Electric power standard knowledge extraction method based on large language model and related device

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an electric power standard knowledge extraction method based on a large language model and a related device. The method comprises the steps of obtaining a to-be-extracted power standard file; extracting a target sentence in the to-be-extracted power standard file, and filling a prompt template based on the target sentence; inputting the filled prompt template into the fine-tuned model of the large language model to obtain a target sentence type; based on the target sentence and the target sentence type, filling an entity attribute extraction template; and inputting the entity attribute extraction template into the fine-tuning large oracle model, and extracting entities contained in the target sentence. According to the method, the general large language model and the electric power professional scene are precisely combined, the problems that a traditional model is insufficient in electric power text semantic understanding and low in professional knowledge extraction pertinence are solved, the normalization and accuracy of entity extraction in the electric power standard are improved, and a solid foundation is laid for subsequent entity verification and knowledge base updating.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a method and related apparatus for extracting power standard knowledge based on a large language model. Background Technology

[0002] Knowledge extraction is a very traditional and widely used task in Natural Language Processing (NLP). Its goal is to extract structured knowledge from unstructured text, including entity extraction, relation extraction, and event extraction. Entity extraction is the most mature task in knowledge extraction. From HMM (Hidden Markov Model) and CRF (Conditional Random Field) to BiLSTM (Bidirectional Long Short-Term Memory), and then to Lattice LSTM (Lattice Long Short-Term Memory) and TENNER (Adapting Transformer Encoder for Named Entity Recognition), entity recognition has made great progress on standard datasets. With the emergence of pre-trained models such as BERT (Bidirectional Encoder Representations from Transformers), entity recognition has been further improved. Relation extraction generally uses supervised or remotely supervised classification models. Before the advent of deep learning, it mainly relied on manual features and kernel-based models. In the deep learning era, mainstream models include CNN (Convolutional Neural Network) and PCNN (Piecewise Convolutional Neural Network). Networks (such as segmented convolutional neural networks) are used for knowledge extraction. Event extraction is the most challenging task, and its development is similar to that of relation extraction, with DMCNN (Dynamic Multi-pooling Convolutional Neural Network) being a representative example. Overall, research on knowledge extraction has made some progress, but problems such as difficulty in obtaining samples, complex text semantics, poor generalization ability, and poor practicality still exist.

[0003] With the rapid development of deep learning technology, large language models such as GPT (Generative Pre-trained Transformer) and LLaMA (Large Language Model Meta AI) possess powerful understanding and generative capabilities. The general intelligence capabilities of pre-trained models provide new ideas for knowledge extraction tasks. However, existing large language models still suffer from problems in power standard scenarios, such as a disconnect between general and specialized knowledge, high fine-tuning costs, and insufficient controllability of extraction results. A mature solution for power standard knowledge extraction has yet to be formed. Summary of the Invention

[0004] The purpose of this invention is to provide a method and related apparatus for extracting knowledge of power standards based on a large language model, to construct a knowledge extraction framework adapted to the characteristics of power standard texts, to improve the accuracy and completeness of professional terminology recognition, complex relationship parsing, and implicit event extraction, and to solve the problem of insufficient professional adaptability of traditional models.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for extracting power standard knowledge based on a large language model, comprising: Obtain the power standard documents to be extracted; Extract the target sentence from the power standard file to be extracted, and fill the prompt template based on the target sentence; input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type; fill the entity attribute extraction template based on the target sentence and the target sentence type; input the entity attribute extraction template into the fine-tuned large oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power domain prompt template, the power domain labeled dataset is transformed into a three-element fine-tuning dataset with instruction prompts → text input → standard output; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

[0006] A further improvement of the present invention is that the prompt template is specifically "Based on the given sentence {}, please determine whether the sentence belongs to type {}, where the types include ={equipment, fault, operation and maintenance, scheduling}".

[0007] A further improvement of this invention is that: the entity attribute extraction template is "The input sentence {} belongs to type {}, please extract all entities and entity types from it and output them in the format <entity, entity type>, where entity type = {PER, ORG, PEM, ET}, where <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>"Indicates electrical technology."

[0008] A further improvement of the present invention is that it also includes: Based on the entities contained in the extracted target sentences, the entity linking method is used to check whether the entity exists in the preset power knowledge base. If it exists, the result is added to the candidate set; if it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuning big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0009] A further improvement of this invention is that the step of searching for the existence of entities in a preset power knowledge base based on the entities contained in the extracted target sentence using an entity linking method specifically includes: Based on the extraction results of the fine-tuned big oracle model, a similarity matching method is used to search whether the entity exists in the preset power knowledge base. If the entity exists in the preset power knowledge base and the similarity score is higher than the set threshold, the entity is added to the candidate set. If it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0010] A further improvement of the present invention is that it also includes: based on the candidate set, updating entities that do not exist in the preset power knowledge base to the preset power knowledge base, and outputting the information extraction results of the sentence in a preset format.

[0011] A further improvement of this invention is that the quantization low-rank adaptive method specifically includes: A quantized model is obtained by quantizing a large language model using 4-bit normal floating-point numbers and double quantization. Low-rank decomposition matrices A and B are introduced as a bypass in the quantization model, and the quantized parameters... Update only adapter parameters ,in The rank of the decomposition matrix is ​​the weight increment. d and k are the row and column dimensions of the weight matrix of the large language model, respectively; The model performance is evaluated by combining objective indicators and subjective scoring, and the qualified models are saved as fine-tuning large prediction models.

[0012] Secondly, the present invention provides a power standard knowledge extraction device based on a large language model, comprising: The acquisition module is used to acquire the power standard file to be extracted; The extraction module is used to extract the target sentence from the power standard file to be extracted, fill the prompt template based on the target sentence; input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type; fill the entity attribute extraction template based on the target sentence and the target sentence type; input the entity attribute extraction template into the fine-tuned large oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power domain prompt template, the power domain labeled dataset is transformed into a three-element fine-tuning dataset with instruction prompts → text input → standard output; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

[0013] A further improvement of the present invention is that the prompt template is specifically "Based on the given sentence {}, please determine whether the sentence belongs to type {}, where the types include ={equipment, fault, operation and maintenance, scheduling}".

[0014] A further improvement of this invention is that: the entity attribute extraction template is "The input sentence {} belongs to type {}, please extract all entities and entity types from it and output them in the format <entity, entity type>, where entity type = {PER, ORG, PEM, ET}, where <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>"Indicates electrical technology."

[0015] A further improvement of the present invention is that it also includes: The judgment module is used to check whether the entity contained in the extracted target sentence exists in the preset power knowledge base through the entity linking method. If it exists, the result is added to the candidate set; if it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuning big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0016] A further improvement of this invention is that the step of searching for the existence of entities in a preset power knowledge base based on the entities contained in the extracted target sentence using an entity linking method specifically includes: Based on the extraction results of the fine-tuned big oracle model, a similarity matching method is used to search whether the entity exists in the preset power knowledge base. If the entity exists in the preset power knowledge base and the similarity score is higher than the set threshold, the entity is added to the candidate set. If it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0017] A further improvement of the present invention is that it also includes: The update module is used to update entities that are not in the preset power knowledge base to the preset power knowledge base based on the candidate set, and output the information extraction results of the sentence in a preset format.

[0018] A further improvement of this invention is that the quantization low-rank adaptive method specifically includes: A quantized model is obtained by quantizing a large language model using 4-bit normal floating-point numbers and double quantization. Low-rank decomposition matrices A and B are introduced as a bypass in the quantization model, and the quantized parameters... Update only adapter parameters ,in The rank of the decomposition matrix is ​​the weight increment. d and k are the row and column dimensions of the weight matrix of the large language model, respectively; The model performance is evaluated by combining objective indicators and subjective scoring, and the qualified models are saved as fine-tuning large prediction models.

[0019] Thirdly, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement a power standard knowledge extraction method based on a large language model as described above.

[0020] Fourthly, the present invention provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the aforementioned method for extracting power standard knowledge based on a large language model.

[0021] Compared with the prior art, the present invention has the following unexpected technical effects: This invention provides a method for extracting power standard knowledge based on a large language model, comprising: acquiring a power standard file to be extracted; extracting a target sentence from the power standard file to be extracted, and filling a prompt template based on the target sentence; inputting the filled prompt template into a fine-tuned model of the large language model to obtain the target sentence type; filling an entity attribute extraction template based on the target sentence and the target sentence type; inputting the entity attribute extraction template into a fine-tuned large oracle model to extract entities contained in the target sentence; the fine-tuned large oracle model is obtained through the following steps: constructing a power domain annotation dataset, the annotation dataset containing three types of annotation task data: entity, relation, and event; using a preset power domain prompt template, converting the power domain annotation dataset into a three-element structure fine-tuning dataset of "instruction prompt → text input → standard output"; using a quantized low-rank adaptive method to perform supervised fine-tuning of the large language model, evaluating and saving the fine-tuned large oracle model. This invention constructs a structured extraction framework adapted to power standard text. This method precisely combines a general-purpose large language model with power industry-specific scenarios, solving the problems of insufficient semantic understanding of power text and weak targeting of professional knowledge extraction in traditional models. It improves the standardization and accuracy of entity extraction in power standards, laying a solid foundation for subsequent entity verification and knowledge base updates. This invention enhances the comprehensive extraction capability of the large-scale predictive model for power standard knowledge, meeting practical business needs such as power knowledge graph construction and intelligent standard applications. This invention proposes a large-scale language model fine-tuning scheme adapted to the power industry. It constructs a power industry dataset containing three types of annotation tasks: entities, relations, and events, and transforms it into a ternary structure fine-tuning dataset of "instruction prompt → text input → standard output," employing a quantized low-rank adaptive method for supervised fine-tuning. This scheme solves the problems of the disconnect between general-purpose large language models and power industry professional knowledge, and the high cost of fine-tuning. It achieves accurate model adaptation to power industry knowledge, significantly improving the model's professional performance in power standard knowledge extraction tasks, laying a model foundation for subsequent efficient extraction.

[0022] Furthermore, this invention provides clear and consistent guidance for determining the type of a target sentence by explicitly defining the prompt template as "Based on the given sentence {}, please determine whether the sentence belongs to type {}, where all types include ={equipment, fault, operation and maintenance, scheduling}". This template avoids ambiguity or deviation from the power industry scenario when large language models determine sentence type, ensuring the consistency and accuracy of sentence type classification. This, in turn, provides accurate scenario positioning for subsequent targeted entity attribute extraction, improving the coherence and reliability of the entire knowledge extraction process.

[0023] Furthermore, this invention defines the entity attribute extraction template as "The input sentence {} belongs to type {}. Please extract all entities and entity types from it and output them in the format <entity, entity type>, where entity type = {PER, ORG, PEM, ET}, where..." <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>It represents "electric power technology"; it clarifies the entity type classification (PER, ORG, PEM, ET) and their corresponding meanings, and limits the output format; this design allows the large language model to accurately focus on the core entities in the power field, avoids the problems of entity type confusion and messy output format, and greatly improves the accuracy and structuring of power professional entity extraction, which facilitates the further processing and application of the extraction results.

[0024] Furthermore, this invention adds an entity candidate set screening step, employing a dual mechanism of "entity link search in the knowledge base + model back-questioning verification" to validate the extracted entities. This mechanism effectively filters out non-entity extraction results, ensuring the authenticity and relevance of entities within the candidate set. It addresses the issues of insufficient controllability and the presence of invalid or erroneous entities in the extraction results from large language models, improving the reliability of power standard knowledge extraction results and providing a high-quality source of entities for subsequent knowledge base updates.

[0025] Furthermore, this invention refines the specific implementation of entity link search, explicitly employing a similarity matching method and threshold judgment rules. This refined design makes the entity existence verification process more operable. By quantifying the judgment standard through similarity thresholds, it avoids biases caused by subjective judgments. Simultaneously, combined with a model-based verification mechanism, it further improves the entity screening logic, ensuring that the candidate set entities not only conform to the existing specifications of the power knowledge base but also effectively identify newly added reasonable entities, thus balancing the accuracy and completeness of the extraction results.

[0026] Furthermore, this invention clarifies the specific steps of the quantization low-rank adaptive method. Model quantization is performed using 4-bit normal floating-point numbers (NF4) and dual quantization (DQ), and adapter parameters are updated using a low-rank decomposition matrix, while freezing the original model weights. This design significantly reduces the computational cost and memory usage of model fine-tuning while ensuring no performance degradation, solving the problems of high resource consumption and low efficiency in fine-tuning large models. Simultaneously, multi-dimensional evaluation metrics ensure that the fine-tuned model meets the standards, improving the reliability and effectiveness of model fine-tuning and providing technical support for the lightweight and efficient application of large models in the power sector. Attached Figure Description

[0027] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart for extracting power standard knowledge based on a large model; Figure 2 A schematic diagram for fine-tuning the low-rank adaptation; Figure 3 This is a flowchart illustrating a method for extracting power standard knowledge based on a large language model, according to an embodiment of the present invention. Figure 4 The attached diagram shows the structure of a power standard knowledge extraction device based on a large language model according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0029] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0030] Example 1 Please see Figure 1 As shown in the figure, this embodiment of the invention proposes a method for extracting power standard knowledge based on a large language model, including the following steps: Step 1: Construct standard initial data This invention, based on massive amounts of power industry standard texts—including technical standards, historical maintenance plans, work orders, and other data related to the power industry—builds an equipment maintenance knowledge base using a question-and-answer format. The knowledge base primarily includes: Step 1-1: Data Collection and Classification: Collect texts covering national, industry, and enterprise standards, technical specifications, standard revision instructions, etc., and classify and archive them according to standard type (technical standards, management standards, safety standards, test standards), voltage level (low voltage, medium voltage, high voltage, ultra-high voltage), and professional direction to obtain an initial dataset; by establishing a corpus classification label system, scenario-based data support will be provided for subsequent model adaptation.

[0031] Steps 1-2: Data Cleaning: Considering the format characteristics of power standard texts (such as chapter / clause structure, table / formula embedding, and symbol annotation), a multimodal preprocessing workflow is designed to clean the initial dataset: Regular expressions are used to clean redundant symbols (such as numbers, punctuation, and annotations); table recognition algorithms are used to convert structured tables into text sequences to ensure corpus integrity; professional terminology is standardized (e.g., "SF6 circuit breaker" and "sulfur hexafluoride circuit breaker" are standardized), and sentence structures are normalized (excessively long compound sentences exceeding a set threshold are split, and omitted components are supplemented), resulting in a cleaned dataset.

[0032] Steps 1-3: Construction of the Power Industry Annotated Dataset: Define and annotate entity categories (e.g., equipment entities, parameter entities, technical entities, organizational entities), relationship categories (e.g., "parameter-equipment" association, "technology-applicable scenario" correspondence, "standard-reference" association), and event categories (e.g., "technical requirements," "test methods," "safety thresholds," "violation handling") to ensure annotation consistency. Select typical power standard texts from the cleaned dataset for manual fine-tuning to construct a seed annotation set. Based on the power industry knowledge graph and professional dictionary, automatically annotate the large-scale corpus in the cleaned dataset using remote supervision methods to reduce annotation costs. Utilize a pre-trained model to pre-annotate unannotated corpus, followed by manual review and correction to improve annotation efficiency. Finally, form a power standard knowledge extraction and annotation dataset containing three types of tasks: entities, relationships, and events. In one specific implementation, the model used for preliminary annotation training can be YOLO or CNN.

[0033] Step 2: Model Training Step 2-1: Model Deployment: Before starting the large language model, it is necessary to prepare teacher models, including but not limited to OpenAI's o1 model, DeepSeek-R1 model, Guangming Power large model (Guangming-32B), etc., and set up corresponding data transmission interfaces to efficiently transmit power standard data to teacher models for processing. Step 2-2: Fine-tuning Dataset Construction: For the three types of tasks—entity extraction, relation extraction, and event extraction—a "term type + contextual hint" prompt template is designed based on the characteristics of power industry terminology. For example: "Please extract equipment entities, parameter entities, and technical entities from the following power standard text, in the format {entity type: entity name}: [text content]". Based on the relation types in the power industry, a "entity pair + relation category hint" prompt template is designed. For example: "Given that the power standard text contains entity A ([entity A name]) and entity B ([entity B name]), please determine the relation type between them (optional: parameter-equipment, technology-applicable scenario, standard-reference). If the above relation does not exist, output 'none'." This invention transforms ordinary labeled datasets into a format suitable for fine-tuning Large Language Models (LLM) by designing prompt templates specific to the power industry, enabling the model to accurately learn entity extraction, relation extraction, and event extraction tasks in the power industry, thus forming a fine-tuned dataset.

[0034] Core objective: To adapt labeled datasets to fine-tune large models; Typical power knowledge extraction and annotation datasets are usually in a structured format of "text + annotation results" (e.g., text annotated with [equipment entity: transformer] [relationship: parameter-equipment]). This format is more suitable for traditional machine learning models, such as the BERT-BiLSTM-CRF joint model; however, it does not conform to the "instruction-output" training paradigm of large models. The purpose of designing the prompt template is to transform the dataset into a three-part structure of "instruction prompt → text input → standard output," which perfectly matches the training data format required for fine-tuning large models.

[0035] The purpose of the two types of prompt templates: The key to fine-tuning large models is to make the model understand the task requirements. The special characteristics of the power field are that the terminology is highly specialized (e.g., SF6 circuit breaker, instrument transformer, zero-sequence current) and the relationship types are fixed (e.g., parameter-equipment, technology-applicable scenario, standard-reference). Therefore, the templates must be designed in accordance with these characteristics. The "Terminology Type + Context Hint" template is applicable to entity extraction tasks (extracting three types of entities: equipment, parameters, and technology). The essence of the template is to clearly define the "task instructions + output format" for the model. Example breakdown: Instruction section: Please extract the equipment entity, parameter entity, and technical entity from the following power standard text, in the format {entity type: entity name}; Input section: [Text content] (Electricity standard text already marked in steps 1-3); Output section: Entity annotation results from the original annotation data (e.g., {Equipment entity: SF6 circuit breaker} {Parameter entity: Rated voltage 10kV}). The modified data format is: instruction + text → labeled results, which conforms to the learning logic of the large model: "input instruction, output answer".

[0036] The "Entity Pair + Relationship Category Hint" template is used for relationship extraction tasks (determining the relationship type between two entities). Its essence is to limit the range of relationship categories, preventing the model from outputting irrelevant relationships outside the power domain. Example breakdown: Instruction section: Given that the power standard text contains entity A ([Entity A name]) and entity B ([Entity B name]), please determine the relationship type between them (optional: parameter-device, technology-applicable scenario, standard-reference). If no such relationship exists, output 'None'. Input: Entity A + Entity B + corresponding text fragment; Output: Relationship results from the original labeled data (e.g., parameter-device or none); Define the set of relationship types in the power industry to address the issues of excessive generalization and insufficient specialization in large models.

[0037] The core of fine-tuning large models is learning the "mapping relationship between instructions and outputs", rather than simply learning the labels.

[0038] Ordinary labeled datasets only have a "text → label" mapping, which large models cannot understand what task the label represents. After adding prompt templates, the dataset becomes a "command + text → label" mapping, which is equivalent to providing large models with a large number of training samples with task descriptions. The model can learn to output the corresponding entity / relationship results when it sees such power task instructions.

[0039] Step 2-3 Model Fine-Tuning: Using the fine-tuning dataset generated in Step 2-2, supervised fine-tuning (SFT) is performed on large models (OpenAI's O1 model, DeepSeek-R1 model) using quantized low-rank adaptive tuning. The basic principle is to quantize the full-parameter model using 4-bit normal floating-point numbers (NF4) and double quantization (DQ), freeze the model weights, and introduce an additional low-rank decomposition matrix in the side path of the model. The introduced parameters are then trained and updated to adapt to new data and tasks. Figure 2 As shown, while ensuring that the performance of the pre-trained model does not degrade, it can effectively improve the flexibility of model application, while saving a lot of training time and cost.

[0040] Steps 2-3 specifically include: Step 2-3-1: Quantize the original pre-trained model using 4-bit Normal Float (NF4) and Double Quantization (DQ). The specific implementations of NF4 and DQ are shown in equations (1) and (2).

[0041] (1)

[0042] In the formula These are the weights of the original model. For the quantization of the first quantiles, The quantile function of the standard normal distribution. This represents the number of bits in the final quantization model.

[0043] (2)

[0044] In the formula For the original 16-bit floating-point model weights, These are the quantization constants and quantization weights after k-bit NormalFloat quantization. It is the quadratic quantization constant, that is Secondary quantization acts on the primary quantization constant, which can further reduce memory usage.

[0045] Step 2-3-2: Add a bypass to the quantization model bypass by introducing a low-rank decomposition matrix. and Initialization is performed using a random Gaussian matrix and a zero matrix respectively, ensuring that the initial state fits the parameter matrix. This does not affect the final output. The quantized parameters are frozen during training. Update only adapter parameters The specific update can be expressed as equation (3).

[0046] (3) The rank of the decomposition matrix .

[0047] Step 2-3-3: Evaluate the model performance comprehensively by using objective quantitative indicators (such as validation loss, precision, recall, etc.) and human subjective scoring, and save the model file to obtain the fine-tuned big oracle model.

[0048] Step 3: Standard Knowledge Extraction Step 3-1: Combining the prompt template from Step 2-2, input the prompt template into the fine-tuning large oracle model to obtain the target sentence type Sentence_type; The design of the information prompt template specifically refers to "Based on the given sentence {}, please determine whether the sentence belongs to type {}, where the types include = {equipment, fault, operation and maintenance, scheduling}". Step 3-2: Based on the sentence type Sentence_type obtained in Step 3-1, construct the entity attribute extraction template ie_prompt according to the pre-designed schema of the corresponding type, and extract the entity attribute elements contained in the sentence.

[0049] Step 3-2 specifically refers to: Step 3-2-1: Entity attribute extraction templates specifically refer to setting different prompt templates according to different information extraction tasks, and then filling the input sentence into specific slots in the template. The goal of the relation extraction task is to extract triples from the text, and the corresponding prompt template is "The input sentence {} belongs to type {}. Please extract entity triples from it and output them, where the triple format is <head entity, relation, tail entity>". The named entity recognition task aims to extract entities from the text, and the corresponding prompt template is "The input sentence {} belongs to type {}. Please extract all entities and entity types from it and output them in the format <entity, entity type>, where entity type = {PER, ORG, PEM, ET}, where..." <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>The task of extracting events from text is titled "Power technology". The corresponding prompt template is "The input sentence {} belongs to the {} type. Please extract the time, place, people, main content, etc. involved in the event and output it in JSON format". Step 3-2-2: Input the prompt template obtained in Step 2-1 into the fine-tuning large oracle model to obtain the information extraction task results; Step 3-3: Based on the entity obtained in Step 3-2, check whether the entity exists in the existing power knowledge base using the entity linking method. If it exists, add the result to the candidate set; if it does not exist, design a result verification prompt template to ask whether the pre-trained model has extracted an entity. If it is an entity, add it to the candidate set; otherwise, discard the entity. Step 3-3 specifically includes: Step 3-3-1: Entity lookup in the existing power knowledge base using the entity linking method involves retrieving the existence of the entity in the knowledge base based on the extraction results of the pre-trained model and using a similarity matching method. If the entity exists in the knowledge base and its similarity score is higher than a set threshold, it is added to the candidate set; otherwise, a result verification prompt template is designed, and the pre-trained model is used again to determine whether it is an entity. If the answer is "yes," the result is added to the candidate set; otherwise, it is discarded. Step 3-4: Based on the candidate set obtained in Step 3-3, update the entities in the set that do not exist in the knowledge base to the power knowledge base, and output the information extraction results of the sentences in JSON format.

[0050] Example 2 This invention proposes a method for extracting power standard knowledge based on a large language model, comprising the following steps: Step 1: Construct standard initial data This invention, based on massive amounts of data related to the power industry and standards, historical maintenance plans, work orders, etc., constructs an equipment maintenance knowledge base using a question-and-answer format. The main tasks include: Step 1-1: Data Collection and Classification: Collect texts including national, industry, and enterprise standards, technical specifications, and standard revision instructions. Classify and archive these texts according to standard type (technical standards, management standards, safety standards, testing standards), voltage level (low voltage, medium voltage, high voltage, ultra-high voltage), and professional direction. Establish a corpus classification and labeling system to provide scenario-based data support for subsequent model adaptation.

[0051] Steps 1-2: Data Cleaning: A multimodal preprocessing workflow is designed to address the format characteristics of power standard texts (such as chapter / clause structure, table / formula embedding, and symbol annotation). Regular expressions are used to clean up redundant symbols (such as numbers, punctuation, and annotations). Table recognition algorithms are used to convert structured tables into text sequences, ensuring the integrity of the corpus. Standardized terminology is implemented (e.g., standardizing "SF6 circuit breaker" and "sulfur hexafluoride circuit breaker"), and sentence structures are standardized (breaking down overly long compound sentences and supplementing omitted components).

[0052] Steps 1-3: Construction of the Power Industry Annotated Dataset: Define and annotate entity categories (e.g., equipment entities, parameter entities, technical entities, organizational entities), relationship categories (e.g., "parameter-equipment" association, "technology-applicable scenario" correspondence, "standard-reference" association), and event categories (e.g., "technical requirements," "test methods," "safety thresholds," "violation handling") to ensure annotation consistency. Select typical power standard texts for manual fine-tuning to construct a seed annotation set. Based on the power industry knowledge graph and professional dictionary, automatically annotate large-scale corpora using remote supervision methods to reduce annotation costs. Use pre-trained models to pre-annotate unannotated corpora, which are then manually reviewed and corrected to improve annotation efficiency. Finally, form a power standard knowledge extraction and annotation dataset containing three types of tasks: entities, relationships, and events.

[0053] Step 2: Model Training Step 2-1: Model Deployment: Before starting the large language model, it is necessary to prepare teacher models, including but not limited to OpenAI's o1 model, DeepSeek-R1 model, etc., and set up corresponding data transmission interfaces to efficiently transmit power standard data to the teacher models for processing. Step 2-2: Fine-tuning the Dataset Construction: For the three tasks of entity extraction, relation extraction, and event extraction, and combining the characteristics of power industry terminology, a "Term Type + Context Hint" template is designed, for example: "Please extract equipment entities, parameter entities, and technical entities from the following power standard text, in the format {Entity Type: Entity Name}: [Text Content]". Based on the relation types in the power field, a "Entity Pair + Relationship Category Hint" template is designed, for example: "Given that the power standard text contains entity A ([Entity A Name]) and entity B ([Entity B Name]), please determine the relationship type between them (optional: parameter-equipment, technology-applicable scenario, standard-reference). If the above relationship does not exist, output 'None'". Combined with the standard data in Step 1-3, a fine-tuning dataset is formed.

[0054] Step 2-3 Model Fine-tuning: Using the power standard thinking chain dataset generated in Step 2-2, supervised fine-tuning (SFT) of the large power industry model is performed using quantized low-rank adaptive modeling. The basic principle is to quantize the full-parameter model using 4-bit normal floating-point numbers (NF4) and double quantization (DQ), freeze the model weights, and introduce an additional low-rank decomposition matrix in the side path of the model. The introduced parameters are then trained and updated to adapt to new data and tasks. Figure 2 As shown, while ensuring that the performance of the pre-trained model does not degrade, it can effectively improve the flexibility of model application, while saving a lot of training time and cost.

[0055] Steps 2-3 specifically include: Step 2-3-1: Quantize the original pre-trained model using 4-bit Normal Float (NF4) and Double Quantization (DQ). The specific implementations of NF4 and DQ are shown in equations (1) and (2).

[0056] (1)

[0057] In the formula These are the weights of the original model. For the quantization of the first quantiles, The quantile function of the standard normal distribution. This represents the number of bits in the final quantization model.

[0058] (2)

[0059] In the formula For the original 16-bit floating-point model weights, These are the quantization constants and quantization weights after k-bit NormalFloat quantization. It is the quadratic quantization constant, that is Secondary quantization acts on the primary quantization constant, which can further reduce memory usage.

[0060] Step 2-3-2: Add a bypass to the quantization model bypass by introducing a low-rank decomposition matrix. and Initialization is performed using a random Gaussian matrix and a zero matrix respectively, ensuring that the initial state fits the parameter matrix. This does not affect the final output. The quantized parameters are frozen during training. Update only adapter parameters The specific update can be expressed as equation (3).

[0061] (3) The rank of the decomposition matrix .

[0062] Step 2-3-3: Evaluate the model performance and save the model file by combining objective quantitative indicators (validation loss, precision and recall indicators) and subjective human scoring.

[0063] Step 3: Standard Knowledge Extraction Sentence: The power grid is a part composed of transformers, power lines and other equipment that transform, transmit and distribute electrical energy.

[0064] Hint template: "Based on the given sentence {The power grid is a part composed of transformers, power lines and other equipment for transforming, transmitting and distributing electrical energy}, please determine whether the sentence belongs to the {} type, where the type includes {equipment, fault, operation and maintenance, dispatch}.

[0065] Based on the given sentence, it belongs to the "equipment" category. This is because the sentence mentions that the power grid is composed of equipment such as transformers and power lines that transform, transmit, and distribute electrical energy; these devices are important components of the power grid.

[0066] Entity Attribute Extraction Template: "The input sentence {The power grid is a part composed of transformers, power lines, and other equipment for transforming, transmitting, and distributing electrical energy} belongs to the {Equipment} type. Please extract all entities and entity types from it and output them in the format <Entity, Entity Type>, where Entity Type = {PER, ORG, PEM, ET}, where..." <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>"Indicates electrical technology."

[0067] Based on the sentence you provided, here are the entities it contains and their corresponding entity types: <Power grid, power equipment (PEM)>; Transformers, Power Equipment (PEM); <Power lines, electrical equipment (PEM)>; <Transformation, Electrical Technology (ET)> <Transmission, Electricity Technology (ET)>; <Distribution, Electric Technology (ET)>.

[0068] Results: <Power grid, PEM>, <Transformer, PEM>, <Power line, PEM>, <Transformation, ET>, <Transmission, ET>, and <Distribution, ET>.

[0069] Example 3 Please see Figure 3 As shown, this invention provides a method for extracting power standard knowledge based on a large language model, including: S100, Obtain the power standard document to be extracted; S200. Extract the target sentence from the power standard file to be extracted, and fill the prompt template based on the target sentence; input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type; fill the entity attribute extraction template based on the target sentence and the target sentence type; input the entity attribute extraction template into the fine-tuned large oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power field prompt template, the power field labeled dataset is transformed into a three-element fine-tuning dataset of "command prompt → text input → standard output"; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

[0070] In one specific implementation, the prompt template is specifically "Based on the given sentence {}, please determine whether the sentence belongs to type {}, where the types all include ={equipment, fault, operation and maintenance, scheduling}".

[0071] In one specific implementation, the entity attribute extraction template is: "The input sentence {} belongs to type {}. Please extract all entities and entity types from it and output them in the format <entity, entity type>, where entity type = {PER, ORG, PEM, ET}, where..." <per>Indicates a person's name. <org>Indicates the name of the organization. <pem>Indicates electrical equipment, <et>"Indicates electrical technology."

[0072] In one specific embodiment, it further includes: Based on the entities contained in the extracted target sentences, the entity linking method is used to check whether the entity exists in the preset power knowledge base. If it exists, the result is added to the candidate set; if it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuning big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0073] In one specific implementation, the step of searching for the existence of entities in a preset power knowledge base based on the entities contained in the extracted target sentence using an entity linking method specifically includes: Based on the extraction results of the fine-tuned big oracle model, a similarity matching method is used to search whether the entity exists in the preset power knowledge base. If the entity exists in the preset power knowledge base and the similarity score is higher than the set threshold, the entity is added to the candidate set. If it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0074] In one specific implementation, the method further includes: based on the candidate set, updating entities that do not exist in the preset power knowledge base to the preset power knowledge base, and outputting the information extraction results of the sentence in a preset format.

[0075] In one specific implementation, the quantization low-rank adaptive method specifically includes: A quantized model is obtained by quantizing a large language model using 4-bit normal floating-point numbers and double quantization. Low-rank decomposition matrices A and B are introduced as a bypass in the quantization model, and the quantized parameters... Update only adapter parameters ,in The rank of the decomposition matrix is ​​the weight increment. d and k are the row and column dimensions of the weight matrix of the large language model, respectively; The model performance is evaluated by combining objective indicators and subjective scoring, and the qualified models are saved as fine-tuning large prediction models.

[0076] Example 4 Please see Figure 4 As shown, this invention provides a power standard knowledge extraction device based on a large language model, comprising: The acquisition module is used to acquire the power standard file to be extracted; The extraction module is used to extract the target sentence from the power standard file to be extracted, fill the prompt template based on the target sentence; input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type; fill the entity attribute extraction template based on the target sentence and the target sentence type; input the entity attribute extraction template into the fine-tuned large oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power field prompt template, the power field labeled dataset is transformed into a three-element fine-tuning dataset of "command prompt → text input → standard output"; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

[0077] In one specific embodiment, it further includes: The judgment module is used to check whether the entity contained in the extracted target sentence exists in the preset power knowledge base through the entity linking method. If it exists, the result is added to the candidate set; if it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuning big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

[0078] In one specific embodiment, it further includes: The update module is used to update entities that are not in the preset power knowledge base to the preset power knowledge base based on the candidate set, and output the information extraction results of the sentence in a preset format.

[0079] Example 5 Please see Figure 5 As shown, this embodiment of the invention provides an electronic device 100 for implementing a power standard knowledge extraction method based on a large language model; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0080] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the above-mentioned method for extracting power standard knowledge based on a large language model by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0081] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0082] The memory 101 in the electronic device 100 stores multiple instructions to implement a power standard knowledge extraction method based on a large language model, and the processor 102 can execute the multiple instructions to achieve the following: Used to obtain the power standard documents to be extracted; The method is used to extract the target sentence from the power standard file to be extracted, fill the prompt template based on the target sentence; input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type; fill the entity attribute extraction template based on the target sentence and the target sentence type; input the entity attribute extraction template into the fine-tuned large oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power field prompt template, the power field labeled dataset is transformed into a three-element fine-tuning dataset of "command prompt → text input → standard output"; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

[0083] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0084] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.< / et> < / pem> < / org> < / per> < / et> < / pem> < / org> < / per> < / et> < / pem> < / org> < / per> < / et> < / pem> < / org> < / per> < / et> < / pem> < / org> < / per> < / et> < / pem> < / org> < / per>

Claims

1. A method for extracting power standard knowledge based on a large language model, characterized in that, include: Obtain the power standard documents to be extracted; Extract the target sentence from the power standard file to be extracted, and fill the prompt template based on the target sentence; The filled-in prompt template is input into the fine-tuned model of the large language model to obtain the target sentence type; Based on the target sentence and target sentence type, fill in the entity attribute extraction template; input the entity attribute extraction template into the fine-tuned big oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power domain prompt template, the power domain labeled dataset is transformed into a three-element fine-tuning dataset with instruction prompts → text input → standard output; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

2. The method for extracting power standard knowledge based on a large language model according to claim 1, characterized in that, Also includes: Based on the entities contained in the extracted target sentences, the entity linking method is used to check whether the entities exist in the preset power knowledge base. If they exist, the results are added to the candidate set. If it does not exist, the design result verification prompt template will ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it will be added to the candidate set; otherwise, the entity will be discarded.

3. The method for extracting power standard knowledge based on a large language model according to claim 2, characterized in that, The step of searching for the existence of entities in a preset power knowledge base based on the entities contained in the extracted target sentence using an entity linking method specifically includes: Based on the extraction results of the fine-tuned big oracle model, a similarity matching method is used to search whether the entity exists in the preset power knowledge base. If the entity exists in the preset power knowledge base and the similarity score is higher than the set threshold, the entity is added to the candidate set. If it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

4. The method for extracting power standard knowledge based on a large language model according to claim 2, characterized in that, Also includes: Based on the candidate set, entities that do not exist in the preset power knowledge base are updated to the preset power knowledge base, and the information extraction results of the sentences are output in a preset format.

5. The method for extracting power standard knowledge based on a large language model according to claim 1, characterized in that, The quantization low-rank adaptive method specifically includes: A quantized model is obtained by quantizing a large language model using 4-bit normal floating-point numbers and double quantization. Low-rank decomposition matrices A and B are introduced as a bypass in the quantization model, and the quantized parameters... Update only adapter parameters ,in The weight increment, the rank of the decomposition matrix d and k are the row and column dimensions of the weight matrix of the large language model, respectively; The model performance is evaluated by combining objective indicators and subjective scoring, and the qualified models are saved as fine-tuning large prediction models.

6. A power standard knowledge extraction device based on a large language model, characterized in that, include: The acquisition module is used to acquire the power standard file to be extracted; The extraction module is used to extract the target sentence from the power standard file to be extracted, fill the prompt template based on the target sentence, and input the filled prompt template into the fine-tuned model of the large language model to obtain the target sentence type. Based on the target sentence and target sentence type, fill in the entity attribute extraction template; input the entity attribute extraction template into the fine-tuned big oracle model to extract the entities contained in the target sentence; The fine-tuned large oracle model is obtained through the following steps: A power industry annotation dataset is constructed, which includes three types of annotation task data: entities, relations, and events. Using a preset power domain prompt template, the power domain labeled dataset is transformed into a three-element fine-tuning dataset with instruction prompts → text input → standard output; A quantized low-rank adaptive method is used to perform supervised fine-tuning of a large language model, and the resulting fine-tuned large oracle model is evaluated and saved.

7. The power standard knowledge extraction device based on a large language model according to claim 6, characterized in that, Also includes: The judgment module is used to check whether the entity contained in the extracted target sentence exists in the preset power knowledge base through the entity linking method. If it exists, the result is added to the candidate set; if it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuning big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

8. The power standard knowledge extraction device based on a large language model according to claim 7, characterized in that, The step of searching for the existence of entities in a preset power knowledge base based on the entities contained in the extracted target sentence using an entity linking method specifically includes: Based on the extraction results of the fine-tuned big oracle model, a similarity matching method is used to search whether the entity exists in the preset power knowledge base. If the entity exists in the preset power knowledge base and the similarity score is higher than the set threshold, the entity is added to the candidate set. If it does not exist, a result verification prompt template is designed to ask whether the entity extracted by the fine-tuned big oracle model is an entity. If it is an entity, it is added to the candidate set; otherwise, the entity is discarded.

9. The power standard knowledge extraction device based on a large language model according to claim 7, characterized in that, Also includes: The update module is used to update entities that are not in the preset power knowledge base to the preset power knowledge base based on the candidate set, and output the information extraction results of the sentence in a preset format.

10. A power standard knowledge extraction device based on a large language model according to claim 6, characterized in that, The quantization low-rank adaptive method specifically includes: A quantized model is obtained by quantizing a large language model using 4-bit normal floating-point numbers and double quantization. Low-rank decomposition matrices A and B are introduced as a bypass in the quantization model, and the quantized parameters... Update only adapter parameters ,in The weight increment, the rank of the decomposition matrix d and k are the row and column dimensions of the weight matrix of the large language model, respectively; The model performance is evaluated by combining objective indicators and subjective scoring, and the qualified models are saved as fine-tuning large prediction models.

11. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement a power standard knowledge extraction method based on a large language model as described in any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements a power standard knowledge extraction method based on a large language model as described in any one of claims 1 to 5.

Images