AI Knowledge Graph Maps the Hidden Story Network of Tianjin’s Grand Canal Heritage

The Grand Canal is not a single monument but a thread of stories…

The methodological choices made in this study reflect a broader shift in how digital humanities researchers approach cultural heritage documentation. Traditional heritage inventories typically record sites as isolated entries with attribute lists: construction dates, architectural styles, protection statuses. Such formats, while useful for administrative management, struggle to capture the relational essence of linear heritage, where the significance of any single lock, temple, or wharf depends heavily on its position within flows of people, goods, and stories. By treating narrative as a first-class object—modeling events with hierarchical structure and linking them to actors, locations, and time periods—the knowledge graph approach allows each of the 98 Tianjin sites to carry not just its own description but its web of connections to every other site mentioned in the same stories.

The use of large language models for extraction deserves particular attention. Historical texts about the Grand Canal exist in heterogeneous forms: local gazetteers, chronicles, administrative records, travel writings, and modern interpretive materials. Manually coding these sources into structured data would require enormous scholarly labor, and consistency across coders is notoriously difficult to maintain. LLM-assisted extraction offers a way to accelerate this process, though it introduces its own challenges. Models can misread archaic phrasing, conflate similarly named places, or hallucinate relationships not actually supported by the source text. The study’s emphasis on traceable narrative evidence suggests an awareness of this risk: rather than treating machine output as ground truth, the pipeline appears designed so that extracted relations can be traced back to their textual origins, allowing human scholars to verify and correct the graph. This human-in-the-loop posture is increasingly seen as essential for responsible AI use in heritage informatics, where errors are not merely technical glitches but potential distortions of cultural memory.

Event hierarchy modeling is another component worth unpacking. In narrative theory, events are not flat occurrences but nested structures: a flood control campaign contains sub-events like dike repairs, which contain further sub-events like specific construction episodes. Encoding this hierarchy into a graph means that queries can operate at multiple granularities. A curator preparing an exhibition on hydraulic engineering could retrieve not only the major historical interventions but also the fine-grained actions that constituted them, distributed across dozens of sites along the canal. Similarly, a tour planner could ask which sites share participation in the same event cluster, producing thematically coherent routes rather than geographically convenient but narratively arbitrary sequences.

The projection from narrative graph to spatial network represents a conceptual bridge between two analytical traditions. Social network analysis, developed in sociology, provides metrics like centrality, clustering, and connectivity that describe how nodes relate to one another abstractly. Spatial analysis, rooted in geography and urban studies, describes how phenomena distribute across physical space. Linear cultural heritage sits precisely at the intersection of these traditions: its nodes are physically ordered along a waterway, yet its meanings flow through stories that may jump across distances. The Tianjin study’s identification of narrative corridors—paths through the network where story connections align with geographic adjacency—illustrates what this intersection can yield. A corridor is not simply a stretch of canal with many sites; it is a stretch where the sites themselves participate in shared narratives, making the physical journey also a narrative journey.

The findings about Tianjin’s spatial-narrative structure resonate with what is known about the city’s historical role. The Old City and the Sanchahekou area, where the Grand Canal meets the Hai River, formed the historic heart of Tianjin, a convergence point that made the city the gateway between the canal system and maritime trade. That this core emerges as the most narratively comprehensive zone in the analysis is consistent with centuries of documentation density: places that hosted more administrative, commercial, and ceremonial activity naturally accumulated more textual records. Peripheral canal sections, by contrast, show specialized hydraulic units—structures dedicated to specific functions like water regulation, flood control, or cargo transfer. Their narratives are narrower but functionally critical, and the medium- and weak-strength relations identified in the network serve as the connective tissue that prevents these specialized nodes from becoming isolated islands.

Yangliuqing’s pattern of locally embedded memory offers a different lesson. This historic town, famous for its folk art traditions including New Year woodblock prints, developed a narrative profile that is dense internally but comparatively thin in its external connections. Such a pattern is common in heritage networks: settlements with strong local identities often generate rich community-level storytelling that does not translate into region-wide textual prominence. For heritage managers, this suggests that different sites require different interpretive strategies. Core nodes with high narrative visibility can anchor citywide or regional narratives, while locally embedded sites may benefit from approaches that celebrate and preserve their distinctiveness rather than forcing them into generic linear narratives where they play minor roles.

The implications extend to public communication as well. Knowledge graphs are inherently queryable structures, which means the same dataset that supports scholarly analysis can power interactive experiences for visitors. A digital interface built on such a graph could let a user select a theme—say, flood management or canal-side commerce—and see which sites participate in that theme, how the events unfold chronologically, and how a walking or boating route might connect them. Because every displayed relationship carries traceable evidence, such interfaces can also teach source literacy, showing visitors that historical narratives are constructed from documents with their own perspectives and gaps. This transparency distinguishes evidence-based digital interpretation from the more free-form storytelling found in many popular heritage apps.

Comparative potential is substantial. The Grand Canal was inscribed on the UNESCO World Heritage List in 2014, and its linear character—spanning multiple provinces and centuries of operation—makes it a natural testbed for relational heritage approaches. The Tianjin section is one segment of a much longer system, and the methods demonstrated here could in principle be applied to other sections, with the resulting graphs merged to reveal inter-regional narrative patterns. Similar logic applies to other linear heritage types: pilgrimage routes, trade roads, defensive lines, and railway corridors all share the property that individual components derive meaning from participation in a larger connected whole. The spatio-narrative network framework, if generalized, could become a standard analytical layer in such contexts, complementing the point-based assessments that dominate current heritage evaluation practice.

Limitations acknowledged in work of this kind typically concern source coverage and extraction fidelity. Textual records skew toward literate, administrative, and elite perspectives; vernacular memories, oral traditions, and the experiences of laborers who actually operated the canal are underrepresented in surviving documents. A narrative graph built from texts therefore inherits these biases, and its picture of the heritage network reflects what was written down rather than everything that happened. The authors’ framing of results as evidence-based rather than exhaustive is a fitting hedge: the network reveals structure in the available narratives, which is a necessary but not sufficient condition for understanding the heritage landscape in full. Future integration of oral history collections, archaeological reports, and material culture databases could progressively broaden the evidentiary base.

There is also a maintenance dimension to consider. Heritage knowledge graphs are not static artifacts; new documents surface, interpretations evolve, and site statuses change with conservation decisions. A graph that cannot be updated risks becoming a snapshot frozen at publication time. Sustainable operation therefore requires pipelines that can ingest new sources incrementally, re-run extraction with versioned models, and propagate updates through the network analyses. The study’s modular design—separating extraction, modeling, and database construction—positions it well for such ongoing curation, and the open access publication of the underlying research supports the community scrutiny that long-term datasets require.

For practitioners in heritage interpretation, the most immediately actionable takeaway may be the differentiation of node roles. Endpoint roles identified in the network analysis—sites that serve as narrative origins or destinations—can guide the placement of interpretive gateways, visitor centers, and route termini. Thematic clusters suggest exhibition groupings and content series. The mapping of medium- and weak-strength relations highlights where narrative linkage is thin and where new research, documentation, or programming could strengthen the connective story. In this way, the analytical apparatus of network science translates into concrete curatorial decisions, closing the loop between computational modeling and the everyday work of presenting heritage to the public.

Finally, the study contributes to an ongoing conversation about what counts as heritage knowledge. By formalizing narratives as networks of events, actors, places, and times, it treats stories not as decorative color applied to physical remains but as structural elements of heritage itself—as real and as analyzable as masonry or hydrology. This ontological commitment has consequences: it legitimizes investment in narrative documentation, justifies computational methods in a field sometimes skeptical of them, and reframes linear heritage as something that exists simultaneously in water, stone, and telling. The Tianjin Grand Canal, with its dense layering of imperial administration, commercial life, and folk tradition, proves an apt case for demonstrating how these layers can be woven into a single, queryable, and interpretable whole.

Subject of Research: Knowledge graph analysis of narrative connections among Tianjin Grand Canal heritage sites

Article Title: From fragmented narratives to spatio-narrative networks: a knowledge graph approach to Tianjin Grand Canal heritage

Article References: Zhang, B., Tian, J., Meng, L., Li, J., & Cao, L. (2026). From fragmented narratives to spatio-narrative networks: a knowledge graph approach to Tianjin Grand Canal heritage. npj Heritage Science. https://doi.org/10.1038/s40494-026-03060-y

Image Credits: AI Generated

DOI: 10.1038/s40494-026-03060-y

Keywords: knowledge graph, Grand Canal, Tianjin, cultural heritage, linear heritage, narrative network, spatial analysis, digital humanities, LLM, heritage science, fragmented, narratives