MuseKG: Unifying Museum Collections with Knowledge Graphs

Digitisation efforts in the cultural heritage sector have produced vast repositories of artefact information, yet these collections remain frustratingly fragmented across text-based catalogues, image archives, and multimedia documentation. The core isn't just the sheer volume of data but the semantic disconnection that prevents holistic search and interpretation. Objects, their relationships, and contextual metadata often exist in isolation within heterogeneous systems, limiting cross-collection and deeper reasoning. This fragmentation has long hindered museums from unlocking the full potential of their digital assets, leaving researchers and the public with incomplete or siloed views of cultural history. Recent advances at the intersection of knowledge graphs and large language models offer a promising pathway, but existing approaches fall short of providing an end-to-end solution that seamlessly integrates data collection, graph construction, and user-friendly interaction.

To address these limitations, researchers from the University of Melbourne developed MuseKG, an end-to-end knowledge graph framework that unifies structured and unstructured museum data through symbolic-neural integration. The system constructs a typed property graph linking entities such as objects, people, organisations, and visual or textual labels, based on a schema that includes node types like object, person, image_label, and concept. MuseKG's ology begins with raw JSON records from museum collections, which undergo a rigorous normalisation process where textual fields are lowercased, whitespace is collapsed, and punctuation is stripped to ensure stable matching. Entity identification then creates nodes for each unique identifier, such as opacObjectId or relatedRecordId, with node types assigned accordingly. Edges are formed by mapping raw relationship identifiers to a fixed vocabulary of relations, such as has_primary_producer, followed by deduplication and schema validation to maintain data integrity. This pipeline transforms disparate records into a coherent, queryable graph that captures the rich interconnections within museum data.

In empirical evaluations, MuseKG demonstrated robust performance across a benchmark of 150 questions categorised into attribute lookup, relation retrieval, and one-hop reasoning tasks. The framework consistently outperformed baseline s, including large-language-model zero-shot, few-shot, and SPARQL-prompt approaches, with accuracy scores reaching up to 0.84 for attribute queries and 0.50 for relation-based questions when using models like Gemma-3-12B. For instance, in a complex query asking for the accession number of an entity related to a 'Certificate of Passing First Year of Bachelor of Laws', MuseKG successfully navigated the graph to retrieve the correct value, MHM06682, by identifying the object, traversing the has_entity relation, and extracting the target attribute. Error analysis revealed that while baseline s often failed due to hallucinations or parsing issues, MuseKG's errors were minor, such as slight inaccuracies in dates, underscoring the benefits of its retrieval-augmented generation approach that grounds responses in factual KG context.

Of MuseKG extend beyond academic research, offering museums a scalable tool for enhancing public engagement and scholarly access. By supporting natural language queries, it lowers the barrier for non-technical users to explore complex relationships, such as tracing an artefact's provenance or identifying all objects produced by a specific organisation. This could revolutionise digital heritage management, enabling institutions to integrate multimodal data—like images and unstructured documents—into a unified knowledge base that supports interpretable reasoning. Moreover, the framework's efficiency, with query latencies as low as 0.36 seconds, makes it practical for real-time applications in virtual exhibitions or educational platforms, potentially fostering greater interoperability across global cultural heritage databases.

Despite its strengths, MuseKG has limitations, including its current reliance on a single institution's dataset and support for only one-hop reasoning queries. The benchmark, while comprehensive, covers only 150 questions, and scaling to larger, more diverse collections may reveal s in handling noisy or incomplete data. Future work could focus on extending the system to multi-hop queries and collaborating with multiple museums to test its adaptability. Nonetheless, MuseKG represents a significant step toward web-scale integration of digital heritage knowledge, highlighting the critical role of symbolic grounding in achieving accurate, scalable AI systems for cultural preservation.