Project
Tang Poetry Imagery Universe
An interactive digital-humanities site built from imagery co-occurrence across 57,607 Tang poems.
Python · Jieba · NetworkX · PyEcharts · ECharts
Product Snapshot
- Role
- Data analysis and interactive visualization
- Users
- Literary researchers, digital-humanities students, and readers exploring imagery relations in Tang poetry.
- Stage
- Data pipeline, methodology, and multi-view interactive site completed.
- Focus
- Large-scale segmentation, imagery extraction, co-occurrence networks, centrality, and poet comparison.
- Validation
- Built a 28-node, 342-edge imagery network from 29,760 active windows and compared structures across Li Bai, Du Fu, and Wang Wei.
- Public Proof
- Processing outputs for 57,607 poems, CSV/JSON, NetworkX analysis, interactive pages, and methodology documentation.
Next Step
From a corpus to a relational universe
Counting how often moon, wind, or mountain appears only gives frequency. This project asks which images repeatedly share a context and how those relationships structure Tang poetry.
I processed 57,607 poems, built co-occurrence relations from local text windows, and turned the network into an exploratory website.
Data and method
The pipeline covers cleaning, Jieba segmentation, imagery-dictionary matching, windowed co-occurrence, NetworkX graph construction, and centrality analysis. It retained 29,760 active windows and produced 28 nodes with 342 edges.
Moon and wind formed the strongest edge with weight 764. Moon also had the highest degree and betweenness centrality, suggesting that it connects several imagery contexts rather than merely appearing often.
Interactive expression
The site separates overview, node relations, strongest links, centrality, and poet comparison into distinct views. Readers can compare Li Bai, Du Fu, and Wang Wei to see how the same image occupies different structural roles.
Verifiable delivery
The repository preserves processing logic, network CSV, frontend JSON, summary metrics, methodology, and interaction-design notes. Interpretations can therefore return to data and rules instead of ending at the visualization.
What I learned
Cultural visualization should do more than look attractive. It should reveal how nodes and links were produced and help readers move back from a graph to a textual question.
Configure the public Giscus environment variables to open discussion on the public site.