TL;DR
This paper presents a novel approach to enhance traffic safety management by constructing knowledge graphs from both structured and unstructured data. They developed two distinct knowledge graphs: one for structured traffic safety management data and another for unstructured traffic accident reports, leveraging Neo4j for storage and Cypher for querying.
What Problem Does This Paper Address?
The ever-increasing volume of traffic safety information, while crucial for sustainable transport development, is often complex, scattered, and lacks a unified framework for integration and analysis. This makes it challenging for safety managers and decision-makers to effectively utilize this data for proactive safety measures and risk mitigation. The paper aims to bridge this gap by creating structured, queryable knowledge representations of this information.
Methods
The authors employed a two-pronged approach:
- Structured Data Knowledge Graph: For structured data (e.g., legal norms, organization information, road data), they designed a knowledge graph with 54 node entity types and 14 relationship entity types. This graph was built directly in Neo4j using
py2neo, focusing on information linkage and search services. - Unstructured Data Knowledge Graph (Traffic Accidents): For unstructured data like traffic accident reports, they focused on extracting entities and relationships using deep learning models.
- Named-Entity Recognition (NER): They compared four models (Bert, Bert-CRF, Bert-BiLSTM, and Bert-BiLSTM-CRF) for entity extraction from accident reports.
- Relationship Extraction: The Bert model was used for extracting relationships between entities.
- The traffic accident knowledge graph was designed with 21 node entity types and 22 relationship entity types, centered around accidents, involved vehicles, and people/institutions.
Both knowledge graphs utilize Neo4j for storage and support query functions. For the structured graph, they implemented queries using Cypher and an automatic Q&A function based on rule matching.
Datasets
The paper utilizes real-world data for its construction and evaluation:
- Structured Data: Information was collected from government apparatus, enterprises, and media, covering six categories: illegal acts, vehicle failure, emergency response, legal norms, organization information, and road-related information.
- Unstructured Data: Traffic accident reports were crawled from safety information websites and official emergency management bureau websites. The dataset size was 758 KB.
- Annotation: The "Baidu Brain EasyData" platform was used for annotating entities and relationships in the unstructured data.
Key Insights and Ideas
- Unified Knowledge Representation: The core idea is to transform disparate and complex traffic safety information into a structured, interconnected knowledge graph, making it more accessible and actionable.
- Dual Approach for Data Types: Recognizing the different nature of structured and unstructured data, the paper proposes distinct but complementary strategies for building knowledge graphs, enhancing the comprehensiveness of the safety analysis.
- Deep Learning for Unstructured Data: The application of advanced deep learning models (especially Bert-BiLSTM-CRF for NER) is crucial for effectively extracting meaningful information from the noisy and varied text of accident reports.
- Neo4j as a Robust Platform: The choice of Neo4j as the graph database provides a scalable and efficient solution for storing and querying complex relational data, essential for knowledge graphs.
- Cypher for Powerful Queries: The use of Cypher enables flexible and expressive querying, allowing for detailed analysis of traffic safety patterns and accident causes.
Key Results
- Structured Knowledge Graph: Successfully constructed a traffic safety management knowledge graph with 54 node types and 14 relationship types, covering six key information categories. This graph supports query functions via Cypher and rule-based Q&A.
- Traffic Accident Knowledge Graph: Developed a traffic accident knowledge graph with 21 node types and 22 relationship types.
- NER Model Performance: The Bert-BiLSTM-CRF model achieved the best performance in named-entity recognition for traffic accident reports, outperforming other tested models in precision, recall, and F1 scores.
- Bert-BiLSTM-CRF: Precision 86.93%, Recall 91.20%, F1 89.01%.
- Relationship Extraction: The Bert model was effective for relationship extraction from accident reports.
- Data Integration: The traffic accident knowledge graph can structurally display accident information and condense accident reports, reducing reading burden and integrating key safety elements. It was populated with 9318 nodes and 14,273 relationships.
Limitations and Future Work
- Model Performance: While Bert-BiLSTM-CRF performed best, the paper notes that relationship extraction is generally easier than entity extraction, suggesting room for improvement in entity recognition for complex accident reports.
- Data Augmentation: The paper mentions using data augmentation for relationship extraction due to a lack of annotation for some relationships, indicating that data scarcity can be a challenge.
- Generalizability: The structured knowledge graph was built using data from Beijing. While the principles are generalizable, specific entity and relationship types might need adaptation for other regions.
- Future Work: The authors suggest that safety knowledge graphs can be applied to more fields, combining safety management models and data to store safety information tailored to industry characteristics. They also highlight the potential for graphs to display structured information and facilitate safety analysis.
Summary updated