A method for automatically constructing an unstructured document knowledge graph in the field of natural resource investigation and monitoring based on multi-agent cooperation

Abstract

The application discloses a kind of natural resource investigation monitoring field unstructured document knowledge graph automatic construction method based on multi-agent cooperation. The method generates semantic vector using natural resource field special Embedding model;Through the extraction agent, the candidate entity and the relationship triple of binding original text evidence interval, data year, source subject, statistical caliber and initial confidence are generated;Through ontology maintenance agent, the semantic distance discrimination, normalization processing and Schema dynamic expansion are carried out to undefined domain concept or relationship;Through weighted voting mechanism, the final adoption value is determined and the non-adoption source is reserved as traceability record to process multi-source conflict item. The method does not depend on pre-complete fixed ontology, can reduce the relationship mismatch, time sequence misjudgment and conflict coverage problem caused by cross-year, cross-source, cross-statistical caliber and cross-space level information confusion, improve the traceability and reviewability of natural resource investigation monitoring knowledge graph construction result.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and natural resource management, specifically a method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration. Background Technology

[0002] In the field of natural resource surveys and monitoring, basic documents are dynamic, multi-source, and policy-sensitive, which fundamentally distinguishes them from conventional unstructured text. One of the most significant characteristics of documents in this field is the deep temporal interweaving of multi-source data. Taking the *China Natural Resources Bulletin* as a typical example, the release of data such as land use change surveys, major mineral resource reserves, and water resources exhibits a systematic year-over-year lag mechanism. Statistical facts from different years naturally coexist within the same document, requiring knowledge representation to have built-in precise temporal tagging capabilities to bind facts to their effective time. However, the predefined ontologies relied upon by traditional knowledge graphs typically only model static categories of entity types and relationships. Temporal information is at most treated as ordinary attribute literals, not elevated to a first-class modeling element. This makes it impossible for the system to semantically distinguish the temporal context of "2023 cultivated land area" from "2024 total water resources" when encountering cross-year data coexistence. A large number of facts are stored mixed together without temporal anchoring, leading to errors in subsequent temporal queries and evolutionary analysis.

[0003] The documentation in this field contains extremely dense cross-departmental data references. The same objective fact is often described by multiple sources, such as the Ministry of Natural Resources, the Ministry of Agriculture and Rural Affairs, the Ministry of Ecology and Environment, and the Ministry of Industry and Information Technology, using different interpretations. Statistical units and publishing standards are inconsistent, making data conflicts from multiple sources the norm rather than the exception. Traditional knowledge graph construction processes generally lack mechanisms for recording and resolving conflicts related to data sources, statistical interpretations, and credibility. When faced with different descriptions of the same cultivated land compensation area from different departments, a simple overlay or direct merging strategy is usually adopted, failing to faithfully preserve the full picture of multi-source knowledge. This not only loses the potential value of complementary information but also results in inconsistent and unreliable assertions hidden within the knowledge base, seriously impairing the interpretability and decision support capabilities of the knowledge.

[0004] Meanwhile, documents in this field often use highly flexible and non-standardized language to characterize complex spatial causal chains. For example, in a typical statement like "storm surges and floods in coastal cities during periods of high sea level have had a significant impact on Fujian, Guangdong, Guangxi, and Hainan," the causal relationships between spatial entities, disaster-causing factors, and affected objects are implicit and multi-dimensional, far exceeding the set of concepts that any static predefined ontology can enumerate. When systems relying on predefined ontology process such expressions, they either discard the information directly because the corresponding relationships and roles are not within the schema scope, or they map it to a semantically distorted generalized relation. This non-standardized spatial causal logic often crosses sentence and even paragraph boundaries. However, relation extraction models based on supervised learning, constrained by their high dependence on large-scale manually labeled data, struggle to perceive and connect such long-distance, implicit causal chains. When faced with each new causal association style, they need to re-accumulate labeled samples, and their inherent generalization inertia makes it almost impossible to effectively cover this type of dynamically open and complex semantics.

[0005] New concepts and relational expressions are constantly entering documents, introducing not only new entity types but also unprecedented relational patterns. Traditional knowledge graph construction paradigms with fixed schemas essentially assume the world is closed and pre-defined. Under this assumption, the system cannot proactively identify and accept new relations never defined in the schema, let alone extract and structure them. This leads to widespread information silence with each round of policy updates. Furthermore, existing graph construction generally adopts a one-time, static deployment model, lacking the adaptive dynamic evolution capability of the schema and instance data. When new survey and monitoring data are continuously input annually, or when policy discourse shifts, the graph structure cannot adjust itself to the changing information environment. Domain experts must intervene again, performing manual ontology revisions and large-scale re-extraction and re-annotation. This offline reconstruction maintenance method contradicts the continuous streaming nature of natural resource survey and monitoring data generation and the constant evolution of new relations. Summary of the Invention

[0006] The purpose of this invention is to provide an automated method for constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, comprising the following steps:

[0007] Step 1: Obtain unstructured documents related to natural resource surveys and monitoring, and dynamically segment them to generate multiple sub-document blocks. The dynamic segmentation is constrained by semantic continuity, and uses at least one of the following as segmentation constraints: data year, source entity, statistical caliber, spatial range, chapter level, table boundaries, and enumeration boundaries. Each sub-document block is associated with and stores block-level metadata, which includes at least the data year, source entity, statistical caliber, and chapter position.

[0008] Step 2: Use a special embedding model for the natural resources field to perform semantic recognition on the sub-document blocks, generate semantic vectors, and store the block-level metadata in association with the semantic vectors;

[0009] Step 3: Define the extraction agent, the ontology maintenance agent, and the auditing agent; the extraction agent is used to extract candidate entities and candidate relation triples from the sub-document blocks.

[0010] The ontology maintenance agent is used to normalize or expand candidate entities and candidate relationships.

[0011] The auditing agent is used to verify the consistency of evidence, timing, statistical scope, spatial hierarchy, and unit of measurement of candidate relation triples.

[0012] Step 4: Extract candidate entity and relation triples based on the semantic vector, corresponding sub-document blocks, block-level metadata, context window, and generated graph state. Each candidate entity and relation triple includes at least a head entity, relation type, tail entity, original evidence interval, data year, source subject, statistical caliber, and initial confidence level. If there are undefined domain concepts or relations, proceed to step 5; otherwise, proceed to step 6.

[0013] Step 5: The ontology maintenance agent performs entity normalization, relation normalization, candidate node generation or relation expansion operations based on semantic similarity, and updates the candidate entity and relation triplets.

[0014] When the semantic distance between an undefined domain concept or relation and an existing set of entities or relations is less than a first distance threshold, a normalization operation is performed.

[0015] When the semantic distance is greater than the second distance threshold, candidate node generation or relationship expansion operations are performed; when the semantic distance is between the first and second distance thresholds, a secondary judgment is performed by combining the original text evidence interval, contextual semantics, and business definition.

[0016] Step 6: The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples, and marks candidate entities or relations that are missing original evidence, inconsistent data years, inconsistent statistical standards, mismatched spatial levels, inconsistent units of measurement, relation directions that do not conform to the business rules of natural resource survey and monitoring, or whose confidence scores are lower than the preset threshold as conflict items.

[0017] If confidence score If the conflict occurs, the reflection mechanism is triggered, returning to step 4, and candidate entity and relation triples are re-extracted based on the original text evidence interval corresponding to the conflict item, the sub-document block in which it is located, the adjacent sub-document block, the relevant chapter, or the full text range. The preset threshold;

[0018] Step 7: Use a weighted voting mechanism to process conflicting items, determine the final accepted value, and retain unaccepted sources as source tracing records, thereby generating and updating the final natural resource survey and monitoring knowledge graph; wherein, the weighted voting mechanism calculates the acceptance score of candidate values ​​based at least on source authority, time decay factor, statistical caliber matching degree, and spatial range matching degree, and the source tracing records include at least the unaccepted candidate value, source subject, data year, statistical caliber, and original evidence interval.

[0019] Compared to general multi-agent knowledge graph frameworks, the key improvements of this invention are: explicit inclusion of data year, source entity, statistical caliber, and spatial range in the dynamic segmentation stage; requirement for candidate triples to be bound to original text evidence intervals in the extraction stage; normalization or expansion only when semantic distance, business definition, and contextual evidence meet the conditions in the ontology maintenance stage; and joint verification of temporal sequence, dimensionality, spatial hierarchy, and source authority in the audit stage. These limitations transform multi-agent collaboration from a general process into a verifiable technical process for natural resource survey and monitoring documents.

[0020] Furthermore, unstructured documents related to natural resource surveys and monitoring include farmland survey and monitoring reports, urban territorial spatial planning monitoring reports, and natural resource bulletins;

[0021] Among them, the farmland survey and monitoring report includes the farmland quality grade evaluation report, the basic farmland protection special report, and the annual report on farmland occupation and compensation balance;

[0022] The urban territorial spatial planning monitoring report includes an urban territorial spatial health check report, a territorial spatial planning implementation evaluation report, and a construction land supply statistics report;

[0023] The natural resources bulletin includes the China Land Resources Bulletin, the China Mineral Resources Report, and the China Natural Resources Bulletin.

[0024] Furthermore, the steps for dynamic block partitioning are as follows:

[0025] The block boundaries are determined primarily by semantic continuity, combined with heading level, table or enumeration boundaries, changes in entity type, data year, data source, and changes in statistical caliber.

[0026] When the semantic similarity between adjacent sentence segments is higher than the segmentation threshold, the current segmentation window is expanded to preserve the complete semantic chain.

[0027] When the semantic similarity between adjacent sentence segments is lower than the segmentation threshold, or when there is a switch in chapter, indicator caliber, data year, or source subject, the segment is automatically truncated and a new sub-document block is generated. At the same time, the data year, source, chapter position, and statistical caliber are saved as block-level metadata.

[0028] Furthermore, the determination of undefined domain concepts or relationships is completed collaboratively by the extraction agent, the ontology maintenance agent, and the auditing agent;

[0029] The judgment steps are as follows: The extraction agent first combines the current sub-document block, the context window and the generated graph state to identify whether the candidate entity or relation belongs to the expression of an existing concept or whether a new semantic orientation has appeared. Candidate entities or relations that do not belong to the expression of an existing concept or have a new semantic orientation are transmitted to the ontology maintenance agent.

[0030] The ontology maintenance agent uses a large language model to semantically discriminate the names, definitions, hierarchical relationships, applicable scenarios, and original text evidence of candidate entities or relations, and transmits the candidate entities or relations as new concepts or new relations to the audit agent.

[0031] The audit agent reviews the original evidence, contextual consistency, and business logic, and outputs undefined domain concepts or relationships.

[0032] Furthermore, the steps for logical verification are as follows:

[0033] The audit agent checks the evidence consistency, spatiotemporal constraints, units of measurement, business rules, and ontology constraints in sequence. It verifies whether the triples have original textual basis, whether the entity level and spatial scope match, whether the numerical units and time years are consistent, whether the causal or subordinate direction conforms to the common sense of natural resource management, and whether the relationship type falls within the defined schema or the schema to be expanded. If any check fails, the audit agent generates a REVISE reason and triggers the reflection mechanism, requiring the extraction agent to re-extract the problematic triples and related text.

[0034] Furthermore, the embedding model specifically designed for the natural resources domain has undergone fine-tuning through comparative learning based on the InfoNCE loss function;

[0035] The InfoNCE loss function is shown below:

[0036]

[0037] in, and These are the vector representations of the anchor text and the positive sample, respectively. For temperature hyperparameters, Let be the cosine similarity.

[0038] Furthermore, the extraction agent models the extraction of candidate entities and relation triples as a conditional probability generation problem;

[0039] Generate candidate entity and relation triples The probabilities are as follows:

[0040]

[0041] in, For input text, Representing head entities, relations, and tail entities. These are parameters for a large language model.

[0042] Confidence scores of candidate entity and relation triples As shown below:

[0043]

[0044] in, This represents the semantic fit between entities and relations. Representation and prior knowledge base Consistency test value, These are the weighting coefficients.

[0045] The weighted voting mechanism is as follows:

[0046]

[0047] in, Weighting based on the authority of the data source. For the observation time difference, This is the time decay factor.

[0048] Furthermore, the steps for performing normalization, candidate relation generation, or relation expansion operations include:

[0049] Calculate unknown relation types Time and existing relationship set Calculate the semantic distance of each relation in the set, and identify the minimum distance, denoted as the minimum semantic distance. ;

[0050] Among them, unknown relation types Time and existing relationship set China-US relations The semantic distance is shown below:

[0051]

[0052] Where v is the vector representation of the relation;

[0053] If the minimum distance is less than the set minimum value, it is determined to be a synonym relationship, and the relationship normalization operation is performed.

[0054] When the minimum distance falls between a set minimum value and a set threshold, the ontology maintenance agent first enters the candidate judgment state. It then performs a secondary judgment based on the relation definition, contextual evidence, and frequently co-existing entities. If the semantic direction aligns with an existing relation, it is normalized to the closest relation; otherwise, it is added as a candidate relation and written into the system after confirmation by the auditing agent. ;

[0055] If the minimum distance is greater than the set threshold, it is determined to be a new mechanism, and an expansion operation is performed, that is, let .

[0056] Furthermore, the re-extraction process includes:

[0057] judge If the triplet corresponds to a paragraph that involves cross-chapter references, then the extraction scope will be expanded to include the relevant chapters; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after it, and are then re-extracted.

[0058] judge Does the triplet have a global temporal conflict? If so, extend it to the entire text; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after them, and are then re-extracted.

[0059] An automated knowledge graph construction system for unstructured documents in the field of natural resource survey and monitoring based on the method described above, comprising a user terminal and a server terminal;

[0060] The user terminal acquires unstructured documents related to natural resource surveys and monitoring, and dynamically divides them into blocks to generate multiple sub-document blocks;

[0061] The user terminal uses a special embedding model for the natural resources field to perform semantic recognition on sub-document blocks, generate semantic vectors, and transmit them to the server.

[0062] The server-side storage includes extraction agents, ontology maintenance agents, and audit agents;

[0063] The extraction agent extracts candidate entity and relation triples from the semantic vector and transmits them to the auditing agent;

[0064] When the reflection mechanism is triggered, the extraction agent re-extracts candidate entity and relation triples from the semantic vector and transmits them to the audit agent;

[0065] The ontology maintenance agent performs normalization, candidate node generation, or relation expansion operations on undefined entities or relations based on semantic similarity, and transmits the results to the auditing agent.

[0066] The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples; if the confidence score is... This triggers the reflection mechanism;

[0067] The auditing agent uses a weighted voting mechanism to resolve logical conflicts between multi-source unstructured data, generating and updating the final natural resource survey and monitoring knowledge graph.

[0068] An electronic device includes a memory, a processor, and a data bus. The memory stores all machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the data bus, and the machine-readable instructions are executed by the processor to perform the method described thereon.

[0069] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to perform the method.

[0070] The beneficial effects of this invention are as follows:

[0071] (1) Breaking through the limitation of complete predefined ontology: This invention does not require the exhaustive definition of all entity types, relationship types and business constraints in the field of natural resource survey and monitoring. Instead, based on the initial basic schema or empty schema, it extracts intelligent agents to identify undefined concepts or relationships, and uses the ontology maintenance intelligent agent to make inferences and judgments by combining original evidence, contextual semantics and business definitions. After being reviewed by the auditing intelligent agent, it is automatically written into the schema, realizing the dynamic construction and continuous evolution of the knowledge graph of natural resource survey and monitoring.

[0072] (2) Improve the understanding of complex semantics: Through the multi-agent adversarial cooperation mechanism, it effectively solves problems such as cross-paragraph causal chains, temporal conflicts, and dimensional differences in unstructured text, and reduces the problems of entity relationship mismatch and conflict coverage caused by cross-paragraph causal chains, temporal conflicts, and dimensional differences.

[0073] (3) Support dynamic deduction of resource evolution: automatically capture the spatiotemporal and causal evolutionary relationships between events, and provide a knowledge base with logical depth for the natural resource retrieval enhancement generation system.

[0074] Furthermore, through a closed-loop process of "block-level metadata retention—candidate triple evidence binding—audit agent review—weighted voting record keeping," the present invention enables the graph results to not only be automatically generated, but also to explain their source, year, scope, and reasons for acceptance, thereby enhancing the feasibility and verifiability of the technical effects that the review focuses on. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating the overall process of the method provided in this embodiment of the invention.

[0076] Figure 2 For large-scale multi-agent collaborative architecture and data flow diagram;

[0077] Figure 3 This is a functional node diagram of the multi-agent system of the present invention;

[0078] Figure 4 To upload document content;

[0079] Figure 5 Cut the document into chunks;

[0080] Figure 6 For vectorization;

[0081] Figure 7 To extract knowledge graphs;

[0082] Figure 8 This is the result of a knowledge graph query. Detailed Implementation

[0083] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0084] Example 1:

[0085] See Figures 1-8 A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration includes the following steps:

[0086] Step 1: Obtain unstructured documents related to natural resource surveys and monitoring, and dynamically segment them to generate multiple sub-document blocks. The dynamic segmentation is constrained by semantic continuity, and uses at least one of the following as segmentation constraints: data year, source entity, statistical caliber, spatial range, chapter level, table boundaries, and enumeration boundaries. Each sub-document block is associated with and stores block-level metadata, which includes at least the data year, source entity, statistical caliber, and chapter position.

[0087] Step 2: Use a special embedding model for the natural resources field to perform semantic recognition on the sub-document blocks, generate semantic vectors, and store the block-level metadata in association with the semantic vectors;

[0088] Step 3: Define the extraction agent, the ontology maintenance agent, and the auditing agent; the extraction agent is used to extract candidate entities and candidate relation triples from the sub-document blocks.

[0089] The ontology maintenance agent is used to normalize or expand candidate entities and candidate relationships.

[0090] The auditing agent is used to verify the consistency of evidence, timing, statistical scope, spatial hierarchy, and unit of measurement of candidate relation triples.

[0091] Step 4: Extract candidate entity and relation triples based on the semantic vector, corresponding sub-document blocks, block-level metadata, context window, and generated graph state. Each candidate entity and relation triple includes at least a head entity, relation type, tail entity, original evidence interval, data year, source subject, statistical caliber, and initial confidence level. If there are undefined domain concepts or relations, proceed to step 5; otherwise, proceed to step 6.

[0092] Step 5: The ontology maintenance agent performs entity normalization, relation normalization, candidate node generation or relation expansion operations based on semantic similarity, and updates the candidate entity and relation triplets.

[0093] When the semantic distance between an undefined domain concept or relation and an existing set of entities or relations is less than a first distance threshold, a normalization operation is performed.

[0094] When the semantic distance is greater than the second distance threshold, candidate node generation or relationship expansion operations are performed; when the semantic distance is between the first and second distance thresholds, a secondary judgment is performed by combining the original text evidence interval, contextual semantics, and business definition.

[0095] Step 6: The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples, and marks candidate entities or relations that are missing original evidence, inconsistent data years, inconsistent statistical standards, mismatched spatial levels, inconsistent units of measurement, relation directions that do not conform to the business rules of natural resource survey and monitoring, or whose confidence scores are lower than the preset threshold as conflict items.

[0096] If confidence score If the conflict occurs, the reflection mechanism is triggered, returning to step 4, and candidate entity and relation triples are re-extracted based on the original text evidence interval corresponding to the conflict item, the sub-document block in which it is located, the adjacent sub-document block, the relevant chapter, or the full text range. The preset threshold;

[0097] Step 7: Use a weighted voting mechanism to process conflicting items, determine the final accepted value, and retain unaccepted sources as source tracing records, thereby generating and updating the final natural resource survey and monitoring knowledge graph; wherein, the weighted voting mechanism calculates the acceptance score of candidate values ​​based at least on source authority, time decay factor, statistical caliber matching degree, and spatial range matching degree, and the source tracing records include at least the unaccepted candidate value, source subject, data year, statistical caliber, and original evidence interval.

[0098] Example 2:

[0099] An automated construction method for knowledge graphs of unstructured documents in the field of natural resource survey and monitoring based on multi-agent collaboration is proposed. The technical content is the same as in Example 1. Furthermore, the unstructured documents in the natural resource survey and monitoring category include cultivated land survey and monitoring reports, urban territorial spatial planning monitoring reports, and natural resource bulletins.

[0100] Among them, the farmland survey and monitoring report includes the farmland quality grade evaluation report, the basic farmland protection special report, and the annual report on farmland occupation and compensation balance;

[0101] The urban territorial spatial planning monitoring report includes an urban territorial spatial health check report, a territorial spatial planning implementation evaluation report, and a construction land supply statistics report;

[0102] The natural resources bulletin includes the China Land Resources Bulletin, the China Mineral Resources Report, and the China Natural Resources Bulletin.

[0103] Example 3:

[0104] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of embodiments 1-2, further comprising the following dynamic segmentation steps:

[0105] The block boundaries are determined primarily by semantic continuity, combined with heading level, table or enumeration boundaries, changes in entity type, data year, data source, and changes in statistical caliber.

[0106] When the semantic similarity between adjacent sentence segments is higher than the segmentation threshold, the current segmentation window is expanded to preserve the complete semantic chain.

[0107] When the semantic similarity between adjacent sentence segments is lower than the segmentation threshold, or when there is a switch in chapter, indicator caliber, data year, or source subject, the segment is automatically truncated and a new sub-document block is generated. At the same time, the data year, source, chapter position, and statistical caliber are saved as block-level metadata.

[0108] Example 4:

[0109] A method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, with the same technical content as any one of embodiments 1-3. Furthermore, the judgment of undefined domain concepts or relationships is completed collaboratively by the extraction agent, the ontology maintenance agent, and the audit agent.

[0110] The judgment steps are as follows: The extraction agent first combines the current sub-document block, the context window and the generated graph state to identify whether the candidate entity or relation belongs to the expression of an existing concept or whether a new semantic orientation has appeared. Candidate entities or relations that do not belong to the expression of an existing concept or have a new semantic orientation are transmitted to the ontology maintenance agent.

[0111] The ontology maintenance agent uses a large language model to semantically discriminate the names, definitions, hierarchical relationships, applicable scenarios, and original text evidence of candidate entities or relations, and transmits the candidate entities or relations as new concepts or new relations to the audit agent.

[0112] The audit agent reviews the original evidence, contextual consistency, and business logic, and outputs undefined domain concepts or relationships.

[0113] Example 5:

[0114] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of embodiments 1-4, further comprising the following logical verification steps:

[0115] The audit agent checks the evidence consistency, spatiotemporal constraints, units of measurement, business rules, and ontology constraints in sequence. It verifies whether the triples have original textual basis, whether the entity level and spatial scope match, whether the numerical units and time years are consistent, whether the causal or subordinate direction conforms to the common sense of natural resource management, and whether the relationship type falls within the defined schema or the schema to be expanded. If any check fails, the audit agent generates a REVISE reason and triggers the reflection mechanism, requiring the extraction agent to re-extract the problematic triples and related text.

[0116] Example 6:

[0117] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with the same technical content as any one of Examples 1-5. Furthermore, the embedding model for the natural resource field has been fine-tuned through comparative learning based on the InfoNCE loss function.

[0118] The InfoNCE loss function is shown below:

[0119]

[0120] in, and These are the vector representations of the anchor text and the positive sample, respectively. For temperature hyperparameters, Let be the cosine similarity.

[0121] Example 7:

[0122] An automated construction method for unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with the same technical content as any one of embodiments 1-6, further wherein the extraction agent models the extraction of candidate entities and relation triples as a conditional probability generation problem;

[0123] Generate candidate entity and relation triples The probabilities are as follows:

[0124]

[0125] in, For input text, Representing head entities, relations, and tail entities. These are parameters for a large language model.

[0126] Example 8:

[0127] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of Examples 1-7, including confidence scoring of candidate entity and relation triples. As shown below:

[0128]

[0129] in, This represents the semantic fit between entities and relations. Representation and prior knowledge base Consistency test value, These are the weighting coefficients.

[0130] Example 9:

[0131] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of embodiments 1-8, wherein the weighted voting mechanism is as follows:

[0132]

[0133] in, Weighting based on the authority of the data source. For the observation time difference, This is the time decay factor.

[0134] Example 10:

[0135] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of embodiments 1-9, further comprising the steps of performing entity normalization, relation normalization, candidate node generation, or relation expansion operations, including:

[0136] Calculate unknown relation types Time and existing relationship set Calculate the semantic distance of each relation in the set, and identify the minimum distance, denoted as the minimum semantic distance. ;

[0137] Among them, unknown relation types Time and existing relationship set China-US relations The semantic distance is shown below:

[0138]

[0139] Where v is the vector representation of the relation;

[0140] If the minimum distance is less than the set minimum value, it is determined to be a synonym relationship, and the relationship normalization operation is performed.

[0141] When the minimum distance falls between a set minimum value and a set threshold, the ontology maintenance agent first enters the candidate judgment state. It then performs a secondary judgment based on the relation definition, contextual evidence, and frequently co-existing entities. If the semantic direction aligns with an existing relation, it is normalized to the closest relation; otherwise, it is added as a candidate relation and written into the system after confirmation by the auditing agent. ;

[0142] If the minimum distance is greater than the set threshold, it is determined to be a new mechanism, and an expansion operation is performed, that is, let .

[0143] Example 11:

[0144] A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, with technical content identical to any one of embodiments 1-10, further comprising the following re-extraction steps:

[0145] judge If the triplet corresponds to a paragraph that involves cross-chapter references, then the extraction scope will be expanded to include the relevant chapters; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after it, and are then re-extracted.

[0146] judge Does the triplet have a global temporal conflict? If so, extend it to the entire text; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after them, and are then re-extracted.

[0147] Example 12:

[0148] An automated method for constructing knowledge graphs of unstructured documents in the field of natural resource survey and monitoring based on multi-agent collaboration includes the following steps:

[0149] Step 1: Obtain unstructured documents related to natural resource surveys and monitoring, and dynamically divide them into blocks to generate multiple sub-document blocks;

[0150] Dynamic segmentation addresses the issue of information mixing across chapters, years, and statistical definitions in long documents related to natural resource surveys and monitoring. Fixed-length segmentation can easily break up causal chains or statistical indicators within the same space, and may also merge data from different years or sources into the same text block, leading to omissions, mismatches, or temporal conflicts in subsequent entity relationship extraction. Dynamic segmentation prioritizes semantic continuity, combining changes in title level, table or enumeration boundaries, entity type, data year, data source, and statistical definition to determine segmentation boundaries. When the semantic similarity between adjacent segments consistently exceeds the segmentation threshold, the current segmentation window is expanded to preserve the complete semantic chain. When the semantic similarity between adjacent segments falls below the segmentation threshold, or when there is a chapter switch, indicator definition switch, data year switch, or source subject switch, the document is automatically truncated and a new sub-document block is generated, while the data year, source, chapter position, and statistical definition are saved as block-level metadata. This makes the sub-document blocks of the input extraction agent relatively complete in semantics and relatively consistent in temporal sequence and source, thereby improving the accuracy and traceability of candidate entity and relation triple extraction.

[0151] Step 2: Use a natural resources-specific embedding model to perform semantic recognition on the sub-document blocks and generate semantic vectors;

[0152] Step 3: Define the extraction agent, the ontology maintenance agent, and the audit agent;

[0153] The extraction agent, ontology maintenance agent, and audit agent here are not fixed models "fine-tuned from a certain model." Essentially, these three types of agents are prompt word engineering and process control units for different tasks, each composed of role definitions, task instructions, input / output formats, tool call interfaces, context memory, and constraint rules. They can be connected to any large language model with natural language understanding, structured generation, and tool call capabilities as a driving foundation. Fine-tuning is optional. If improved adaptability to the natural resource domain is needed in a private deployment scenario, domain-specific instruction fine-tuning or efficient parameter fine-tuning can be performed on the driving large language model, such as LoRA or QLoRA. The fine-tuning object is the driving model, not the agent itself. Fine-tuning data includes natural resource survey and monitoring texts, manually annotated triples, schema evolution examples, and audit correction examples. After fine-tuning, different agents assume extraction, ontology maintenance, and auditing responsibilities respectively through different prompt words and tool permissions.

[0154] Step 4: Extract candidate entity and relation triples from semantic vectors using the intelligent agent;

[0155] If there are undefined domain concepts or relationships, proceed to step 5; otherwise, proceed to step 6.

[0156] The determination of undefined domain concepts or relationships is completed collaboratively by three agents: the extraction agent first combines the current sub-document block, context window, and the generated graph state to identify whether candidate entities or relationships belong to conventional expressions and whether new semantic references appear; the ontology maintenance agent then utilizes the domain understanding and reasoning capabilities of the large language model to perform semantic discrimination on the name, definition, hierarchical relationship, applicable scenario, and original text evidence of candidate concepts or relationships, determining whether they should be classified into the existing schema, normalized as synonymous expressions, or extended as new concepts or relationships; the auditing agent finally reviews the results based on original text evidence, contextual consistency, and business logic. The above judgments do not require a pre-set ontology or a pre-set local prior knowledge base as necessary premises; the domain knowledge internalized by the large language model, the current document context, and the historical state of the constructed graph all serve as the basis for reasoning.

[0157] Step 5: The ontology maintenance agent performs entity normalization, relation normalization, candidate node generation or relation expansion operations based on semantic similarity, and updates the candidate entity and relation triplets.

[0158] Step 6: The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples.

[0159] Logical verification is performed by the audit agent in the following order: evidence consistency, spatiotemporal constraints, units of measurement, business rules, and ontology constraints. Specifically, this includes verifying whether the triples have original textual basis, whether the entity hierarchy and spatial scope match, whether the numerical units and time years are consistent, whether the causal or subordinate direction conforms to common sense in natural resource management, and whether the relation type falls within a defined schema or a schema to be expanded. If any item fails, the audit agent generates a REVISE reason and triggers a reflection mechanism, requiring the extraction agent to re-extract the problematic triples and their associated text.

[0160] If confidence score If this happens, the reflection mechanism is triggered, and the process returns to step 4 to re-extract candidate entity and relation triples. The preset threshold;

[0161] Re-extraction does not default to re-running the entire text; instead, it prioritizes localization. The triples correspond to the original text paragraph, the sub-document block, and the adjacent blocks before and after as the re-extraction scope; only when the low confidence item involves cross-chapter references or global temporal conflicts is the scope expanded to the relevant chapter or the entire text.

[0162] Step 7: Use a weighted voting mechanism to resolve logical conflicts between multi-source unstructured data, and generate and update the final natural resource survey and monitoring knowledge graph.

[0163] Entity logic conflicts are first identified by logic verification in step 6. The audit intelligence will compare the attribute values, time tags, spatial scope, and data sources of the same entity in different sources; if there are inconsistencies in the numerical values, years, statistical scope, or source credibility of the same attribute, it will be marked as a conflict item. Then, in step 7, a weighted voting mechanism will be used to determine the final accepted value, and the unaccepted sources will be retained as traceability records.

[0164] Example 13:

[0165] An automated knowledge graph construction system for unstructured documents in the field of natural resource survey and monitoring, based on the method described in any one of Examples 1-12, includes a user terminal and a server terminal;

[0166] The user terminal acquires unstructured documents related to natural resource surveys and monitoring, and dynamically divides them into blocks to generate multiple sub-document blocks;

[0167] The user terminal uses a special embedding model for the natural resources field to perform semantic recognition on sub-document blocks, generate semantic vectors, and transmit them to the server.

[0168] The server-side storage includes extraction agents, ontology maintenance agents, and audit agents;

[0169] The extraction agent extracts candidate entity and relation triples from the semantic vector and transmits them to the auditing agent;

[0170] When the reflection mechanism is triggered, the extraction agent re-extracts candidate entity and relation triples from the semantic vector and transmits them to the audit agent;

[0171] The ontology maintenance agent performs normalization, candidate node generation, or relation expansion operations on undefined entities or relations based on semantic similarity, and transmits the results to the auditing agent.

[0172] The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples; if the confidence score is... This triggers the reflection mechanism;

[0173] The auditing agent uses a weighted voting mechanism to resolve logical conflicts between multi-source unstructured data, generating and updating the final natural resource survey and monitoring knowledge graph.

[0174] Example 14:

[0175] An electronic device includes: a memory, a processor, and a data bus, wherein the memory stores all machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the data bus, and the machine-readable instructions are executed by the processor to perform the method described in any one of embodiments 1-12.

[0176] Example 15:

[0177] A computer-readable storage medium storing a computer program that, when executed by a processor, performs the method described in any one of embodiments 1-12.

[0178] Example 16:

[0179] A method for generating unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration includes the following steps:

[0180] Step S1, Document Preprocessing and Vectorization: Dynamically segment unstructured documents related to natural resource surveys and monitoring, such as farmland survey and monitoring reports, urban land space planning monitoring reports, and natural resource bulletins, and generate semantic vectors using a special embedding model for the natural resource field;

[0181] Step S2, Multi-agent role initialization: Instantiate the extraction agent ExtractorAgent, the ontology maintenance agent OntologyAgent, and the audit agent AuditAgent in the system;

[0182] Step S3, Collaborative Knowledge Extraction and Verification: Utilizing the thinking chain capability of the large language model, the extraction agent extracts candidate entities and relation triples from the segmented text, and the auditing agent performs logical verification and confidence evaluation on the candidate triples.

[0183] Step S4, Schema Dynamic Evolution: When the extraction agent extracts undefined domain concepts or relations, the ontology maintenance agent determines whether to perform relation normalization, candidate relation generation or relation expansion based on semantic similarity calculation, thereby achieving adaptive evolution that does not depend on a pre-complete fixed ontology.

[0184] Step S5, Conflict Resolution and Knowledge Graph Fusion: Based on a weighted voting mechanism, logical conflicts between multi-source unstructured data are resolved, and the final knowledge graph for natural resource survey and monitoring is generated and updated.

[0185] In step S1, the unstructured documents for natural resource survey and monitoring include: (a) farmland survey and monitoring reports, including farmland quality grade evaluation reports, basic farmland protection special reports, and farmland occupation and compensation balance annual reports; (b) urban territorial spatial planning monitoring reports, including urban physical examination reports, territorial spatial planning implementation evaluation reports, and construction land supply statistics reports; (c) natural resource bulletins, including China Land Resources Bulletin, China Mineral Resources Report, and China Natural Resources Bulletin. The above documents have common characteristics such as coexistence of multi-source data time series, cross-departmental data references, dense spatial entity descriptions, and non-standardized relational expressions. The Embedding model is fine-tuned for the above characteristics.

[0186] In step S1, the natural resources-specific embedding model is fine-tuned using contrastive learning based on the InfoNCE loss function, which is defined as:

[0187]

[0188] in, and These are the vector representations of the anchor text and the positive sample, respectively. is the temperature hyperparameter, and sim is the cosine similarity.

[0189] In step S3, candidate entity and relation triples are generated. The probabilities are as follows:

[0190]

[0191] in, For input text, Representing head entities, relations, and tail entities. These are parameters for a large language model.

[0192] In step S3, the auditing agent calculates the confidence score Sconf(tᵢ) of the candidate triplet tᵢ using the following formula:

[0193]

[0194] in, This represents the semantic fit between entities and relations. Representation and prior knowledge base Consistency test value, These are the weighting coefficients; if If this occurs, the reflection mechanism is triggered, requiring the extracted agent to be extracted again.

[0195] In step S4, the ontology-maintaining agent handles unknown relationship types. When calculating its relation to the existing relation set Rexist, calculate its relation to the relation in the Rexist set. semantic distance:

[0196]

[0197] Where v is the vector representation of the relation; if the minimum distance is less than the set minimum value (preferred range [0.05, 0.15]), it is determined to be a synonym and normalized; if the minimum distance is greater than the set threshold (preferred range [0.4, 0.6]), it is determined to be a new mechanism and an expansion operation is performed. .

[0198] In step S5, the weighted voting mechanism combines the authority of the data source with a time decay factor, and the attribute scoring formula for conflicting entities is as follows:

[0199]

[0200] in, Weighting based on the authority of the data source. For the observation time difference, The time decay factor is preferably set within the range of [0.1, 0.3]; the data source authority weight is also specified. With time weight The preferred ratio is 2:1.

[0201] Example 17:

[0202] This embodiment uses the "2024 China Natural Resources Bulletin" (hereinafter referred to as the "Bulletin," the full text of which is publicly available) officially released by the Ministry of Natural Resources in March 2025 as the input document to demonstrate the complete execution process of each step of the present invention. The Bulletin covers ten chapters, including land, minerals, forests, grasslands, wetlands, oceans, ecological restoration, and geological hazards. Its data description clearly states that "the annual data for land use change surveys, major mineral resource reserves, water resources, and state-owned natural resource assets use 2023 data and are released every other year," while the remaining data are preliminary statistics for 2024—this characteristic makes the Bulletin a natural ideal test document for verifying the conflict resolution mechanism of the present invention.

[0203] 1. Document preprocessing and construction of a natural resource-specific embedding model

[0204] This invention employs contrastive learning to construct a domain-specific vectorized model, collects national standards and monitoring specifications to build a corpus, and fine-tunes the basic model using an InfoNCE-based loss function, enabling the model to have higher semantic discrimination capabilities for natural resource domain terms (land category codes, reserve units, administrative levels, etc.). Subsequently, dynamic block segmentation is performed, and the model calculates the semantic coherence of adjacent sentences, truncating at semantic transition points to ensure the integrity of causal relationships within each semantic block.

[0205] S1 Step Implementation Demonstration – Taking the “I. Overview of Natural Resources – Land Resources” section of the bulletin as an example

[0206] Enter the original text:

[0207] "The 2023 National Land Use Change Survey results show that the country has a total of 128.6088 million hectares of cultivated land, 19.6109 million hectares of orchards, 283.6957 million hectares of forest land, 263.2157 million hectares of grassland, 23.5198 million hectares of wetlands, 36.1037 million hectares of urban and rural land and industrial and mining land, 10.4291 million hectares of transportation land, and 36.2648 million hectares of water area and water conservancy facilities land."

[0208] Dynamic block segmentation result: The system recognizes that the text block is a pure statistical enumeration structure ("Land Use Classification → Area Value"), with high semantic cohesion. It does not perform internal segmentation and outputs a complete semantic block, and marks key metadata: Data Year = 2023 (not 2024), Source = Land Change Survey, Publisher = Ministry of Natural Resources - At this time, the sequential labeling will play a key role in the conflict resolution in step S5.

[0209] S1 Step Implementation Demonstration – Taking the "VI. Disasters and Prevention: Geological Disasters" chapter of the bulletin as an example

[0210] "In 2024, a total of 5,719 geological disasters occurred nationwide, including 3,316 landslides, 1,579 collapses, 727 debris flows, and 97 ground subsidences, resulting in 187 deaths, 10 missing persons, and 49 injuries, with direct economic losses of 2.539 billion yuan... Throughout the year, 622 geological disasters were successfully averted, potentially affecting 10,235 people and preventing direct economic losses of 725 million yuan."

[0211] Dynamic segmentation result: The system identified that this segment contains two semantic units. "Disaster Occurrence Statistics" (Disaster Type → Casualties → Economic Losses) and "Disaster Avoidance Effectiveness" (Defense Response) Successfully avoided (Reduce losses), identify semantic transitions at "Successfully avoided throughout the year", and output 2 semantic blocks.

[0212] 2. Definition of Agent Roles and Description of System Nodes

[0213] The system defines a set of intelligent agents A = {Aext, Aonto, Aaudit}, which aims to be used given a context C and a history graph. Generate the optimal set of triples under the following conditions The functions, interfaces, and triggering conditions of each agent node are shown in Table 1.

[0214] Table 1. Functional Description of Three-Agent Nodes

[0215] 3. Multi-agent collaborative workflow

[0216] 3.1 Extracting the Aext mathematical model of the intelligent agent

[0217] Responsible for extracting candidate triples and generating conditional probabilities based on cue engineering modeling:

[0218]

[0219] The system Prompt specifies that its zero-sample open-ended sampling must output in standard JSON format and include the data year markers and original source information.

[0220] 3.2 Aaudit Confidence Model for Auditing Agents

[0221] The confidence score function is:

[0222]

[0223] In practice, Because the field of natural resources places greater emphasis on consistency with prior physical common sense and policy logic, it sets... In this embodiment, threshold .

[0224] 3.3 Ontology-Maintaining Agent Aonto Semantic Distance Model

[0225] Calculate semantic distance when processing new relations rnew:

[0226]

[0227] The preferred range for the minimum threshold is [0.05, 0.15] (tolerating slight semantic fluctuations in synonyms); the preferred range for the expanded threshold is [0.4, 0.6]. When The system exhibits the best discovery rate for novel evolutionary relationships. In this embodiment, the minimum threshold is set to 0.10, and the expansion threshold is set to 0.50.

[0228] 3.4 Complete Implementation Demonstration – Three-Round Collaborative Process, based on the "Ecological Protection and Restoration" section on page 5 of the communique.

[0229] Input semantic block:

[0230] "Four batches of 63 marine ecological protection and restoration projects were implemented, restoring approximately 400 kilometers of coastline and 31,000 hectares of coastal wetlands. Mangrove area increased to 30,300 hectares, making my country one of the few countries in the world with a net increase in mangrove area. The carbon storage survey and assessment of mangrove, seagrass bed, and salt marsh blue carbon ecosystems were also advanced, continuously consolidating and enhancing marine carbon sequestration capacity."

[0231] First round (Drafting) – Aext extraction: Output candidate triples (JSON format):

[0232] Triad 1:{"head":"Marine Ecological Protection and Restoration Project","relation":"Restoration Achieved","tail":"31,000 Hectares of Coastal Wetlands",

[0233] "evidence":"Approximately 31,000 hectares of coastal wetlands have been restored and rehabilitated","confidence":0.92,"datayear":}

[0234] Triad 2:{"head":"Marine ecological restoration","relation":"Promoting growth","tail":"30,300 hectares of mangrove land",

[0235] "evidence":"Mangrove area increased to 30,300 hectares","confidence":0.89,"datayear":"As of 2024"}

[0236] Triad 3:{"head":"Investigation of carbon storage in blue carbon ecosystems","relation":"Consolidation and enhancement","tail":"Marine carbon sink capacity",

[0237] "evidence":"Continuously consolidate and enhance marine carbon sequestration capacity","confidence":0.74,"datayear":"2024"}

[0238] Triad 4:{"head":"Mangroves, seagrass beds, salt marshes","relation":"Blue carbon ecosystem carbon sink linkage","tail":"Marine carbon sink capacity",

[0239] "evidence":"Promoting the investigation and assessment of carbon storage in mangrove, seagrass bed, and salt marsh blue carbon ecosystems","confidence":0.68,"datayear":"2024"}

[0240] Second round (Debate) – Aaudit review:

[0241] For triplets 1 and 2: Dimensions match (units in ten thousand hectares are consistent), spatiotemporal logic is correct, and common sense regarding business is reasonable (the causal direction of repair project → area growth is correct). ①Sconf=0.35×0.92+0.65×0.94=0.933>τ, pass. ②Sconf=0.35×0.89+0.65×0.90=0.897>τ, pass.

[0242] For triple 3: the relationship "consolidation and enhancement" exists in the existing schema, but Aaudit found that the original text supports "investigation and assessment" and requested Aext to add the intermediate link. After Aext added the link, the output was re-output, and triple 3 was split into two (investigation and assessment → revealing carbon reserves → supporting carbon sink capacity), with the confidence level increased to 0.85, and it was approved.

[0243] For triple 4: the relationship "blue carbon ecosystem carbon sink linkage" is not defined in the existing Rexist database, so it is marked as a new relationship and transferred to Aonto for processing. The confidence level is 0.68 < =0.70, Aext is requested to provide more original supporting evidence. Aext supplemented the data, Sconf=0.72> ,pass.

[0244] The third round (Evolution) – Aonto intervention (schema dynamic evolution):

[0245] The new relationship "Blue Carbon Ecosystem Carbon Sink Linkage" is entered into Aonto: The semantic distance to existing relationships in Rexist is calculated: Dist("Blue Carbon Ecosystem Carbon Sink Linkage", "Promote Growth") = 0.41 Dist("Blue Carbon Ecosystem Carbon Sink Linkage", "Enhance Capacity") = 0.38 Dist("Blue Carbon Ecosystem Carbon Sink Linkage", "Consolidate and Enhance") = 0.35 Minimum semantic distance = 0.35 > expansion threshold 0.50? No, 0.35 < 0.50. However, 0.35 > the minimum threshold 0.10, which does not satisfy the MERGE condition. Since this relationship has a significant ecosystem coupling meaning, Aonto semantically judges that although the distance 0.35 is less than the expansion threshold, its semantic direction (multi-factor synergy generates carbon sink benefits) is different from existing relationships, and therefore, CREATE is executed. Aonto output: [CREATE:"Blue Carbon Ecosystem Carbon Sink Linkage", def:"Coupling gain relationship in which blue carbon ecosystems such as mangroves, seagrass beds, and salt marshes contribute to marine carbon sink capacity through synergistic effects"] Schema expansion: A new node type, "Blue Carbon Ecosystem," has been added.

[0246] Supplementary Demonstration – Schema Evolution in Section III, Territorial Spatial Planning of the Bulletin

[0247] "The Ministry of Natural Resources and the Shanghai Stock Exchange jointly released the 'China Marine Economy Stock Price Index' (code: 932056)... The release of this index is of great significance for creating a barometer for my country's marine economic development, guiding capital towards the marine economy, and promoting the effective allocation of financial capital in the marine real economy."

[0248] Aext extracts a new relation: "Guiding Capital Aggregation" (Financial Market Index → ​​Marine Industry Investment). Aonto calculates the minimum semantic distance with all relations in Rexist = 0.61, which is greater than the expansion threshold of 0.50, thus classifying it as a completely new mechanism. Aonto outputs: [CREATE:"Guiding Capital Aggregation", def:"Policy-driven relation that guides social capital to specific industrial sectors through the release of financial instruments such as indices"]

[0249] 4. Conflict resolution mechanism

[0250] To address conflicting descriptions of multi-source data, a weighted voting mechanism is employed.

[0251]

[0252] Parameter settings: : =2:1, =0.15 (the weight decays by half in about 4.6 years).

[0253] S5 Step-by-Step Implementation Demonstration – Resolving Multiple Temporal Conflicts in Farmland Area

[0254] Regarding the core entity "total area of ​​cultivated land nationwide," the system detected the following multi-source description conflicts in the bulletins: Source A: Page 3 of the bulletin (Ministry of Natural Resources, released in March 2025, data year 2023): Cultivated land 128.6088 million hectares; Source B: "2022 National Cultivated Land Quality Grade Bulletin" (Ministry of Natural Resources, released in 2023, data year 2022): National cultivated land 127.951 million hectares (reference value in another article); Source C: A provincial land survey report (local natural resources department, released in 2024, data year 2023, not covering the whole country): Local differences arise due to different data calibers. Taking March 2025 as the current time node, the comprehensive score of each source is calculated: Source A (national authoritative, (2023 data, release interval approximately 1.5 years): Score = (2 / 3) × 0.98 + (1 / 3) × exp(-0.15 × 1.5) ≈ 0.653 + 0.257 = 0.910 Source B (National authoritative source, 2022 data, release interval approximately 2.5 years): Score = (2 / 3) × 0.96 + (1 / 3) × exp(-0.15 × 2.5) ≈ 0.640 + 0.237 = 0.877 Source C (Provincial authoritative source, partial scope, not applicable to national statistics, Scredibility = 0.60): Score = (2 / 3) × 0.60 + (1 / 3) × exp(-0.15 × 1.0) ≈ 0.400 + 0.291 = 0.691

[0255] Conflict resolution results: The final map adopts source A's 128,608,800 hectares (highest score), while retaining historical version trajectories and supporting time-series knowledge queries. The system automatically labels this node with "Data base year = 2023, next update expected = 2026 (released the following year)," providing a basis for subsequent conflict early warning.

[0256] S5 Step-by-Step Supplementary Demonstration – Resolving Conflicts in Multi-Year Comparison Data on Geological Disasters

[0257] Page 22 of the bulletin contains a line graph showing "Direct Economic Losses from Marine Disasters 2020-2024." The figures for each year are from the same publisher (Ministry of Natural Resources) but released at different times. The system establishes time-series nodes for each year's data, without handling "conflicts," but instead constructs a time-series evolutionary relationship chain.

[0258] "Direct economic losses from marine disasters" → "Time-series changes" → [8.3, 30.7, 24.1, 25.1, 109.3] billion yuan (2020-2024). Simultaneously, the causal triplet is extracted: "The sea level along the coast in 2024 was 96 mm higher than the average" → "Exacerbated" → "Direct economic losses from storm surge disasters amounted to 10.496 billion yuan". This time-series causal chain fully demonstrates the core capability of this invention in handling the problem of "coexistence of multiple data sources".

[0259] 5. Global Loss Function and Optimization

[0260] During the fine-tuning phase of the private model, the goal is to minimize the structural loss. :

[0261]

[0262] The behavior matrix of the constrained agent, Magent, approximates the decision-making pattern of the human expert, Mhuman.

[0263] 6. System Interface and Agent Node Description

[0264] The multi-agent system of the present invention includes the following functional module interfaces:

[0265] (1) Document upload and preprocessing interface: Supports uploading natural resource survey documents in PDF, Word, TXT and other formats (such as directly uploading the PDF of "China Natural Resources Bulletin"); Visualizes dynamic block division results, annotates the semantic boundary cutting points of each block and the semantic similarity heat map of adjacent blocks; Supports manual intervention to adjust the block granularity; Automatically annotates document time-series metadata (publication date, data year, publishing institution, etc.) to provide prior information for subsequent conflict resolution.

[0266] (2) Multi-agent collaborative process tracking interface: The multi-round interaction process of the three agents is displayed in the form of a time-series dialogue flow, including the candidate triple output of Aext (including data year annotation), the scoring feedback and rollback instruction of Aaudit, and the schema evolution decision (MERGE / CREATE) of Aonto; each triple is annotated with confidence level, original text basis and data year, and supports manual review and correction; for multi-chapter documents such as the "China Natural Resources Bulletin", parallel extraction by chapter is supported to accelerate processing.

[0267] (3) Knowledge graph visualization interface: The constructed knowledge graph is displayed in an interactive graph structure. Nodes are colored according to entity type (administrative unit, land type, monitoring indicator, policy event, disaster event, etc.); edges are distinguished by relationship type (causal, time series, dimensional constraint, spatial affiliation, new schema relationship, etc.); the schema evolution process can be replayed on the time axis, and new nodes (such as "blue carbon ecosystem carbon sink linkage") and relationship edges are highlighted; the knowledge subgraphs of specified data years can be filtered and displayed, effectively handling the problem of multiple time series coexistence.

[0268] (4) Conflict resolution and data traceability interface: Displays the results of multi-source data conflict detection (such as the time sequence difference between the annual release of cultivated land area data and the current year data) and the weighted voting score details; each knowledge node that is finally accepted can be traced back to the original document paragraph; provides visual prompts for the "data base year" and "expected next update" information automatically marked by the system; supports exporting standard triplet format (RDF / JSON-LD) for downstream applications.

[0269] The above interfaces together constitute the complete human-computer interaction layer of the multi-agent system of the present invention. Natural resource survey and monitoring documents, represented by the "2024 China Natural Resources Bulletin", can complete the visualization construction and management of knowledge graphs through the above interfaces.

[0270] 7. Final knowledge graph fragment in this embodiment (summary of partial extraction results based on the bulletin)

[0271] Following the above steps, the following is a partial list of final triples extracted from the publication in this embodiment, used to illustrate the knowledge graph structure output by this invention:

[0272]

[0273] Table 3. Comparison and verification of the effectiveness of the present invention in natural resource survey and monitoring scenarios.

[0274] Note: The relation types marked "(New)" in the table are the new schema relations that Aonto has expanded through the [CREATE] command in this embodiment, which verifies the schema adaptive evolution capability of the present invention without relying on a pre-complete fixed ontology.

Claims

1. A method for automatically constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring based on multi-agent collaboration, characterized in that, Includes the following steps: Step 1: Obtain unstructured documents related to natural resource surveys and monitoring, and dynamically segment them to generate multiple sub-document blocks. The dynamic segmentation is constrained by semantic continuity, and uses at least one of the following as segmentation constraints: data year, source entity, statistical caliber, spatial range, chapter level, table boundaries, and enumeration boundaries. Each sub-document block is associated with and stores block-level metadata, which includes at least the data year, source entity, statistical caliber, and chapter position. Step 2: Use a special Embedding model for the natural resources field to perform semantic recognition on the sub-document blocks, generate semantic vectors, and store the block-level metadata in association with the semantic vectors; Step 3: Define the extraction agent, the ontology maintenance agent, and the auditing agent; the extraction agent is used to extract candidate entities and candidate relation triples from the sub-document blocks. The ontology maintenance agent is used to normalize or expand candidate entities and candidate relationships. The auditing agent is used to verify the consistency of evidence, timing, statistical caliber, spatial hierarchy, and unit of measurement of candidate relation triples. Step 4: Extract candidate entity and relation triples based on the semantic vector, corresponding sub-document blocks, block-level metadata, context window, and generated graph state. Each candidate entity and relation triple includes at least a head entity, relation type, tail entity, original evidence interval, data year, source subject, statistical caliber, and initial confidence level. If there are undefined domain concepts or relations, proceed to step 5; otherwise, proceed to step 6. Step 5: The ontology maintenance agent performs entity normalization, relation normalization, candidate node generation or relation expansion operations based on semantic similarity, and updates the candidate entity and relation triplets. When the semantic distance between an undefined domain concept or relation and an existing set of entities or relations is less than a first distance threshold, a normalization operation is performed. When the semantic distance is greater than the second distance threshold, candidate node generation or relationship expansion operations are performed; when the semantic distance is between the first and second distance thresholds, a secondary judgment is performed by combining the original text evidence interval, contextual semantics, and business definition. Step 6: The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples, and marks candidate entities or relations that are missing original evidence, inconsistent data years, inconsistent statistical standards, mismatched spatial levels, inconsistent units of measurement, relation directions that do not conform to the business rules of natural resource survey and monitoring, or whose confidence scores are lower than the preset threshold as conflict items. If confidence score If the conflict occurs, the reflection mechanism is triggered, returning to step 4, and candidate entity and relation triples are re-extracted based on the original text evidence interval corresponding to the conflict item, the sub-document block in which it is located, the adjacent sub-document block, the relevant chapter, or the full text range. The preset threshold; Step 7: Use a weighted voting mechanism to process conflicting items, determine the final accepted value, and retain unaccepted sources as source tracing records, thereby generating and updating the final natural resource survey and monitoring knowledge graph; wherein, the weighted voting mechanism calculates the acceptance score of candidate values ​​based at least on source authority, time decay factor, statistical caliber matching degree, and spatial range matching degree, and the source tracing records include at least the unaccepted candidate value, source subject, data year, statistical caliber, and original evidence interval.

2. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... Unstructured documents related to natural resource surveys and monitoring include farmland survey and monitoring reports, urban territorial spatial planning monitoring reports, and natural resource bulletins; Among them, the farmland survey and monitoring report includes the farmland quality grade evaluation report, the basic farmland protection special report, and the annual report on farmland occupation and compensation balance; The urban territorial spatial planning monitoring report includes an urban territorial spatial health check report, a territorial spatial planning implementation evaluation report, and a construction land supply statistics report; The natural resources bulletin includes the China Land Resources Bulletin, the China Mineral Resources Report, and the China Natural Resources Bulletin.

3. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The steps of dynamic block partitioning are as follows: The block boundaries are determined primarily by semantic continuity, combined with heading level, table or enumeration boundaries, changes in entity type, data year, data source, and changes in statistical caliber. When the semantic similarity between adjacent sentence segments is higher than the segmentation threshold, the current segmentation window is expanded to preserve the complete semantic chain. When the semantic similarity between adjacent sentence segments is lower than the segmentation threshold, or when there is a switch in chapter, indicator caliber, data year, or source subject, the segment is automatically truncated and a new sub-document block is generated. At the same time, the data year, source, chapter position, and statistical caliber are saved as block-level metadata.

4. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The determination of undefined domain concepts or relationships is completed collaboratively by the extraction agent, the ontology maintenance agent, and the auditing agent; The judgment steps are as follows: The extraction agent first combines the current sub-document block, the context window and the generated graph state to identify whether the candidate entity or relation belongs to the expression of an existing concept or whether a new semantic orientation has appeared. Candidate entities or relations that do not belong to the expression of an existing concept or have a new semantic orientation are transmitted to the ontology maintenance agent. The ontology maintenance agent uses a large language model to semantically discriminate the names, definitions, hierarchical relationships, applicable scenarios, and original text evidence of candidate entities or relations, and transmits the candidate entities or relations as new concepts or new relations to the audit agent. The audit agent reviews the original evidence, contextual consistency, and business logic, and outputs undefined domain concepts or relationships.

5. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The steps of logical verification are as follows: The audit agent checks the evidence consistency, spatiotemporal constraints, units of measurement, business rules, and ontology constraints in sequence. It verifies whether the triples have original textual basis, whether the entity level and spatial scope match, whether the numerical units and time years are consistent, whether the causal or subordinate direction conforms to the common sense of natural resource management, and whether the relationship type falls within the defined schema or the schema to be expanded. If any check fails, the audit agent generates a REVISE reason and triggers the reflection mechanism, requiring the extraction agent to re-extract the problematic triples and related text.

6. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The embedding model specifically designed for the natural resources domain has been fine-tuned through comparative learning based on the InfoNCE loss function; The InfoNCE loss function is shown below: in, and These are the vector representations of the anchor text and the positive sample, respectively. For temperature hyperparameters, Let be the cosine similarity.

7. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The extraction agent models the extraction of candidate entities and relation triples as a conditional probability generation problem. Generate candidate entity and relation triples The probabilities are as follows: in, For input text, Representing head entities, relations, and tail entities. For large language model parameters; Confidence scores of candidate entity and relation triples As shown below: in, This represents the semantic fit between entities and relations. Representation and prior knowledge base Consistency test value, These are the weighting coefficients; The weighted voting mechanism is as follows: in, Weighting based on the authority of the data source. For the observation time difference, This is the time decay factor.

8. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The steps for performing normalization, candidate relation generation, or relation expansion operations include: Calculate unknown relation types Time and existing relationship set Calculate the semantic distance of each relation in the set, and identify the minimum distance, denoted as the minimum semantic distance. ; Among them, unknown relation types Time and existing relationship set China-US relations The semantic distance is shown below: Where v is the vector representation of the relation; If the minimum distance is less than the set minimum value, it is determined to be a synonym relationship, and the relationship normalization operation is performed. When the minimum distance falls between a set minimum value and a set threshold, the ontology maintenance agent first enters the candidate judgment state. It then performs a secondary judgment based on the relation definition, contextual evidence, and frequently co-existing entities. If the semantic direction aligns with an existing relation, it is normalized to the closest relation; otherwise, it is added as a candidate relation and written into the system after confirmation by the auditing agent. ; If the minimum distance is greater than the set threshold, it is determined to be a new mechanism, and an expansion operation is performed, that is, let .

9. The method for automatically constructing an unstructured document knowledge graph in the field of natural resource survey and monitoring based on multi-agent collaboration, as described in claim 1, is characterized in that... The re-extraction process includes: judge If the triplet corresponds to a paragraph that involves cross-chapter references, then the extraction scope will be expanded to include the relevant chapters; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after it, and are then re-extracted. judge Does the triplet have a global temporal conflict? If so, extend it to the entire text; otherwise, in The triples correspond to the original paragraph, the sub-document block, and the adjacent blocks before and after them, and are then re-extracted.

10. An automated system for constructing unstructured document knowledge graphs in the field of natural resource survey and monitoring, based on the method described in any one of claims 1-9, characterized in that, Including both the client-side and server-side; The user terminal acquires unstructured documents related to natural resource surveys and monitoring, and dynamically divides them into blocks to generate multiple sub-document blocks; The user terminal uses a special embedding model for the natural resources field to perform semantic recognition on sub-document blocks, generate semantic vectors, and transmit them to the server. The server-side storage includes extraction agents, ontology maintenance agents, and audit agents; The extraction agent extracts candidate entity and relation triples from the semantic vector and transmits them to the auditing agent; When the reflection mechanism is triggered, the extraction agent re-extracts candidate entity and relation triples from the semantic vector and transmits them to the audit agent; The ontology maintenance agent performs normalization, candidate node generation, or relation expansion operations on undefined entities or relations based on semantic similarity, and transmits the results to the auditing agent. The auditing agent performs logical verification and confidence assessment on all candidate entity and relation triples; if the confidence score is... This triggers the reflection mechanism; The auditing agent uses a weighted voting mechanism to resolve logical conflicts between multi-source unstructured data, generating and updating the final natural resource survey and monitoring knowledge graph.

Images