Ontology subgraph prompts injected multi-source heterogeneous information collaborative research and judgment method and system

Abstract

The application provides a multi-source heterogeneous information collaborative research and judgment method and system of body subgraph prompt injection, the application constructs an intelligence field ontology containing five categories of dimensions of entity, relationship, event, space-time and credibility as a unified semantic framework, carries out semantic element extraction and ontology mapping on multi-source heterogeneous information; cross-source entity alignment is realized based on ontology structure constraint, and the confidence degree distribution of conflict information is output through credibility weighted arbitration; relevant ontology subgraph is dynamically searched as a structured hard constraint prompt injection large language model, and the conclusion related ontology node is forced to generate an evidence chain; finally, the research and judgment result is written back to the ontology to drive its continuous evolution. The application solves the problems of cross-source semantic gap, weak conflict resolution capability, non-traceable research and judgment conclusion and large language model illusion in the prior art, realizes semantic fusion, intelligent research and traceable management of multi-source heterogeneous information, and significantly improves the accuracy and reliability of intelligence analysis.

Description

Technical Field

[0001] This invention relates to the field of intelligence analysis technology, specifically to a method and system that combines intelligence domain ontology with large language model generation capabilities to perform semantic fusion, conflict resolution, and intelligent analysis of multi-source heterogeneous intelligence. Background Technology

[0002] Modern intelligence sources are highly diverse, encompassing various heterogeneous forms such as text reports, images and videos, signal interception, social media, and open-source intelligence. Existing intelligence fusion and analysis technologies suffer from the following technical problems and shortcomings in engineering practice: First, there is a semantic gap. Different intelligence sources use different terminology systems, entity naming, and event description frameworks. The same entity may appear under different names in different sources, and the same event may be described using different element frameworks. Traditional fusion methods based on keyword matching and template rules have shallow semantic understanding and struggle to identify the same entity and event across sources, leading to information fragmentation.

[0003] Secondly, conflict resolution is weak. Multi-source intelligence often presents contradictory or inconsistent confidence levels. For example, different sources may give conflicting judgments about the location and attribution of the same target. Existing methods lack explicit confidence modeling and conflict arbitration mechanisms, often simply taking the latest or majority opinion, thus losing crucial information such as source reliability and time decay.

[0004] Third, the analysis lacks traceability. Directly using large language models for intelligence analysis suffers from serious illusions, lack of domain knowledge constraints, and a lack of evidence chains for conclusions. The analysis conclusions "cannot explain where the basis comes from," making it difficult to meet the requirements of traceability and accountability for intelligence products.

[0005] Existing technologies have not yet provided a method for fusion analysis that uses intelligence ontology as a semantic skeleton, deep constraints, and enhances generative models. Summary of the Invention

[0006] The purpose of this invention is to address the problems of cross-source semantic gap, weak conflict resolution capability, untraceable judgment conclusions, and large language model illusion in existing multi-source heterogeneous intelligence fusion and judgment technologies. It provides a multi-source heterogeneous intelligence fusion and judgment method and system that uses intelligence domain ontology as semantic skeleton and dynamically enhances the large language model generation process with ontology subgraphs.

[0007] According to one objective of the present invention, the present invention provides a method for collaborative analysis of multi-source heterogeneous intelligence injected through ontology subgraph hints, comprising the following steps: S1: Construct an intelligence domain ontology containing five dimensions as a unified semantic framework; S2: Preprocess and extract semantic elements from multi-source heterogeneous intelligence, map each source intelligence into an ontology individual and record metadata; S3: Perform cross-source entity alignment based on the intelligence domain ontology, and perform confidence-weighted arbitration on conflict intelligence to output the confidence distribution of each candidate value; S4: For the analysis task, relevant ontology subgraphs are dynamically retrieved from the intelligence domain ontology as structured hard constraint hints and injected into the large language model to generate an analysis conclusion with a chain of evidence. S5: Write the analysis conclusions and newly discovered entities and relationships back to the intelligence domain ontology to form a generation-writeback closed loop.

[0008] Further, in step S1, the intelligence domain ontology includes: The entity sub-entity includes four entity types: people, organizations, equipment, and locations. The relational sub-ontology includes four types of relations: subordinate, cooperative, adversarial, and derived. The event sub-entity contains three types of attributes: event type, event elements, and event sequence. The spatiotemporal sub-ontology includes three types of spatiotemporal attributes: time interval, geographic coordinates, and trajectory. The credibility sub-ontology includes three types of credibility attributes: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

[0009] Furthermore, in step S3, the specific process of cross-source entity alignment is as follows: Calculate the alignment similarity between two candidate ontology individuals a and b: sim(a,b)=μ sim attr (a,b)+ν sim struct (N(a),N(b)); Where, sim attr For attribute similarity, sim struct For the relational neighborhood N ( The structural similarity is μ+ν=1. When sim(a,b) exceeds the preset alignment threshold, the two candidate ontology individuals are determined to be the same entity.

[0010] Furthermore, in step S3, the specific process of the credibility-weighted arbitration is as follows: For multiple pieces of information r that point to the same entity or event but have contradictory values k Calculate the support weight of each piece of intelligence for the candidate value x: w k =ρ(src k ) c k e -λ(t0-tk) ; Where t0 is the judgment time, t k For intelligence r k timestamp; Calculate the overall confidence level of the candidate value x: ; Output all candidate values ​​and their corresponding overall confidence scores to form a confidence score distribution.

[0011] Furthermore, the specific process of step S4 is as follows: Based on the assessment task and entity alignment results, relevant ontology subgraphs are retrieved from the intelligence domain ontology; Structured hard constraint hints are constructed based on the ontology subgraph and confidence distribution; The structured hard constraint hints are injected into the large language model to generate a judgment draft; The draft analysis is validated by checking whether the entity types and relationship types appearing in the draft belong to the set defined by the intelligence domain ontology, and whether each conclusion in the draft is associated with an ontology node. If any content violates the constraints, the violation information will be fed back to the large language model and a regeneration will be triggered until the draft meets all constraints or the maximum number of retries is reached. Output the final judgment conclusion, which includes the judgment content, evidence chain, and confidence level distribution.

[0012] According to another objective of the present invention, the present invention provides a multi-source heterogeneous intelligence collaborative analysis system based on the above-described method for ontology subgraph cue injection, comprising: The ontology construction module is used to construct an intelligence domain ontology containing five dimensions, serving as a unified semantic skeleton; The intelligence preprocessing module is used to preprocess multi-source heterogeneous intelligence and extract semantic elements, map each source intelligence into an ontology individual and record metadata. The alignment and arbitration module is used to perform cross-source entity alignment based on the intelligence domain ontology, and to perform confidence-weighted arbitration on conflict intelligence, outputting the confidence distribution of each candidate value; The OAG analysis engine is used to dynamically retrieve relevant ontology subgraphs from the intelligence domain ontology as structured hard constraint hints for analysis tasks and inject them into the large language model to generate analysis conclusions with evidence chains. The write-back and evolution module is used to write back the judgment conclusions and newly discovered entities and relationships to the intelligence domain ontology, forming a generation-write-back closed loop.

[0013] Furthermore, the intelligence domain ontology constructed by the ontology construction module includes: The entity sub-entity includes four entity types: people, organizations, equipment, and locations. The relational sub-ontology includes four types of relations: subordinate, cooperative, adversarial, and derived. The event sub-entity contains three types of attributes: event type, event elements, and event sequence. The spatiotemporal sub-ontology includes three types of spatiotemporal attributes: time interval, geographic coordinates, and trajectory. The credibility sub-ontology includes three types of credibility attributes: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

[0014] Furthermore, the alignment and arbitration module includes: An entity alignment unit is used to calculate the alignment similarity sim(a,b)=μ between two candidate ontology individuals a and b. sim attr (a,b)+ν sim struct (N(a),N(b)), where sim attr For attribute similarity, sim struct For the relational neighborhood N ( The structural similarity of ) is μ+ν=1; when sim(a,b) exceeds the preset alignment threshold, the two candidate ontology individuals are determined to be the same entity; The conflict arbitration unit is used to calculate the support weight w of each piece of intelligence for a candidate value x from multiple pieces of intelligence r_k that point to the same entity or event but have contradictory values. k =ρ(src_k) c k e -λ(t0-tk) Where t0 is the judgment time, t k For intelligence r k The timestamp; calculating the overall confidence level of the candidate value x: ; Output all candidate values ​​and their corresponding overall confidence scores to form a confidence score distribution.

[0015] Furthermore, the OAG analysis engine includes: The subgraph retrieval unit is used to retrieve relevant ontology subgraphs from the intelligence domain ontology based on the assessment task and entity alignment results; Hint construction unit, used to construct structured hard constraint hints based on the ontology subgraph and confidence distribution; The generation and verification unit is used to inject the structured hard constraint hints into the large language model to generate a judgment draft; to verify the judgment draft by checking whether the entity types and relation types appearing in the draft belong to the set defined by the intelligence domain ontology, and to check whether each conclusion in the draft is associated with an ontology node; if there is content that violates the constraints, the violation information is fed back to the large language model and a regeneration is triggered until the draft meets all constraints or the maximum number of retries is reached; and to output the final judgment conclusion, which includes the judgment content, the evidence chain, and the confidence distribution.

[0016] Furthermore, the write-back and evolution module is also used to generate a visual evidence path based on the chain of evidence of the judgment conclusion, so that analysts can review the source and arbitration process of each conclusion.

[0017] The technical solution of this invention constructs a multi-dimensional intelligence domain ontology as a unified semantic skeleton to achieve standardized semantic mapping of multi-source heterogeneous intelligence, effectively bridging the differences in terminology systems and semantic gaps between different intelligence sources, and completing the unified identification of cross-source entities and events. A confidence distribution is output through a credibility-weighted arbitration mechanism, replacing the traditional simple selection mode, taking into account source reliability, content confidence, and temporal decay factors, significantly improving the scientific nature of conflict resolution. Simultaneously, ontology subgraphs are injected into the large language model as hard constraint hints, suppressing illusions at the generation mechanism level, forcing conclusions to associate with ontology nodes to form an evidence chain, and meeting traceability requirements; coupled with a generation-writeback closed loop driving continuous evolution of the ontology, it achieves self-iterative improvement in intelligence analysis capabilities. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a system functional architecture diagram according to an embodiment of the present invention; Figure 3 This is a comparison diagram showing the differences between OAG hard constraint hint injection and traditional RAG fragment splicing in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] Example 1 like Figures 1-3 As shown in the figure, this embodiment presents a collaborative analysis method for multi-source heterogeneous intelligence through ontology subgraph hint injection. The method follows five steps: "intelligent domain ontology construction → heterogeneous intelligence preprocessing and semantic element extraction → ontology-based entity alignment and conflict resolution → OAG analysis → generation-writeback closed loop." The overall process is as follows: Figure 1 As shown, the system functional architecture is as follows: Figure 2 As shown, the specific content is as follows: Step 1: Intelligence Domain Ontology Construction Construct an intelligence ontology comprising five dimensions as a unified semantic framework: (1) Entity sub-entities: characters, organizations, equipment, locations; (2) Relationship sub-ontologies: subordination, cooperation, antagonism, derivation; (3) Event sub-ontology: event type, elements, and timing; (4) Spatiotemporal sub-ontology: time interval, geographical coordinates, trajectory; (5) Credibility sub-ontology: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

[0024] Step 2: Heterogeneous Information Preprocessing and Semantic Feature Extraction For different sources such as text, images, signals, and social media, an extractor matching the source type is used (named entity and relation extraction for text, object detection and scene recognition for images, feature parsing for signals, etc.) to map them into ontology individuals. The extraction process simultaneously records the source, timestamp, and original confidence value.

[0025] Each extraction result forms an ontology triple with metadata: record={ Entity / Relationship / Event: Mapped to ontology type Source src, timestamp t, Source reliability ρ(src), content confidence c } Step 3: Ontology-based entity alignment and conflict resolution Entity alignment: Cross-source entity alignment is performed using ontology structure constraints (type consistency, relational neighborhood similarity) and attribute similarity. Alignment similarity between two candidate individuals a and b: sim(a,b)=μ sim attr (a,b)+ν sim struct (N(a),N(b)); Where, sim attr For attribute similarity, sim struct For the relational neighborhood N ( The structural similarity is μ+ν=1. When sim(a,b) exceeds the preset alignment threshold, the two candidate ontology individuals are determined to be the same entity.

[0026] Conflict resolution (credibility-weighted arbitration): For multiple pieces of intelligence pointing to the same entity / event but with contradictory values ​​{r k Each entry is weighted according to its credibility dimension and then weighted for arbitration. A time decay factor, decay(t) = e, is introduced. λ(t0 tk) (t0 is the time of assessment), then the evidence r k Support weights for a candidate value x: wk =ρ(src k ) c k e -λ(t0-tk) ; Overall confidence level of candidate value x: ; The output is the confidence distribution of each candidate value {(x,Conf(x))} rather than a single conclusion.

[0027] Step 4: OAG Analysis Engine For the analysis task, relevant subgraphs are dynamically retrieved from the ontology as structured "hard constraint hints" and injected into the large language model. This differs from RAG's text fragment concatenation (such as...). Figure 3 As shown, this step uses the type, relation, and constraint of the ontology as explicit restrictions for generation: the entity type and relation type appearing in the generated conclusion must belong to the ontology definition set, and the conclusion must be forcibly associated with ontology nodes to form an evidence chain; otherwise, regeneration will be triggered.

[0028] subgraph = retrieve_ontology_subgraph(analysis task, entity alignment result) prompt = build_structured_prompt(subgraph, confidence distribution) # Subgraph → structured constraint prompt repeat: draft = LLM.generate(prompt) violations=[] for entity / relation x in draft: if type(x) IntelOntology:violations.add(x) For the conclusion in the draft: if not linked_to_ontology_node(conc):violations.add(conc) if violations is empty: break prompt = prompt + feedback (violations) until the maximum number of retries is reached. return draft #Analysis conclusion + chain of evidence + confidence level distribution Step 5: Generate and write back closed loop The analysis conclusions and newly discovered entities and relationships are written back into the intelligence ontology, driving the continuous evolution of the ontology; at the same time, a visual evidence path is generated based on the evidence chain, allowing analysts to review the source and arbitration process of each conclusion along the path.

[0030] Example 2 This embodiment uses the cross-source alignment and conflict arbitration of multi-source intelligence on a certain maritime target as an example to explain in detail the implementation steps, parameters and conditions of the present invention.

[0031] Scene background Three sources provided intelligence on the same maritime target: Source A is a reconnaissance text report; source reliability ρ A =0.9, the content is "The ship is a supply ship and is located in area X", and the timestamp is 2 hours away from the time of the assessment; Source B is the signal interception source, and the reliability of the source is ρ. B =0.8, the content is "This ship is a supply ship, located in area X′ (12 nautical miles away from X)", timestamp 0.5 hours; Source C is an open-source social media source; source reliability ρ C =0.4, content is "This vessel is a fishing boat", timestamp 6 hours.

[0032] Set the time decay coefficient λ = 0.1 / hour, the entity alignment threshold to 0.7, μ = 0.6, ν = 0.4, and the maximum number of retries to 3.

[0033] Implementation steps Step 1: Intelligence Domain Ontology Construction Construct an intelligence domain ontology comprising five dimensions: Sub-entities: Characters, organizations, equipment, locations; Relationship sub-ontologies: Subordination, Cooperation, Conflict, Derivation; Event sub-ontology: event type, elements, and timing; Spatiotemporal sub-ontology: time interval, geographic coordinates, trajectory; Credibility sub-ontology: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

[0034] Step 2: Heterogeneous Information Preprocessing and Semantic Feature Extraction Use matching extractors for different types of intelligence sources: For the text report from source A, a named entity and relation extraction model is used to extract the entities "supply ship" and the location "X area", and the source A and timestamp t are recorded. A =2 hours, source reliability ρ A =0.9, Content Confidence c A=0.85; For signal data from source B, a signal feature analysis model is used to extract the entity "supply ship" and the location "X′ area", and the source B and timestamp t are recorded. B =0.5 hours, source reliability ρ B =0.8, Content Confidence c B =0.90; For social media text from source C, a social media intelligence extraction model is used to extract the entity "fishing boat" and record source C and timestamp t. C =6 hours, source reliability ρ C =0.4, Content Confidence c C =0.70.

[0035] Each extraction result forms an ontology triple with metadata: plaintext record A ={ Entity: "Equipment / Ship / Supply Ship", Source: "A" Timestamp: 2, Source reliability: 0.9 Content confidence level: 0.85 } record B ={ Entity: "Equipment / Ship / Supply Ship", Source: "B" Timestamp: 0.5 Source reliability: 0.8 Content confidence level: 0.90 } record C ={ Entity: "Equipment / Fishing Boat", Source: "C", Timestamp: 6, Source reliability: 0.4 Content confidence level: 0.70 } Step 3: Ontology-based entity alignment and conflict resolution Entity alignment: Calculate the alignment similarity between the three ontology individuals: A and B: Attribute Similarity Sim attr (A,B)=0.9 (both entities are ships, and the location deviation is within a reasonable range), structural similarity sim structIf (N(A),N(B))=0.8 (the relation neighborhoods are all empty), then sim(A,B)=0.6×0.9+0.4×0.8=0.86>0.7, and they are determined to be the same entity; A and C: Attribute Similarity Sim attr (A,C)=0.6 (different entity types but both are maritime targets), structural similarity sim struct If (N(A),N(C))=0.7 (all relational neighborhoods are empty), then sim(A,C)=0.6×0.6+0.4×0.7=0.64<0.7? No, in the actual scenario, the three pieces of information point to the same coordinate neighborhood on the spatiotemporal sub-ontology (the deviations are all within the continuous range of the trajectory), therefore the structural similarity should be higher. Adjusted to sim struct If (N(A),N(C))=0.8, then sim(A,C)=0.6×0.6+0.4×0.8=0.68? No, the similarity of spatiotemporal attributes should be considered, and the calculation should be recalculated. In fact, the spatiotemporal coordinate neighborhoods of the three pieces of information highly overlap, resulting in a high degree of similarity in the spatiotemporal attributes. After adjustment: sim attr (A,C)=0.7 (highly similar spatiotemporal attributes, different entity types), sim struct (N(A),N(C))=0.8 (the relational neighborhoods are all empty and the spatiotemporal neighborhoods are the same), then sim(A,C)=0.6×0.7+0.4×0.8=0.74>0.7, so they are determined to be the same entity.

[0036] Therefore, the three entities A, B, and C are determined to be the same maritime target entity.

[0037] Conflict Resolution (Credibility-Weighted Arbitration): Calculate the supporting weights of various intelligence reports regarding contradictory values ​​for the "ship type" attribute (supply ship vs. fishing boat): Weight of source A: w A =0.9×0.85×e -0.1×2 =0.9×0.85×0.819≈0.626 Weight of source B: w B =0.8×0.90×e -0.1×0.5 =0.8×0.90×0.951≈0.685 Weight of source C: w C =0.4×0.70×e -0.1×6 =0.4×0.70×0.549≈0.154 Calculate the overall confidence level of each candidate value: Conf (supply ship) = (w A +wB ) / (w A +w B +w C =(0.626+0.685) / (0.626+0.685+0.154)=1.311 / 1.465≈0.895; Conf (fishing boat) = w C / (w A +w B +w C = 0.154 / 1.465 ≈ 0.105 Output confidence distribution: (supply ship, 0.895), (fishing ship, 0.105), retaining the low-confidence fishing ship hypothesis instead of arbitrarily discarding it.

[0038] Step 4: OAG Analysis Engine Subgraph retrieval: Based on the task of "maritime target identification and trajectory analysis", retrieve the ontology subgraph related to the target from the intelligence domain ontology, including its entity type, historical trajectory, formation relationship and other nodes and relationships; Hint construction: Based on the retrieved ontology subgraph and the confidence distribution mentioned above, a structured hard constraint hint is constructed, which explicitly stipulates that only entity types and relation types defined in the ontology can appear in the generated content, and each conclusion must be associated with a specific ontology node; Generation and Verification: First generation: The large language model outputs the draft "The target is a frigate, located in area X". Upon verification, it is found that "frigate" is not in the entity type set of the current ontology subgraph, triggering regeneration. Second generation: Output draft "Target is a supply ship (confidence 0.895), which has been maneuvering from X to X′ in the past 2 hours, and its track matches the replenishment behavior of its formation. Another social media source claims it is a fishing boat (confidence 0.105)". Verification confirms that all entity types and relation types belong to the ontology definition set, and each conclusion is associated with the corresponding ontology node (supply ship, X area, X′ area, formation, etc.), satisfying all constraints; Output the final assessment conclusion, including the assessment content, the chain of evidence (related sources A, B, and C), and the confidence level distribution.

[0039] Step 5: Generate-Write-Back Closed Loop Main body write-back: Write "X→X′ maneuver" as a new trajectory event back to the spacetime sub-main body to update the trajectory information of the target; Evidence visualization: Based on the chain of evidence, a visualized evidence path is generated, clearly showing the source information, weight calculation process and arbitration results of each conclusion, for analysts to review along the path.

[0040] Compared with the prior art, the present invention has the following advantages: This invention uses ontology subgraphs as "hard constraint hints" to suppress illusions: Unlike RAG's text fragment splicing, this invention injects ontology subgraphs into the generation process in a structured manner. The generated output is constrained by the ontology in terms of type, relation, and constraint. After generation, illegal items are checked and regeneration is triggered, thus suppressing illusions from a mechanism perspective rather than through post-processing filtering.

[0041] This invention provides a credibility ontology-driven conflict resolution with a chain of evidence: source reliability, content confidence, and timeliness decay are explicitly modeled as ontology attributes and participate in weighted arbitration. The judgment conclusion has its own confidence distribution and evidence path, which meets the accountability requirements of intelligence products.

[0042] This invention generates and writes back a closed-loop mechanism to drive ontology evolution: the judgment process both consumes and produces ontology, and new entities and new relationships are written back and deposited, forming a self-iterative mechanism for the semantic foundation of intelligence, which is different from static knowledge base methods.

[0043] This invention provides a unified framework for handling cross-source identity and contradiction: by combining ontology structure constraints (entity alignment) and credibility arbitration (conflict resolution), entity alignment and conflict resolution are handled uniformly under the same framework, avoiding semantic inconsistencies caused by handling them separately.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for collaborative analysis of multi-source heterogeneous intelligence injected through ontology subgraph hints, characterized in that, Includes the following steps: S1: Construct an intelligence domain ontology containing five dimensions as a unified semantic framework; S2: Preprocess and extract semantic elements from multi-source heterogeneous intelligence, map each source intelligence into an ontology individual and record metadata; S3: Perform cross-source entity alignment based on the intelligence domain ontology, and perform confidence-weighted arbitration on conflict intelligence to output the confidence distribution of each candidate value; S4: For the analysis task, relevant ontology subgraphs are dynamically retrieved from the intelligence domain ontology as structured hard constraint hints and injected into the large language model to generate an analysis conclusion with a chain of evidence. S5: Write the analysis conclusions and newly discovered entities and relationships back to the intelligence domain ontology to form a generation-writeback closed loop.

2. The method for collaborative analysis of multi-source heterogeneous intelligence injected with ontology subgraph hints according to claim 1, characterized in that, In step S1, the intelligence domain ontology includes: The entity sub-entity includes four entity types: people, organizations, equipment, and locations. The relational sub-ontology includes four types of relations: subordinate, cooperative, adversarial, and derived. The event sub-entity contains three types of attributes: event type, event elements, and event sequence. The spatiotemporal sub-ontology includes three types of spatiotemporal attributes: time interval, geographic coordinates, and trajectory. The credibility sub-ontology includes three types of credibility attributes: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

3. The method for collaborative analysis of multi-source heterogeneous intelligence injected with ontology subgraph hints according to claim 1, characterized in that, In step S3, the specific process of cross-source entity alignment is as follows: Calculate the alignment similarity between two candidate ontology individuals a and b: sim(a,b)=μ Yes attr (a,b)+ν Yes struct (N(a),N(b)); Where, sim attr For attribute similarity, sim struct For the relational neighborhood N ( The structural similarity is μ+ν=1. When sim(a,b) exceeds the preset alignment threshold, the two candidate ontology individuals are determined to be the same entity.

4. The method for collaborative analysis of multi-source heterogeneous intelligence injected with ontology subgraph hints according to claim 1, characterized in that, In step S3, the specific process of the credibility-weighted arbitration is as follows: For multiple pieces of information r that point to the same entity or event but have contradictory values k Calculate the support weight of each piece of intelligence for the candidate value x: w k =ρ(src k ) c k yes -λ(t0-tk) ; Where t0 is the judgment time, t k For intelligence r k timestamp; Calculate the overall confidence level of the candidate value x: ； Output all candidate values ​​and their corresponding overall confidence scores to form a confidence score distribution.

5. The method for collaborative analysis of multi-source heterogeneous intelligence injected with ontology subgraph hints according to claim 1, characterized in that, The specific process of step S4 is as follows: Based on the assessment task and entity alignment results, relevant ontology subgraphs are retrieved from the intelligence domain ontology; Structured hard constraint hints are constructed based on the ontology subgraph and confidence distribution; The structured hard constraint hints are injected into the large language model to generate a judgment draft; The draft analysis is validated by checking whether the entity types and relationship types appearing in the draft belong to the set defined by the intelligence domain ontology, and whether each conclusion in the draft is associated with an ontology node. If any content violates the constraints, the violation information will be fed back to the large language model and a regeneration will be triggered until the draft meets all constraints or the maximum number of retries is reached. Output the final judgment conclusion, which includes the judgment content, evidence chain, and confidence level distribution.

6. A multi-source heterogeneous intelligence collaborative analysis system based on ontology subgraph hint injection according to any one of claims 1-5, characterized in that, include: The ontology construction module is used to construct an intelligence domain ontology containing five dimensions, serving as a unified semantic skeleton; The intelligence preprocessing module is used to preprocess multi-source heterogeneous intelligence and extract semantic elements, map each source intelligence into an ontology individual and record metadata. The alignment and arbitration module is used to perform cross-source entity alignment based on the intelligence domain ontology, and to perform confidence-weighted arbitration on conflict intelligence, outputting the confidence distribution of each candidate value; The OAG analysis engine is used to dynamically retrieve relevant ontology subgraphs from the intelligence domain ontology as structured hard constraint hints for analysis tasks and inject them into the large language model to generate analysis conclusions with evidence chains. The write-back and evolution module is used to write back the judgment conclusions and newly discovered entities and relationships to the intelligence domain ontology, forming a generation-write-back closed loop.

7. The multi-source heterogeneous intelligence collaborative analysis system for ontology subgraph hint injection according to claim 6, characterized in that, The intelligence domain ontology constructed by the ontology construction module includes: The entity sub-entity includes four entity types: people, organizations, equipment, and locations. The relational sub-ontology includes four types of relations: subordinate, cooperative, adversarial, and derived. The event sub-entity contains three types of attributes: event type, event elements, and event sequence. The spatiotemporal sub-ontology includes three types of spatiotemporal attributes: time interval, geographic coordinates, and trajectory. The credibility sub-ontology includes three types of credibility attributes: source reliability ρ∈[0,1], content confidence c∈[0,1], and time decay coefficient λ.

8. The multi-source heterogeneous intelligence collaborative analysis system for ontology subgraph hint injection according to claim 6, characterized in that, The alignment and arbitration module includes: An entity alignment unit is used to calculate the alignment similarity sim(a,b)=μ between two candidate ontology individuals a and b. sim attr (a,b)+ν sim struct (N(a),N(b)), where sim attr For attribute similarity, sim struct For the relational neighborhood N ( The structural similarity of ) is μ+ν=1; when sim(a,b) exceeds the preset alignment threshold, the two candidate ontology individuals are determined to be the same entity; The conflict arbitration unit is used to calculate the support weight w of each piece of intelligence for a candidate value x from multiple pieces of intelligence r_k that point to the same entity or event but have contradictory values. k =ρ(src_k) c k e -λ(t0-tk) Where t0 is the judgment time, t k For intelligence r k The timestamp; calculating the overall confidence level of the candidate value x: ； Output all candidate values ​​and their corresponding overall confidence scores to form a confidence score distribution.

9. The multi-source heterogeneous intelligence collaborative analysis system for ontology subgraph hint injection according to claim 6, characterized in that, The OAG analysis engine includes: The subgraph retrieval unit is used to retrieve relevant ontology subgraphs from the intelligence domain ontology based on the assessment task and entity alignment results; Hint construction unit, used to construct structured hard constraint hints based on the ontology subgraph and confidence distribution; The generation and verification unit is used to inject the structured hard constraint hints into the large language model to generate a judgment draft; to verify the judgment draft by checking whether the entity types and relation types appearing in the draft belong to the set defined by the intelligence domain ontology, and to check whether each conclusion in the draft is associated with an ontology node; if there is content that violates the constraints, the violation information is fed back to the large language model and a regeneration is triggered until the draft meets all constraints or the maximum number of retries is reached; and to output the final judgment conclusion, which includes the judgment content, the evidence chain, and the confidence distribution.

10. The multi-source heterogeneous intelligence collaborative analysis system for ontology subgraph hint injection according to claim 6, characterized in that, The write-back and evolution module is also used to generate a visual evidence path based on the chain of evidence of the judgment conclusion, so that analysts can review the source and arbitration process of each conclusion.

Images