Method for dynamically constructing and evolving astronomical knowledge graph based on multi-agent

Abstract

The application relates to the technical field of artificial intelligence application, and discloses an astronomical knowledge graph dynamic construction and evolution method based on multiple agents, which obtains original data from multiple-source heterogeneous astronomical data sources in parallel, carries out format analysis and structured reconstruction, and generates standardized text data in a unified format; carries out multi-dimensional parallel information extraction on the standardized text data, and generates a candidate knowledge triple set; carries out adaptive semantic alignment and disambiguation processing on entity reference items in the candidate knowledge triples based on a pre-constructed seed knowledge base, and generates standardized triple data; carries out conflict detection on the standardized triple data and existing knowledge in an existing knowledge graph, and completes dynamic evolution of the knowledge graph; and constructs a dual-modal knowledge graph supporting structured multi-hop query and semantic similarity retrieval. Through division and cooperation of the agents, the application establishes a conflict resolution criterion based on confidence weighting, and realizes time-effectiveness adaptive updating of the graph.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, especially the combination of artificial intelligence and astronomy, and in particular, to a method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems. Background Technology

[0002] With the advent of the era of full-spectrum and multi-messenger astronomy, large-scale sky surveys are continuously generating massive amounts of observational data, enabling astronomy to fully enter a data-intensive research paradigm. However, these valuable knowledge assets are highly fragmented. The latest theoretical derivations and physical models are mainly stored in unstructured databases such as arXiv and the Astrophysical Data System (ADS). Meanwhile, precise physical parameters of celestial objects, such as coordinates, parallax, and spectral types, are stored in semi-structured dedicated star catalog databases such as Simbad and VizieR. This "data silo" distribution means that researchers still need to spend a significant amount of time on cross-database searches and manual comparisons when studying target celestial objects, and the efficiency of knowledge acquisition lags far behind the speed of data growth.

[0003] Currently, the automated construction of knowledge graphs for unstructured text mainly follows two technical paradigms. The first is the traditional pipeline-based information extraction method, which separates named entity recognition and relation extraction into independent serial modules. This type of method is prone to the cascading propagation of upstream identification errors to downstream methods when dealing with entity aliases commonly found in astronomical literature (such as the Crab Nebula also being referred to as "NGC1952" or "PKS0531+21") and nested long and complex sentences. The second is the end-to-end generative extraction method based on large language models. While this method has the advantage of semantic generalization, it is prone to producing "factual illusions" when dealing with physical parameters with multiple numerical dependencies. For example, for conditional descriptions such as "Satellite A measured diameter X, Satellite B measured diameter Y," generative models often forcibly merge multiple source values ​​or misattribute them, failing to accurately restore the strict correspondence between "numerical value - observation equipment - physical attribute." This makes the extracted triples scientifically unreliable and unable to support subsequent precise numerical reasoning.

[0004] In addition, astronomy is an observation-driven, dynamically evolving discipline, with new transient source discoveries and more precise corrections to the Hubble constant being made daily. However, existing knowledge graph construction systems generally follow a static model of "one-time construction and fixed storage," severely lacking the ability to agilely incorporate incremental knowledge. More critically, when newly ingested observational data conflicts with existing records in the graph in terms of physical properties (e.g., updated celestial distances obtained through higher-precision measurements), traditional systems lack automated conflict resolution mechanisms. They cannot determine whether the discrepancy stems from semantic noise or represents an improvement in scientific accuracy, nor can they decide whether to perform an overwrite update or allow the versions to coexist. This gap in mechanisms leads to a long-term accumulation of outdated or contradictory information in the graph, significantly diminishing the practical value of knowledge graphs in assisting astronomical research. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems. It aims to overcome the bottleneck of extraction accuracy, establish a conflict resolution criterion based on confidence weighting, and achieve timely adaptive updates of the graph through the professional division of labor and pipeline-style collaboration of agents. Ultimately, it aims to build a high-fidelity, evolvable structured knowledge base for intelligent question answering and complex reasoning in the field of astronomy.

[0006] Based on the first main aspect of the present invention, a method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems is provided, comprising the following steps:

[0007] The data reconnaissance agent acquires raw data in parallel from multiple heterogeneous astronomical data sources, performs format parsing and structure reconstruction on the raw data, and generates standardized text data in a unified format.

[0008] The information extraction agent performs multi-dimensional parallel information extraction on the standardized text data to identify astronomical entities and their static attributes, dynamic relationships and time-series events, and generates a set of candidate knowledge triples.

[0009] The knowledge auditing agent performs adaptive semantic alignment and disambiguation on the entity references in the candidate knowledge triples based on a pre-built seed knowledge base, mapping non-standard entity references to standard terminology identifiers and generating normalized triple data.

[0010] The graph architect agent performs conflict detection between the normalized triplet data and existing knowledge in the existing knowledge graph, calculates a comprehensive confidence score based on the authority of the data source and the timestamp, and determines whether to perform addition, update or archive operations based on the comprehensive confidence score, thus completing the dynamic evolution of the knowledge graph.

[0011] The updated knowledge graph stores the topological relationships of entities in a graph database and stores the semantic feature vectors of entities and relationships in a vector database, thus constructing a bimodal knowledge graph that supports structured multi-hop queries and semantic similarity retrieval.

[0012] As a whole concept, this invention is the first to systematically integrate the entire process of multi-source heterogeneous data acquisition, semantic disambiguation, conflict resolution and incremental evolution from a methodological perspective. It provides a holistic implementation framework for solving technical problems such as the difficulty in accurately extracting divergent terminology and deep logical connections, the lack of a basis for resolving multi-source spatiotemporal numerical conflicts, and the inability to dynamically evolve the fixed graph architecture.

[0013] Optionally, the information extraction agent adopts a parallel operation mode, simultaneously performing the following four independent extraction tasks on the standardized text data:

[0014] Multidimensional entity recognition task, used to identify entities of celestial objects, observation facilities, and physical concept categories;

[0015] Static attribute extraction task, used to extract the physical parameter attribute values ​​of entities;

[0016] The dynamic relationship building task is used to identify relationship triples and their conditional attributes between entities;

[0017] The time-series event capture task is used to extract event information that includes time anchors and evolution results.

[0018] The above scheme overcomes the shortcomings of traditional pipeline-style extraction methods that suffer from error propagation due to serial cascading by limiting the operation mode of the information extraction agent to four parallel and independent tasks: multi-dimensional entity recognition, static attribute extraction, dynamic relationship construction, and temporal event capture. At the same time, it improves the accuracy of capturing deep physical logic and temporal dynamic information by parallel decoupling, addressing the multiple conditional dependencies of "numerical value-observation equipment-physical properties" that are common in astronomical literature. Thus, it provides a direct local optimization solution for the technical challenge of missing extraction accuracy and scientific logic.

[0019] Furthermore, the adaptive semantic alignment and disambiguation processing includes the following sub-steps:

[0020] The candidate entity references to be processed are encoded into semantic vectors, and the top-K standard terms with the highest semantic similarity are retrieved from the vector index of the seed knowledge base as candidate alignment items.

[0021] The original context text corresponding to the candidate entity reference, the standard term name of the candidate alignment item and its authoritative definition are jointly input into the large language model. The large language model performs triangulation to determine whether the semantics of the candidate entity reference item in the current context is consistent with the authoritative definition of the candidate alignment item.

[0022] If the determination result is consistent, a merge instruction is generated to map the candidate entity reference to the standard term identifier of the candidate alignment item; if the determination result is inconsistent, the candidate entity reference is marked as a new term and registered to the seed knowledge base.

[0023] The above scheme constructs a two-stage hybrid alignment mechanism of vector coarse screening and large language model triangulation verification. By inputting the original context of the candidate entity, the retrieved standard term name and its authoritative definition into the large language model for cross-discrimination, it not only ensures the breadth of recall but also achieves accurate normalization. At the same time, through the branch logic of merging when the discrimination is consistent and registering as a new term when it is inconsistent, a dynamic acceptance channel is reserved for unknown new concepts, fundamentally solving the problem of low knowledge quality caused by ambiguity in entity reference.

[0024] Furthermore, in the conflict resolution and incremental update steps, the comprehensive confidence score... Calculate using the following formula:

[0025]

[0026] in, This represents the authority score of the data source, which is a preset constant. Different data source types correspond to different scores. value; This represents a time decay function based on timestamps. This represents the data publication timestamp carried in the normalized triplet data;

[0027] The time decay function is expressed by the formula. Calculation, where Indicates the current system time. Indicates the data publication timestamp. Indicates the preset attenuation coefficient. It is a natural constant;

[0028] If the overall confidence score of the new data is higher than that of the existing data, the new data will be updated to the current value and the existing data will be archived as a historical version.

[0029] If the difference between the combined confidence scores of the new data and the existing data is within the preset error range, then both will be retained as coexisting viewpoints.

[0030] The above scheme provides a scientific quantitative basis for determining numerical conflicts in multi-source heterogeneous data over time. By introducing a comprehensive confidence calculation formula that integrates the authority score of the data source with an exponential decay function based on timestamps, the subjective conflict resolution that originally relied on human experience is transformed into an automatically executed objective mathematical comparison. When the confidence of new data is higher than that of old data, an update is performed to cover it; when the difference between the two is within the error range, it is retained as a coexisting version. This accurately distinguishes between improved measurement accuracy and semantic noise interference, completely solving the persistent problem of static solidification of outdated and contradictory information in the graph.

[0031] Furthermore, the graph database adopts an index-free adjacency architecture to store entity nodes, relation edges, and attribute key-value pairs, and supports multi-hop path traversal queries.

[0032] The vector database is used to store entity name vectors, entity definition vectors, and relationship description vectors encoded by the embedding model, and supports semantic retrieval based on cosine similarity.

[0033] By employing a dual-modal storage scheme—a graph database using an index-free adjacency architecture and a vector database storing semantic vectors encoded by an embedded model—the updated knowledge graph receives underlying physical support that combines logical rigor and semantic generalization. The graph database, with its O(1) adjacency traversal efficiency, lays the performance foundation for structured multi-hop path queries, while the vector database compensates for the shortcomings of exact matching in long-tail knowledge retrieval through cosine similarity retrieval. Together, they ensure that the high-fidelity knowledge graph constructed upstream can be efficiently and accurately consumed and utilized by the downstream inference system.

[0034] Furthermore, the seed knowledge base is integrated from the Unified Astronomical Terminology (UAT), which includes standard term names and corresponding authoritative scientific definitions; the vector index of the seed knowledge base is dynamically updated as new terms are registered, enabling the seed knowledge base to have adaptive growth capabilities.

[0035] The statistical data of the knowledge graph includes: the number of entity nodes, the number of relation edges, the number of attribute key-value pairs, the number of time-series event records, and the number of source documents; the entity nodes at least cover the categories of stars, galaxies, astronomical events, observation instruments, and physical concepts; the attribute key-value pairs include celestial coordinates, spectral type, magnitude, parallax, proper motion, and radial velocity physical parameters.

[0036] The above scheme ensures the long-term authority and continuous expansion capability of the terminology standardization benchmark by integrating a unified astronomical thesaurus (UAT) at the seed knowledge base level and limiting its adaptive growth mechanism to dynamically update with the registration of new terms. Simultaneously, by specifying the concrete limitations of entity nodes (covering stars, galaxies, astronomical events, etc.) and attribute key-value pairs (covering coordinates, spectral type, magnitude, parallax, etc.) in the knowledge graph, the granularity of knowledge coverage and parameter richness of the graph for the astronomical vertical field are clarified, thereby ensuring that the structured knowledge output by the system can fully meet the stringent application requirements of scientific research-level numerical computation and multi-hop source reasoning.

[0037] Based on the second main aspect of the present invention, a multi-agent-based dynamic construction and evolution system for astronomical knowledge graphs is provided for implementing the aforementioned method, comprising the following components:

[0038] A data reconnaissance intelligent agent is used to acquire raw data in parallel from multiple heterogeneous astronomical data sources, perform format parsing and structure reconstruction on the raw data, and generate standardized text data in a unified format.

[0039] An information extraction agent is used to perform multi-dimensional parallel information extraction on the standardized text data, identify astronomical entities and their static attributes, dynamic relationships and time-series events, and generate a set of candidate knowledge triples.

[0040] A knowledge review agent is used to perform adaptive semantic alignment and disambiguation processing on entity references in the candidate knowledge triples based on a pre-built seed knowledge base, mapping non-standard entity references to standard terminology identifiers and generating normalized triple data.

[0041] The knowledge graph architect agent is used to detect conflicts between the normalized triplet data and existing knowledge in the existing knowledge graph, calculate the comprehensive confidence score based on the authority of the data source and the timestamp, and determine whether to perform addition, update or archive operations based on the comprehensive confidence score to complete the dynamic evolution of the knowledge graph.

[0042] The dual-modal storage module is used to store the entity topological relationships in the updated knowledge graph to the graph database and the semantic feature vectors of entities and relationships to the vector database, thereby constructing a dual-modal knowledge graph that supports structured multi-hop queries and semantic similarity retrieval.

[0043] The above solution solidifies the core method steps of this invention into a system structure in the form of logical modules: a data reconnaissance agent, an information extraction agent, a knowledge verification agent, a graph architect agent, and a dual-modal storage module. It clarifies the functional boundaries and interaction relationships of each component. This enables the invention to be industrially implemented as a complete hardware and software system.

[0044] Furthermore, the data reconnaissance agent, information extraction agent, knowledge verification agent, and graph architect agent adopt a pipelined microservice deployment architecture;

[0045] The dynamic construction and evolution system of the astronomical knowledge graph also includes a central scheduling engine. The central scheduling engine constructs a standardized data bus through message queues, triggers the execution of each intelligent agent according to a preset pipeline sequence, and defines the data contract between each intelligent agent through a standardized JSON Schema.

[0046] The data reconnaissance agent is deployed as an I / O-intensive containerized service instance, the information extraction agent is deployed as a compute-intensive containerized service instance, and the knowledge review agent and the graph architect agent are deployed as highly reliable containerized service instances.

[0047] The dynamic construction and evolution system of the astronomical knowledge graph also includes a distributed memory cache and a distributed object storage. Each intelligent agent synchronizes the intermediate state of the task by sharing the distributed memory cache and synchronizes standardized text data by sharing the distributed object storage.

[0048] The central scheduling engine has a built-in dynamic load balancer. When the amount of data to be processed exceeds a preset threshold, the central scheduling engine automatically instantiates multiple information extraction agent work nodes to extract text fragments in parallel using a producer-consumer model. The knowledge review agent then performs global normalization merging on the local candidate triple set.

[0049] The dynamic construction and evolution system for the astronomical knowledge graph also includes a unified large language model API gateway. The large language model API gateway encapsulates multiple large language model interfaces, supports the dynamic configuration of different base models for different intelligent agents, and performs automatic routing and failover based on the gateway load.

[0050] The above solutions address efficiency and scalability bottlenecks in system engineering deployment. By defining a series of specific deployment schemes, they upgrade single-point serial processing to a horizontally scalable parallel pipeline operation. In particular, by automatically instantiating multiple information extraction agent worker nodes using a producer-consumer model through a load balancer to process large-scale text fragments, the system throughput can increase nearly linearly with the growth of computing resources. At the same time, a unified LLM gateway enables flexible plug-and-play and automatic routing failover of different base models, providing a practical engineering guarantee for the high reliability and high efficiency of the system in resource-constrained edge environments or ultra-large-scale astronomical data processing tasks.

[0051] Based on a third key aspect of the present invention, an electronic device is provided, comprising one or more processors;

[0052] Storage device for storing one or more programs;

[0053] When one or more programs are executed by one or more processors, the one or more processors implement the aforementioned method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems.

[0054] Based on the fourth principal aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, implements the aforementioned method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems.

[0055] Compared with existing technologies, this invention improves the scientific rigor and micro-level extraction quality of astronomical knowledge graph construction through a multi-agent collaborative architecture. Experimental data shows that, under an automated evaluation system verified by double-blind testing, this invention achieves near-perfect scores in four micro-dimensions: accuracy, context independence, granularity, and domain relevance. Furthermore, it achieves 76.42% on the MINE metric, which measures lossless information compression capability, representing a significant performance improvement compared to traditional pipelined methods and end-to-end generative baselines.

[0056] Addressing the persistent problem of terminological ambiguity caused by numerous entity aliases and the proliferation of non-standard abbreviations in the astronomical field, this invention employs a unique hybrid alignment strategy combining vector coarse screening with language model triangulation verification, achieving a significant balance between accuracy and open adaptability. In standard entity link evaluation, this invention achieved a 75.0% Hit@1 accurate alignment rate while maintaining a high rejection rate of 70.0% for unknown new concepts not yet included in the knowledge base. This effect means that the system can accurately normalize the diverse entity references from massive amounts of heterogeneous literature into authoritative standard terms, and can also keenly identify cutting-edge knowledge and proactively register it as new seeds. This fundamentally solves the dilemma of traditional methods where forced merging leads to the "contamination" or "swallowing" of new knowledge, clearing obstacles for the continuous expansion of the knowledge graph in the open world.

[0057] This invention, based on an incremental conflict resolution mechanism using confidence-weighted and exponential time decay functions, has achieved the expected technical results in resolving spatiotemporal numerical conflicts in multi-source heterogeneous data. The system can automatically perform objective judgments based on the authority and timeliness of the data source. For example, when newly ingested data has a higher confidence level, it automatically overwrites and updates the old version and archives it; when the differences are within the error range, it retains the data from a coexistence perspective to reflect scientific controversy. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0059] Figure 1 The following is an execution flowchart of a method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems, according to one embodiment of the present invention.

[0060] Figure 2 This diagram illustrates a framework diagram of a multi-agent-based dynamic construction and evolution system for astronomical knowledge graphs, according to one embodiment of the present invention.

[0061] Figure 3 This paper illustrates a roadmap for adaptive terminology normalization and conflict resolution techniques in one embodiment of the present invention. Detailed Implementation

[0062] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0063] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0064] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0065] The following are the main technical terms that may be involved in the embodiments of this invention and their meanings:

[0066] Multi-agent collaboration: The task of building an astronomical knowledge graph is decoupled into four types of intelligent agents with clearly defined functions: data reconnaissance, information extraction, knowledge verification, and graph architecture. A distributed architecture mode is used to complete the fully automated construction process through standardized one-way data flow.

[0067] Astronomical Knowledge Graph: A structured semantic knowledge base for the astronomical field stored in a graph structure. Its core consists of entities such as celestial bodies, observation facilities, and physical concepts, the relationships between entities, and triples of entity attributes. It supports structured precise queries and multi-hop logical reasoning.

[0068] Terminology standardization: The process of mapping aliases, non-standard abbreviations, and variant expressions in astronomical literature to unique identifiers of authoritative standard terms, in order to eliminate redundant nodes and semantic ambiguities in the spectrum.

[0069] Triangular semantic adjudication: The core decision-making mechanism for entity alignment, which simultaneously inputs three types of information: the original context of candidate terms, candidate standard terms, and the scientific definition of standard terms. It completes multi-dimensional semantic cross-validation through a large language model and outputs the alignment result.

[0070] Dual-modal storage architecture: It adopts a hybrid storage scheme that uses a graph database to store the entity relationship topology and a vector database to store terms and text semantic vectors, while taking into account both structured precise query and unstructured semantic fuzzy retrieval capabilities.

[0071] Incremental graph update: No need to fully reconstruct the graph, only perform conflict detection and merging on newly ingested knowledge, and archive the dynamic evolution mechanism of historical versions, which reduces the computing power cost of updating while ensuring the timeliness of knowledge.

[0072] Knowledge confidence: A quantitative value of knowledge credibility calculated by combining the authority score of the comprehensive data source and the time decay factor. It is used to determine the priority and version replacement decisions when multiple sources of knowledge conflict.

[0073] Information extraction: The core technology for automatically identifying and extracting structured knowledge such as entities, attributes, relationships and time-series events from unstructured astronomical literature and semi-structured observational data is the upstream foundational step in map construction.

[0074] Please see Figure 1 The present invention provides an embodiment of a method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems, including steps S100-S500 executed by a computer system:

[0075] Step S100: The data reconnaissance agent acquires raw data in parallel from multiple heterogeneous astronomical data sources, performs format parsing and structure reconstruction on the raw data, and generates standardized text data in a unified format.

[0076] Step S200: The information extraction agent performs multi-dimensional parallel information extraction on the standardized text data, identifies astronomical entities and their static attributes, dynamic relationships and time-series events, and generates a set of candidate knowledge triples.

[0077] Step S300: The knowledge auditing agent performs adaptive semantic alignment and disambiguation processing on the entity references in the candidate knowledge triples based on the pre-built seed knowledge base, maps non-standard entity references to standard terminology identifiers, and generates standardized triple data.

[0078] Step S400: The graph architect agent performs conflict detection between the normalized triplet data and the existing knowledge in the existing knowledge graph, calculates the comprehensive confidence score based on the authority of the data source and the timestamp, and determines whether to perform the addition, update or archive operation based on the comprehensive confidence score to complete the dynamic evolution of the knowledge graph.

[0079] Step S500: Store the entity topological relationships in the updated knowledge graph to the graph database, and store the semantic feature vectors of entities and relationships to the vector database to construct a bimodal knowledge graph that supports structured multi-hop queries and semantic similarity retrieval.

[0080] The method for dynamic construction and evolution of astronomical knowledge graph based on multi-agent systems developed in this invention is not an isolated technological breakthrough, but rather is built on the foundation of the current paradigm shift in astronomical research and a series of cutting-edge technological developments in the field of natural language processing. It requires the following multi-layered prior knowledge as the starting point and support for technological cognition.

[0081] Firstly, regarding the understanding of the astronomical data environment, the implementation of this invention relies on a deep understanding of the data-intensive research paradigm in modern astronomy. Current astronomy has fully entered the era of full-band and multi-messenger systems, with numerous ground-based and space-based survey facilities (such as the Guo Shoujing Telescope, Gaia satellite, and FAST) continuously producing massive amounts of observational data. Meanwhile, the latest theoretical results are mainly distributed in unstructured databases such as arXiv and astrophysical data systems, while precise physical parameters of celestial bodies are stored in semi-structured dedicated star catalog databases such as Simbad and VizieR. This multi-source, heterogeneous, and highly fragmented knowledge distribution pattern directly necessitates the architectural design requirement of this invention for cross-platform parallel acquisition capabilities of the data reconnaissance agent. Therefore, implementers must fully understand the distribution characteristics of astronomical data sources, differences in access interfaces, and the specifications for physical parameter values.

[0082] Secondly, regarding the technical characteristics of large language models, implementers must grasp the basic principles and limitations of current mainstream large language models. While general-purpose large models (such as GPT, DeepSeek, and Qwen) demonstrate powerful semantic understanding capabilities in the field of natural language processing, their autoregressive generation mechanism based on probabilistic prediction makes them prone to "fact illusions" when dealing with long-tailed numerical parameters in astronomy. For example, when faced with queries about precise physical quantities such as the surface temperature of Sirius B, the model tends to generate approximate rounded values ​​based on common numerical distributions in the training corpus rather than actual observation results. Furthermore, the end-to-end reasoning process lacks interpretability and traceability.

[0083] Therefore, this invention does not directly rely on the intrinsic knowledge of the large model, but rather uses it as a "cognitive engine" for information extraction, semantic discrimination and conflict resolution, supplemented by external knowledge bases for constraint. The establishment of this technical route stems from the prior insight into the capability boundaries of the large model.

[0084] Furthermore, at the knowledge graph construction technology level, implementing this invention requires understanding the core defects and causes of existing information extraction paradigms. Traditional pipeline-style extraction methods separate named entity recognition and relation extraction into serial modules, where upstream entity recognition errors irreversibly propagate downstream. End-to-end joint extraction models based on instruction fine-tuning exhibit extremely unstable adherence to structured constraints and are prone to illusions when faced with multi-conditional dependencies in astronomical literature, such as "numerical values-observation equipment-physical properties." Open information extraction methods, on the other hand, are entirely constrained by the surface structure of long and complex sentences, resulting in extracted triples with serious fallacies at the level of physical common sense and extreme fragmentation. Implementers must have a clear understanding of the failure mechanisms of these existing paradigms.

[0085] Furthermore, at the level of retrieval enhancement generation and agent collaboration paradigms, implementers should also understand the macro trend and basic principles of the current RAG technology's evolution from pure text vector retrieval to structured knowledge graph retrieval. Existing text retrieval based on dense vector similarity is prone to "semantic drift" and breakage of key evidence chains when handling multi-hop logical associations, while structured retrieval based on knowledge graphs is limited by insufficient graph coverage and difficulties in entity disambiguation.

[0086] Meanwhile, the maturation of the ReAct (Reasoning and Action Alternation) paradigm has endowed large models with tool-calling and autonomous planning capabilities. Multi-agent systems, through role-playing and standardized communication protocols, decouple complex cognitive tasks into well-defined sub-tasks, which has been proven to significantly reduce the contextual burden of a single model call and decrease factual illusions in long-chain engineering tasks. This invention is based on these technological evolution trends, upgrading large language models from passive text generators to collaborative work clusters with perception, planning, and verification capabilities. This design approach must be grounded in an accurate understanding of the compatibility and limitations of these cutting-edge paradigms.

[0087] like Figure 2 The diagram shows the overall framework of a multi-agent-based dynamic construction and evolution system for astronomical knowledge graphs. This invention employs a pipelined multi-agent collaborative architecture. The entire system is coordinated by a central scheduling engine, executing in a preset order of "data reconnaissance layer → information extraction layer → knowledge verification layer → construction layer." Agents exchange data via standardized JSON format, with each agent focusing solely on the core task of its current stage. This effectively reduces the complexity of prompts in a single large language model call, ensuring the stability and controllability of long text information extraction. The central scheduling engine maintains a global task state machine, using message queues (e.g., RabbitMQ or Kafka) to achieve decoupling and asynchronous communication between agents.

[0088] The following embodiments illustrate specific implementations of the data reconnaissance agent.

[0089] During the system startup phase, the data reconnaissance agent automatically identifies and classifies the data source type based on the input task instructions, and acquires raw data in parallel from multiple heterogeneous astronomical data sources. Specifically:

[0090] For academic literature data sources, the data reconnaissance agent calls professional PDF parsing tools (such as MinerU) to perform in-depth layout analysis, accurately identify two-column layouts, embedded mathematical formulas, and cross-page tables, and reconstructs them into Markdown format text that retains the original logical structure, thereby avoiding the loss of scientific parameters due to format conversion.

[0091] For the latest astronomical news, the data reconnaissance agent integrates search engines such as BochaAI to conduct extensive searches, and filters internet noise through a built-in authoritative domain whitelist and keyword density scoring mechanism, extracting only high-confidence text fragments for downstream processing.

[0092] For specific astrophysical parameter queries, the data reconnaissance agent bridges directly with authoritative databases such as SIMBAD through standard application programming interfaces (APIs) to obtain accurate metadata; at the same time, it supplements the background descriptions of general concepts through the Wikipedia API.

[0093] After the above processing, the data reconnaissance agent generates standardized text data in a unified format and outputs it to the information extraction agent.

[0094] The following embodiments illustrate specific implementations of the information extraction agent.

[0095] After receiving the standardized text data, the information extraction agent abandons the single-instruction full-extraction method and instead adopts a "step-by-step decoupled" parallel job mode. Using the batch processing interface of the large language model, it simultaneously executes the following four independent extraction tasks:

[0096] Multi-dimensional entity recognition: Accurately locate and classify key entities such as celestial objects (e.g., stars, galaxies, asteroids), observation facilities (e.g., IRAS satellite, Akari satellite), and physical concepts (e.g., spectral type, albedo) from unstructured text.

[0097] Static attribute extraction: Following the "fact list" principle, focus on extracting inherent physical attributes of celestial bodies such as mass, radius, spectral type, diameter, albedo, etc., and force the model to ignore non-factual descriptions such as research team introductions to ensure the objective purity of the data.

[0098] Dynamic relationship construction: Identify the association logic between entities, extract the "subject-relationship-object" triples, and simultaneously capture the conditional attributes of the existence of the relationship. For example, for the statement "According to the observations of the Infrared Astronomical Satellite IRAS, the diameter of Belopolskya is 71.60 km", the information extraction agent outputs the triple <Belopolskya, diameter, 71.60 km> and attaches the conditional attribute {observation device: IRAS satellite} to retain the boundary conditions of scientific conclusions.

[0099] Temporal event capture: For dynamic processes with a time span such as supernova explosions, gravitational wave detections, and orbital maneuvers, focus on capturing event information containing clear time anchors, participating entities, and evolution results, and format it into independent JSON objects to support subsequent temporal reasoning.

[0100] The information extraction agent aggregates the results of the above four parallel tasks into a set of candidate knowledge triples and transmits them to the knowledge review agent.

[0101] In the following embodiments, the specific implementation method of the knowledge review agent, namely the adaptive term normalization method, is described.

[0102] After receiving the set of candidate knowledge triples, the knowledge review agent performs adaptive semantic alignment and disambiguation processing on each candidate entity reference term based on the seed knowledge base. The seed knowledge base is integrated from the Unified Astronomy Thesaurus (UAT), contains standard term names and their corresponding authoritative scientific definitions, and the entire seed library is encoded as high-dimensional vectors and stored in a vector database.

[0103] As Figure 3 shown, the adaptive term normalization specifically includes the following sub-steps:

[0104] Vector retrieval sub-step: When the information extraction agent extracts a new candidate entity designation (e.g., "TypeIa SNe"), the system first encodes the candidate word into a semantic vector using an embedding model, and then retrieves the Top-K standard terms with the highest semantic similarity from the vector index of the seed knowledge base as candidate alignment items (e.g., "Type IaSupernovae", "Supernova"). In this embodiment, the value of K is preferably between 5 and 10.

[0105] The large language model triangulation verification sub-step: The knowledge auditing agent simultaneously inputs the following three elements into the large language model: (1) the original context text corresponding to the candidate entity reference; (2) the candidate entity reference itself; and (3) the standard term name and authoritative definition of the retrieved candidate alignment item. The large language model, as a semantic discriminator, determines whether the meaning of the candidate entity reference in the current context completely matches the authoritative definition of the candidate alignment item. For example, for the abbreviation "LMC", if its context is "the LMC is a dwarf galaxy", the triangulation verification determines that it should be mapped to "LargeMagellanic Cloud"; if the context is "LMC parameters in laboratory experiments", it is determined to be a different entity.

[0106] The normalized sequence generation sub-step involves the following steps: If the large language model determines that a candidate word is semantically consistent with a candidate alignment item, a merging instruction is generated, mapping the candidate entity reference to its corresponding standard term identifier (i.e., UAT standard ID). If the candidate word cannot be merged with any Top-K seed (e.g., a completely new celestial body type is discovered), it is marked as a new term (NEW_ENTRY), a new definition description is generated, and this new term is written back to the vector database as a new seed. Through this dynamic normalization mechanism, the seed knowledge base possesses adaptive growth capabilities.

[0107] like Figure 3 As shown, the conflict resolution and incremental update methods for the graph architect agent are as follows:

[0108] After receiving the normalized triplet data processed by the knowledge auditing agent, the graph architect agent performs conflict detection between the new data and existing knowledge in the existing knowledge graph, calculates a comprehensive confidence score based on the authority of the data source and the timestamp, and determines whether to perform an add, update, or archive operation based on the comprehensive confidence score.

[0109] Specifically, when the newly ingested triplet Compared with the existing triplet in the diagram When attribute value conflicts occur on the same relation, the system triggers a conflict resolution process. This represents the head entity, the main body in the knowledge graph, i.e., the astronomical object being described. For example, in the scenario of describing the properties of an asteroid, It is known as "Belopolskya" (asteroid) or "Betelgeuse" (star). This indicates a relationship, representing the attribute type or association type between the head and tail entities. Examples include "diameter," "albedo," "redshift value," or "discoverer." This represents a new tail entity, indicating a newly ingested object or attribute value to be written into the graph. Represents old tail entities, indicating entities that already exist in the graph and are related to... and The corresponding original object or attribute value.

[0110] Overall confidence score Calculate using the following formula:

[0111]

[0112] in, This represents the authority score of the data source, which is a preset constant. Different data source types correspond to different scores. value; This represents a time decay function based on timestamps. This represents the data publication timestamp carried in the normalized triplet data;

[0113] In this embodiment, different data source types The preferred values ​​are: 1.0 for the SIMBAD database, 0.9 for peer-reviewed academic papers, 0.8 for Wikipedia, and 0.7 for general online news.

[0114] The time decay function is calculated using the following formula:

[0115]

[0116] in Indicates the current system time. Indicates the data publication timestamp. Indicates the preset attenuation coefficient. It is a natural constant;

[0117] In this embodiment The optimal value is 0.1. Through this exponential decay function, the interference of outdated data on current decisions is effectively reduced, and the system can keenly capture the latest observation corrections.

[0118] The graph architect agent is based on the comprehensive confidence score. The following decision logic will be executed:

[0119] like The new data will be updated to the current value, and the old data will be marked as a history version and archived, thus realizing the iteration of knowledge.

[0120] like (in If a preset error threshold is set (preferably 0.05 in this embodiment), then both viewpoints will be retained simultaneously and provided to users for reference during subsequent reasoning, thus truly reflecting the controversies existing in the scientific community.

[0121] like If the old data is retained, the new data will be marked as a low-confidence candidate.

[0122] The following embodiments illustrate specific implementations of the dual-modal storage module.

[0123] The "golden knowledge," after layers of cleaning and verification, is finally gathered in the dual-modal storage module. This module adopts a "graph + vector" dual-modal storage architecture.

[0124] exist Figure 3 In the illustrated embodiment, Neo4j is used as the graph database. It employs an index-free adjacency architecture, where each node directly stores the physical address pointing to its neighboring nodes. During multi-hop queries or path exploration, the system can traverse the relational network with a time complexity close to O(1). This graph database stores entity nodes (containing entity name, category, and unique identifier ID), relation edges (containing relation type and conditional attributes), and attribute key-value pairs (containing physical parameters such as celestial coordinates, spectral type, magnitude, parallax, proper motion, and radial velocity), supporting structured multi-hop path traversal queries.

[0125] The vector database chosen is ChromaDB, which stores entity name vectors, entity definition vectors, and relationship description vectors encoded by the embedding model, supporting semantic retrieval based on cosine similarity. In this embodiment, the embedding model preferably uses NV-Embed-V2.

[0126] Graph databases and vector databases are cross-linked through unique entity IDs, thereby providing upper-layer applications with retrieval capabilities that combine precise logic and fuzzy semantics.

[0127] The following embodiments provide a detailed description of the specific deployment methods of each intelligent agent in this invention. For example... Figure 2 As shown, the data reconnaissance agent, information extraction agent, knowledge review agent, and graph architect agent adopt a pipelined microservice deployment architecture.

[0128] (a) Central Dispatch Engine

[0129] The system is equipped with a central scheduling engine. The central scheduling engine builds a standardized data bus through message queues (such as Kafka or RabbitMQ), triggers the execution of each agent according to a preset pipeline sequence, and defines the input / output data contract between each agent through a standardized JSONSchema.

[0130] (ii) Containerized deployment and resource isolation

[0131] Each agent is deployed as an independent containerized service instance (based on Docker and Kubernetes). Specifically:

[0132] The data reconnaissance agent is deployed as an I / O-intensive service instance, configured with high network bandwidth resources and a multi-source API connection pool to cope with the large-scale concurrent collection needs of multi-source heterogeneous data.

[0133] The information extraction agent is deployed as a compute-intensive service instance, configured with high GPU resources to support batch parallel decoding and multi-task parallel extraction of standardized text data.

[0134] The knowledge auditing agent and the graph architect agent are deployed as high-reliability service instances, configured with redundant backups and transaction rollback mechanisms to ensure that terminology standardization adjudication and conflict resolution operations can be automatically recovered in abnormal situations.

[0135] (iii) Shared state storage

[0136] Each agent synchronizes intermediate task states by sharing a distributed memory cache (such as Redis) and synchronizes standardized text data and original documents by sharing a distributed object storage (such as MinIO or Amazon S3), thus avoiding direct transmission of large-scale text data between agents.

[0137] (iv) Dynamic load balancing and parallel scaling

[0138] The central scheduling engine incorporates a dynamic load balancer. When the amount of standardized text data to be processed exceeds a preset threshold (e.g., more than 1000 text fragments), the engine automatically instantiates multiple information extraction agent worker nodes to extract text fragments in parallel using a producer-consumer model. Each worker node independently outputs a local set of candidate triples, which are then globally normalized and merged by the knowledge review agent in subsequent stages, thereby achieving near-linear throughput scaling.

[0139] (v) Large Language Model API Gateway

[0140] The system also includes a unified large language model API gateway. This gateway encapsulates multiple commercial and open-source large language model interfaces (such as GPT-4o-mini, DeepSeek-V3, Qwen3-Max, etc.). During deployment, system administrators can dynamically specify different base models for different agents through configuration files. For example, the information extraction agent uses a cost-effective open-source model, while the knowledge review agent uses a flagship model with stronger inference capabilities. The gateway also performs automatic routing and failover based on the current load to ensure high availability.

[0141] The following embodiments illustrate specific implementations of the overall framework of the present invention in handling complex astronomical context extraction tasks.

[0142] To further illustrate the technical effects of the present invention, this embodiment takes a typical complex numerical condition description in astronomical literature as an example to describe in detail the working process of the overall framework of the present invention.

[0143] Consider the following excerpt from English literature describing the physical characteristics of asteroid Belopolskya: "According to the surveys carried out by the Infrared Astronomical Satellite IRAS and the Japanese Akari satellite, Belopolskya measures 71.60 and 79.83 kilometers indiameter, and its surface has an albedo of 0.0348 and 0.028, respectively. The Collaborative Asteroid Lightcurve Link adopts the shorter diameter obtained by IRAS."

[0144] The difficulty in analyzing this passage lies in the fact that the diameter and albedo of the same celestial body have two parallel numerical versions due to different observation instruments, and the syntax heavily relies on the word "respectively" to construct the cross-correspondence logic.

[0145] In a specific implementation of this invention, after receiving the standardized text, the information extraction agent executes four extraction tasks in parallel. For the cross-correspondence logic indicated by "respectively," the information extraction agent uses a large language model to identify conditional dependencies and outputs the following structured candidate triples:

[0146] <Belopolskya, Diameter, 71.60 km, {Observation Equipment: IRAS Satellite}>

[0147] <Belopolskya, Diameter, 79.83 km, {Observation Equipment: Akari Satellite}>

[0148] <Belopolskya, Albedo, 0.0348, {Observation Equipment: IRAS Satellite}>

[0149] <Belopolskya, Albedo, 0.028, {Observation Equipment: Akari Satellite}>

[0150] <Asteroid Light Curve Collaboration Database, Adopted Diameter, 71.60 km, {Data Source: IRAS Satellite}>

[0151] Subsequently, the knowledge review agent retrieves the entity reference terms "Belopolskya", "IRAS", "Akari", etc. in the candidate triples one by one from the seed knowledge base vector and verifies them through the large language model triangle, mapping them to the UAT standard term identifiers. The graph architect agent detects conflicts between the above normalized triples and the existing graph, and writes the new data into the bimodal storage module.

[0152] This embodiment overcomes the defects of the subject-object inversion and fragmentation of traditional open information extraction methods, and also avoids the problem of scientific logic breakage caused by the inability of the end-to-end generative framework to parse the mapping logic indicated by "respectively". Finally, this embodiment successfully constructs high-fidelity triples with a strict correspondence relationship of "numerical value - observation equipment - physical property", which can support basic traceability queries such as "which equipment measured 79.83 kilometers", fully reflecting the technical effects of the present invention.

[0153] In the following embodiments, the specific implementation methods of the term normalization and entity linking of the present invention are described, especially the preferred implementation parameters of the adaptive term normalization mechanism in the present invention are described in detail.

[0154] Seed Knowledge Base Construction: When the system is initialized, all standard terms (about 18,000) are imported from the Unified Astronomy Thesaurus (UAT). Each term contains the standard name, the text description of the definition, and the corresponding unique identifier (URI). The system uses the NV-Embed-V2 embedding model to encode the concatenated name and definition of each term into a 768-dimensional semantic vector, stores it in the ChromaDB vector database, and establishes an HNSW (Hierarchical Navigable Small World Graph) index to accelerate approximate nearest neighbor retrieval.

[0155] Vector retrieval parameters: For each candidate entity reference to be processed, the system retrieves the top-10 most semantically similar candidate alignment items from the seed knowledge base. The retrieval uses cosine similarity measurement, with a preset similarity threshold of 0.65; candidates below this threshold will be directly filtered out and will not enter the large language model triangulation verification stage.

[0156] Large Language Model Triangle Validation: The knowledge review agent concatenates the original context (up to 512 tags) of the candidate entity references, the standard name of the candidate alignment item, and the authoritative definition into a prompt word, and inputs it into the large language model. The prompt word is designed to require the model to output one of the following three types of results: (1) Align to a specific standard ID; (2) Marked as "Unable to align, it is recommended to register a new term"; (3) Marked as "Insufficient information, do not process for now".

[0157] Dynamic updates: When the system determines that the alignment is successful, it stores the mapping relationship between the candidate entity reference and the standard term in the mapping cache table for quick reuse of the same reference in the future. When the system determines that it is a new term, the knowledge review agent calls the large language model to automatically generate the term definition text. After confirmation by the manual review node (optional), the new term and its definition are encoded into a vector and appended to the seed knowledge base to achieve adaptive growth of the terminology base.

[0158] Through the above specific implementation methods, the terminology normalization module of the present invention can maintain a high accuracy alignment of known standard terms (Hit@1 reaches 75.0%) while maintaining a high rejection rate and active registration rate of 70.0% for unknown new terms, thus ensuring the continuous expansion capability of the knowledge graph in open world scenarios.

[0159] The following examples illustrate preferred embodiments for verifying the system performance of the present invention.

[0160] Testing environment: The present invention was deployed on a computing cluster containing 8 NVIDIA A100 (80GB) GPUs. The large language model base was DeepSeek-V3, and the embedded model was NV-Embed-V2.

[0161] Test corpus: Three astronomical topics were selected: Betelgeuse, Andromeda Galaxy and GW170817 gravitational wave event. Layered sampling was performed to construct a test corpus containing approximately 600 texts with an average length of 800 tags.

[0162] Evaluation metrics: The LLM-as-a-Judge paradigm was adopted, with Gemini-3.0-Pro as the automated judging model. To verify the reliability of the automated judging, a double-blind test was first conducted: 30 triples were randomly selected from the results of this invention and the baseline method, and combined to form a blind test set of 90 samples. After anonymizing and desensitizing the human experts, domain experts independently scored the samples using a 0-2 subscale. Subsequently, the Pearson correlation coefficient between the automated judging scores and the human scores was calculated. The test results showed that the Pearson correlation coefficient between the automated judging scores of this invention and the human expert scores reached 0.964, demonstrating the high reliability of the automated evaluation results.

[0163] Experimental Results: Under the above test environment, this invention achieved the following scores in the micro-level extraction quality assessment: accuracy 1.93, context independence 1.83, granularity 1.86, and domain relevance 1.81 (all out of 2). It also achieved 76.42% on the macro-level lossless information compression index MINE, comprehensively outperforming baseline methods such as OpenIE, OneKE, and KGGEN. Ultimately, the knowledge graph constructed by this invention contains 144,373 entity nodes, 230,044 relation edges, 2,180,843 attribute key-value pairs, and 646,472 temporal event records, covering 132,476 source documents.

[0164] It should be understood that the program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0165] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0166] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this invention does not impose any limitations on them.

[0167] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0168] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems, characterized in that, Includes the following steps: The data reconnaissance agent acquires raw data in parallel from multiple heterogeneous astronomical data sources, performs format parsing and structure reconstruction on the raw data, and generates standardized text data in a unified format. The information extraction agent performs multi-dimensional parallel information extraction on the standardized text data to identify astronomical entities and their static attributes, dynamic relationships and time-series events, and generates a set of candidate knowledge triples. The knowledge auditing agent performs adaptive semantic alignment and disambiguation on the entity references in the candidate knowledge triples based on a pre-built seed knowledge base, mapping non-standard entity references to standard terminology identifiers and generating normalized triple data. The graph architect agent performs conflict detection between the normalized triplet data and existing knowledge in the existing knowledge graph, calculates a comprehensive confidence score based on the authority of the data source and the timestamp, and determines whether to perform addition, update or archive operations based on the comprehensive confidence score, thus completing the dynamic evolution of the knowledge graph. The updated knowledge graph stores the topological relationships of entities in a graph database and stores the semantic feature vectors of entities and relationships in a vector database, thus constructing a bimodal knowledge graph that supports structured multi-hop queries and semantic similarity retrieval.

2. The method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 1, characterized in that, The information extraction agent adopts a parallel operation mode, simultaneously performing the following four independent extraction tasks on the standardized text data: Multidimensional entity recognition task, used to identify entities of celestial objects, observation facilities, and physical concept categories; Static attribute extraction task, used to extract the physical parameter attribute values ​​of entities; The dynamic relationship building task is used to identify relationship triples and their conditional attributes between entities; The time-series event capture task is used to extract event information that includes time anchors and evolution results.

3. The method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 2, characterized in that, The adaptive semantic alignment and disambiguation process includes the following sub-steps: The candidate entity references to be processed are encoded into semantic vectors, and the top-K standard terms with the highest semantic similarity are retrieved from the vector index of the seed knowledge base as candidate alignment items. The original context text corresponding to the candidate entity reference, the standard term name of the candidate alignment item and its authoritative definition are jointly input into the large language model. The large language model performs triangulation to determine whether the semantics of the candidate entity reference item in the current context is consistent with the authoritative definition of the candidate alignment item. If the determination result is consistent, a merge instruction is generated to map the candidate entity reference to the standard term identifier of the candidate alignment item; If the determination result is inconsistent, the candidate entity designation is marked as a new term and registered in the seed knowledge base.

4. The method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 3, characterized in that, The comprehensive confidence score Calculate using the following formula: in, This represents the authority score of the data source, which is a preset constant. Different data source types correspond to different scores. value; This represents a time decay function based on timestamps. This represents the data publication timestamp carried in the normalized triplet data; The time decay function is expressed by the formula. Calculation, where Indicates the current system time. Indicates the data publication timestamp. This indicates the preset attenuation coefficient. It is a natural constant; If the overall confidence score of the new data is higher than that of the existing data, the new data will be updated to the current value and the existing data will be archived as a historical version. If the difference between the combined confidence scores of the new data and the existing data is within the preset error range, then both will be retained as coexisting viewpoints.

5. The method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 4, characterized in that, The graph database adopts an index-free adjacency architecture to store entity nodes, relation edges, and attribute key-value pairs, and supports multi-hop path traversal queries. The vector database is used to store entity name vectors, entity definition vectors, and relationship description vectors encoded by the embedding model, and supports semantic retrieval based on cosine similarity.

6. The method for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 5, characterized in that, The seed knowledge base is integrated from the Unified Astronomical Terminology (UAT) and includes standard term names and corresponding authoritative scientific definitions. The vector index of the seed knowledge base is dynamically updated as new terms are registered, enabling the seed knowledge base to have adaptive growth capabilities. The statistical data of the knowledge graph includes: the number of entity nodes, the number of relation edges, the number of attribute key-value pairs, the number of time-series event records, and the number of source documents; The entity nodes at least cover categories such as stars, galaxies, astronomical events, observation instruments, and physical concepts; the attribute key-value pairs include physical parameters such as celestial coordinates, spectral type, magnitude, parallax, proper motion, and radial velocity.

7. A multi-agent-based dynamic construction and evolution system for astronomical knowledge graphs to implement the method of any one of claims 1-6, characterized in that, It includes the following components: A data reconnaissance intelligent agent is used to acquire raw data in parallel from multiple heterogeneous astronomical data sources, perform format parsing and structure reconstruction on the raw data, and generate standardized text data in a unified format. An information extraction agent is used to perform multi-dimensional parallel information extraction on the standardized text data, identify astronomical entities and their static attributes, dynamic relationships and time-series events, and generate a set of candidate knowledge triples. A knowledge review agent is used to perform adaptive semantic alignment and disambiguation processing on entity references in the candidate knowledge triples based on a pre-built seed knowledge base, mapping non-standard entity references to standard terminology identifiers and generating normalized triple data. The knowledge graph architect agent is used to detect conflicts between the normalized triplet data and existing knowledge in the existing knowledge graph, calculate the comprehensive confidence score based on the authority of the data source and the timestamp, and determine whether to perform addition, update or archive operations based on the comprehensive confidence score to complete the dynamic evolution of the knowledge graph. The dual-modal storage module is used to store the entity topological relationships in the updated knowledge graph to the graph database and the semantic feature vectors of entities and relationships to the vector database, thereby constructing a dual-modal knowledge graph that supports structured multi-hop queries and semantic similarity retrieval.

8. The system for dynamic construction and evolution of astronomical knowledge graphs based on multi-agent systems according to claim 7, characterized in that, The data reconnaissance agent, information extraction agent, knowledge review agent, and graph architect agent adopt a pipelined microservice deployment architecture. The dynamic construction and evolution system of the astronomical knowledge graph also includes a central scheduling engine. The central scheduling engine constructs a standardized data bus through message queues, triggers the execution of each intelligent agent according to a preset pipeline sequence, and defines the data contract between each intelligent agent through a standardized JSON Schema. The data reconnaissance agent is deployed as an I / O-intensive containerized service instance, the information extraction agent is deployed as a compute-intensive containerized service instance, and the knowledge review agent and the graph architect agent are deployed as highly reliable containerized service instances. The dynamic construction and evolution system of the astronomical knowledge graph also includes a distributed memory cache and a distributed object storage. Each intelligent agent synchronizes the intermediate state of the task by sharing the distributed memory cache and synchronizes standardized text data by sharing the distributed object storage. The central scheduling engine has a built-in dynamic load balancer. When the amount of data to be processed exceeds a preset threshold, the central scheduling engine automatically instantiates multiple information extraction agent work nodes to extract text fragments in parallel using a producer-consumer model. The knowledge review agent then performs global normalization merging on the local candidate triple set. The dynamic construction and evolution system for the astronomical knowledge graph also includes a unified large language model API gateway. The large language model API gateway encapsulates multiple large language model interfaces, supports the dynamic configuration of different base models for different intelligent agents, and performs automatic routing and failover based on the gateway load.

9. An electronic device, characterized in that, Includes one or more processors; Storage device for storing one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the method for dynamic construction and evolution of astronomical knowledge graphs based on multiple agents as described in any one of claims 1-6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed, the computer program implements the method for dynamic construction and evolution of astronomical knowledge graphs based on multiple agents as described in any one of claims 1-6.

Images