Dynamic and static dual-state knowledge graph generation system

Abstract

The invention relates to the technical field of knowledge graph generation, in particular to a dynamic and static dual-state knowledge graph generation system, which is characterized in that an object is used for replacing an entity as a node, a traditional triple is upgraded into a tetrad containing a timestamp, and a state tetrad is newly added; quadruple elements (including timestamp mining) are extracted through a semantic parser, a parameter space is maintained, a construction engine constructs a quadruple, stores the quadruple into a graph database and executes four-item rationality detection, a vacancy filling engine calls a fine-tuned large language model, and an intelligent agent screens high-credibility answers to complement knowledge. Traditional static representation limitation is broken through, object dynamic description is realized, accurate and complete atlas is guaranteed, and an important logic basis is provided for next-generation artificial intelligence interpretable reasoning.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of knowledge graph generation, and particularly relates to a dynamic and static knowledge graph generation system. BACKGROUND

[0002] The reasoning level of artificial intelligence depends on the completeness of the knowledge representation of the knowledge graph supporting the reasoning, and the completeness depends on the scientificity of the underlying knowledge framework. For a long time, Europe and the United States have dominated the development process of artificial intelligence, and their knowledge frameworks are mostly based on western philosophical ontology: in 1968, J.R. Quillian proposed the semantic network theory, which describes the concepts of things and their relationships through the structure of nodes and arcs, wherein the nodes represent the concepts themselves, and the arcs represent the relationships between the concepts; in 1974, Marvin Minsky proposed the "frame" model, which realizes the description of structured knowledge by means of slots and fillers, and is used to define the attribute characteristics or association relationships of entities; in 2004, W3C proposed the network ontology language OWL (Web Ontology Language), which further standardized the logical expression form of the attribute slot through the definition of data type attributes and object attributes; in 2012, Google proposed the concept of knowledge graph, which added the "entity-attribute name-attribute value" triple form on the basis of the semantic network "entity1-relation-entity2" triple structure, established the knowledge representation system for describing entities from the two independent dimensions of attributes and relationships, and the subsequent natural language processing (NLP) technology and knowledge graph generation technology are all based on this framework. On June 30, 2025, the Chinese national standard GB / T 45949-2025 "Chinese news semantic structured annotation (SSA)" was implemented, which reconstructed the AI underlying knowledge framework from the perspective of Chinese and Western integration, introduced the time dimension and the state dimension, realized the representation transition from the static "relation-attribute" to the dynamic "relation-attribute-state", and provided a complete and landable symbolic infrastructure for the rigorous reasoning of the next generation of artificial intelligence. The application is based on the innovative new knowledge graph automatic generation system of the national standard.

[0003] The Western knowledge framework based on the triplet, supported by ontology, lacks a temporal dimension and therefore cannot incorporate a "state" dimension, thus it is not a complete knowledge framework. Knowledge graphs built upon this framework are primarily suitable for representing static relationships between entities, supporting only association-based applications and being relatively insufficient in semantic representation of the dynamic evolution of things. Consequently, the derived Natural Language Processing (NLP) and knowledge graph generation technologies also have inherent limitations in describing dynamic things, making it difficult to support the construction of high-level machine cognitive capabilities and failing to meet the demands of next-generation artificial intelligence for interpretable reasoning. Therefore, in response to the above situation, there is an urgent need to develop a dynamic and static dual-state knowledge graph generation system to overcome the shortcomings in current practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic and static dual-state knowledge graph generation system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A dynamic-static dual-state knowledge graph generation system includes a dynamic-static dual-state semantic parser, a dynamic-static dual-state knowledge graph construction engine, and a dynamic-static dual-state knowledge graph knowledge gap filling engine;

[0007] The dynamic and static dual-state semantic parser uses "things" as node semantic units to identify and extract semantic units, semantic unit parameters, other semantic elements and timestamps from massive text data, and creates and maintains a parameter space.

[0008] The dynamic and static dual-state knowledge graph construction engine constructs timestamp-containing quadruples based on the extracted semantic elements and stores them in the database, while performing a rationality check.

[0009] The dynamic and static dual-state knowledge graph knowledge gap filling engine calls a large language model to fill in and correct missing or incorrect knowledge in the graph, thereby achieving a comprehensive representation of the static characteristics and dynamic evolution process of things.

[0010] As a further aspect of the present invention: the quadruple includes four types of structures:

[0011] The 4-tuple of Thing-Relationship Name-Relationship Value-Timestamp contains four elements: Thing, Relationship Name, Relationship Value, and Timestamp.

[0012] The subject-behavior-object-timestamp quadruple contains four elements: subject, behavior, object, and timestamp.

[0013] The 4-tuple of Thing-Attribute Name-Attribute Value-Timestamp contains four elements: Thing, Attribute Name, Attribute Value, and Timestamp.

[0014] The four-tuple of Transaction-StateName-StateValue-Timestamp contains four elements: Transaction, StateName, StateValue, and Timestamp.

[0015] In this context, both the subject and the object belong to the things described, and the timestamp records the first moment or the effective time when the corresponding semantics and interrelationships become facts;

[0016] The first moment refers to the moment when the corresponding semantics and interrelationships of the timestamp record become a fact; the valid time refers to the time within the validity period of the timestamp record.

[0017] As a further aspect of the present invention: the timestamp acquisition strategy of the dynamic-static dual-state semantic parser includes:

[0018] Extract the first moment directly from the raw data;

[0019] When the first moment is missing, extract the event reference time that can be used as the valid time.

[0020] When no reference event time is available, the author's article publication time should be used.

[0021] When the author's publication time is unclear, the first publication time of the information source should be used;

[0022] The timestamps are labeled according to their credibility level, and high-credibility timestamps are continuously updated through offline data collection.

[0023] As a further aspect of the present invention: the parameter space stores data pairs corresponding to parameter names and parameter values, wherein the parameter names include relation names, attribute names, and status names, and its creation and maintenance methods include:

[0024] The system automatically calculates the co-occurrence probability of parameters based on massive amounts of text data; allows users to input, modify, or delete parameters through a user-friendly interface; and utilizes a large language model to restore incomplete parameter relationships to expand the parameter data.

[0025] As a further aspect of the present invention: the rationality detection includes logical consistency detection, semantic rationality detection, time series rationality detection, and state constraint rationality detection;

[0026] The semantic rationality detection relies on the built-in semantic collocation dictionary, and the state constraint rationality detection sets constraints based on the objective laws of the development of things.

[0027] As a further aspect of the present invention, it also includes a semantic unit classification manager, which supports automatic or manual classification and management of node semantic units according to any dimension or combination of behavior, attribute, relationship and state.

[0028] As a further aspect of the present invention: the knowledge gap-filling engine completes missing or incorrect semantic units of nodes in the following ways:

[0029] Based on neighbor nodes and relationship types, the representation of missing nodes is inferred through formulas; node-related elements are input as prompt words into a large language model to generate the representation of missing nodes.

[0030] As a further aspect of the present invention: the knowledge gap filling engine corrects missing or incorrect edges in the following way:

[0031] When the parameter value is known, the corresponding parameter name is first retrieved from the parameter space; if the retrieval fails, the parameter name is generated through the large language model, and the new parameter name and the corresponding data pair are added to the parameter space while completing the edge.

[0032] As a further aspect of the present invention, the knowledge gap-filling engine discovers new knowledge in the following ways:

[0033] When the semantic parser extracts an item that is not in the semantic unit classification manager, the large language model is activated to query the item's attributes, relationships, behaviors, and state information. Valid information is extracted to construct a new quadruple and incorporate it into the knowledge graph.

[0034] As a further aspect of the present invention: the large language model in the knowledge gap filling engine is fine-tuned to adapt to the knowledge framework of the system;

[0035] It also includes task management and answer credibility assessment and screening agents. These agents send the same knowledge request to multiple large language models, calculate the answer credibility based on the model weights, and select the answer with the highest credibility as supplementary knowledge.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] This invention replaces traditional "entities" with "things" that have both static and dynamic characteristics as nodes in the knowledge graph, upgrades the triplet structure to a quadruple structure that includes timestamps, and introduces a "state" dimension to add exclusive quadruples. This breaks through the limitations of traditional knowledge frameworks that only describe entities from the two dimensions of attributes and relations, and realizes the leap from static description of "relationship-attribute" to dynamic description of "relationship-attribute-state". This makes the knowledge representation more complete and can fully represent the static characteristics and dynamic evolution process of things, providing an important logical foundation for the interpretable reasoning of next-generation artificial intelligence.

[0038] The timestamp recognition and mining module uses a multi-scenario timestamp acquisition strategy (first moment extraction, effective time extraction, author publication time replacement, and information source publication time supplementation) to ensure that all four-tuples can obtain timestamps. It also continuously collects data offline to update high-reliability timestamps, effectively capturing the dynamic characteristics of things, ensuring the accuracy and completeness of the knowledge graph's time information, and providing key support for dynamic semantic representation.

[0039] The creation and maintenance of the parameter space, by standardizing parameter name definitions (unifying synonyms / near-synonyms) and solidifying the correspondence between parameter names and parameter values, facilitates parameter querying, comparison, and completion, improves the standardization of knowledge graph construction, and provides data support for subsequent knowledge gap filling (such as parameter name restoration when edges are missing), ensuring the consistency of the knowledge graph.

[0040] The dynamic and static dual-state knowledge graph construction engine has four major functions: logical consistency detection, semantic rationality detection (based on a semantic collocation dictionary), time series rationality detection, and state constraint rationality detection. These functions verify the quadruplets from the dimensions of logic, semantics, time, and state constraints, effectively avoiding problems such as logical loops, semantic errors, temporal logic confusion, and violations of state change rules, thus ensuring the rigor and accuracy of the knowledge graph.

[0041] The knowledge gap filling engine employs node missing inference (combining formula calculation with a large language model), edge missing correction (combining parameter space retrieval with large language model generation), and new knowledge mining (large language model query). Leveraging a finely tuned large language model and a multi-model answer credibility filtering mechanism (agent weight allocation calculation), it effectively avoids the intelligent illusion problem inherent in single large language models, ensuring the accuracy of supplemented knowledge. It also replaces the traditional method of knowledge completion relying on raw data collection, significantly improving the efficiency and convenience of knowledge completion and continuously driving the improvement of the knowledge graph. Simultaneously, the semantic unit classification manager supports automatic or manual classification management of nodes across multiple dimensions, replacing traditional classification directory indexes and enhancing the management efficiency and retrieval convenience of the knowledge graph.

[0042] The overall system achieves automated construction and continuous optimization of dynamic and static dual-state knowledge graphs through the operation process of "semantic element extraction - quadruple construction and detection - knowledge gap filling - iterative optimization". It can process massive text data, continuously expand the coverage of knowledge graphs, support the construction of high-level machine cognitive capabilities, and meet the application needs of next-generation artificial intelligence. Attached Figure Description

[0043] Figure 1 This is a diagram illustrating the overall framework of the dynamic and static dual-state knowledge graph generation system in this embodiment of the invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0045] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0046] Please see Figure 1 The present invention provides a dynamic and static dual-state knowledge graph generation system, which improves the underlying knowledge framework of artificial intelligence and combines natural language processing (NLP) technology, neural network deep learning technology and timestamp recognition and mining technology to realize the automatic generation and intelligent completion of knowledge graphs, providing an important logical foundation for the next generation of interpretable reasoning in artificial intelligence.

[0047] I. Implementation of the Static-Dynamic Bi-state Semantic Parser

[0048] The core function of the dynamic-static dual-state semantic parser is to identify and extract semantic elements in text that meet the requirements of a dynamic-static dual-state knowledge graph, and to create and maintain a parameter space. The specific implementation process is as follows:

[0049] 1. Construction of a knowledge framework

[0050] First, we define "things" as the semantic units of nodes in the knowledge graph, replacing "entities" in traditional knowledge graphs (redefining "entities" as "objects" with only static features); around "things," we construct four four-tuple structures to form the knowledge framework of a dynamic and static dual-state knowledge graph, specifically including:

[0051] Item-Relationship Name-Relationship Value-Timestamp: This describes the association between items and the time when the relationship was established;

[0052] Subject-Behavior-Object-Timestamp: This term describes the behavior of a thing and the time information when the behavior occurs, where both "subject" and "object" belong to the category of "thing".

[0053] Item-Attribute Name-Attribute Value-Timestamp: Used to describe the attribute characteristics of an item and the time information of the attribute value;

[0054] Thing-StateName-StateValue-Timestamp: This is used to specifically describe the state of a thing and the time when the state takes effect. The "state value" covers scenarios such as the establishment of relationships between things, the values ​​of attributes, the occurrence and development of events.

[0055] The timestamp in each quadruple records the exact moment (first moment) when the semantics represented by the first three elements of the quadruple and their interrelationships become facts, or the time within the validity period (valid time). By introducing timestamps and state dimensions, the limitations of traditional triplet knowledge frameworks, which lack time data and cannot represent dynamic features, are overcome. This achieves a leap from static description of "relation-attribute" to dynamic description of "relation-attribute-state", making knowledge representation more complete.

[0056] 2. Implementation of the semantic element recognition and extraction module

[0057] The semantic element recognition and extraction module uses neural network deep learning technology to identify and extract all semantic elements related to the above four quadruples from the massive amount of raw text data, including things, relation names, relation values, subjects, behaviors, objects, attribute names, attribute values, state names, state values, and timestamps.

[0058] Among these, the identification, extraction, and mining of timestamps are key to capturing the dynamic characteristics of things. The specific implementation strategies are as follows:

[0059] If the original data contains a first moment, that first moment is directly extracted as a timestamp and included in the quadruple. The reliability of this timestamp is 100%, ensuring the accuracy of the time information.

[0060] If the first moment is missing, find the time of things that can prove the validity of the relationship and include it in the quadruple as a reference timestamp. For example, "the master and apprentice in Shanghai in 1985", use "1985" as the reference timestamp of the master-apprentice relationship to prove that the relationship was established and valid in 1985 and before, thus making up for the lack of the first moment.

[0061] If a timestamp is missing and there is no reference event time, the time when the author published the article will be used as the reference timestamp, assuming that the time is when the description of the event is truly valid, thus ensuring the availability of the timestamp;

[0062] If the timestamp is missing, there is no reference event time, and the author's publication time is unclear, the time when the target article was first published by the information source will be used as the reference timestamp. It is also assumed that this time is when the description of the thing is truly valid, so as to ensure that the timestamp can be obtained in all scenarios.

[0063] Meanwhile, the timestamp identification and mining module adds specific identifiers to all reference timestamps to distinguish their credibility levels (first moment > event reference time > author publication time > information source publication time), and continuously collects data related to missing timestamps offline, or searches for missing timestamp information from the new data each time the data is updated. Once the first moment or a more credible valid timestamp is found, it immediately replaces the corresponding reference timestamp in the graph, so that the time information of the knowledge graph is continuously improved.

[0064] 3. Creation and maintenance of parameter space

[0065] The parameter space is used to store data pairs corresponding to parameter names and parameter values, including relation names and relation values, attribute names and attribute values, and status names and status values. Its creation and maintenance are achieved through the following methods:

[0066] Deploy a program to automatically calculate the co-occurrence probability of parameter names and values, mine high-frequency parameter combinations based on massive text data, and provide basic data for the parameter space;

[0067] It provides an interactive interface for manually supplementing parameters, supporting manual input, modification, or deletion of parameter data to meet parameter standardization requirements in special scenarios;

[0068] Deploy a program that uses a large language model to generate parameters, restore incomplete parameter relationships through the large language model, and expand the coverage of the parameter space.

[0069] The establishment of a parameter space can standardize the definition of parameter names, unify synonyms or near-synonyms (such as "color" and "color"), facilitate computers to identify the same attributes, and improve the efficiency of parameter query and comparison; at the same time, it solidifies the correspondence between parameter names and parameter values, supports reverse lookup of parameter names through parameter values, and vice versa, and provides support for the subsequent construction and filling of knowledge graphs.

[0070] II. Implementation of the Dynamic and Static Dual-State Knowledge Graph Construction Engine

[0071] The dynamic and static dual-state knowledge graph construction engine constructs four major quadruplets based on the semantic elements output by the semantic parser and stores them in the graph database. At the same time, it performs multiple detection functions to ensure the rationality and accuracy of the graph. The specific implementation is as follows:

[0072] 1. Construction and storage of quadruples

[0073] The engine builds a directed acyclic network graph structure with the "thing" in the quadruple as the head node, the parameter name (which can represent behavior, relation name, attribute name, state name) as the edge (with timestamps marked on the edge), and the parameter value (which can represent another thing, attribute value, state value, etc.) as the tail node. The four quadruples are stored in a graph database (or object-oriented database).

[0074] This storage method breaks away from the traditional distributed storage model of triplet graphs, realizing the integrated representation of the attributes, relationships, behaviors, states, and time information of things, making the structure of the knowledge graph more unified and facilitating subsequent retrieval and reasoning.

[0075] 2. Implementation of the four major detection functions

[0076] Logical consistency check: The algorithm checks whether the dynamic and static dual-state knowledge graph is a directed acyclic graph, avoiding reasoning errors caused by logical loops and ensuring the logical rigor of the graph;

[0077] Semantic rationality detection: The built-in dictionary of commonly used semantic collocations allows for the query of reasonable collocations in the dictionary when a certain semantic element is known. If the collocation matches the collocation in the data, the verification is successful; otherwise, the verification fails. This avoids situations where the syntax is correct but the semantics are incorrect, thus improving the semantic accuracy of the graph. The semantic collocation dictionary can be continuously updated through manual maintenance or by using a large model to restore incomplete semantic collocation relationships.

[0078] Time series rationality check: Verify whether the timestamps in the quadruple conform to objective laws such as unidirectional time flow, avoid time logic errors (such as the time of a change in the state of a thing being earlier than the time of the establishment of its relationship), and ensure the consistency of the time dimension of the graph;

[0079] State constraint rationality test: Based on objective laws, set constraints on state changes (such as temperature can only rise or fall), and verify whether the dynamic characteristics of things captured in the graph meet these constraints to ensure that the state description conforms to the objective logic of the development of things.

[0080] 3. Semantic Unit Classification Management

[0081] Deploy a dynamic and static dual-state knowledge graph semantic unit classification manager, supporting the classification and management of node semantic units (things) in the graph based on any one or a combination of four dimensions: behavior, attribute, relationship, and state. This function can replace the traditional classification directory index, with the system automatically completing the classification. It also supports manual expansion and maintenance, improving the classification flexibility and management efficiency of the knowledge graph, and facilitating the rapid location of target semantic units.

[0082] In practice, training and testing sets of artificial neural networks are constructed by manually annotating according to the knowledge framework defined in this invention, and specialized small models are trained to accurately identify the elements required to generate the four major quadruplets in massive text data, thus solving the problem of semantic element extraction efficiency under massive data.

[0083] III. Implementation of the Knowledge Gap Filling Engine for Dynamic and Static Dual-State Knowledge Graph

[0084] The knowledge gap completion engine addresses missing semantic elements, unrecognized new knowledge, and error information during the generation of quadruples by calling multiple large language models to complete the knowledge gaps. The specific implementation is as follows:

[0085] 1. Identification and completion of missing (or incorrect) semantic units in nodes.

[0086] When there are missing or incorrect semantic units in the graph, the representation of the missing node is inferred by jointly using neighboring nodes and relation types, using the following formula:

[0087]

[0088] in, Here is the representation of node v at layer l; Let u be the set of neighbors of node v that is missing under relation r; u is any neighbor node; R is the set of relation types (i.e., parameter names); A relation-specific weight matrix (learnable); Normalization constant (e.g., degree normalization) σ is the activation function (such as ReLU).

[0089] The application of this formula is based on the relation type R and its weights. The known assumptions are supported by the parameter space. Simultaneously, it is also possible to... , , R and As prompt words, they are input into a large language model to generate The objective is to fill in missing nodes. Both methods can accurately infer the representation of missing nodes, compensate for missing or incorrect semantic units of nodes, and improve the completeness of the graph.

[0090] 2. Identification and correction of missing (or incorrect) edges

[0091] Missing (or incorrect) edges mainly refer to the situation where parameter names are missing in a quadruple. The correction process is as follows:

[0092] Given the parameter values, first search the parameter space for the parameter name and parameter value data pair corresponding to the parameter value. If the search is successful, directly restore the corresponding parameter name to the knowledge graph and fill in the missing edges.

[0093] If no relevant record is found in the parameter space, the parameter value is input into the knowledge gap filling big language model to generate the corresponding parameter name as the task objective. Based on the output of the big language model, a new parameter name and parameter value data pair is constructed. On the one hand, the newly generated parameter name is used to fill in the missing edges in the graph, and on the other hand, the new parameter is added to the parameter space to provide support for subsequent graph construction.

[0094] This process combines parameter space retrieval with large language model generation to efficiently restore missing parameter names, correct edge errors, and ensure the complete and accurate relationships between nodes in the graph.

[0095] 3. Discovery and exploration of new knowledge

[0096] When the semantic parser extracts an item that is not in the semantic unit classification manager, it is determined that the item is new knowledge that the system cannot recognize. At this time, the knowledge gap filling big language model is directly launched, and it is asked for relevant information about the new knowledge (including attributes, relationships, behaviors and states, etc.). Valid information is extracted from the answer of the big language model, new quadruplets are constructed and incorporated into the knowledge graph, so as to realize the supplementation of new knowledge and continuously expand the coverage of the knowledge graph.

[0097] 4. Implementation of the knowledge completion module for the large language model

[0098] The large language model used for knowledge gap filling is fine-tuned to adapt it to the knowledge framework of this invention, familiarize it with relevant terminology and expression paradigms, and improve the accuracy of knowledge gap filling.

[0099] Deploy a task management and answer credibility assessment and screening agent, which uniformly executes the knowledge gap filling task: for query needs in scenarios such as missing nodes, missing edges and new knowledge mining, make the same knowledge request to multiple large language models;

[0100] Based on the fundamental assumption that "the same or similar answers given by most models have high credibility," the agent first calculates the average semantic similarity of the answers from each major model as the total credibility:

[0101]

[0102] in, For the number of models, are the vectors of the i / j-th model's response, and Sim is the cosine similarity function.

[0103] Then, set the overall credibility. A similar threshold (e.g., 0.8) is used to aggregate large model answers with answers greater than or equal to the threshold. A text fusion model (e.g., BART, T5) is used to deduplicatize and integrate this answer group, retaining the core semantics and optimizing the expression. Finally, it is output as supplementary knowledge.

[0104] The supplemented knowledge is returned to the semantic element recognition and extraction module to enter the next round of the main graph generation process to complete the knowledge completion; the knowledge completion process can be repeated until the dynamic and static dual-state knowledge graph reaches the preset completeness requirements.

[0105] This module effectively avoids the intelligent illusion problem that may exist in a single large language model through multi-model query and credibility screening mechanism, ensuring the accuracy of knowledge supplementation. At the same time, it replaces the traditional supplementation method that relies on raw data collection by leveraging the massive knowledge reserves of large language models, greatly improving the efficiency and convenience of knowledge supplementation.

[0106] In summary, the overall operation flow of the dynamic and static dual-state knowledge graph generation system of the present invention is as follows:

[0107] Step 1: Input massive text data into the semantic element recognition and extraction module. This module extracts the semantic elements required by the four major quadruplets based on the knowledge framework. The timestamp recognition and mining sub-module simultaneously extracts and supplements the timestamps.

[0108] Step 2: If there are missing parameters (especially missing parameter names) during the extraction process, they are filled in using the parameter space. The parameter space is continuously expanded through word frequency statistics, manual supplementation, and generation by a large language model.

[0109] Step 3: The extracted semantic elements are passed to the dynamic and static dual-state knowledge graph construction engine to build four quadruplets and store them in the graph database. At the same time, logical consistency detection, time series rationality detection, state constraint rationality detection and semantic rationality detection are performed.

[0110] Step 4: The detected quadruplets are stored in the knowledge graph (repository) and incorporated into the semantic unit classification manager, completing one round of the main graph generation process;

[0111] Step 5: The knowledge gap filling engine detects the generated graph, identifies missing (or incorrect) nodes and missing (or incorrect) edges, starts the large language model gap filling process, completes the knowledge and returns it to the semantic element recognition and extraction module, and enters the next round of the main process;

[0112] Step 6: Semantic elements (new knowledge) that cannot be recognized by the semantic parser are restored by the large language model and returned to the semantic element recognition and extraction module to enter the next round of the main process, generating new quadruplets to achieve knowledge supplementation.

[0113] Through the above process, the system realizes the automated construction, continuous completion and optimization of dynamic and static dual-state knowledge graphs. The generated knowledge graphs can comprehensively represent the static characteristics and dynamic evolution process of things, providing solid knowledge support for high-level cognition and interpretable reasoning in artificial intelligence.

[0114] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A dynamic and static dual-state knowledge graph generation system, characterized in that, This includes a dynamic-static dual-state semantic parser, a dynamic-static dual-state knowledge graph construction engine, and a dynamic-static dual-state knowledge graph knowledge gap filling engine; The dynamic and static dual-state semantic parser uses "things" as node semantic units to identify and extract semantic units, semantic unit parameters, other semantic elements and timestamps from massive text data, and creates and maintains a parameter space. The dynamic and static dual-state knowledge graph construction engine constructs timestamp-containing quadruples based on the extracted semantic elements and stores them in the database, while performing a rationality check. The dynamic and static dual-state knowledge graph knowledge gap filling engine calls a large language model to fill in and correct missing or incorrect knowledge in the graph, thereby achieving a comprehensive representation of the static characteristics and dynamic evolution process of things.

2. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The quadruple includes four types of structures: The 4-tuple of Thing-Relationship Name-Relationship Value-Timestamp contains four elements: Thing, Relationship Name, Relationship Value, and Timestamp. The subject-behavior-object-timestamp quadruple contains four elements: subject, behavior, object, and timestamp. The 4-tuple of Thing-Attribute Name-Attribute Value-Timestamp contains four elements: Thing, Attribute Name, Attribute Value, and Timestamp. The four-tuple of Transaction-StateName-StateValue-Timestamp contains four elements: Transaction, StateName, StateValue, and Timestamp. In this context, both the subject and the object belong to the things described, and the timestamp records the first moment or the effective time when the corresponding semantics and interrelationships become facts; The first moment refers to the moment when the corresponding semantics and interrelationships of the timestamp record become a fact; the valid time refers to the time within the validity period of the timestamp record.

3. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The timestamp acquisition strategy of the dynamic and static dual-state semantic parser includes: Extract the first moment directly from the raw data; When the first moment is missing, extract the event reference time that can be used as the valid time. When no reference event time is available, the author's article publication time should be used. When the author's publication time is unclear, the first publication time of the information source should be used; The timestamps are labeled according to their credibility level, and high-credibility timestamps are continuously updated through offline data collection.

4. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The parameter space stores data pairs corresponding to parameter names and parameter values. Parameter names include relation names, attribute names, and status names. Its creation and maintenance methods include: The system automatically calculates the co-occurrence probability of parameters based on massive amounts of text data; allows users to input, modify, or delete parameters through a user-friendly interface; and utilizes a large language model to restore incomplete parameter relationships to expand the parameter data.

5. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The rationality detection includes logical consistency detection, semantic rationality detection, time series rationality detection, and state constraint rationality detection; The semantic rationality detection relies on the built-in semantic collocation dictionary, and the state constraint rationality detection sets constraints based on the objective laws of the development of things.

6. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, It also includes a semantic unit classification manager, which supports automatic or manual classification and management of node semantic units based on any dimension or combination of behavior, attribute, relationship and state.

7. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The knowledge gap-filling engine includes the following methods for completing missing or incorrect semantic units of nodes: Based on neighbor nodes and relationship types, the representation of missing nodes is inferred through formulas; node-related elements are input as prompt words into a large language model to generate the representation of missing nodes.

8. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The knowledge gap filling engine corrects missing or incorrect edges in the following way: When the parameter value is known, the corresponding parameter name is first retrieved from the parameter space; if the retrieval fails, the parameter name is generated through the large language model, and the new parameter name and the corresponding data pair are added to the parameter space while completing the edge.

9. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The knowledge gap filling engine discovers new knowledge in the following ways: When the semantic parser extracts an item that is not in the semantic unit classification manager, the large language model is activated to query the item's attributes, relationships, behaviors, and state information. Valid information is extracted to construct a new quadruple and incorporate it into the knowledge graph.

10. The dynamic and static dual-state knowledge graph generation system according to claim 1, characterized in that, The large language model in the knowledge gap filling engine is fine-tuned to fit the knowledge framework of the adaptation system; It also includes task management and answer credibility assessment and screening agents. These agents send the same knowledge request to multiple large language models, calculate the answer credibility based on the model weights, and select the answer with the highest credibility as supplementary knowledge.

Images