize concurrent execution. We implemented the optimizations in a new backend engine called konverter and benchmarked the engine with an RML frontend against two comparable engines, Morph-KGC and FlexRML. Findings: Empirical results show that constant-folding improves performance by approximately 7.4% and heuristic scheduling by approximately 14.7% compared to a worst case scenario, with minimal additional memory overhead. Overall, konverter outperforms the current state-of-the-art in performance FlexRML, reducing execution time by 61.5% and peak memory usage by 25.1%, though it currently only supports CSV files, a limitation we aim to address in future work. Value: The proposed framework and optimizations provide a formal and practically validated approach to optimizing the execution of declarative mappings for KG construction. The konverter engine demonstrates the potential for building high-performance, language-agnostic engines for enterprise KG construction.
Purpose: The heterogeneity of scholarly information in knowledge graphs (KGs) and unstructured textual sources poses challenges in building robust Scholarly Question Answering (SQA) systems. Existing datasets and models typically address a narrow spectrum, focusing exclusively on KGs or unstructured sources and limiting evaluation to simple factoid questions. This gap leaves current systems unable to answer complex, hybrid scholarly questions that require integrating evidence from multiple heterogeneous data sources. Methodology: We introduce HySQA (Hybrid Scholarly Question Answering), a large-scale benchmarking dataset containing hybrid questions over scholarly KGs and Wikipedia text. HySQA contains complex questions that need to traverse facts across structured and unstructured sources. We also develop a baseline model that adaptively decomposes each question into sub-questions, identifies their answer sources, retrieves relevant information from SKGs and Wikipedia, and generates an answer using a hybrid augmented answer generation framework. Findings: The experimental results show that integrating static and adaptive decomposition methods is more effective than static decomposition alone. Value: Introducing HySQA provides the community with resources for evaluating the advancements in scholarly QA research.
