Method for constructing target time sequence knowledge graph facing space-based data

Abstract

This invention relates to a method for constructing a target temporal knowledge graph for space-based data. The method includes: S100, constructing a target ontology, a radar ontology, a region ontology, and an event ontology; S200, extracting knowledge based on the target ontology, the radar ontology, the region ontology, and the event ontology; S300, fusing the extracted knowledge to obtain a target knowledge graph; S400, embedding a temporal model into the target knowledge graph; and S500, constructing the temporal knowledge graph according to the structure of step S400. This invention adds temporal and spatial processing mechanisms to the traditional knowledge graph, enabling the extraction and modeling of multi-source heterogeneous data, and providing better support for intelligence analysis, activity pattern analysis, decision support, and subsequent target identification.

Description

Technical Field

[0001] This invention relates to, specifically to, a method for constructing a target temporal knowledge graph for space-based data. Background Technology

[0002] In the context of big data, the analysis results of space-based reconnaissance intelligence are of great significance for battlefield situation assessment and decision-making. In high-tech warfare, various battlefield sensors and monitoring equipment are widely used, resulting in an explosive growth in various types of combat and intelligence data. Furthermore, information on various combat targets is characterized by large data volumes and rapid transmission. Analyzing and identifying enemy information captured in the battlefield environment can effectively determine the types of target equipment, thereby estimating enemy combat actions and intentions, and assisting our command and decision-making.

[0003] However, current space-based reconnaissance intelligence analysis and identification knowledge bases are lacking, and most identification requires manual intervention. Due to the limitations of human ability and energy, it is impossible to make accurate judgments and analyses in the face of massive amounts of data. Therefore, in modern warfare, the main factor causing the "fog of war" is no longer a lack of information, but an overabundance of information. How to provide commanders with various types of operational target information quickly, efficiently, and accurately, and prevent them from being overwhelmed by massive amounts of information, is one of the core issues of operational target information service support and decision-making assistance.

[0004] Knowledge graphs are one approach to addressing these challenges. They can abstract, organize, and restructure massive amounts of data, presenting knowledge to users in a more rational, orderly, and intuitive way, thus enhancing understanding and cognition and more effectively guiding decision-making. A knowledge graph constitutes a vast semantic network, where nodes represent entities or concepts, and edges between nodes represent attributes and relationships. Therefore, the basic building blocks of a knowledge graph are the triples of "entity-attribute," "entity-attribute-entity," and "entity-relationship-entity." Entities are interconnected through relationships, forming a network of knowledge structures. This network, composed of numerous nodes and edges, provides a complete and clear description of knowledge. When large amounts of knowledge graphs are integrated and organized systematically, a knowledge base is formed.

[0005] Current research focuses on static knowledge graphs where facts do not change over time. For example, the paper "Construction and Application of Operational Target Knowledge Graph" proposes a basic architecture for an operational target knowledge graph and identifies the technical challenges it faces in application. Additionally, Chinese patent CN113806513A discloses a method for constructing a question-answering system based on a military domain knowledge graph. This method stores military domain knowledge in the form of a military domain knowledge graph and uses natural language questions as input to quickly and accurately recommend battlefield situational knowledge from the military knowledge graph.

[0006] It is evident that the aforementioned solutions and current research in this field focus on static knowledge graphs where facts do not change over time. They model the target itself and the associated past events, failing to accurately describe the real-time battlefield situation and the dynamic trends of the target. This leads to numerous problems in practical applications, while the temporal dynamics of knowledge graphs have been rarely studied.

[0007] However, in the intelligence field, time is a crucial factor influencing command and decision-making. Because much structured knowledge is only valid within a specific timeframe, and events unfold in chronological order, expressing causal relationships, which aligns more closely with reality, temporal knowledge graphs will have wider applications in the military. Furthermore, since space-based data is often multi-source and heterogeneous, and largely contains temporal and spatial information, processing this information is also a vital part of constructing temporal knowledge graphs. Summary of the Invention

[0008] In view of this, the present invention aims to propose a method for constructing a target temporal knowledge graph for space-based data. Based on the traditional knowledge graph, it adds processing mechanisms related to time and space, which can extract and model multi-source heterogeneous data, and can better support intelligence analysis, activity pattern analysis, decision support and subsequent target identification.

[0009] This invention provides a method for constructing a target time-series knowledge graph for space-based data, the method comprising:

[0010] S100, constructing target ontology, radar ontology, area ontology and event ontology;

[0011] S200, knowledge extraction is performed based on the target entity, the radar entity, the region entity, and the event entity;

[0012] S300, perform knowledge fusion on the knowledge extraction results to obtain the target knowledge graph;

[0013] S400, embed a time model into the target knowledge graph;

[0014] S500: Construct a time-series knowledge graph based on the results of step S400.

[0015] In a preferred embodiment of the present invention, in step S100:

[0016] A local ontology is constructed based on the characteristics of the data, and a global ontology is constructed based on the commonalities among the data.

[0017] In a preferred embodiment of the present invention, in step S200:

[0018] The knowledge extraction includes entity extraction, attribute extraction, and relation extraction.

[0019] In a preferred embodiment of the present invention, in step S200:

[0020] For structured data, a direct mapping method is used to map linked data or database data into a resource description framework diagram;

[0021] For unstructured data, the unstructured data is divided into image data and text data. The image data is processed using a target recognition algorithm to convert the image data into semantic information. The text data is processed using a natural language processing algorithm to extract entities and relationships from the intelligence file.

[0022] For semi-structured data, induction, manual extraction, and automatic extraction are packaged into modules for the extraction of entities and relationships.

[0023] In a preferred embodiment of the present invention, in step S300, the knowledge fusion includes entity linking and knowledge merging.

[0024] In a preferred embodiment of the present invention, the entity linking includes determining whether there are entities with the same meaning in the existing knowledge graph library, obtaining entity objects through entity extraction, and mapping the entities to the correct entities in the knowledge graph library;

[0025] The knowledge merging includes selecting relationship paths and retaining those potentially relevant to the predicted target relationship, calculating feature values ​​for each entity, and training a classification model for each target based on training examples and the feature values.

[0026] In a preferred embodiment of the present invention, in step S400:

[0027] The temporal model is fused into the entity or relation representation vector of the sequence learning model to estimate the authenticity of triples with temporal information, which is then used to complete the temporal knowledge graph.

[0028] In a preferred embodiment of the present invention, the sequence learning model includes a relation sequence composed of relations and timestamps. The features of the relations and timestamps are mapped into vectors of the same dimension through a linear layer function and encoded through a long short-term memory network to learn a relation representation vector that integrates time information.

[0029] In a preferred embodiment of the present invention, the relationship and the timestamp information are processed by an LSTM model to form a relationship representation vector that integrates time information, and the triple information is completed according to the scoring function of the TransE model. The triple information is: s, pseq, o; where s is the head entity, o is the tail entity, and pseq is the relationship.

[0030] In a preferred embodiment of the present invention, in step S500, a time-series knowledge graph is constructed based on the completion and the scoring function.

[0031] The method for constructing a target time-series knowledge graph for space-based data in this embodiment of the invention can: (1) model multi-source heterogeneous intelligence data and construct a knowledge graph; (2) introduce time-series information into the knowledge graph, which can better describe intelligence information and assist command decision-making; (3) convert traditional static knowledge graphs into time-series knowledge graphs, thereby enhancing the semantic description capability of the target knowledge graph; and (4) use the representation sequence learning method of long short-term memory networks to convert time-series information into time vectors. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the construction method according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic block diagram illustrating the principle of the construction method according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram illustrating knowledge extraction in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the time model embedding in an embodiment of the present invention. Detailed Implementation

[0037] The description of the embodiments in this specification should be taken in conjunction with the accompanying drawings, which should form part of the complete specification. In the drawings, the shape or thickness of the embodiments may be exaggerated and may be indicated in a simplified or convenient manner. Furthermore, parts of the various structures in the drawings will be described separately; it is worth noting that elements not shown in the figures or not described in words are in a form known to those skilled in the art.

[0038] The descriptions of the embodiments herein, including any references to directions and orientations, are for ease of description only and should not be construed as limiting the scope of the invention. The following description of preferred embodiments involves combinations of features, which may exist independently or in combination; the invention is not particularly limited to the preferred embodiments. The scope of the invention is defined by the claims.

[0039] like Figures 1-2 As shown, the method for constructing a target temporal knowledge graph for space-based data according to an embodiment of the present invention includes:

[0040] S100 constructs the target ontology, radar ontology, area ontology, and event ontology.

[0041] In this embodiment, target ontology, radar ontology, region ontology, and event ontology are constructed based on different types of space-based data and actual business needs. Local ontology is constructed based on the characteristics of the data, while global ontology is constructed based on the commonalities among the data. That is, a target local ontology is constructed based on the characteristics of the target data, and a target global ontology is constructed based on the commonalities among the target data. Similarly, local and global ontologs are constructed for the radar ontology, region ontology, and event ontology, respectively.

[0042] For example:

[0043] A target ontology is constructed based on internet knowledge and expert knowledge bases, including a classification structure for ship types and classes. A radar ontology is constructed based on internet encyclopedias and professional books such as the "World Radar Handbook," including the hierarchical structure of radar systems. Based on internet data segmentation and expert knowledge, geographical regions and sea areas are hierarchically divided, including sea area segmentation and port segmentation for different countries, constructing a regional ontology. Based on semi-structured and unstructured data files, events described in intelligence files are extracted using word segmentation and clustering algorithms, constructing an event ontology.

[0044] (1) Define the hierarchical structure of the domain:

[0045] Ship, ship type, ship class;

[0046] Radar system, technical system;

[0047] Country name, port name;

[0048] Sea area name, event name, formation name.

[0049] (2) Define the attributes and facets of the ontology:

[0050] Based on the characteristics of the entity itself and combined with actual needs, extract the entity attributes;

[0051] Ships: length, width, Chinese name, English name, nickname, draft, etc.;

[0052] Radar: Transmission parameters, frequency, mode, swath width, etc.;

[0053] Sea area: Sea area coordinates;

[0054] Ports: Port coordinates, port type, etc.;

[0055] At the same time, the domain and range of each attribute are defined.

[0056] (3) Define the relations in the ontology:

[0057] Based on the basic semantic relations of the ontology (part-of, kind-of, instance-of, attribute-of), as well as the category settings and specific characteristics of the domain itself, we summarize and categorize semantic relation types (such as loading, opposition, alliance, and other specific relations).

[0058] Using the Protégé software, create an ontology graph and import the completed ontology data into the Neo4j graph database.

[0059] S200, knowledge extraction is performed based on the target entity, the radar entity, the region entity, and the event entity.

[0060] like Figure 3 As shown, in this embodiment, knowledge extraction is performed on multi-source heterogeneous space-based data, including attribute extraction, entity extraction, and relation extraction. For space-based multi-source intelligence data, there are mainly structured, semi-structured, and unstructured data sources.

[0061] Given the characteristics of structured data sources, a direct mapping method is selected to directly map linked data or database data into an RDF (Resource Description Frame) diagram.

[0062] For unstructured data, it is divided into remote sensing image data and text data. Remote sensing image data is processed using target recognition algorithms to transform the image data into semantic information; text data is processed using natural language processing algorithms to extract entities and relationships from intelligence documents.

[0063] For semi-structured data, the main sources are open-source web information and interaction information from various information systems. For these types of data sources, based on different needs and tasks, a wrapper approach is used to package summarization, manual extraction, and automatic extraction into a single module for entity and relation extraction.

[0064] S300, perform knowledge fusion on the results of the knowledge extraction to obtain the target knowledge graph.

[0065] In this embodiment, knowledge fusion includes entity linking and knowledge merging.

[0066] Entity linking transforms data into an entity-annotated format (i.e., text), establishing a connection between the text and the knowledge base. Specifically, it maps all identical entities obtained to the same correct entity in the knowledge base. This includes the following steps:

[0067] a1. Determine whether there are identical or related entities in the existing knowledge base. Identical entities refer to entities that are exactly the same, while related entities refer to entities that are not exactly the same but have the same meaning.

[0068] a2. Obtain entity objects through entity extraction;

[0069] a3. Map the entity to the correct entity in the knowledge base.

[0070] Rule mining uses a path sorting method, which includes the following steps:

[0071] b1. Feature Selection: Relationship paths are selected, retaining those potentially useful for predicting the target relationship. To ensure efficiency, a random walk-based feature selection method is used. For a given relationship path, the accuracy and coverage of the path are calculated based on the random walk.

[0072] b2. Feature Calculation: After selecting useful relational paths, calculate feature values ​​for each entity.

[0073] b3. Relationship Classification: Based on training examples and their features, a classification model is trained for each target relationship. Using the trained model, the existence of a specific relationship between any two entities in the knowledge graph is predicted.

[0074] S400, embed a time model into the target knowledge graph and perform knowledge processing, completion and reasoning.

[0075] In this embodiment, time information is embedded into the knowledge graph to complete the knowledge graph obtained in the previous step that did not contain time information. Specifically:

[0076] The sequence learning-based approach first designs a sequence learning model to integrate temporal information into the representation vectors of entities or relations. Then, it uses an existing representation learning model to estimate the veracity of triples with temporal information, thereby completing the task of completing the temporal knowledge graph.

[0077] The features of the relation and timestamp constitute a relation sequence. Through a linear layer function, the relation and timestamp features are mapped into vectors of the same dimension. Then, the sequence vector is input into a long short-term memory (LSTM) network for encoding, and the relation representation vector that integrates time information is learned.

[0078] like Figure 4 As shown in the diagram, "Dock" represents a relation, and 220810 indicates August 10, 2022. The relation and time information are processed by an LSTM model to form a relation representation vector that incorporates time information. Finally, the triple (s, pseq, o) is completed using the scoring function of the TransE model. Here, s represents the head entity, o represents the tail entity, pseq represents the relation, and o represents the tail entity. esq Let represent a vector.

[0079] S500: Construct a time-series knowledge graph based on the results of step S400.

[0080] In this embodiment, information is completed and a time-series knowledge graph is constructed based on the completion and scoring functions in step S400.

[0081] A temporal knowledge graph is a set of quadruples. Formally speaking, given any set of entities E, a set of relations R, and a finite set of timestamps T, a temporal knowledge graph is a subset of the Cartesian product E×R×E×T.

[0082] In summary, the method for constructing a target time-series knowledge graph for space-based data in this embodiment of the invention can: (1) model multi-source heterogeneous intelligence data and construct a knowledge graph; (2) introduce time-series information into the knowledge graph, which can better describe intelligence information and assist command decision-making; (3) convert traditional static knowledge graphs into time-series knowledge graphs, thereby enhancing the semantic description capability of the target knowledge graph; and (4) use the representation sequence learning method of long short-term memory networks to convert time-series information into time vectors.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a target timing knowledge graph for space-based data, characterized in that, The construction method comprises: S100, constructing a target ontology, a radar ontology, a region ontology and an event ontology; S200, performing knowledge extraction according to the target ontology, the radar ontology, the region ontology and the event ontology; S300, performing knowledge fusion on a result of the knowledge extraction to obtain a target knowledge graph; S400, embedding a time model in the target knowledge graph; S500, constructing a time-series knowledge graph according to a result of step S400; In step S400: The time model is fused into a representation vector of an entity or a relation of a sequence learning model to estimate a real degree of a triad with time information for time knowledge graph completion; The sequence learning model comprises a relation sequence composed of a relation and a timestamp, the features of the relation and the timestamp are mapped into a same dimension vector through a linear layer function, and a long short-term memory network is used for encoding to learn a relation representation vector fused with time information; The information of the relation and the timestamp is formed into a relation representation vector fused with time information through an LSTM model, and a score function of a TransE model is used for completing triad information, the triad information being: s, pseq, o; wherein s is a head entity, o is a tail entity, and pseq is a relation. 2.The method of claim 1, wherein, In step S100: Local ontologies are constructed according to characteristics among data, and global ontologies are constructed according to commonalities among data.

3. The method of claim 1, wherein, In step S200: The knowledge extraction comprises entity extraction, attribute extraction and relation extraction.

4. The method of claim 3, wherein, In step S200: For structured data, a direct mapping method is adopted to map data of link data or a database into a resource description framework graph; For unstructured data, the unstructured data is divided into image data and text data, the image data is processed by using a target recognition algorithm to convert the image data into semantic information, and the text data is processed by using a natural language processing algorithm to extract entities and relations in an intelligence file; For semi-structured data, induction, manual extraction and automatic extraction are packaged into modules to extract entities and relations.

5. The method of claim 1, wherein, In step S300, the knowledge fusion comprises entity linking and knowledge merging.

6. The method of claim 5, wherein, The entity linking comprises judging whether an entity in an existing knowledge graph library has an entity with the same meaning, obtaining an entity object through entity extraction, and corresponding the entity to a correct entity in the knowledge graph library; The knowledge merging comprises selecting a relation path and retaining a relation path potentially related to a predicted target relation, calculating a feature value for each entity, and training a classification model for each target based on a training sample and the feature value.

7. The method of claim 1, wherein, In step S500, a time-series knowledge graph is constructed according to the completion and the score function.

Images