Fiscal Policy Knowledge Graph & Search Engine

Project Background:

Designed and led the development of a fiscal intelligence system for Zhong Jin Suo, focusing on policy document retrieval and intelligent recommendation.
Objective: Solve inefficiencies in policy document search and enable deep policy analysis via knowledge graph technology.

Data Collection & Processing:

Built a policy crawler system for scheduled acquisition of fiscal and tax policy documents.
Structured storage in MySQL and MongoDB; full-text indexing in Elasticsearch.
Entity & relation extraction:
- Word2Vec + K-Means for entity discovery.
- N-Gram + top-down taxonomy construction for category-label hierarchy.
- LSTM-CRF for fine-grained entity-relation extraction.

Knowledge Graph Construction:

Ontology modeling (services, industries, regions) based on OWL.
Implemented ontology query with SPARQL; developed region recognition interface for geographic tagging.
Knowledge graph stored and visualized in Neo4j.

Modeling & Optimization:

Multi-level policy association graphs:
- 2-level association graph for related policies.
- Tree-structured hierarchy graph (central, provincial, municipal levels).
Recommendation:
- Similarity-based: Jaccard similarity.
- Hot policy discovery: PageRank centrality.
- Community detection: Louvain algorithm for policy domain clustering.
Path discovery:
- Maximum Spanning Tree (MST) for strong relationship tracing.
- Dijkstra algorithm for shortest policy association path analysis.
Policy Q&A:
- RASA-based intent recognition and slot filling.
- Query templates mapped to Elasticsearch/MySQL/Neo4j.

Workflow:

Project Outcome:

Successfully deployed the Fiscal Intelligence System, commercialized and sold to government and enterprise clients.
Reduced policy retrieval time drastically, enabling policy analysts to discover associations and insights efficiently.
Significantly enhanced fiscal policy research and decision-making processes through intelligent search and graph-based analytics.