Large model extraction knowledge graph construction method and device oriented to field of oil exploration and development, and electronic equipment

Abstract

The invention provides an oil exploration and development field-oriented large model extraction knowledge graph construction method and apparatus, and an electronic device, and is completed by applying an intelligent batch processing architecture. The method comprises the following steps of: performing semantic segmentation on a document in the field of oil exploration and development, including a knowledge extraction process of named entity recognition, event extraction, relation reasoning and entity standardization, so as to obtain a standardized sub-graph; the entity standardization is realized by adopting a three-layer progressive relation normalization algorithm; all the extracted and standardized sub-graphs are fused and loaded into a graph database, and a knowledge graph oriented to the field of oil exploration and development is formed; in the whole knowledge graph construction process, input and output Token consumption and knowledge extraction reasoning paths are recorded through a log system. The new knowledge graph intelligent construction normal form is advanced in technology, feasible in economy, controllable in process and credible in result.

Description

Technical Field

[0001] This application relates to the field of knowledge graph construction technology, and in particular to a method, apparatus and electronic equipment for constructing a knowledge graph from a large model for the field of oil exploration and development. Background Technology

[0002] In recent years, the emergence of Large Language Models (LLMs) has provided unprecedented opportunities for automating the processing of massive amounts of unstructured text. Academia and industry have begun exploring the use of LLMs' powerful understanding capabilities to automatically construct knowledge graphs. Against this backdrop, a number of knowledge graph construction frameworks based on LLMs have emerged, such as the existing technology represented by the KAG (Knowledge Augmented Generation) framework of OpenSPG (Open Semantic Augmented Programmable Graph).

[0003] Frameworks like KAG provide an advanced, schema-based paradigm for automated knowledge extraction. They allow developers to first define a domain-specific knowledge schema, and then use LLM as a "knowledge extractor" to automatically identify and extract entities, attributes, and relationships that conform to the schema definition from unstructured text.

[0004] Compared to traditional NLP (Natural Language Processing) techniques that rely on large amounts of labeled data and complex model training, this method demonstrates significant advantages in flexibility, rapid deployment, and zero-shot extraction capabilities.

[0005] However, in practice, it has been found that when general, prototyping knowledge extraction frameworks like KAG are directly applied to the highly specialized, rigorous, and cost- and stability-critical industrial field of oil exploration and development, their incompatibility quickly becomes apparent, specifically manifested in the following ways: 1. Risks of uncontrolled costs and performance bottlenecks: Reports in the oil industry are lengthy and densely packed with information. Frameworks like KAG typically employ a simple asynchronous concurrent call model, which generates massive API call requests when dealing with a corpus of tens of thousands of documents from an oil field. This not only easily triggers the API rate limit of LLM service providers, leading to numerous task failures and retries, causing the overall system processing efficiency to decrease rather than increase, and stability to be compromised; but also, the cost of these API calls can reach hundreds of thousands or even millions of yuan, which is completely unacceptable in an industry that strives for cost reduction and efficiency improvement.

[0006] 2. The Consistency Between "Illusions" and Facts: When lacking strict domain knowledge constraints, general-purpose LLMs may "create" facts that contradict geological laws or common engineering sense. For example, incorrectly connecting two geographically distant oil wells or generating a non-existent stratigraphic age. While KAG provides schema constraints, its underlying Prompt engineering design still struggles to completely avoid extraction "illusions" arising from complex semantics, which is fatal for the highly rigorous oil industry.

[0007] 3. Lack of knowledge standardization capabilities: The petroleum field contains numerous complex semantic relationships, including synonyms, near-synonyms, and hierarchical relationships (such as "porosity" and "porosity"). Existing frameworks like KAG mostly rely on the implicit normalization capabilities of LLMs or simple string matching. This approach is ineffective and costly in addressing large-scale, deep-seated knowledge heterogeneity and inconsistency within the field, lacking a robust normalization mechanism that balances cost and effectiveness.

[0008] 4. Process "Black Box" and Unreliable Results: Frameworks like KAG present a "black box" to users during extraction, as their internal decision-making processes (e.g., why the LLM makes a certain judgment) remain unknown. When the LLM returns an extraction result, we have no way of knowing its reasoning process. If a critical "fault closure" relation is extracted incorrectly, it could mislead subsequent drilling decisions, causing tens of millions of yuan in economic losses. Research largely focuses on evaluating the final extraction results, lacking documentation of the decision-making paths during knowledge formation. For example, how a relation is standardized (exact matching, fuzzy matching, semantic matching), and its confidence level—this process information is crucial for building a highly reliable knowledge graph. The lack of process traceability and verifiability makes it difficult for business experts to trust and adopt the results. Summary of the Invention

[0009] The purpose of this application is to provide a method, device, and electronic equipment for extracting knowledge graphs from large models in the field of petroleum exploration and development. This is a new paradigm for intelligent knowledge graph construction that is technologically advanced, economically feasible, process-controllable, and results-reliable.

[0010] Firstly, this application provides a method for constructing a large-scale knowledge graph for the field of oil exploration and development. The method includes: applying an intelligent batch processing architecture to complete the following steps in a multi-task parallel manner: based on the domain knowledge system that conforms to the operational logic of oil exploration and development, which is loaded during initialization, performing semantic segmentation and knowledge extraction processes on documents in the oil exploration and development field to obtain standardized subgraphs; wherein, the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning, and entity standardization; entity standardization is implemented using a three-layer progressive relation normalization algorithm; all extracted and standardized subgraphs are merged and loaded into a graph database to form a unified, queryable, and analyzable knowledge graph for the field of oil exploration and development, and providing a visual query and analysis interface; throughout the entire knowledge graph construction process, the input and output token consumption and knowledge extraction reasoning path are recorded through a log system.

[0011] Secondly, this application also provides a knowledge graph construction device for large-scale model extraction in the field of oil exploration and development. The device comprises: a knowledge extraction module, used to apply an intelligent batch processing architecture to complete the following steps in a multi-task parallel manner: based on an initialized domain knowledge system conforming to the operational logic of oil exploration and development, performing semantic segmentation and knowledge extraction processes on documents in the oil exploration and development field to obtain standardized subgraphs; wherein the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning, and entity standardization; entity standardization is implemented using a three-layer progressive relation normalization algorithm; a graph construction module, used to merge all extracted and standardized subgraphs and load them into a graph database to form a unified, queryable, and analyzable knowledge graph for the oil exploration and development field, and providing a visual query and analysis interface; and a log recording module, used to record input / output token consumption and knowledge extraction reasoning paths through a log system throughout the entire knowledge graph construction process.

[0012] Thirdly, this application also provides an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the method described in the first aspect.

[0013] The method, apparatus, and electronic equipment for extracting knowledge graphs from large models in the field of oil exploration and development provided in this application have the following four advantages: (1) Architectural innovation from asynchronous concurrency to intelligent synchronous batch processing: The original asynchronous concurrent call model was abandoned, and a synchronous sequential intelligent batch processing architecture was fully adopted. Independent and configurable batch splitting algorithms were designed for different task characteristics such as named entity recognition, relation extraction, event extraction, and entity standardization. This architecture can intelligently decompose large-scale extraction tasks into multiple small batches, realize fine-grained management of large language model calls, and significantly improve the overall efficiency and throughput of data processing while ensuring stable system operation.

[0014] (2) Full-process Token Consumption Statistics and Cost Control Mechanism: This application introduces for the first time a cost quantification and monitoring system that runs through the entire extraction process. Through a series of methods such as built-in consumption estimation, process recording, and summary reports, it realizes the pre-estimation and accurate post-estimation of the input and output tokens for each model call. This mechanism transforms the API call cost from an uncontrollable "black box" state into a transparent indicator that is measurable, analyzable, and optimizable, providing key data support for system cost control and performance tuning.

[0015] (3) A multi-layered relation normalization algorithm that balances cost and effectiveness: To address the challenge of knowledge standardization, this invention innovatively designs and implements a three-layer progressive relation normalization algorithm consisting of "fast lexical matching → efficient semantic matching → deep intelligent adjudication". This algorithm prioritizes low-cost matching methods (such as string and fuzzy matching), and only when low-cost methods fail to meet the requirements does it progressively enable higher-cost semantic vector matching, and finally, large-scale model semantic adjudication. This cost-sensitive design ensures high normalization recall while minimizing unnecessary computational resource overhead.

[0016] (4) Verifiable and Traceable Knowledge Extraction Log System: To address the issue of opaque extraction processes, this invention establishes a fine-grained structured log system. This system records in detail the process of each key decision in a machine-readable format (such as CSV), especially the complete path of relation normalization, including the original extraction results, normalized standard knowledge, the matching method used, similarity scores, and key contextual information. This provides a complete and verifiable data loop for subsequent knowledge quality assessment, error tracing, and iterative optimization of the domain knowledge system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for extracting knowledge graphs from large models in the field of oil exploration and development, provided in an embodiment of this application; Figure 2 A diagram illustrating the effect of defining the schema portion of a domain knowledge system as provided in this application embodiment; Figure 3 This is a schematic diagram illustrating the extraction effect provided in an embodiment of this application; Figure 4 A schematic diagram of an overall technical process provided for an embodiment of this application; Figure 5 A system initialization flowchart provided for an embodiment of this application; Figure 6 A multi-level relationship normalization flowchart is provided for embodiments of this application; Figure 7 A flowchart of a log monitoring system provided in this application embodiment; Figure 8 A structural block diagram of a knowledge graph extraction and construction device for the field of oil exploration and development, provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] This application addresses the core engineering challenges faced by existing technologies (such as KAG frameworks) in the petroleum industry, including uncontrolled costs, performance bottlenecks, weak knowledge standardization capabilities, and the "black box" nature of the extraction process with unreliable results. This application provides a method, apparatus, and electronic device for constructing a knowledge graph from a large model in the petroleum exploration and development field. Its fundamental motivation is not simply to apply LLM to the petroleum industry, nor to replicate the functionality of existing frameworks like KAG, but rather to directly address and solve the core engineering challenges of industrializing such advanced frameworks. This embodiment aims to explore a new paradigm for intelligent knowledge graph construction that is technologically advanced, economically feasible, process-controllable, and results-reliable. Through systematic architectural reshaping and algorithmic innovation of the underlying extraction components, a prototype extraction tool (such as KAG) is upgraded into a truly intelligent, automated knowledge graph construction system for the petroleum industry, suitable for industrial production environments.

[0021] To facilitate understanding of this embodiment, a detailed description of a method for extracting knowledge graphs from large models in the field of petroleum exploration and development, as disclosed in this application, will be provided first. Figure 1 A flowchart of a method for extracting knowledge graphs from large models in the field of oil exploration and development, provided in this application embodiment, is included in the following steps: Step S102: Apply an intelligent batch processing architecture and complete the following steps in a multi-task parallel manner: Based on the domain knowledge system that conforms to the operation logic of oil exploration and development, load it during initialization, perform semantic segmentation and knowledge extraction processes on documents in the oil exploration and development domain to obtain standardized subgraphs; wherein, the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning and entity standardization; entity standardization is implemented using a three-layer progressive relation normalization algorithm; The aforementioned domain knowledge system, which conforms to the operational logic of petroleum exploration and development, includes the following definitions: concept types, entity types, entity attributes, and entity relationships. Concept types include: problem classification, cause classification, solution classification, development stage classification, indicator change classification, reservoir classification, heterogeneity classification, sedimentary facies classification, structural unit classification, and well network classification, well classification, and extraction method classification related to engineering parameters. Entity types include: geological and engineering unit entities, development indicator and parameter entities, and dynamic analysis and decision-making entities. Entity relationships include: static structural relationships, indicator attribution relationships, and dynamic causal relationships.

[0022] The semantic segmentation described above uses LLM-driven semantic segmentation to divide the text into knowledge blocks with complete semantic units.

[0023] The above three-layer progressive relation normalization algorithm includes the following three strategies in descending order of priority: fast lexical matching strategy, efficient semantic matching strategy, and deep intelligent adjudication strategy; if any strategy fails to normalize, the next priority strategy will be used for normalization processing.

[0024] Specifically, the aforementioned fast lexical matching strategy includes: cleaning the terms to be normalized; performing precise matching between the cleaned terms and a standard knowledge base; the standard knowledge base consists of standard terms defined in the domain knowledge system; if the match is successful, normalization is performed with the highest confidence; if the match is unsuccessful, fuzzy matching is performed by calculating the similarity with the standard terms; if fuzzy matching is unsuccessful, normalization is performed using an efficient semantic matching strategy; the efficient semantic matching strategy includes: calling a vectorization model to convert the terms to be normalized into semantic vectors, calculating the similarity between the semantic vectors and the standard semantic vectors in the vectorization model, determining a specified number of candidate vectors with the highest similarity ranking, and performing normalization based on the vectors among the candidate vectors whose similarity exceeds the confidence threshold; if there are no vectors whose similarity exceeds the set confidence threshold, a deep intelligent adjudication strategy is used for normalization; the deep intelligent adjudication strategy includes: dynamically constructing a complex instruction containing the terms to be normalized, the context corpus, and the candidate standard terms selected by the efficient semantic matching strategy, requiring the large language model to make a final decision based on a full understanding of the context semantics.

[0025] Step S104: All extracted and standardized subgraphs are merged and loaded into the graph database to form a unified, queryable, and analyzable knowledge graph for the field of oil exploration and development, and a visual query and analysis interface is provided; throughout the knowledge graph construction process, the input and output token consumption and knowledge extraction reasoning path are recorded through the log system.

[0026] The steps described above for recording input and output token consumption through a log system include: estimating the token consumption for the upcoming API call based on the task type and input text length before calling the large language model; recording the actual number of input tokens and output tokens consumed in each call after each call to the large language model; and automatically generating a cost summary report after all batches of tasks are completed, which calculates the total token consumption of this build task at each stage.

[0027] The steps described above for recording the knowledge extraction and reasoning path through a log system include: initializing the log system when starting to build the knowledge graph, creating a unique log file with a timestamp, and configuring the console processor and the file processor simultaneously; recording each key knowledge decision process in a machine-readable format during the execution of the construction process; the knowledge decision process includes the complete path of the normalization process.

[0028] The process of determining the domain knowledge system that conforms to the operational logic of oil exploration and development is described as follows: To overcome the "illusion" problem of Large Language Models (LLMs) and ensure the professionalism, consistency, and factual accuracy of the extraction results, the primary task is to construct a rigorous, comprehensive, and logically consistent domain knowledge schema. This schema serves as the "blueprint" and "framework" constraining the behavior of the large language model and forms the foundation and basis for all subsequent intelligent extraction tasks. The domain knowledge schema defined in this application is as follows: Figure 2 As shown, it deeply integrates the experience of experts from multiple disciplines such as oilfield development geology, engineering and dynamic analysis, and covers the core elements of the entire life cycle of oilfield development.

[0029] The construction of this system begins with the definition of concept types, which are used to abstract macro-level classifications within the domain. These concept types themselves are not usually concrete nodes in the graph, but rather exist as "classification" attributes of entity nodes, providing a high-level index for knowledge organization and retrieval. The system defined in this embodiment covers classification standards for the entire lifecycle of oil development, such as defining "problem classification" (used to summarize problems such as "declining production," "increasing water cut," "insufficient energy," "wellbore failure," "inter-layer contradictions," and "water injection problems"), "cause classification" (used to summarize causes such as "geological factors," "engineering factors," "fluid factors," "development factors," and "equipment factors"), "solution classification" (used to summarize technical means such as "production enhancement measures," "water shut-off and profile adjustment," "well network adjustment," "process optimization," and "water injection adjustment"), and "development stage classification" (used to divide by time or state, such as "rising production period," "stable production period," and "declining production period," or by water cut, such as "waterless period," "low water cut period," "medium water cut period," "high water cut period," and "ultra-high water cut period"). In addition, the conceptual system also includes "indicator change classification" to describe dynamic monitoring data patterns, as well as a series of conceptual systems related to geological characteristics such as "reservoir classification", "heterogeneity classification", "sedimentary facies classification", "tectonic unit classification" and engineering parameters such as "well network classification", "well classification", and "mining method classification".

[0030] Building upon the conceptual framework, this embodiment further defines specific entity types. These entity types are the basic building blocks of the knowledge graph (i.e., Figure 2 and Figure 3 The entity types defined in this embodiment (the "points") take into account both the static geological description and dynamic production analysis of oilfield development. These entity types can be mainly divided into several categories: First, geological and engineering unit entities, which serve as the physical basis of the oilfield, such as defining geological and engineering units like "well area," "production well," "injection well," "stratum," "structural unit," "basin," and "well group"; Second, development index and parameter entities, which serve as quantitative data describing the state of the oilfield, such as defining key development index entities like "reserves," "oil production," "water production," "liquid production," "water cut," "pressure," and "water injection index," as well as "fluid parameters" (such as crude oil viscosity and density) and "operating parameters" (such as pump strokes and nozzle diameter); Third, dynamic... The state analysis and decision-making entity is the core of the knowledge system in this embodiment, used to realize the knowledge closed loop from "phenomenon" to "diagnosis" and then to "decision". This embodiment innovatively defines "problem" (to carry specific problems, such as "insufficient energy in the S12d well area"), "cause" (to carry diagnostic conclusions, such as "low formation pressure"), "solution" (to carry countermeasures, such as "increase the water injection volume of injection well C11d"), "indicator change" (to carry dynamic change phenomena, such as "rapid increase in water cut"), "diagnostic method" (to carry analytical techniques) and "problem manifestation" (to describe observable phenomena).

[0031] This knowledge system not only defines the categories of entities but also strictly stipulates the attributes that each entity should possess, used to describe the characteristics of the entity in detail. For example, the "stratum" entity should possess attributes such as "porosity," "permeability," "effective thickness," "lithology," "sedimentary facies," "mud content," and "oil-water interface"; the "production well" entity should possess attributes such as "well number," "completion date," "perforated section," "lifting method," "well status," "pump type," and "pump depth"; the "oil production" entity should possess attributes such as "daily oil production per well," "cumulative oil production," "decline rate," "oil production rate," and "oil production index"; the "problem" entity should possess attributes such as "description," "severity level," "date of occurrence," and "impact on production"; and the "development well network" entity should possess attributes such as "number of production wells," "number of water injection wells," "injection-production ratio," "well network density," and "injection-production well spacing."

[0032] More importantly, the knowledge system defined in this embodiment explicitly specifies the legal relationships between entities (i.e. Figure 2 and Figure 3(The "edges" in the text). These relationships constitute the logical framework of knowledge and are key to achieving intelligent reasoning and question answering. The relationships defined in this embodiment mainly include: static structural relationships, used to describe physical and attribution connections, such as the "location of a layer" relationship between "producing well" and "layer", the "belonging to a well group" relationship between "producing well" and "well group", the "containing producing well" relationship between "well area" and "producing well", the "adjacent well area" relationship between "well area" and "well area", and the "belonging to a basin" relationship between "tectonic unit" and "basin"; index attribution relationships, used to attach quantified indicators to corresponding units, such as the "oil production index" relationship between "producing well" and "oil production", and the "water content" relationship between "producing well" and "water-bearing". The relationships between indicators include: the relationship between "well area" and "reserves" ("reserve indicators"), and the relationship between "injection well" and "injection indicators" ("injection indicators"); the dynamic causal relationship (PRS model), which is the core of intelligent diagnosis in this embodiment, includes the "caused by cause" relationship between "problem" and "cause", the "problem-solving" relationship between "solution" and "problem", and the "cause-targeting" relationship between "solution" and "cause"; as well as the relationship between phenomena and triggers, such as the "problem-causing" relationship between "indicator change" and "problem" (e.g., "rapid increase in water cut" leading to "flood risk problem").

[0033] Figure 2 This embodiment demonstrates the partial schema logical architecture for the petroleum industry. This architecture serves as a "blueprint" for extracting from a large model, constraining entity attributes by defining "meta" concepts (such as problem classification and cause classification) and constructing a core PRS (Problem-Reason-Solution) inference chain. The diagram clearly illustrates how "indicator changes" perceived from "production wells" and "strata" point to specific "problems" through "cause / effect" relationships, ultimately achieving logical closure through "solutions." This rigorous, structured definition provides hard, factual constraints for addressing the "illusion" problem of large models.

[0034] In summary, this rigorous four-in-one knowledge system of "concept-entity-attribute-relationship" provides overall guidance for subsequent structured information extraction: First, as the core basis for all Prompt design, it "translates" domain knowledge into natural language that the large language model can understand, clearly informing the model which entities, attributes, and relationships should be extracted; second, it sets clear boundaries and goals for the model, enabling the LLM to output results within the established framework in a "fill-in-the-blank" manner, thereby significantly improving the accuracy and reliability of extraction; finally, because each step strictly follows the same system, the entire knowledge graph maintains a high degree of consistency and standardization in structure, avoiding data redundancy and inconsistency, and becoming the quality cornerstone of the automated construction system.

[0035] Based on the rigorously defined domain knowledge schema described above, this embodiment implements an automated knowledge extraction and graph construction pipeline, namely the "Petroleum Knowledge Graph Construction Pipeline." The overall technical process of this pipeline is as follows: Figure 4 As shown, it is an end-to-end automated process.

[0036] First, the system receives a massive amount of documents from the oil exploration and development field as input, including reports in various formats such as PDF, Word, and TXT. Due to the length and density of the original documents, the system first employs semantic segmentation technology to intelligently divide long documents into "knowledge blocks" containing complete contextual semantics. This step is the cornerstone of ensuring the quality of subsequent extraction. Traditional text segmentation methods, such as segmenting by a fixed number of tokens (e.g., 512 tokens) or by fixed paragraphs, have serious shortcomings when dealing with highly structured and logically rigorous reports in the oil field (such as drilling daily reports and well logging interpretation reports). For example, a crucial "fracking operation parameter table" might be forcibly cut in the middle, resulting in the loss of contextual information; or, a complete argumentation process describing the "cause of water flooding" (which may span multiple paragraphs) might be split into two or more "knowledge blocks," making it impossible for the Large Language Model (LLM) to obtain the complete information chain in any single "knowledge block," thus leading to failed relation extraction or factual errors.

[0037] The semantic segmentation technology used in this embodiment abandons this mechanical segmentation method. It is more like "intelligent sentence segmentation," with the core goal of maintaining "knowledge integrity." Using LLM-driven semantic segmentation: first, anchor points are generated by integrating heading levels, pauses such as periods, and semantic coherence; then, ultra-long segments are recursively subdivided, ultimately resulting in self-contained "knowledge blocks" (such as complete geological strata evaluation or construction procedures). Each block retains its hierarchical ID, ensuring contextual integrity and appropriate granularity, making subsequent NER, relation, and event extraction more accurate and stable, and facilitating subsequent knowledge fusion. These "knowledge blocks" are then fed into an extraction process driven by a Large Language Model (LLM) and composed of multiple core modules. This process transforms unstructured text into structured subgraphs through a series of highly collaborative decoupling tasks. These tasks specifically include: 1. Named Entity Recognition: Based on the domain knowledge system, identify all entities in the text that meet the definition.

[0038] 2. Event Extraction: Identify complex dynamic events described in the text.

[0039] 3. Relational Reasoning: Based on the knowledge system, reason and establish legitimate relationship connections between identified entities.

[0040] 4. Entities: Normalize the extracted entities and relations to resolve issues such as synonyms, near-synonyms, and heterogeneity, ensuring the uniqueness and consistency of knowledge.

[0041] Entity and Event Extraction Instructions: Entity and event extraction instructions explicitly direct the Large Language Model (LLM) to act as an "expert in named entity recognition and graph knowledge extraction in the oilfield development field." These instructions define the entity types (such as WellArea, Problems) and their required attributes (such as name, desc, porosity) in the domain knowledge schema as mandatory output formats, requiring the LLM to find all instances that match the definitions from the input text and return them strictly according to a predefined JSON list format. This instruction paradigm ensures that the LLM's output is no longer arbitrary natural language, but rather structured data that can be directly parsed by the machine. Furthermore, by incorporating rules such as "strict schema matching" and "prohibiting unknown types" into the instructions, the behavior of the LLM is effectively constrained, significantly reducing the occurrence of "illusions."

[0042] Relation Extraction Instruction: For relational reasoning, this embodiment employs a more rigorous "constraint-extraction" paradigm. This instruction recognizes that in specialized domains, the relationships between entities are finite and explicit (already defined in the schema). Therefore, this instruction provides the large language model with the entity list identified in the previous step and all valid relationships in the knowledge system as context. The instruction requires the LLM to "connect" between "given entities" and "given relationships," extracting all valid relationships that conform to the definition of a five-tuple (subject entity, subject type, predicate, object entity, object type). For example, when the text contains "6 layers long...implementing measures to increase water injection wells...solved the problem of low overall yield," the system will extract the corresponding triple based on the valid relationship definition of "solution-problem-problem." This design transforms open relation discovery (which is prone to illusion) into a constrained, context-based linking task, greatly improving the accuracy and consistency of the extraction results.

[0043] Ultimately, all extracted and standardized "subgraphs" are merged and loaded into the graph database to form a unified, queryable, and analyzable domain knowledge graph, providing a visual query and analysis interface.

[0044] See Figure 5 As shown, before the knowledge extraction process is completed through multi-task parallelism in the above-mentioned intelligent batch processing architecture, the following system initialization steps are also included: (1) Schema Relationship Extraction and Mapping Construction: The system automatically parses the predefined .schema file to extract the relationship definitions and attribute definitions between all entities. A bilingual (Chinese and English) relationship mapping table is constructed to ensure that the system has the underlying capability to process bilingual corpora.

[0045] (2) Structured assembly of relation lists: The extracted relations are classified and stored according to the logical structure of subject, predicate, and object to form standardized English relation lists, Chinese relation lists, and full relation sets, which serve as the "standard answer library" for subsequent normalization.

[0046] (3) Embedding pre-computation and index construction: This is the most critical computational task in the initialization. The system calls the vectorized model to pre-convert all standard relation names into high-dimensional semantic vectors. These vectors are computed and cached in batch processing (BatchSize is 64), and a semantic index is built in system memory to achieve high-speed semantic matching in the runtime phase (phase two), thereby avoiding repeated calls to the Embedding interface during the extraction process and reducing latency. Among them, Embedding pre-computation can be implemented in parallel and fast computation through an intelligent batch processing architecture.

[0047] (4) Initialization of the logging and monitoring system: Create CSV log files with unique timestamps and set the maximum size of the file processor (e.g., 50MB) and the recording level. At this time, more than 15 core log fields (e.g., timestamp, source_relation, score, etc.) will be predefined to ensure that every decision during the construction process can be recorded in real time.

[0048] The following explains the collaborative relationship between initialization, runtime phase, log monitoring, and batch processing: To clearly demonstrate the operating mechanism of this system, the collaborative relationship between the system initialization and operation phases is described as follows: Initialization is a process of "scaling" and "preloading": it is completed at the beginning of system startup, and its purpose is to generate semantic indexes and mapping tables based on the schema definition. The intelligent batch processing architecture at this stage is mainly used for pre-computation of embeddings, ensuring that large-scale standard lexicons can be quickly converted into vectors.

[0049] The runtime phase is a process of "extraction" and "pipeline": based on the "blueprint" loaded during initialization, the system begins to execute tasks such as semantic segmentation, NER, and relation extraction. The intelligent batch processing architecture plays a core role in this phase, breaking down large-scale chunk tasks into smaller batches and sending them synchronously and sequentially to the LLM, thus solving the rate limiting problem caused by asynchronous concurrency.

[0050] Log monitoring is a "synchronous daemonization" process: the log system (such as...) Figure 7As shown, the system operates completely synchronously with the runtime phase. When the batch processing architecture executes each LLM call or three-layer normalization decision, the logging system captures its input / output token consumption, matching scores, and inference paths in real time. This synchronization ensures that if an exception occurs during the build process, the system can immediately record the error point through the exception stack trace, achieving verifiable and traceable processes.

[0051] The following explains the multi-level normalization algorithm that balances cost and effectiveness. Figure 6 As shown: Simply extracting the original terms (such as "porosity," "unsatisfactory output," and "cause") is insufficient to construct a unified and integrated knowledge graph. The core challenge lies in addressing the aforementioned problem of "heterogeneity and inconsistency in knowledge fusion." Therefore, the entity standardization module in this embodiment's pipeline (such as...) Figure 4 As shown in the figure, an original "three-layer progressive" normalization algorithm that balances cost and effectiveness is employed. The detailed process is as follows: Figure 6 As shown. The algorithm is activated when the system extracts a term to be normalized from the text (such as a non-standard entity or relation name): Layer 1 (Phase 1): Fast Lexical Matching. This is a low-cost, local computation. The system first performs simple cleaning on the terms to be normalized (e.g., removing spaces, converting to lowercase, etc.), and then performs an exact match against the standard knowledge base (i.e., the standard terms defined in the schema). If the match is successful, normalization is completed with the highest confidence (1.0). If unsuccessful, further fuzzy matching is attempted, such as using algorithms like edit distance or Jaccaro similarity, to calculate the lexical similarity with the standard terms. This step can quickly recall most simple, consistent terms or terms with minor spelling errors, and incurs almost no API call cost.

[0052] The second layer (phase two): Efficient semantic matching (Embedding). If the first-layer lexical matching fails (i.e., the terms are synonyms, near-synonyms, or abbreviations, such as "porosity" vs. "pore volume"), the system initiates the second-layer semantic matching. This embodiment employs an efficient vectorization model, which has pre-converted all standard terms in the knowledge system into high-dimensional semantic vectors and established an index. At runtime, the system calls this model to convert the term to be normalized (such as "porosity") into a semantic vector as well, and then finds the Top-K candidates with the closest semantics in the vector space by calculating cosine similarity, etc. (e.g., "porosity" scores 0.98, "permeability" scores 0.65). The system sets a high confidence threshold (e.g., 0.75). If the highest-scoring candidate exceeds this threshold, normalization is automatically completed. This step has a moderate cost (only one Embedding call) and can efficiently handle a large number of heterogeneous problems with different lexical structures but similar semantics.

[0053] The third layer (stage three): Deep Intelligent Decision Making (LLM). This embodiment only activates the most expensive and powerful third layer, "Deep Intelligent Decision Making," when the second-layer semantic matching fails (all candidate scores are below the threshold) or multiple high-scoring candidates are returned (e.g., "well pattern adjustment" and "well spacing adjustment" have similar scores, indicating semantic ambiguity). At this point, the system dynamically constructs a complex instruction containing the original terms, the context corpus, and the candidate standard terms selected from the second layer. This instruction requires the Large Language Model (LLM) to make a final decision based on a thorough understanding of the contextual semantics, choosing one out of N options. For example, the instruction might ask: "In the context of '...the well has poor physical properties, with a porosity of only 8%...', which of the following standard concepts best matches the term 'porosity': [1. porosity, 2. permeability, 3. reserve abundance]?" This cost-sensitive, progressive design ensures that only the most complex normalization tasks consume valuable large model computational resources, maximizing API call overhead savings while maintaining high normalization recall.

[0054] Fourth layer (phase four): Failure handling and retention. For example... Figure 4 As shown in "4. Check", if all three levels of normalization fail, business experts are allowed to decide whether to "completely discard" the unrecognizable knowledge (to ensure high accuracy) or "retain the original terminology" and put it into the database (to ensure high recall and leave it for subsequent manual review).

[0055] To ensure that the complex extraction and normalization process described above can operate stably, efficiently, economically, and reliably in an industrial production environment, this embodiment implements a series of key engineering innovations in the pipeline's underlying architecture to address the limitations of frameworks like KAG in the background art.

[0056] (1) Intelligent Synchronous Batch Processing Architecture (Solving Performance Bottlenecks and Stability Issues) Addressing the risks of API rate limits and system crashes easily triggered by large-scale concurrent requests mentioned in the background technology, this embodiment abandons the traditional high-concurrency asynchronous model. Instead, it adopts a synchronous sequential intelligent batch processing architecture. This architecture incorporates configurable batch processing splitting algorithms tailored to different task characteristics. Independent batch processing configuration for the four core tasks: The system provides independent batch size configuration parameters for the four key steps of knowledge extraction—entity recognition, event extraction, relation extraction, and entity standardization. This means that the batch processing granularity of each step can be flexibly adjusted according to the computational complexity and API consumption characteristics of different tasks. For example, for computationally intensive event extraction and relation extraction tasks requiring complex reasoning, a smaller batch size than entity recognition can be set to ensure the stability and success rate of each API call. Synchronous sequential execution ensures stability: Unlike traditional high-concurrency models, this embodiment executes these tasks in batches synchronously. Whether it's entity recognition, event extraction, relation reasoning, or entity standardization, the system intelligently breaks down the large amount of data to be processed (such as "knowledge blocks," relations to be extracted, and entities to be standardized) into multiple small batches. Then, the system processes each batch sequentially, waiting for the previous batch to successfully return its result before proceeding to the next. Intelligent resource optimization: In tasks requiring contextual information, such as relation extraction, this architecture also introduces an entity filtering optimization mechanism. When processing relation extraction tasks in batches, the system can choose to only pass the entity information involved in the current batch of relations to the model, instead of the entire entity list. This significantly reduces the token consumption for each API call, lowering costs while improving processing efficiency.

[0057] (2) Full-Process Token Consumption Statistics (Solving the Problem of Uncontrolled Costs) Addressing the "black box" and "uncontrolled" API call costs mentioned in the background technology, this embodiment introduces a cost quantification and monitoring system that runs through the entire process. This system includes three stages: pre-process, during-process, and post-process. Pre-process: The system estimates the token consumption for upcoming API calls based on the task type and input text length. During-process: After each call to the large language model (whether extraction or adjudication), the system accurately records the actual number of input and output tokens consumed in that call. Post-process: After all batches of tasks are completed, the system automatically generates a detailed cost summary report, accurately calculating the total token consumption and estimated costs for each stage of the construction task, including NER, relationships, and standardization. This mechanism transforms API call costs from an uncontrollable "black box" state into a measurable, analyzable, and optimizable transparent indicator, providing crucial data support for system cost control and performance tuning (e.g., routing simple tasks to cheaper smaller models).

[0058] (3) Verifiable and Traceable Knowledge Extraction Log System (Solving the "Black Box" and Unreliability Issues) Addressing the fatal flaws of the LLM process being a "black box" and the results being "unreliable" as mentioned in the background technology, this embodiment establishes a fine-grained structured log system, the operation of which is as follows: Figure 7 As shown.

[0059] like Figure 7 As shown, when "Start Building the Knowledge Graph," the system "Initializes the Log System," creating a unique log file with a timestamp and simultaneously configuring "Set Console Processor" (to display INFO-level progress information to the operator in real time) and "Set File Processor" (to record the most detailed debug information at the DEBUG level, with a size limit such as 50MB to prevent the log file from becoming too large). During the "Execute the Build Process," the core value of this log system lies in its ability to record every key knowledge decision in detail in a machine-readable format (such as CSV or JSONL).

[0060] Especially for Figure 6 The normalization process shown in the diagram is meticulously recorded by the logging system, detailing its entire path. For instance, when relation normalization fails or a low-scoring semantic matching result is used, the logging system records the following: [Timestamp], [Original Relation]: "Location", [Normalized Relation]: "Located in", [Method Used]: "Efficient Semantic Matching (Embedding)", [Similarity Score]: 0.6, [Subject]: S12d Well Area, [Object]: XX Construction Zone, [Context]: ... When the process is "successfully constructed", the system calculates the total time and records the success information; if an exception is "captured" during the process, the system records the exception stack trace. This complete and verifiable data loop provides a comprehensive and verifiable data loop for subsequent knowledge quality assessment, manual review, error tracing, and iterative optimization of the domain knowledge system, forming the cornerstone for building a highly reliable industrial-grade knowledge graph.

[0061] After the above "extraction-standardization-engineering assurance" process, the system will assemble the entities, events, and relationships extracted from the "knowledge chunks" (Chunks), conforming to the schema definition, and normalized, into a structured "subgraph" in memory. For example... Figure 4 As shown, the final step is to fuse all the "subgraphs" generated from the documents and load them into the graph database to form a unified, queryable, and analyzable domain knowledge graph.

[0062] The effectiveness of this pipeline is reflected on two levels. First, when extracting from a single document or a group of related documents, the system can construct a logically rigorous and clearly defined "subgraph," such as... Figure 3 As shown in the example sub-diagram, the poor performance of the oil wells in the "6-layer long" (problem entity, pink) is caused by the "excessive injection-production well spacing" and "insufficient displacement pressure" (cause entity, pink), and the corresponding "adjusting the existing well network" (solution entity, green) is proposed to solve the problem.

[0063] Secondly, after the massive amount of documents has been processed and integrated, the system constructs a comprehensive knowledge graph covering a wide range of knowledge in oil exploration and development, such as... Figure 3 As shown, this comprehensive map illustrates how tens of thousands of knowledge points are organically connected into a network, with core entities such as "well areas," "problems," "measures," and "layers" serving as hubs. Different colors in the map represent different entity types (such as entity type, concept type, and event type), clearly demonstrating the macro-structure of the knowledge system. This map provides a solid data foundation for subsequent intelligent retrieval, dynamic analysis, and decision support.

[0064] Based on the above method embodiments, this application also provides a large-scale model extraction and knowledge graph construction device for the oil exploration and development field. See [link to relevant documentation]. Figure 8 As shown, the device includes: a knowledge extraction module 82, which uses an intelligent batch processing architecture to complete the following steps in a multi-task parallel manner: based on the domain knowledge system that conforms to the operation logic of oil exploration and development, it performs semantic segmentation and knowledge extraction processes on documents in the oil exploration and development domain to obtain standardized subgraphs; wherein, the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning and entity standardization; entity standardization is implemented using a three-layer progressive relation normalization algorithm; a graph construction module 84, which integrates all extracted and standardized subgraphs and loads them into a graph database to form a unified, queryable and analyzable knowledge graph for the oil exploration and development domain, and provides a visual query and analysis interface; and a log recording module, which records the input and output token consumption and knowledge extraction reasoning path through a log system throughout the entire knowledge graph construction process.

[0065] Furthermore, the aforementioned domain knowledge system, which conforms to the operational logic of petroleum exploration and development, includes the following multiple definitions: concept types, entity types, entity attributes, and entity relationships. Concept types include: problem classification, cause classification, solution classification, development stage classification, indicator change classification, reservoir classification, heterogeneity classification, sedimentary facies classification, structural unit classification, and well network classification, well classification, and extraction method classification related to engineering parameters. Entity types include: geological and engineering unit entities, development indicator and parameter entities, and dynamic analysis and decision-making entities. Entity relationships include: static structural relationships, indicator attribution relationships, and dynamic causal relationships.

[0066] Furthermore, before completing the following steps through multi-task parallelism, the aforementioned intelligent batch processing architecture also includes: schema relationship extraction and mapping construction, structured assembly of the relationship list, embedding pre-computation and index construction, and log and monitoring system initialization; embedding pre-computation can be achieved through parallel and fast computation via the intelligent batch processing architecture.

[0067] Furthermore, the semantic segmentation described above adopts LLM-driven semantic segmentation, which divides the text into knowledge blocks with complete semantic units.

[0068] Furthermore, the aforementioned three-layer progressive relation normalization algorithm includes the following three strategies in descending order of priority: fast lexical matching strategy, efficient semantic matching strategy, and deep intelligent adjudication strategy; if any strategy fails to normalize, the next priority strategy will be used for normalization processing.

[0069] Furthermore, the aforementioned fast lexical matching strategy includes: cleaning the terms to be normalized; performing precise matching between the cleaned terms and a standard knowledge base; the standard knowledge base consists of standard terms defined in the domain knowledge system; if the match is successful, normalization is performed with the highest confidence; if the match is unsuccessful, fuzzy matching is performed by calculating the similarity with the standard terms; if fuzzy matching is unsuccessful, normalization is performed using an efficient semantic matching strategy; the efficient semantic matching strategy includes: calling a vectorization model to convert the terms to be normalized into semantic vectors, calculating the similarity between the semantic vectors and the standard semantic vectors in the vectorization model, determining a specified number of candidate vectors with the highest similarity ranking, and performing normalization based on the vectors among the candidate vectors whose similarity exceeds the confidence threshold; if there are no vectors whose similarity exceeds the set confidence threshold, a deep intelligent adjudication strategy is used for normalization; the deep intelligent adjudication strategy includes: dynamically constructing a complex instruction containing the terms to be normalized, the context corpus, and the candidate standard terms selected by the efficient semantic matching strategy, requiring the large language model to make a final decision based on a full understanding of the context semantics.

[0070] Furthermore, the steps described above for recording input and output token consumption through a log system include: estimating the token consumption for the upcoming API call based on the task type and input text length before calling the large language model; recording the actual number of input tokens and output tokens consumed in each call after each large language model call; and automatically generating a cost summary report after all batch tasks are completed, which calculates the total token consumption of this build task at each stage.

[0071] Furthermore, the steps described above for recording the knowledge extraction and reasoning path through a log system include: initializing the log system when starting to build the knowledge graph, creating a unique log file with a timestamp, and simultaneously configuring and setting the console processor and the file processor; recording each key knowledge decision process in a machine-readable format during the execution of the construction process; the knowledge decision process includes the complete path of the normalization process.

[0072] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts of the device embodiment not mentioned can be referred to the corresponding content in the aforementioned method embodiment.

[0073] This application also provides an electronic device, such as... Figure 9 The diagram shows the structure of the electronic device, which includes a processor 91 and a memory 90. The memory 90 stores computer-executable instructions that can be executed by the processor 91, and the processor 91 executes the computer-executable instructions to implement the above-described method.

[0074] exist Figure 9 In the illustrated embodiment, the electronic device further includes a bus 92 and a communication interface 93, wherein the processor 91, the communication interface 93, and the memory 90 are connected via the bus 92.

[0075] The memory 90 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 93 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 92 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 92 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0076] The processor 91 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 91 or by instructions in software form. The processor 91 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory, and the processor 91 reads the information in the memory and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiment.

[0077] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the above-described method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0078] The computer program products of the methods, apparatus, and electronic devices provided in the embodiments of this application include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementations, please refer to the method embodiments, which will not be repeated here.

[0079] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0080] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0082] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for extracting knowledge graphs from large models in the field of oil exploration and development, characterized in that, The method includes: An intelligent batch processing architecture is applied to complete the following steps in a multi-task parallel manner: Based on the domain knowledge system that conforms to the operation logic of oil exploration and development, which is loaded during initialization, semantic segmentation and knowledge extraction processes are performed on documents in the oil exploration and development domain to obtain standardized subgraphs; wherein, the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning and entity standardization; the entity standardization is implemented using a three-layer progressive relation normalization algorithm; All extracted and standardized subgraphs are merged and loaded into the graph database to form a unified, queryable, and analyzable knowledge graph for the field of oil exploration and development, and a visual query and analysis interface is provided. Throughout the knowledge graph construction process, a log system records the input and output token consumption and the knowledge extraction and reasoning path.

2. The method according to claim 1, characterized in that, The domain knowledge system that conforms to the operational logic of petroleum exploration and development includes the following definitions: concept types, entity types, entity attributes, and entity relationships. The concept types include: problem classification, cause classification, solution classification, development stage classification, indicator change classification, reservoir classification, heterogeneity classification, sedimentary facies classification, structural unit classification, and well network classification, well classification, and extraction method classification related to engineering parameters. The entity types include: geological and engineering unit entities, development indicator and parameter entities, and dynamic analysis and decision-making entities. The entity relationships include: static structural relationships, indicator attribution relationships, and dynamic causal relationships.

3. The method according to claim 1, characterized in that, Before completing the following steps using an intelligent batch processing architecture and a multi-task parallel approach, the process also includes: schema relationship extraction and mapping construction, structured assembly of the relationship list, embedding pre-computation and index construction, and log and monitoring system initialization; the embedding pre-computation can be performed in parallel and quickly using the intelligent batch processing architecture.

4. The method according to claim 1, characterized in that, The semantic segmentation adopts LLM-driven semantic segmentation, which divides the text into knowledge blocks with complete semantic units.

5. The method according to claim 1, characterized in that, The three-layer progressive relation normalization algorithm includes the following three strategies in descending order of priority: fast lexical matching strategy, efficient semantic matching strategy, and deep intelligent adjudication strategy; if any strategy fails to normalize, the next priority strategy will be used for normalization processing.

6. The method according to claim 5, characterized in that, The fast lexical matching strategy includes: cleaning the terms to be normalized; performing precise matching between the cleaned terms and a standard knowledge base; the standard knowledge base consists of standard terms defined in the domain knowledge system; if the match is successful, normalization is performed with the highest confidence; if the match is unsuccessful, fuzzy matching is performed by calculating the similarity with the standard terms; if fuzzy matching is unsuccessful, normalization is performed using the efficient semantic matching strategy; the efficient semantic matching strategy includes: calling a vectorization model to convert the terms to be normalized into semantic vectors, calculating the similarity between the semantic vectors and the standard semantic vectors in the vectorization model, determining a specified number of candidate vectors with high similarity ranking, and performing normalization based on the vectors among the candidate vectors whose similarity exceeds the confidence threshold; if there are no vectors whose similarity exceeds the set confidence threshold, a deep intelligent adjudication strategy is used for normalization; the deep intelligent adjudication strategy includes: dynamically constructing a complex instruction containing the terms to be normalized, the context corpus, and the candidate standard terms selected by the efficient semantic matching strategy, requiring the large language model to make a final decision based on a full understanding of the context semantics.

7. The method according to claim 1, characterized in that, The steps involved in consuming input and output tokens are recorded through a logging system, including: Before calling the large language model, the token consumption of the upcoming API call is estimated based on the task type and the length of the input text. After each call to the large language model, the actual number of input tokens and output tokens consumed in that call is recorded. After all batches of tasks are completed, a cost summary report is automatically generated to calculate the total token consumption of this build task at each stage.

8. The method according to claim 1, characterized in that, The steps of the knowledge extraction and reasoning path are recorded through a log system, including: When starting to build the knowledge graph, the logging system is initialized, a unique log file with a timestamp is created, and the console processor and file processor are configured and set simultaneously. During the execution of the building process, each key knowledge decision process is recorded in a machine-readable format. The knowledge decision process includes the complete path of the normalization process.

9. A knowledge graph extraction and construction device for large-scale models in the field of petroleum exploration and development, characterized in that, The device includes: The knowledge extraction module is used to apply an intelligent batch processing architecture and complete the following steps in a multi-task parallel manner: Based on the domain knowledge system that conforms to the operation logic of oil exploration and development, loaded during initialization, semantic segmentation and knowledge extraction processes are performed on documents in the oil exploration and development domain to obtain standardized subgraphs; wherein, the knowledge extraction process includes: named entity recognition, event extraction, relation reasoning and entity standardization; the entity standardization is implemented using a three-layer progressive relation normalization algorithm; The graph construction module is used to merge all extracted and standardized subgraphs and load them into the graph database to form a unified, queryable, and analyzable knowledge graph for the field of oil exploration and development, and provides a visual query and analysis interface. The logging module is used to record the input and output token consumption and knowledge extraction and reasoning paths throughout the entire knowledge graph construction process.

10. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 8.

Images