We present EmbER, a new neuro-symbolic method for embedding-based entity comparison and ranking in text-rich knowledge graphs (KGs) that have frequent textual fields. Traditional symbolic and homogeneous embedding approaches struggle with distinguishing the semantic nuances of diverse attribute data types in text-rich KGs. EmbER addresses this problem by partitioning KG attributes into categories - numerical values, short terms, and long textual descriptions - guided by the KG schema, and applying type-specific embedding strategies to each data type. The resulting embeddings are then combined into a unified entity representation for similarity computation and relevance ranking of entities. To evaluate EmbER, we construct a curated subset of DBpedia entities, focusing on Irish populated places, e.g. towns and counties. We use SMATCH, a graph-based similarity metric commonly used to evaluate semantic structures, along with human evaluation, to build a ground truth of entity similarity and ranking. Results show that EmbER outperforms a symbolic query-driven method and general KG embedding models, and is competitive with strong textual baselines. These results demonstrate the effectiveness of schema-guided data type-specific embeddings for entity representation, which offer the potential to enhance applications such as semantic search, entity linking and recommendation systems.
Property graphs and knowledge graphs are two dominant paradigms for graph-based data modeling, yet they remain structurally and semantically incompatible. Property graphs support richly attributed relationships but lack formal semantics and reasoning. Knowledge graphs, grounded in RDF, RDFS, and OWL, enable inference and interoperability but cannot natively represent edge metadata. This paper introduces a categorical framework that bridges this gap through a functorial transformation from typed, attributed property graphs to RDF-star knowledge graphs enriched with OWL/RDFS semantics. The transformation preserves edge attributes using quoted triples, consolidates repeated relationships, and ensures semantic fidelity through alignment functions. This formalization provides a foundation for integrating the structural expressiveness of property graphs with the logical rigor of knowledge graphs, paving the way for interoperable, ontology-aware graph systems that combine flexible data modeling with semantic reasoning.
Epistolary letter collections are stored in distributed local archives as letters are sent from one place to another. To find and study letters of a particular person or group on a global level, data from different local sources can be aggregated and harmonized into a global knowledge graph (KG). This paper argues that it is important to understand possible quality issues of the global KG that may arise due to the heterogeneity of the local datasets, aggregation process, and mutual linkedness of the local data. For example: In what ways do the local collections enrich each other? How complete is the aggregated dataset? Are there duplicates or misaligned entities and concepts in the aggregate? This paper presents a set of data-analytic tools to address such issues in order to support data literacy in Digital Humanities (DH) research. As a case study, the LetterSampo Finland Linked Open Data (LOD) KG is considered and the results are reported. It aggregates data about almost 1.3 million historical letters sent in the Grand Duchy of Finland (1809–1917) and 118 000 related actors harvested from 18 different archival data sources and 1670 fonds, enriched by data from 12 external databases.
Large Language Models (LLMs) demonstrate impressive capabilities in Question Answering tasks, yet previous research suggests current approaches fail to accurately reflect their real SPARQL query generation capabilities due to memorization effects caused by integrating benchmark datasets into the training data. These effects artificially inflate perceived performance, i.e., the quality of the showing by LLMs cannot actually be achieved on previously unseen queries. This paper presents DynBench – a novel approach to creating high-quality benchmarking datasets that address this challenge in the field of Knowledge Graph Question Answering (KGQA). We develop new datasets based on two datasets – QALD-9-Plus and LC-QuAD – by systematically replacing entities in SPARQL queries with alternatives retrieved from Wikidata and within the corresponding natural-language questions. Our findings confirm that the proposed approach successfully creates new benchmark datasets that can be used for evaluating KGQA systems. Therefore, our approach drastically reduces the risk of memorization effects, thereby increasing the trustworthiness of benchmark results for LLM-based KGQA approaches.
With the growth in the use of knowledge graphs (KGs) in a variety of applications, it is more important than ever to understand how such KGs are constructed. A core component of many construction pipelines is entity alignment, the identification of corresponding entities across KGs. A number of embedding-based approaches have been developed in the literature to address this problem. However, these algorithms are generally based on a few benchmark datasets. These benchmark datasets are not reflective of reality in regard to data quality, given that real-world KG data will often include false triples and incorrect alignments. This paper presents an investigation into how embedding-based approaches can be affected by low data quality, and what this might mean for real-world application. We find that the effects of noise can be both dataset and algorithm specific, but that relative performance remains unchanged. We also show that data cleaning can in some cases be far more effective in improving performance than collecting more data.
