Urban disaster event knowledge graph construction and representation method and device, and storage medium
By constructing a knowledge graph of urban disaster events, combining entity-centric and event-centric knowledge graphs, and employing entity linking and graph attention neural networks, the problem of multimodal representation of urban disaster data was solved, achieving efficient and unified representation and relationship prediction of urban disaster events.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2023-09-06
- Publication Date
- 2026-07-24
Smart Images
Figure CN117194679B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of resilient cities, and in particular to a method, apparatus and storage medium for constructing and representing a knowledge graph of urban disaster events. Background Technology
[0002] A resilient city is a city capable of withstanding (or experiencing) disasters. Urban functions include providing various types of production and living services to its residents. If these functions are disrupted, the inability of the urban system to function properly will affect citizens' lives and urban production. Therefore, the analysis and prediction of urban disasters plays a crucial role in building resilient cities.
[0003] Furthermore, urban disaster perception and monitoring data exhibits new characteristics such as massive volume, high dimensionality, multimodal heterogeneity, and semantic discrepancies. This presents challenges in urban disaster early warning systems, including difficulties in consistent representation of cross-modal perception data, data disturbances or gaps caused by urban uncertainties, and unclear semantic parsing in complex environments. Knowledge graph technology can help address these issues by resolving the multimodal and heterogeneous nature of data, integrating different urban disasters into a single graph structure, extracting the relationships between data to form facts or knowledge, and enabling complex semantic analysis.
[0004] A knowledge graph consists of entities and relations. Entities refer to concepts or things in the physical world, while relations refer to the relationships between these things, such as... Figure 1 As shown, an event knowledge graph is a branch of a knowledge graph, using events and related things as entities. Unlike general knowledge graphs, event knowledge graphs can treat event descriptions as entities and connect related things together through relationships.
[0005] The contributing factors to various urban disasters are numerous. For example, the contributing factors to natural disasters are mostly meteorological and geological; the contributing factors to public health disasters are mostly biological diseases and pests. Furthermore, disasters have special characteristics such as time and location. Therefore, the current design of simply treating disasters as entities in knowledge graphs is clearly unreasonable. Summary of the Invention
[0006] The purpose of this invention is to provide a method, apparatus, and storage medium for constructing and representing an Urban Disaster Event Knowledge Graph (UDKG). First, a central knowledge graph of urban disaster events is established, then a central knowledge graph of entities related to the urban system is established, and finally the two are merged through an entity linking method. The representation of the urban disaster event knowledge graph is learned through a knowledge representation learning model, thus laying the foundation for predicting the impact relationships of disasters.
[0007] The objective of this invention can be achieved through the following technical solutions:
[0008] A method for constructing and representing a knowledge graph of urban disaster events includes the following steps:
[0009] S1. Obtain raw city data;
[0010] S2. Define the ontology of the entity-centric knowledge graph and the event-centric knowledge graph, and process the acquired data through entity mapping and relation extraction techniques;
[0011] S3. Merge the entity-centered knowledge graph and the event-centered knowledge graph based on entity linking to achieve the fusion of the two knowledge graphs and obtain the urban disaster event knowledge graph. In this process, entity linking is improved by increasing training data through adversarial generation to enhance alignment accuracy.
[0012] S4. Perform representation learning on the nodes and relationships of the urban disaster event knowledge graph to obtain a low-dimensional embedding vector representation of the knowledge graph.
[0013] The ontology of the event-centric knowledge graph includes three types of abstractions: event, location, and time.
[0014] The ontology of the entity-centric knowledge graph includes two types of entities: subjects and locations.
[0015] The entity mapping in step S2 is achieved either by data coordination or by manually defining the mapping.
[0016] The relation extraction in step S2 relies on rules. It is achieved by pattern matching based on manually defined semantic rules, and semi-automatic extraction is achieved by modeling relations using machine learning models.
[0017] The relation extraction in step S2 is implemented using a stochastic expectation-maximization algorithm with a Bayesian network structure.
[0018] The merging of knowledge graphs in step S3 is based on the following two rules:
[0019] 1) The location of the disaster is the same as the area where the main infrastructure is located;
[0020] 2) The parties involved in the incident are the same as the civic community or management organization;
[0021] Define a similarity function and measure similarity based on the two rules mentioned above, then merge pairs with high similarity.
[0022] The representation learning in step S4 is achieved by encoding through a graph attention neural network and decoding through relational paths.
[0023] An apparatus for constructing and representing a knowledge graph of urban disaster events includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.
[0024] A storage medium having a program stored thereon, which, when executed, implements the method described above.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] (1) Representing urban disasters and disaster-related urban systems from the perspective of the city center.
[0027] At the city center level, this invention comprises transportation systems, material storage depots, medical facilities, municipal infrastructure (water supply, drainage, power supply), and evacuation shelters, forming a point-line-plane structure. Among these, a collection of land parcels providing a certain basic service function is a "node," such as a power plant or a residential community. The road connections and municipal infrastructure relationships between nodes serve as "edges," and the social networks formed by people on these nodes can also be considered as relationships, such as population flow and organizational management.
[0028] (2) A framework for a unified representation of urban disasters and urban system elements is proposed.
[0029] This invention designs a merged event-centric and entity-centric knowledge graph, enabling simultaneous knowledge graph representation of urban operational elements related to events, disasters, and their impacts. Furthermore, the entity-centric knowledge graph enriches the sparse relationships between disasters. Ontologies for both knowledge graphs are defined, and corresponding entity mappings and relationship extraction methods are proposed. Finally, the two knowledge graphs are merged using entity linking.
[0030] (3) Construct a highly available knowledge graph of urban disaster events
[0031] This invention enriches the relationships within a knowledge graph of urban disaster events by merging knowledge graphs. It incorporates urban elements with high indirect correlation to disasters in resilient city construction into the knowledge graph, rather than simply designing urban system elements directly related to urban disasters. Furthermore, it proposes embedding complex semantic representations into a low-dimensional space; the resulting vector representations facilitate relational reasoning and completion through the design of deep learning models. Attached Figure Description
[0032] Figure 1 This is an example of a knowledge graph;
[0033] Figure 2 This is a schematic diagram of the method flow of the present invention;
[0034] Figure 3This is a schematic diagram of the event-centric knowledge graph ontology.
[0035] Figure 4 This is a schematic diagram of the entity-centric knowledge graph ontology;
[0036] Figure 5 This is a schematic diagram of the knowledge extraction process. Detailed Implementation
[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0038] This embodiment provides a method for constructing and representing a knowledge graph of urban disaster events, such as... Figure 2 As shown. This invention differs from existing technologies by studying disasters as events, treating events as second-order entities, unlike ordinary entities. Specifically, this invention establishes two knowledge graphs: an entity-centric knowledge graph and an event-centric knowledge graph, and then merges them to create an urban disaster event knowledge graph. The event-centric knowledge graph integrates data on urban disaster events. Due to the sparse nature of the relationships between disasters, additional auxiliary information is needed to enhance the factual representation of the knowledge graph. Therefore, the entity-centric knowledge graph is constructed, primarily containing important systems in resilient urban management and planning, such as transportation and healthcare systems.
[0039] Specifically, the method includes the following steps:
[0040] S1. Obtain the original city data.
[0041] S2. Define the ontology of the entity-centric knowledge graph and the event-centric knowledge graph, and process the acquired data through entity mapping and relation extraction techniques.
[0042] ①Ontology definition
[0043] The ontology definition requires separate design for the event-centric knowledge graph and the entity-centric knowledge graph. Firstly, the event-centric knowledge graph ontology mainly includes three types of abstractions: event, location, and time. Secondly, the entity-centric knowledge graph ontology includes two types of entities: subject and location. The specific content is shown in Table 1, and further details need to be designed based on the specific disaster and the key impacts of concern.
[0044] Table 1. Design Concept
[0045]
[0046] Relationships need to be analyzed in more detail based on the specific content defined by the entity subject, time, event, and location, and the relationships between entities within a class and between entities between classes should be given, providing the main categories and content.
[0047] ② Data mapping and extraction
[0048] The data on entities and their attributes is mapped from existing disaster data (relational data in a database) to the content defined in the knowledge graph ontology. Many ready-made methods are available for this step, such as the Data Harmonization and Integration (DETIC) method, or the mapping could be manually defined and the data processed using Python. Disaster data primarily comes from city government databases, while city disaster events mainly originate from news data on websites. Additionally, city community and resource data are obtained through web scraping of resource websites and text data provided by city management departments.
[0049] Relation extraction is a challenging part of data processing, but many models exist for relation extraction in knowledge graphs. The resilient city disaster knowledge graph contains relationships between entity categories and within categories, which are extracted using rule-based extraction. These models can be divided into two categories: those based on manually defined semantic rules using pattern matching, and those using machine learning models to model relationships for semi-automatic extraction. The second category is currently the mainstream approach, and this invention can employ a stochastic expectation-maximization algorithm based on a Bayesian network structure.
[0050] S3. Merge the entity-centered knowledge graph and the event-centered knowledge graph based on entity linking to achieve the fusion of the two knowledge graphs and obtain the urban disaster event knowledge graph. In this process, entity linking is used to increase training data through adversarial generation to improve the alignment accuracy.
[0051] Based on the above steps, the separate construction processes for the two knowledge graphs have been completed. The purpose of step S3 is to merge the event-centric knowledge graph, i.e., the sparse graph with urban disasters as its main content, and the entity-centric knowledge graph, i.e., the graph structure with important urban resident life systems as its main content, to realize the construction of the urban disaster graph. At the same time, it leverages the resident service systems in the city that are closely related to disasters to enrich the sparse disaster relationships.
[0052] Entity linking is a crucial research area in knowledge graphs, used to address the issue of identical entities having different representations within the knowledge graph. Knowledge graph merging is achieved based on the following two rules:
[0053] 1) If the location of the disaster is in the same area as the main body of the relevant infrastructure, then physical links can be established;
[0054] 2) If the participants in the event are the same as the civic community or management organization, then a physical link can be established;
[0055] Define a similarity function and measure similarity based on the two rules mentioned above, then merge pairs with high similarity.
[0056] S4. Representation learning is performed on the nodes and relationships of the urban disaster event knowledge graph. The graph is encoded through a graph attention neural network and decoded through relational paths to obtain a low-dimensional embedding vector representation of the knowledge graph.
[0057] Finally, a model for learning a low-dimensional spatial representation of urban disaster event knowledge graphs is designed. Knowledge graphs are high-dimensional representations of semantic content, which is very difficult for computers to perform knowledge reasoning or knowledge completion. Therefore, learning low-dimensional embeddings of complex semantic content facilitates subsequent predictive research using the topological structure of the knowledge graph. Currently, relevant learning models mainly include translation models, tensor decomposition models, and complex neural network models. This invention designs and uses a graph attention network (GAN) combined with a relation path-based decoding model to achieve a more accurate low-dimensional representation of the knowledge graph.
[0058] In summary, the functions achieved by this invention are as follows:
[0059] 1. Ontology definitions for disaster knowledge graphs and urban system element knowledge graphs are proposed, along with the entities and relationships they may include;
[0060] 2. Process urban disaster data and other system data, and construct knowledge graph instances using entity and relation ontology.
[0061] 3. Propose entity linking to enrich the sparse relationships between disasters using relevant knowledge graphs;
[0062] 4. Perform representation learning on the merged and unified urban disaster event knowledge graph to obtain low-dimensional representations of facts such as the city POIs of the events and their impacts, which will facilitate the next step of relationship prediction.
[0063] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0064] To further illustrate this invention, this embodiment uses COVID-19 as a specific disaster example. Urban disasters are divided into three categories based on different classification standards: coarse-grained, medium-grained, and fine-grained, denoted as C-1, C-2, and C-3 respectively in the ontology definition. The classifications differ at each granularity level. Coarse-grained disasters mainly include three types: natural disasters, accident disasters, and public health disasters. Medium-grained disasters mainly include floods, rainstorms, tornadoes, typhoons, land subsidence, earthquakes, traffic accidents, fires, industrial accidents, lifeline systems, and infectious diseases. Fine-grained disasters include road flooding, road collapses, traffic congestion, regional power outages, regional water outages, regional internet outages, interruptions in essential supplies, and travel restrictions. The COVID-19 pandemic is classified as a public health disaster at the coarse-grained level and an infectious disease at the medium-grained level.
[0065] Taking COVID-19 as an example, the ontology of the event-centric knowledge graph specifically includes entities such as Event_Description, Participating Subjects (Actor), Place, Role, Role_Type, and Core Members (Core), as detailed below. Figure 3 As shown.
[0066] The ontology of the entity-centric knowledge graph needs to be based on urban elements most closely related to COVID-19, such as hospitals, farmers' markets, large supermarkets, bus stops, subway stations, residential areas, and office buildings. See details below. Figure 4 .
[0067] Based on the ontology definition of the knowledge graph mentioned above, entity mapping and relation extraction can be performed on the collected epidemic data. Currently, there are many methods for knowledge extraction from knowledge graphs. Since the data studied in this invention consists of database and text data, mapping these data is sufficient to obtain entities. Relationships, however, require training a SEM-based model for semi-automatic extraction. The specific process is as follows: Figure 5 .
[0068] After instance extraction from the two knowledge graphs, they are merged using an entity linking method. Many models exist for this approach; this patent employs a rule-based similarity metric. Entity linking involves defining a similarity function, which is automatically learned through a machine learning model based on three principles: the similarity between the location of the COVID-19 outbreak and the location of the community; the similarity between COVID-19 participants and the community or organization; and the similarity between the outbreak location and the location of the hospital or testing site. Finally, results with higher similarity scores are merged. This step is subject to error, therefore a more accurate evaluation method is needed to determine the final outcome.
[0069] Finally, the goal is to learn a more accurate low-dimensional representation of the overall COVID-19 pandemic event graph. This process primarily involves information transmission and aggregation through the topology of the complex network, using graph neural attention networks to encode complex semantic knowledge, and then decoding it based on relational paths to enhance the model's learning ability, ultimately resulting in a more accurate and efficient representation.
[0070] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for constructing and representing a knowledge graph of urban disaster events, characterized in that, Includes the following steps: S1. Obtain raw city data; S2. Define the ontology of the entity-centric knowledge graph and the event-centric knowledge graph, and process the acquired data through entity mapping and relation extraction techniques; S3. Merge the entity-centered knowledge graph and the event-centered knowledge graph based on entity linking to achieve the fusion of the two knowledge graphs and obtain the urban disaster event knowledge graph. In this process, entity linking is improved by increasing training data through adversarial generation to enhance alignment accuracy. S4. Perform representation learning on the nodes and relationships of the urban disaster event knowledge graph to obtain a low-dimensional embedding vector representation of the knowledge graph; The merging of knowledge graphs in step S3 is based on the following two rules: 1) The location of the disaster is the same as the area where the main infrastructure is located; 2) The parties involved in the incident are the same as the civic community or management organization; Define a similarity function and measure similarity based on the two rules mentioned above, then merge pairs with high similarity.
2. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The ontology of the event-centric knowledge graph includes three types of abstractions: event, location, and time.
3. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The ontology of the entity-centric knowledge graph includes two types of entities: subjects and locations.
4. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The entity mapping in step S2 is achieved either by data coordination or by manually defining the mapping.
5. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The relation extraction in step S2 relies on rules. It is achieved by pattern matching based on manually defined semantic rules, and semi-automatic extraction is achieved by modeling relations using machine learning models.
6. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The relation extraction in step S2 is implemented using a stochastic expectation-maximization algorithm with a Bayesian network structure.
7. The method for constructing and representing a knowledge graph of urban disaster events according to claim 1, characterized in that, The representation learning in step S4 is achieved by encoding through a graph attention neural network and decoding through relational paths.
8. A device for constructing and representing a knowledge graph of urban disaster events, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.
9. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-7.