An ontology-driven spatiotemporal knowledge graph automatic construction method

Abstract

This invention discloses an ontology-driven automated construction method for spatiotemporal knowledge graphs. The method includes: performing a weighted fusion evaluation of information entropy and entity density on a set of text blocks to obtain a high-entropy text block set; constructing an ontology structure based on the high-entropy text block set using a large language model, including constraint rules such as entity type, entity attributes, and entity relationships; using the ontology structure to construct a prompt word traversal text block set oriented towards the large language model, and extracting a set of entity instances conforming to the ontology structure based on the ontology structure; and constructing a multi-level spatiotemporal knowledge graph using the entity instance set. This invention accurately filters text blocks rich in semantic and entity information through an information entropy-driven intelligent text selection mechanism, and enhances the adaptability and domain generalization ability of the ontology structure by combining it with a large language model, achieving rapid construction of large-scale spatiotemporal knowledge graphs and providing high-quality structured knowledge support for applications such as intelligent question answering and decision support.

Description

Technical Field

[0001] This invention relates to the field of ontology-driven spatiotemporal knowledge graph construction, and more particularly to an ontology-driven automated spatiotemporal knowledge graph construction method. Background Technology

[0002] In the current context of profound digital transformation, massive amounts of unstructured data (text, images, voice, etc.) contain rich knowledge value, a significant proportion of which possess substantial spatiotemporal attributes. Event entities are often accompanied by a clear time dimension (such as time of occurrence and duration), while geographic entities possess specific spatial distribution characteristics (such as location coordinates and geographical extent). Extracting structured knowledge from this type of spatiotemporally correlated data and constructing computable and reasonable spatiotemporal knowledge graphs is an important research direction in the field of knowledge engineering, and a core foundation for realizing data value transformation and supporting intelligent applications.

[0003] In existing technologies, ontology-based knowledge graph construction methods typically require domain experts to predefine a complete ontology structure (including entity types, attributes, relationships, etc.) before extracting entity and relationship knowledge based on this pre-defined structure. However, this construction model has several inherent drawbacks: First, the definition and design of the ontology are highly dependent on expert experience, resulting in high labor costs and a lengthy construction cycle (usually requiring months or even years). Furthermore, expert subjectivity can easily lead to biases in the ontology structure, affecting the accuracy of subsequent knowledge extraction. Second, for the specific domain of spatiotemporal knowledge graphs, existing technologies lack dedicated ontology adaptation mechanisms, making it difficult to effectively express the core elements unique to spatiotemporal knowledge, such as spatial topological relationships (e.g., containment, proximity), temporal evolution rules (e.g., sequential, simultaneous), and spatiotemporal scale constraints (e.g., temporal granularity, spatial precision). This results in the constructed knowledge graph failing to accurately depict the spatiotemporal relationships between entities. Third, existing technologies have not formed a standardized spatiotemporal semantic expression system, making it difficult to support cross-system knowledge sharing and interoperability.

[0004] In the knowledge extraction stage, existing technologies mostly use uniform text segmentation or random sampling to process input text, without fully considering the differences in effective information content among different text fragments. Due to the uneven distribution of information density in unstructured text (e.g., some fragments are rich in entities and concepts, while others are merely redundant descriptions), existing methods cannot guarantee that the selected text fragments can effectively support the extraction of high-quality ontology, thus limiting the knowledge coverage and accuracy of the extraction results and failing to meet the needs of constructing large-scale spatiotemporal knowledge graphs.

[0005] In terms of spatiotemporal semantic standardization, existing technologies mostly use natural language descriptions to express spatiotemporal information, without establishing clear temporal semantic standards (such as not introducing the W3C-developed OWL Time ontology standard), and without converting spatial types (such as points, lines, and surfaces) into universal spatial geometric representations (such as the WKT geometric markup language specified by the OGC GeoSPARQL standard). This results in inconsistent semantics of spatiotemporal knowledge graphs built by different systems, making it impossible to achieve cross-system knowledge interaction and collaborative reasoning.

[0006] In terms of relation extraction, existing technologies mainly focus on general semantic relations between entities (such as "belonging to" and "cooperating"). There is a lack of mature solutions for the extraction of the core temporal relations (such as before, after, simultaneously, during, etc.) and spatial relations (such as containing, adjacent, located, etc.) of spatiotemporal knowledge graphs. This makes it difficult to fully characterize the multidimensional spatiotemporal relationships between entities, resulting in the constructed knowledge graphs only being able to achieve static knowledge storage and unable to support dynamic spatiotemporal reasoning and complex queries.

[0007] Based on the above analysis, existing technologies have significant shortcomings in terms of ontology adaptive construction, improved accuracy of knowledge extraction, standardized expression of spatiotemporal semantics, and joint extraction of spatiotemporal relationships. They cannot meet the needs of constructing large-scale, high-quality spatiotemporal knowledge graphs in practical production applications, and a new technical solution is urgently needed to address these shortcomings. Summary of the Invention

[0008] The purpose of this invention is to provide an ontology-driven method for the automated construction of spatiotemporal knowledge graphs, realizing a complete technical closed loop from unstructured documents to knowledge graphs constrained by ontology specifications, improving the construction efficiency, standardization, and generalization ability of spatiotemporal knowledge graphs, and providing a high-quality structured knowledge foundation for upper-layer intelligent applications.

[0009] The objective of this invention is achieved through the following technical solution: An ontology-driven method for automated construction of spatiotemporal knowledge graphs, comprising the following steps: S1. Divide the input text into blocks to obtain a set of text blocks. Perform information entropy and entity density weighted fusion evaluation on each text block in the set of text blocks and select the n text blocks with the highest evaluation to form a high-entropy text block set. S2. Based on a high-entropy text block set, an ontology structure including entity type, entity attribute, and entity relationship is constructed using a large language model. The entity type includes general fixed class, spatiotemporal fixed class and domain-specific class. The entity attributes are constructed with the entity attributes corresponding to the entity type. The entity relationship is constructed with the entity attributes to build the association relationship. S3. Constructing prompt words for large language models using ontology structures. Traverse the text block set and extract the entity instance set that conforms to the ontology structure based on the ontology structure. The entity instance set includes entity instances and entity relations. Entity instances are entities corresponding to entity types and / or entity attributes. Entity relations include general semantic relations, temporal relations and spatial relations. Build a multi-level spatiotemporal knowledge graph using the entity instance set.

[0010] To better implement this invention, in method S1, a high-entropy text block set... The method to obtain it is as follows: S11, Information Entropy of Text in a Text Block The expression is as follows: , Words that are meaningless words removed from the text The probability of a word's frequency relative to the input text, where m is the total number of words in the text block after removing meaningless words; S12, Entity density in text blocks The expression is as follows: , This represents the number of entities matched by the regular expression within the text block. This represents the total number of characters in the text block. The expansion coefficient is used to correct missed detections in regular expression matching; S13. Perform a weighted fusion evaluation of the information entropy and entity density of the text block, as shown in the following expression: , For the final evaluation score, , These are the normalized information entropy and entity density, respectively. For information entropy weights; S14. Select the top n text blocks with the highest or highest evaluation scores to form a high-entropy text block set. , , Represents a collection of text blocks. Represents text block i.

[0011] Preferably, in method S2, the ontological structure , , For entity type, For entity attributes, For entity relationships; entity type The types are constructed according to a hierarchical architecture, entity types The first level of types includes general fixed classes. Fixed-time and spacetime categories Domain-specific categories General fixed classes include , The second level of events includes two fixed-level classes: spatiotemporal fixed classes, which include the second level of time and space; and domain-specific classes, which are domain-level classes constructed using the Large Language Model (LLM) according to domain classification. Domain-specific classes include several domain levels. Entity types... The set is: .

[0012] Preferably, the entity attributes Including the basic attributes corresponding to the entity type Basic attributes Includes descriptive attributes and group attributes; the spatiotemporal fixed class The basic attributes corresponding to time The necessary attribute fields include time start, time end, time duration, and time type. Time type includes time point, time interval, duration, and general time entity; the spatiotemporal fixed class Basic properties corresponding to the middle space The necessary attribute fields include spatial type and spatial location information. Spatial types include points, lines, surfaces, and complex sets.

[0013] Preferably, the entity relationship The expression is as follows: , Represents a relational description. Represents relational keywords, and These represent the starting entities of the relationship.

[0014] Preferably, in method S2, the ontology structure is automatically extracted from the high-entropy text block set using a large language model (LLM), and type deduplication and attribute field fusion are performed during the ontology structure extraction process.

[0015] Preferably, in method S3, a set of entity instances conforming to the ontology structure is extracted based on ontology structure-driven extraction. The expression is as follows: , This indicates that entity instances are deduplicated and their attributes are merged. For merging processing, Large Language Model (LLM) utilizes cue words Iterate through the collection of text blocks to extract entity instances.

[0016] Preferably, in method S3, the general semantic relations include belonging, participating, and cooperating; the temporal relations include after, before, simultaneously, and during; and the spatial relations include several spatial topological relations, including containing, adjacent, and located.

[0017] Preferably, in method S3, entity instances in the multi-level spatiotemporal knowledge graph are grouped according to the type hierarchy of entity types to form multiple levels, each level having a corresponding entity relationship, and the spatiotemporal knowledge graph sequentially constructs several levels of entity relationships according to the type hierarchy of entity types.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention uses an intelligent text selection mechanism driven by information entropy to accurately filter text blocks rich in semantic and entity information. Combined with the powerful semantic understanding capabilities of the large language model, the ontology structure construction is shortened to "minutes". Experimental verification shows that the ontology completeness rate is over 90%, the entity type coverage rate is about 35% higher than the existing technology, and the knowledge extraction efficiency is over 60%. It can meet the needs of rapid construction of large-scale spatiotemporal knowledge graphs and improve the efficiency of knowledge graph construction.

[0019] (2) This invention has the ability to express multidimensional spatiotemporal relationships and supports dynamic knowledge reasoning: It innovatively realizes a three-dimensional joint extraction framework of "general relationship + temporal relationship + spatial relationship", which can comprehensively depict the multidimensional relationship between entities and meet the complex spatiotemporal query needs such as "events that occur in a specific area within a certain time period" and "the spatial location of an entity changes at different times". It realizes the leap from static knowledge storage to dynamic knowledge reasoning and provides high-quality structured knowledge support for upper-level intelligent applications (such as intelligent question answering and decision support).

[0020] (3) The group mechanism of this invention enhances the knowledge organization capability. It automatically adds classification information to entities through the "group" attribute, and creates exclusive entity instances for high-frequency groups. It constructs the association relationship between groups and entities, making the structure of the knowledge graph clearer and the organization more reasonable. It is convenient for users to quickly retrieve and understand the association between entities, and further improves the practicality of the knowledge graph. This invention realizes the standardization of spatiotemporal semantics and supports cross-system interoperability. It introduces the W3C OWL Time temporal ontology standard and the OGC GeoSPARQL spatial geometry standard into the construction process of spatiotemporal knowledge graph, standardizes the definition of the types and attributes of time and space entities, realizes the formal expression of temporal semantics and spatial semantics, solves the problem of inconsistent spatiotemporal semantics in the existing technology, supports cross-system knowledge sharing and interoperability, and provides a foundation for spatiotemporal reasoning of multi-system collaboration.

[0021] (4) The present invention significantly improves the self - adaptability and domain generalization ability of the ontology structure. It adopts a two - layer ontology mechanism of "forced retention + adaptive expansion". Five core entity types (person, organization, event, time, space) are retained to ensure the stability of the ontology structure. At the same time, the remaining extended types are dynamically discovered through the large - language model, enabling the system to automatically adapt to different - domain and different - type document contents, greatly reducing the dependence on domain experts and enhancing the generalization ability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a method flow chart of the automated construction method of the spatio - temporal knowledge graph of the present invention; Figure 2 It is a schematic diagram of the method principle of the automated construction method of the spatio - temporal knowledge graph in the embodiment; Figure 3 It is a schematic diagram of the principle of the information entropy and entity density weighted fusion evaluation - driven processing in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The present invention will be further described in detail below in conjunction with the embodiments: Embodiment As Figure 1 、 Figure 2 shown, an automated construction method of a spatio - temporal knowledge graph based on ontology drive, the method includes: S1. The input text is segmented (the present invention uses a sliding window for segmentation) to obtain a set of text blocks , and each text block in the set of text blocks is subjected to information entropy and entity density weighted fusion evaluation processing, and n text blocks with the top evaluation are selected to form a high - entropy text block set. The present invention is applicable to the structured knowledge extraction and graph construction of various unstructured data containing spatio - temporal attributes (such as news texts, government affairs files, geographical survey reports, event records, etc.), and can be widely applied to upper - layer application scenarios such as intelligent question - answering, decision - making assistance, spatio - temporal reasoning, and geographical information services. In some embodiments, as Figure 3 shown, the method for obtaining the high - entropy text block set is as follows: S11. The information entropy (used to measure the semantic diversity or information content) of the text in the text block is expressed as follows: , is the word in the text after removing meaningless words (meaningless words in the text, including auxiliary words, modal particles, conjunctions, adverbs, punctuation marks. Examples of auxiliary words are as follows: de, di, de, le, zhe, guo; examples of modal particles are as follows: ne, ma, ba, a, ya, etc.; examples of conjunctions are as follows: he, yu, ji, huo, dan, ze, er, ze, etc.; examples of adverbs are as follows: hen, fei chang, ji qi, zhi, jin, etc.) The word frequency probability (i.e., the number of times the word appears / the total number of words) of the input text or text block, where m is the total number of words in the text block after removing meaningless words. The higher the entropy value, the more evenly the vocabulary distribution in the text, the richer the semantics, the more new concepts / entities it contains, and the greater its information content. The purpose is to facilitate the subsequent screening of text blocks with rich semantics and high information density, and to avoid using low-value text to interfere with ontology extraction.

[0024] S12, Entity density in text blocks The expression (used to measure the richness of entities) is as follows: , This represents the number of entities (such as people, events, time, and space) matched by regular expressions in the text block. This represents the total number of characters in the text block. The expansion coefficient used to correct missed detections in regularity matching (to more accurately reflect entity richness, the example value in this embodiment is [value]). The denser the entities contained in a text block, the more suitable it is for extracting entities, attributes, and relationships from an ontology. This facilitates combining it with information entropy to compensate for its deficiencies and ensures that the selected text blocks are rich in entities and suitable for ontology construction.

[0025] S13. Perform a weighted fusion evaluation of the text block's information entropy (i.e., semantic information content) and entity density (i.e., entity richness). The final evaluation score expression for each text block is as follows: , For the final evaluation score, , These are the normalized information entropy and entity density, respectively, unified to the [0,1] interval to eliminate dimensional differences; The information entropy weight (in this example, the value is taken as...) That is, information entropy accounts for 65% of the weight, and entity density accounts for 35% of the weight, which may be adjusted as needed.

[0026] S14. Select the top n text blocks with the highest or highest evaluation scores to form a high-entropy text block set. , , Represents a collection of text blocks. Let i represent text block. The information entropy-driven strategy of this invention can adaptively identify text blocks rich in semantic information and entity content, ensuring both information richness and efficiency in entity density and ontology structure extraction.

[0027] S2. Based on a high-entropy text block set, an ontology structure including entity types, entity attributes, and entity relationships is constructed using a large language model. Entity types include general fixed classes, spatiotemporal fixed classes, and domain-specific classes. Entity attributes are constructed according to the entity types, and entity relationships are constructed using entity attributes to establish associations. Preferably, as follows... Figure 2 As shown, this invention utilizes a Large Language Model (LLM) to automatically extract ontology structure from high-entropy text block sets, performing type deduplication and attribute field fusion during the extraction process. The ontology structure according to this invention... Constructing dynamic ontology extraction prompts for large language models (The prompts include ontology structure definitions, entity type constraints, attribute constraints, relation constraints, and text block semantic parsing requirements.) Traverse the high-entropy text block set. Guide the large language model to process each text block Perform semantic parsing to extract the corresponding entity type, attributes, and relationships, and then use functions to... The extracted results are then deduplicated by type, merged by attribute fields, and standardized by relational normalization to obtain the complete dynamic ontology structure Ω, as shown in the following expression: , This is the final extracted dynamic ontology structure, used to constrain the subsequent extraction of entity instances and relationships. This represents the ontology processing function, whose core functions include entity type deduplication (removing duplicate entity types), attribute field fusion (integrating different attribute fields of the same entity type to ensure attribute integrity), and relation normalization (standardizing and organizing the extracted relations according to the structure of R). This represents the inference function of a large language model. It takes the prompt word "prompt" and the text block σ as input and outputs the entity type, attributes, and preliminary relationship extraction results corresponding to the text block. This represents the set of all high-entropy text blocks. The extracted results are merged to ensure the integrity of the ontology structure. This invention achieves automated dynamic ontology extraction through a two-layer mechanism of predefined ontology constraints and self-discovery. It ensures the stability of the ontology by forcibly retaining five core entity types, and dynamically expands entity types and attributes through a large language model, taking into account the flexibility of domain knowledge. This solves the problems of rigid ontology structure and poor adaptability in existing technologies.

[0028] In some embodiments, the body structure of the present invention , , For entity type, For entity attributes, This refers to entity relationships. Entity types include general fixed classes, spatiotemporal fixed classes, and domain-specific classes. The types are constructed according to a hierarchical architecture, entity types The first level of types includes general fixed classes. Fixed-time and spacetime categories Domain-specific categories General fixed class include , The second level of the event Fixed-time and spacetime class Including the second level of time and space, , For example: January 1st or 3 days in 2023. For example: No. 777, Zhongguan West Road. Domain-specific category. To utilize the Large Language Model (LLM) to construct a domain hierarchy based on domain classification, domain-specific classes include several domain levels. For example, in the domain of government archives, there are second-level levels such as policies, approval items, and regulatory departments. Entity types. The set is: .

[0029] This invention also constructs an entity attribute set, defining corresponding attribute fields for different types of entities, divided into general attributes (i.e., basic attributes) and spatiotemporal entity-specific attributes. This ensures the standardization and completeness of attribute descriptions, and entity attributes... Including the basic attributes corresponding to the entity type Other exclusive attributes (including attributes specific to spacetime entities), basic attributes Including descriptive attributes and group attributes, , This refers to the description of an entity, such as "the inventor of the technology who is engaged in big data processing research" for a person, or "a research organization that focuses on the research and application of spatiotemporal big data technology" for an organization. This refers to the grouping label of an entity, which is used to construct subsequent entity relationships. For example, the grouping label for an entity of type "time" is "time", and the grouping label for an entity of type "space" is "space".

[0030] Fixed spacetime class Necessary attributes for time-type entities Time type, start time, end time, and duration. , and These represent "start of time" and "end of time" respectively; Indicates "duration of time". The time type is indicated. As shown in Table 1, the time types in this invention adopt the W3C OWL Time ontology standard to define a set of time types (Temporal types), including time points, time intervals, durations, and general time entities. .

[0031]

[0032] Fixed spacetime class Necessary attributes corresponding to medium-space type entities Spatial type and spatial location information, , Indicates "space type"; This represents "spatial coordinates," i.e., spatial location information. As shown in Table 2, this invention uses the WKT geometric markup language, developed by the Open Geospatial Consortium (OGC) GeoSPARQL standard, to define a set of spatial types, including points, lines, polygons, and other complex sets. .

[0033]

[0034] The entity attribute set of this invention also dynamically extracts attribute fields corresponding to extended entity types based on the Large Language Model (LLM). For example, the extended attributes of the "mountain" type are "altitude" and "length across", and the extended attributes of the "policy" type are "release time" and "scope of implementation".

[0035] Entity Relationship The expression is as follows: , Relationship description refers to the textual description of the relationship between entities, such as "Zhang San is located in Ningbo City, Zhejiang Province"; Relational keywords are categorized into three types: general relations, temporal relations, and spatial relations. and These represent the starting entity and the target entity of the relationship, respectively.

[0036] S3, such as Figure 2 As shown, prompt words for large language models are constructed using ontology structures. Traverse the text block set and extract the entity instance set that conforms to the ontology structure based on the ontology structure. The entity instance set includes entity instances and entity relations. Entity instances are entities corresponding to entity types and / or entity attributes. Entity relations include general semantic relations, temporal relations and spatial relations. Build a multi-level spatiotemporal knowledge graph using the entity instance set.

[0037] Extracting a set of entity instances that conform to the ontology structure based on ontology structure. The expression is as follows: , This indicates that entity instances are deduplicated and their attributes are merged. For merging processing, Large Language Model (LLM) utilizes cue words Traversing the collection of text blocks Perform entity instance extraction processing. The extracted spatiotemporal type entity instances are then validated and post-processed as follows: Time entity type validation: Ensure that all attribute field values ​​conform to the OWL Time standard; check the attribute fields of each time entity instance one by one to ensure that each attribute value conforms to the OWL Time standard. For example, INSTANT type time entities must contain an exact time point, INTERVAL type time entities must contain both start and end time information, and DURATION type time entities must contain the duration, and all attribute values ​​must conform to standardized format requirements (e.g., time format is uniformly "YYYY-MM-DD,HH:MM:SS", duration format is uniformly "X days X hours X minutes"). Entity instances that do not conform to the standard are corrected by re-parsed text using a large language model; those that cannot be corrected are discarded.

[0038] Spatial entity type validation process: Ensure that the values ​​of each attribute field conform to the OGC GeoSPARQL standard; check the attribute fields of each spatial entity instance one by one to ensure that the values ​​of each attribute conform to the OGC GeoSPARQL standard. For example, for spatial entities of type POINT... Spatial entities must conform to WKT format "POINT" and POLYGON type. Entities must conform to the WKT format "POLYGON". For non-compliant entity instances, coordinate correction will be performed using geographic information tools; those that cannot be corrected will be removed.

[0039] Spatiotemporal entity type validation processing: This ensures the "Group" field value is correct (limited to "Time" or "Space"), and that the "Group" field value is correct for all spatiotemporal type entity instances. Specifically, the group label for time entities is limited to "Time," and the group label for space entities is limited to "Space," avoiding group label confusion and providing accurate classification criteria for subsequent relationship extraction. Furthermore, this invention analyzes the distribution of all "Group" attribute values ​​in the entity instance set, creating additional entity instances with the same name for "Groups" that appear multiple times. For example, if a group of entities share a certain "Group" value (such as Company A), an entity node named "Group" is created, laying the groundwork for establishing relationships between "Groups" and entities later.

[0040] In some embodiments, entity relationships include general semantic relationships (also known as general relationships), temporal relationships, and spatial relationships. Preferably, general semantic relationships (also known as general relationships) include belonging, participating, and cooperating; temporal relationships include after, before, simultaneously, and during; and spatial relationships include several spatial topological relationships, including containing, adjacent, and located.

[0041] Iterate through the entire collection of text blocks. ,Will and each text block Input a large language model, guide the large language model to extract three types of relational information contained in the text block based on the entity instance set I, and then use a function to... The extracted results are then deduplicated, merged, and standardized to obtain a complete set of entity relations R, expressed as follows: R represents the final set of entity relations extracted. This represents relation processing functions, which include relation deduplication (removing duplicate relations), relation fusion (integrating different relations between the same pair of entities and supplementing the relation description), and relation normalization (standardizing and organizing the extracted relations to ensure the accuracy of relation type, starting entity, and target entity). Represents the relational inference function of a large language model, with input... and text blocks Output the preliminary extraction results of the three relationships contained in the text block; This indicates that the relationship extraction results of all text blocks are merged to ensure the comprehensiveness of the relationships.

[0042] This invention also includes the following relation verification: checking the rationality of each relation R one by one to ensure the starting entity and target entity All belong to entity instance set I, relation type Conforming to the definition of this invention, relationship description Consistent with text semantics. For unreasonable relationships (such as non-existent entities or incorrect relationship types), the text is re-parsed and corrected using a large language model; those that cannot be corrected are discarded to ensure the accuracy of the relationship set. This invention achieves full automation from unstructured text input to spatiotemporal knowledge graph construction through the above steps, forming a logical closed loop of "text segmentation → ontology extraction → entity extraction → relationship extraction." All steps are based on clear technical specifications and formula constraints, ensuring the feasibility and repeatability of the technical solution.

[0043] A multi-level spatiotemporal knowledge graph is constructed using a set of entity instances. Preferably, in the multi-level spatiotemporal knowledge graph, entity instances are grouped according to the type hierarchy of entity types to form multiple levels, each level having a corresponding entity relationship. The spatiotemporal knowledge graph constructs several levels of entity relationships sequentially according to the type hierarchy of entity types.

[0044] This invention improves the efficiency of ontology development: Addressing the issues of existing ontology design relying on domain experts, high labor costs, and long construction cycles, this invention shortens the ontology construction cycle from "monthly" to "minutes" through an automated ontology discovery and entity relation extraction mechanism, while simultaneously increasing the ontology completeness rate to over 90%, reducing reliance on domain experts. This invention achieves spatiotemporal semantic standardization: Addressing the lack of standardized temporal semantic specifications and the failure to convert spatial types into universal geometric representations in existing technologies, this invention introduces the OWL Time temporal ontology standard and the OGC GeoSPARQL spatial geometric standard into the spatiotemporal knowledge graph construction process for the first time, achieving standardized and formalized expression of temporal and spatial semantics, supporting cross-system interoperability and spatiotemporal reasoning computation. This invention enhances system adaptability: Addressing the rigid ontology structures and difficulty in adapting to documents from different domains in existing technologies, this invention, based on the complex semantic understanding capabilities of large language models, creates a flexible adaptive ontology structure pattern, enabling the system to automatically identify and adapt to document content from different domains (such as government affairs, economics, and science and technology), improving the universality and generalization ability of the knowledge graph. This invention can support knowledge-based intelligent applications: by constructing a high-quality, standardized spatiotemporal knowledge graph, it provides a structured knowledge foundation for upper-level applications such as question-answering systems, decision support, and knowledge reasoning, realizing a leap from static knowledge storage to dynamic knowledge reasoning, and meeting the needs of complex spatiotemporal queries and intelligent analysis.

[0045] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An ontology-driven method for automated construction of spatiotemporal knowledge graphs, characterized in that: The methods include: S1. Divide the input text into blocks to obtain a set of text blocks. Perform information entropy and entity density weighted fusion evaluation on each text block in the set of text blocks and select the n text blocks with the highest evaluation to form a high-entropy text block set. S2. Based on a high-entropy text block set, an ontology structure including entity type, entity attribute, and entity relationship is constructed using a large language model. The entity type includes general fixed class, spatiotemporal fixed class and domain-specific class. The entity attributes are constructed with the entity attributes corresponding to the entity type. The entity relationship is constructed with the entity attributes to build the association relationship. S3. Constructing prompt words for large language models using ontology structures. Traverse the text block set and extract the entity instance set that conforms to the ontology structure based on the ontology structure. The entity instance set includes entity instances and entity relations. Entity instances are entities corresponding to entity types and / or entity attributes. Entity relations include general semantic relations, temporal relations and spatial relations. Build a multi-level spatiotemporal knowledge graph using the entity instance set.

2. The method for automatically constructing a spatiotemporal knowledge graph based on ontology-driven approach according to claim 1, characterized in that: In method S1, the high-entropy text block set The method to obtain it is as follows: S11, Information Entropy of Text in a Text Block The expression is as follows: , Words that are meaningless words removed from the text The probability of a word's frequency relative to the input text, where m is the total number of words in the text block after removing meaningless words; S12, Entity density in text blocks The expression is as follows: , This represents the number of entities matched by the regular expression within the text block. This represents the total number of characters in the text block. The expansion coefficient is used to correct missed detections in regular expression matching; S13. Perform a weighted fusion evaluation of the information entropy and entity density of the text block, as shown in the following expression: , For the final evaluation score, , These are the normalized information entropy and entity density, respectively. For information entropy weights; S14. Select the top n text blocks with the highest or highest evaluation scores to form a high-entropy text block set. , , Represents a collection of text blocks. Represents text block i.

3. The method for automatically constructing a spatiotemporal knowledge graph based on ontology-driven approach according to claim 1, characterized in that: In method S2, the ontology structure , , For entity type, For entity attributes, For entity relationships; entity type The types are constructed according to a hierarchical architecture, entity types The first level of types includes general fixed classes. Fixed-time and spacetime categories Domain-specific categories General fixed classes include , The second level of events includes the spatiotemporal fixed class, which includes the second level of time and space; the domain-specific class is the domain level constructed by using the Large Language Model (LLM) according to domain classification; and the domain-specific class includes several domain levels. Entity type The set is: .

4. The ontology-driven automated construction method for spatiotemporal knowledge graphs according to claim 3, characterized in that: The entity attributes Including the basic attributes corresponding to the entity type Basic attributes Includes descriptive attributes and group attributes; the spatiotemporal fixed class The basic attributes corresponding to time The necessary attribute fields include time start, time end, time duration, and time type. Time type includes time point, time interval, duration, and general time entity; the spatiotemporal fixed class Basic properties corresponding to the middle space The necessary attribute fields include spatial type and spatial location information. Spatial types include points, lines, surfaces, and complex sets.

5. The ontology-driven automated construction method for spatiotemporal knowledge graphs according to claim 3 or 4, characterized in that: The entity relationship The expression is as follows: , Represents a relational description. Represents relational keywords, and These represent the starting entities of the relationship.

6. The ontology-driven automated construction method for spatiotemporal knowledge graphs according to claim 3 or 4, characterized in that: In method S2, the large language model LLM is used to automatically extract the ontology structure from the high-entropy text block set. During the extraction of the ontology structure, type deduplication and attribute field fusion are performed.

7. The method for automatically constructing a spatiotemporal knowledge graph based on ontology-driven approach according to claim 2, characterized in that: In method S3, a set of entity instances conforming to the ontology structure is extracted based on ontology structure-driven extraction. The expression is as follows: , This indicates that entity instances are deduplicated and their attributes are merged. For merging processing, Large Language Model (LLM) utilizes cue words Iterate through the collection of text blocks to extract entity instances.

8. The method for automatically constructing a spatiotemporal knowledge graph based on ontology-driven approach according to claim 1, characterized in that: In method S3, the general semantic relations include belonging, participating, and cooperating; the temporal relations include after, before, simultaneously, and during; and the spatial relations include several spatial topological relations, including containing, adjacent, and located.

9. The ontology-driven automated construction method for spatiotemporal knowledge graphs according to claim 3, characterized in that: In method S3, entity instances in the multi-level spatiotemporal knowledge graph are grouped according to the type hierarchy of entity types to form multiple levels. Each level has a corresponding entity relationship. The spatiotemporal knowledge graph constructs several levels of entity relationships in sequence according to the type hierarchy of entity types.

Images