Power industry standard provisions search method and system based on semantic understanding

Abstract

A power industry standard provisions search method and system based on semantic understanding. The method comprises: collecting standard provisions from various professions in the power industry, and pre-processing the collected standard provisions; constructing a deep learning model capable of understanding power industry terms and contextual meaning; performing semantic annotation on the pre-processed power industry standard provisions, and on the basis of the annotation result, constructing an index for the power industry standard provisions; using the deep learning model to perform semantic understanding on a search condition, and querying matching power industry standard provisions; and sorting and displaying the searched standard provisions according to relevance and matching degrees. Compared with previous semantic annotation methods only based on dictionary word segmentation, the semantic annotation method based on semantic understanding realizes a deeper level of semantic annotation, such that annotation is performed on the basis of a certain understanding of the concepts of power industry terms, and the deeper level of semantic annotation provides a foundation for subsequent search and query.

Description

Power industry standard clause search method and system based on semantic understanding

[0001] Cross-reference to Related Applications

[0002] The present application claims priority to the Chinese patent application No. 202410450051.8, filed on April 15, 2024, and entitled "Power industry standard clause search method and system based on semantic understanding", the entire content of which is incorporated herein by reference. TECHNICAL FIELD

[0003] The present application belongs to the field of power industry standard application and implementation, and particularly relates to a power industry standard clause search method and system based on semantic understanding. BACKGROUND

[0004] The power industry standard clauses have always been a series of specifications for technical personnel, involving scattered and miscellaneous content, which is dry and not conducive to memory. Early power industry standards mostly exist in paper publications, although paper materials are convenient for professional personnel to read at any time, but as the number of standard clauses increases, the update speed is also faster and faster, and the scope of involvement is also wider and wider, it becomes more and more difficult for professional personnel to quickly search for target clauses in the massive clauses. With the development of information technology, power industry standard clauses are stored in computers in the form of strings, and there are related technologies for electronic search of massive clauses, mainly in the form of string matching search, but due to the simple search method, the search result is considered as a successful match as long as it contains the search keyword, and the result has no relevance and matching degree sorting, which leads to the fact that this search method cannot achieve the user's goal.

[0005] SUMMARY

[0006] The purpose of the present application is to provide a power industry standard clause search method and system based on semantic understanding, to solve the problem that the string matching form search has no relevance and matching degree sorting in the result, which leads to the fact that this search method cannot achieve the user's goal.

[0007] To achieve the above-mentioned purpose, the present application adopts the following technical solution:

[0008] The power industry standard clause search method based on semantic understanding comprises:

[0009] Collecting standard clauses of each professional in the power industry, and preprocessing the collected standard clauses;

[0010] Constructing a deep learning model capable of understanding power industry terminology and contextual meaning;

[0011] Semantic annotation is performed on the processed power industry standard clauses, and an index of the power industry standard clauses is constructed based on the annotation results;

[0012] A deep learning model is used to understand the semantics of the search conditions, extract key information and key semantic tags, search in the index according to the key information and key semantic tags, and query the matched power industry standard clauses;

[0013] The searched standard clauses are sorted according to relevance and matching degree and displayed.

[0014] Optionally, standard clauses of various professions in the power industry are collected:

[0015] including national standards, industry standards, national policy documents, professional books, equipment manuals, operation procedures and maintenance procedures.

[0016] Optionally, the clause data is preprocessed, including data cleaning, word segmentation, noise removal, part-of-speech tagging, entity recognition, text vectorization, sentence boundary detection, and feature extraction.

[0017] Specifically, data cleaning includes removing special symbols and stop words; word segmentation is to cut sentences into word groups, conjunctions or phrases; noise removal is to remove errors in the clauses; part-of-speech tagging is to assign a part-of-speech tag to each word group; entity recognition is to identify entities in the text and replace them with the same representation; text vectorization is to convert Chinese text into machine code for easy machine deep learning; sentence boundary detection is to identify the sentence boundaries of paragraphs and segment them into independent sentences; and feature extraction is to extract key feature words of sentences.

[0018] Optionally, a deep learning model capable of understanding power industry terminology and contextual meaning is constructed:

[0019] Using a combination of machine learning and natural language processing techniques, using the InstructGPT model based on the Transformer architecture, training the InstructGPT model with a large amount of power industry standard clauses to learn the structure and semantics of the standard clauses; and using reinforcement learning RLHF based on human feedback to continuously fine-tune the pre-trained language model LLM;

[0020] Specifically, the InstructGPT model is trained, a portion of the search conditions submitted by the test user is randomly extracted, and then answers are given for these search conditions, and then the <condition, answer> data is used to Fine-tune the InstructGPT model to make it initially have the ability to understand the intent contained in the human search conditions and give relatively high-quality answers according to this intent.

[0021] Train the InstructGPT model, manually annotate the training data to train the reward model, randomly sample the search conditions submitted by the user, and use the above fine-tuned model to generate K different answers, forming <condition, answer1> <condition, answer2> … <condition, answerK> data; then, the annotators sort the K results according to relevance, information, and harmful information standards to generate sorting result data; next, use this sorting result data for pair-wise learning to rank training mode to train the reward model.

[0022] Train the InstructGPT model to generate candidate answers, and then use semantic reasoning techniques to verify and filter these answers.

[0023] Optionally, the processed power industry standard provisions are semantically annotated:

[0024] Semantic annotation of power industry standard provisions means that the computer uses dictionaries and machine learning algorithm models to annotate power industry standard provisions; a pre-defined label set is used to annotate the provisions;

[0025] The annotation results are used to construct an index of power industry standard provisions:

[0026] Based on the annotation results of the power industry standard provisions, an index of power industry standard provisions is constructed; using this index, each provision is associated with a semantic label;

[0027] Specifically, the index is a data structure that sorts the values of the provisions in the database and stores pointers to the corresponding provisions in the database, recording the occurrence position of the word group; when the standard provision data is updated, the index needs to be automatically updated to maintain consistency between the index and the standard provision data.

[0028] Optionally, a deep learning model is used to understand the semantics of the search conditions, extract key information and key semantic labels, and search for matching power industry standard provisions based on the key information and key semantic labels in the index:

[0029] The user's input search conditions are semantically understood using a deep learning model, and the search is performed based on the synonym relationship and the dependency relationship between the search conditions.

[0030] Optionally, search result display and sorting means sorting the searched standard provisions according to relevance and matching degree and displaying them to the user, while providing the source and publication date of the provisions.

[0031] Specifically, the search results are intervened according to the search of the user input, the relevance of the criterion is judged according to the number of keywords searched, and the sorting is performed according to the relevance and matching degree; the attention of the user is tracked according to the click viewing situation of the user on the search results, the accuracy of the search results is fed back, and the sorting strategy and the deep learning model are optimized.

[0032] In a second aspect, the application provides a power industry standard clause search system based on semantic understanding, comprising:

[0033] A data acquisition module is configured to acquire standard clauses of various professions in the power industry and pre-process the acquired standard clauses;

[0034] A model construction module is configured to construct a deep learning model capable of understanding power industry terms and contextual meanings;

[0035] An index construction module is configured to perform semantic annotation on the pre-processed power industry standard clauses, and construct an index of the power industry standard clauses based on the annotation results;

[0036] A matching display module is configured to perform semantic understanding on a search condition by using the deep learning model, extract key information and key semantic tags, search in the index based on the key information and key semantic tags, query the matched power industry standard clauses, sort the searched standard clauses according to the relevance and matching degree, and display the sorted standard clauses.

[0037] In a third aspect, the application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the power industry standard clause search method based on semantic understanding when executing the computer program.

[0038] In a fourth aspect, the application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program implements the steps of the power industry standard clause search method based on semantic understanding when executed by a processor.

[0039] Compared with the prior art, the application has the following technical effects:

[0040] The application constructs a deep learning model capable of understanding power industry terms and contextual meanings, trains the model by using the collected massive standard clause data, enables the model to achieve semantic understanding, performs semantic annotation on the power industry standard clauses by using the deep learning model, and realizes deeper semantic annotation than the previous semantic annotation method based on only dictionary word segmentation, so that the annotation is performed based on a certain understanding of the power industry term concept, and the deeper semantic annotation provides a basis for subsequent search and query.

[0041] In the user search and query stage, the application uses a deep learning model to understand the user's search conditions semantically, deeply understands the user's intention, makes the search range more specific and possible, and makes the search result more accurate. The previous search method for simple text matching often cannot find the result or the result is not what the user wants. The application can make the search tag range method specific through the subordinate relationship, accurately through the professional field, and make the search engine more like a professional technical expert to help users filter the desired results.

[0042] The deep learning model constructed by the application to understand the power industry terms and context meanings is a combination of machine learning and natural language processing technology (NLP), which improves the accuracy and efficiency of the deep learning model. The deep learning model is guided by the supervision strategy and artificial annotation training data strategy, so that the trained model is more intelligent.

[0043] The dictionary and index constructed by the application have a self-learning function, which can automatically update the dictionary and index content when the data is updated.

[0044] The search result obtained by the application can track the user's attention according to the user's click view of the search result, feedback the accuracy of the search result, and optimize the sorting strategy and the deep learning model. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flowchart of the application.

[0046] Figure 2 is a logic block diagram of the application.

[0047] Figure 3 is a schematic diagram of an embodiment of the application. DETAILED DESCRIPTION

[0048] The following describes the application with reference to the accompanying drawings:

[0049] Please refer to Figures 1 to 3, the purpose of the application is to provide a power industry standard clause search method based on semantic understanding, which provides more accurate and comprehensive search results according to the search conditions of professional technicians.

[0050] The power industry standard clause search method based on semantic understanding includes:

[0051] Collect standard clauses of various professions in the power industry, and preprocess the collected standard clauses;

[0052] Construct a deep learning model that can understand the power industry terms and context meanings;

[0053] Semantic labeling is performed on the processed electric power industry standard clauses, and an index of the electric power industry standard clauses is constructed based on the labeling results;

[0054] A deep learning model is used to understand the semantics of the search conditions, extract key information and key semantic tags, search in the index according to the key information and key semantic tags, and query the matched electric power industry standard clauses;

[0055] The searched standard clauses are sorted according to relevance and matching degree, and displayed.

[0056] Specifically:

[0057] S1, standard clause collection, collecting standard clauses of various professions in the electric power industry, including but not limited to national standards, industry standards, national policy documents, professional books, equipment manuals, operation procedures, maintenance procedures, etc.; the purpose is to collect data covering all aspects of various professions in the electric power industry, providing a data foundation for subsequent machine deep learning.

[0058] S2, preprocessing of clause data, preprocessing the standard clauses collected in S1, including but not limited to data cleaning, word segmentation, noise removal, part-of-speech tagging, entity recognition, text vectorization, sentence boundary detection, feature extraction, etc.

[0059] Specifically, data cleaning in step S2 includes removing special symbols (such as * &! and other useless characters) and removing stop words (de, le, etc.); word segmentation in step S2 is to cut sentences into word groups, conjunctions or phrases, etc. using a dictionary that should include all aspects of various professions in the electric power industry and is a self-learning and growing dictionary; noise removal in step S2 refers to removing errors in the clauses; part-of-speech tagging in step S2 refers to assigning a part-of-speech tag (such as noun, verb, relational word, adjective, etc.) to each word group, which helps to understand the context, especially the relational words commonly used in standard clauses such as "should", "should", "can" and "strictly prohibited"; entity recognition in step S2 refers to identifying entities in the text, such as device nouns, system names, method names, and regulation names, and replacing them with the same representation, similar to synonyms; text vectorization in step S2 refers to converting Chinese text into machine code for easy machine deep learning; sentence boundary detection in step S2 refers to identifying the sentence boundaries of paragraphs and dividing them into independent sentences; feature extraction in step S2 refers to extracting key features of sentences, such as key words (device nouns, system names), relational words, and ownership relationships.

[0060] S3, build a deep learning model that can understand the terms and context of the power industry, use machine learning and natural language processing technology (NLP) to improve the accuracy and efficiency of the deep learning model. Use the InstructGPT model based on the Transformer architecture to train the InstructGPT model with a large amount of power industry standard provisions, learn the structure and semantics of the standard provisions. And use reinforcement learning based on human feedback (RLHF) to continuously fine-tune the pre-trained language model (LLM), the purpose is to make the model better understand the meaning of human commands and instructions. Deep learning model is the core content of semantic understanding, the feedback of the training result marks the good or bad of the deep learning model.

[0061] Specifically, in step S3, the InstructGPT model needs to be trained, and some supervision strategies need to be set to determine whether the generated content is a high-quality result. Randomly extract a part of the search conditions submitted by the test users, and then ask professional technicians to give high-quality answers to these search conditions. Then use these <condition, answer> data to fine-tune the InstructGPT model to make it initially have the ability to understand the intention contained in the human search conditions and give relatively high-quality answers according to this intention.

[0062] Specifically, in step S3, the InstructGPT model needs to be trained, and the training data needs to be manually annotated to train the reward model. Randomly sample the search conditions submitted by the user, and use the above fine-tuned model to generate K different answers, forming <condition, answer1> <condition, answer2> … <condition, answerK> data. Then, the annotator sorts K results according to relevance, information, and harmful information standards to generate sorting result data. Next, the researcher uses this sorting result data to train the reward model in the pair-wise learning to rank training mode.

[0063] Specifically, in step S3, the InstructGPT model can generate candidate answers itself, and then use semantic reasoning technology to verify and filter these answers to improve the accuracy of semantic understanding under semi-automatic conditions.

[0064] S4, semantic annotation of power industry standard clauses, refers to letting the computer automatically annotate the power industry standard clauses in S1 using dictionaries and machine learning algorithm models. The clauses are annotated using a predefined set of labels, which can be entity labels (such as device name, system name, method name, etc.) or relationship labels (such as subordinate relationship, parallel relationship, exception relationship, etc.). The annotation of standard clauses focuses more on the understanding of the deep meaning of the clauses, trying to reveal the concepts, entities and relationships behind the clauses.

[0065] In particular, the power industry standard clauses are carefully written by professional technical experts and are valuable knowledge and experience, but they need to be understood and understood by personnel with certain professional technical foundation. In the annotation process in step S4, the relationship labels should not be annotated on the surface meaning level, but should be annotated deeply, for example, the clause "DCS control system power supply should be redundantly configured", if only "DCS control system power supply" should be redundantly configured, it is considered that this is a shallow semantic annotation, if it is annotated as "DCS control system power supply includes control cabinet power supply, network cabinet power supply, instrument transmitter power supply, relay cabinet power supply, independent device power supply...." should be redundantly configured, it is considered that this is a deep redundant power supply, after effective annotation of the clause, when the user searches "related standard clauses about MFT cabinet power supply should be redundantly configured", it can be accurately matched to the clause. The deep learning model in step S3 is the key to achieving deep annotation of clause annotation.

[0066] S5, constructing an index of power industry standard clauses, refers to constructing an index of power industry standard clauses based on the annotation results of power industry standard clauses in S4. The index is used to associate each clause with semantic labels, making it easier to search and query later. The index helps the system quickly locate the required clauses, thereby avoiding full database search and improving query performance.

[0067] Specifically, the index in step S5 is a data structure that sorts the values of the clauses in the database and stores pointers to the corresponding clauses in the database. It records the occurrence of word groups in which clauses and their positions. It should be noted that when the standard clause data is updated, the index needs to be automatically updated to maintain consistency between the index and the standard clause data.

[0068] S6, user search and query, refers to when the user inputs search conditions, the system first preprocesses the user input search conditions, then uses the deep learning model in S3 to understand the semantics of the search conditions and extract key information and key semantic labels. Then the system searches the index in S5 according to the key information and key semantic labels to query the matched power industry standard clauses.

[0069] Specifically, the semantic understanding of the search condition input by the user in step S6 using the deep learning model is the key to accurate search. For example, when the user searches for "relevant standard provisions about redundant configuration of power supply of MFT cabinet", if the deep learning model is not used, the text is simply segmented into "MFT cabinet", "power supply" and "redundant configuration", and then these three semantic tags are used to search in the index, the searched provisions may be few or even not found. Because the power industry standard provisions are provisions and specifications summarized by many professional and technical experts after deep thinking, the integration is high, and professional knowledge is required. The words used in the process of standard writing are general and not detailed. The deep learning model in S3 can understand the semantics of the search condition input by the user. "MFT cabinet" is a device entity, which belongs to "DCS control system" and "FSSS control hardware", and the near-synonymous word is "MFT control cabinet". The near-synonymous word of "redundant configuration" is "two-way". Using these semantic tags to search in the index can search for relevant standard provisions with a wider range, and the number of provisions and the relevance are higher.

[0070] S7, search result display and sorting refers to sorting the searched standard provisions according to relevance and matching degree, and displaying them to the user. At the same time, some additional information such as the source of the provisions and the publication date can be provided to help the user better understand and use the standard provisions.

[0071] Specifically, in step S7, the search results can be intervened according to the search profession input by the user. For example, the above search is a problem in the thermal engineering profession. If all the keywords are searched in the thermal engineering profession entry, it is considered that the relevance of this standard is the highest (T1); if the keywords are searched in the thermal engineering profession entry, the matching degree is determined to determine the relevance (T2). If all the keywords are searched in other non-thermal engineering profession entries, it is considered that the relevance of this standard is higher (T3); if part of the keywords is searched in other non-thermal engineering profession entries, the matching degree is determined to determine the relevance (T4). According to the relevance and matching degree, the sorting is T1>T3>T2>T4.

[0072] Specifically, the accuracy of the search results can be fed back according to the clicking and viewing of the user on the search results, the attention of the user is tracked, and the sorting strategy and the deep learning model are optimized.

[0073] The application will be described in detail below in combination with FIG. 3 and through embodiments. The following embodiments are an explanation of the application, and the application is not limited to the following embodiments.

[0074] Collection standard clauses refer to collecting standard clauses of various professions in the power industry, including but not limited to national standards, industry standards, national policy documents, professional books, equipment manuals, operation procedures, maintenance procedures, etc.; the purpose is to comprehensively cover all aspects of various professions in the power industry, and to provide a data foundation for subsequent machine deep learning.

[0075] The standard clauses are processed to generate processed standard clause data after clause data preprocessing. Including but not limited to data cleaning, word segmentation, noise removal, part-of-speech tagging, entity recognition, text vectorization, sentence boundary detection, feature extraction, etc.

[0076] The processed standard clause data is used to train a deep learning model, the purpose is to build a deep learning model that can understand the terminology and context of the power industry, and the training process needs to use a supervised strategy and manual annotation method to make the training result more accurate.

[0077] The processed standard clause data is annotated using a deep learning model and a dictionary, and the annotated standard clause is generated. Then, based on the annotation results, an index library of power industry standard clauses is constructed. Semantic annotation of standard clauses focuses more on understanding the deep meaning of the clauses, and tries to reveal the concepts, entities and relationships behind the clauses. For example, the clause "DCS control system power supply should be redundantly configured" is annotated. If only "DCS control system power supply" should be redundantly configured, it is considered to be a shallow semantic annotation. If it is annotated as "DCS control system power supply includes control cabinet power supply, network cabinet power supply, instrument transmitter power supply, relay cabinet power supply, independent device power supply...." should be redundantly configured, it is considered to be a deep-level redundant power supply. After effective annotation of the clause, when the user searches for "related standard clauses about MFT cabinet power supply should be redundantly configured", the clause can be accurately matched.

[0078] The user searches for "related standard clauses about MFT cabinet power supply should be redundantly configured", and the data preprocessing module performs data preprocessing on the input condition. The processed search condition is "MFT cabinet / MFT control cabinet" + "power supply" and "redundant configuration". Next, the deep learning model is used to understand the semantics of the processed search condition, extract keywords and key labels, and obtain the search condition as "MFT cabinet / MFT control cabinet / DCS control system / FSSS control hardware" + "power supply" + "redundant configuration / two-way". It can be seen that the search condition processed by the deep learning model has a wider search range and more specific search content than the search condition processed without the deep learning model. The following compares the differences between search results with and without deep learning model processing:

[0079] According to the search condition without deep learning model processing, the following search results are obtained by searching the standard clause library (only part of the display):

[0080] 9.2.2 The controllers of the distributed control system, the system power supply, the DC power supply for the signal input / output (I / O) module, the communication network (including the field bus form), etc. should adopt completely independent redundant configuration and have non-disturbance switching function. The redundant communication network should have intercommunication function.

[0081] From Twenty-five Key Requirements for Preventing Power Production Accidents (2023 Edition)

[0082] 6.5.1.1 The power supply system redundancy configuration should meet the following requirements:

[0083] a) Distributed control system power supply redundancy configuration:

[0084] 1) The distributed control system should be equipped with reliable two independent power supply sources, and the priority is to consider two independent UPS power sources that can meet the control system capacity needs: during normal operation, each carries 1 / 2 load and works at the same time. Secondly, a UPS and a security power supply are used, and during normal operation, the UPS power supply and the protection power supply supply power to the power supply module respectively: if the security power supply fluctuates greatly, a voltage stabilizer should be added to stabilize the voltage.

[0085] From DL / T 261-2022 Technical Guide for Reliability Evaluation of Thermal Power Plant Automation System

[0086] 6.2.1.1 The controller redundancy and distributed configuration should meet the following requirements:

[0087] a) The controllers of the main control systems such as unit DCS, DEH, MEH, desulfurization, denitrification, and peripheral auxiliary control should be redundantly configured separately, strictly follow the principle of independence of unit important protection and control, and meet the requirements of dispersion.

[0088] From DL / T 261-2022 Technical Guide for Reliability Evaluation of Thermal Power Plant Automation System

[0089] 6.2.2.1 The communication network redundancy and fault tolerance configuration should meet the following requirements:

[0090] a) The network switching equipment (communication interface or communication module) of the main control communication and I / O communication of each main control system such as unit DCS, DEH, desulfurization, and peripheral auxiliary control should select a level equipment and redundantly configure.

[0091] From DL / T 261-2022 Technical Guide for Reliability Evaluation of Thermal Power Plant Automation System

[0092] From the search results, there are few results that exactly match the three keywords "MFT cabinet" + "power supply" + "redundant configuration", most of which are search results for "redundant configuration", followed by search results for "power supply" + "redundant configuration".

[0093] According to the search conditions processed by the deep learning model, the search index library obtains the following search results (only part of them are shown):

[0094] 6.5.1 Power system reliability evaluation

[0095] 6.5.1.1 The power system redundancy configuration should meet the following requirements:

[0096] a) Distributed control system power supply redundancy configuration:

[0097] 1) The distributed control system should be equipped with reliable two independent power supply, priority is given to two independent operation to meet the control system capacity needs of UPS power supply.

[0098] From DL / T261-2022 "Reliability Evaluation Technology Guide for Thermal Automation System of Thermal Power Plant"

[0099] 3.4.3 The power supply of the control and protection system should meet the following requirements:

[0100] 1 The unit distributed control system, steam turbine digital electro-hydraulic control system should have two power supplies, one of which should be from the AC uninterruptible power supply, the other can be from the AC security power supply or the second AC uninterruptible power supply. When the steam turbine digital electro-hydraulic control system and the unit distributed control system use the same hardware, the power supply system can also be unified.

[0101] From DL / T 5455-2012 "Design Technical Regulations for Thermal Power Plant Thermal Power Supply and Gas Supply System"

[0102] 3.1 The configuration of operator station, engineer station, real-time data server and communication network should meet the following requirements:

[0103] a) The operator station, server, communication network and power supply in the distributed control system should adopt reliable redundant configuration.

[0104] From "Reliability Configuration and Accident Pre-control of Thermal Power Plant Thermal Control System"

[0105] 9.1.10 The power supply of important control systems independent of DCS (such as main fuel trip (MFT) control cabinet, emergency trip system (ETS) power supply cabinet, turbine monitoring instrument system (TSI) etc.) should be redundantly configured, and power failure sound and light alarm should be set.

[0106] Twenty-five Key Requirements for Preventing Power Production Accidents (2023 Edition)

[0107] From the search results, the search condition "MFT cabinet / MFT control cabinet / DCS control system / FSSS control hardware" + "power supply" + "redundant configuration / two-way" after semantic understanding by the deep learning model has a wider search range, more search results, and more useful matched results.

[0108] It can be seen that the search results of the method of the present application can better understand the user's intention and are the search results that the user wants.

[0109] Optionally, the user's attention is tracked, and the deep learning model is fed back according to the click frequency and labeling of the user on the search results for model optimization.

[0110] In another embodiment of the present application, a power industry standard clause search system based on semantic understanding is provided, which can be used to implement the power industry standard clause search method based on semantic understanding described above. Specifically, the system comprises:

[0111] A data acquisition module is configured to acquire standard clauses of various professions in the power industry and pre-process the acquired standard clauses.

[0112] A model construction module is configured to construct a deep learning model capable of understanding power industry terminology and contextual meaning.

[0113] An index construction module is configured to perform semantic labeling on the pre-processed power industry standard clauses and construct an index of the power industry standard clauses based on the labeling results.

[0114] A matching display module is configured to perform semantic understanding on a search condition by using the deep learning model, extract key information and key semantic labels, search in the index based on the key information and key semantic labels, query matched power industry standard clauses, sort the searched standard clauses according to relevance and matching degree, and display the sorted standard clauses.

[0115] The division of modules in the embodiments of the present application is illustrative, and is only a logical functional division. In actual implementation, another division manner can be used. In addition, each functional module in each embodiment of the present application can be integrated in one processor, or can be physically separated, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module.

[0116] In another embodiment of the present application, a computer device is provided, which comprises a processor and a memory, the memory is configured to store a computer program, the computer program comprises program instructions, and the processor is configured to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method flow or a corresponding function; the processor in the embodiments of the present application can be used for the operation of the power industry standard clause search method based on semantic understanding.

[0117] In another embodiment of the present application, the present application further provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in a computer device, and is configured to store programs and data. It can be understood that the computer readable storage medium herein can include an internal storage medium in the computer device, and of course can also include an expansion storage medium supported by the computer device. The computer readable storage medium provides a storage space, which stores an operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the power industry standard clause search method based on semantic understanding in the above embodiments.

[0118] Those skilled in the art will appreciate that embodiments of the application can be readily used as a method, a system or a computer program product. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0119] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in combination with the flowchart block or blocks.

[0120] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks or combination thereof.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in combination with the flowchart block or blocks.

[0122] Finally, it should be noted that the above-described embodiments are merely given as examples of the application and are not intended to limit the present application in any manner. Although the present application has been described in detail with reference to the above embodiments, it should be understood that the inventive concept can be implemented in many other ways. Therefore, modifications and / or additions to the above-described embodiments can occur to those skilled in the art without departing from the spirit and scope of the application. Accordingly, the scope of the application should be determined by the following claims and their legal equivalents rather than by the above description.

Claims

1. A method for searching power industry standard clauses based on semantic understanding, characterized in that, Comprise: Collect standard clauses of various professions in the power industry, and preprocess the collected standard clauses; Build a deep learning model that can understand the terminology and contextual meaning of the power industry; Semantic annotation of processed power industry standard clauses, and construction of power industry standard clause index based on annotation results; Use the deep learning model to understand the semantics of the search conditions, extract key information and key semantic tags, and search in the index according to the key information and key semantic tags to query the matching power industry standard clauses; Sort the searched standard clauses according to relevance and matching degree and display. 2.The method of claim 1, wherein, Collect standard clauses of various professions in the power industry: Including national standards, industry standards, national policy documents, professional books, equipment manuals, operation procedures and maintenance procedures. 3.The method of claim 1, wherein, Preprocessing of clause data, including data cleaning, word segmentation, noise removal, part-of-speech tagging, entity recognition, text vectorization, sentence boundary detection, and feature extraction; Specifically, data cleaning includes removing special symbols and stop words; Word segmentation is to cut sentences into word groups, conjunctions or phrases; Noise removal is to remove errors in the text; part-of-speech tagging is to assign a part-of-speech tag to each word group; entity recognition refers to identifying entities in the text and replacing them with the same representation; text vectorization is to convert Chinese text into machine code for easy machine deep learning; Sentence boundary detection is to identify the sentence boundaries of paragraphs and divide them into independent sentences; Feature extraction is to extract key feature words of sentences. 4.The method of claim 1, wherein, Build a deep learning model that can understand the terminology and contextual meaning of the power industry: Using a combination of machine learning and natural language processing techniques, using the InstructGPT model based on the Transformer architecture, training the InstructGPT model with massive power industry standard clauses to learn the structure and semantics of the standard clauses; and using reinforcement learning RLHF based on human feedback to continuously fine-tune the pre-trained language model LLM; Specifically, train the InstructGPT model, randomly extract a portion of the search conditions submitted by the test user, then give answers to these search conditions, and then Fine-tune the InstructGPT model with these <condition, answer> data to make it initially have the ability to understand the intent contained in the human search conditions and give relatively high-quality answers according to this intent; Train the InstructGPT model, manually annotate the training data to train the reward model, randomly sample the search conditions submitted by the user, and use the above Fine-tuned model to generate K different answers, forming <condition, answer1> <condition, answer2> … <condition, answerK> data; Then, the annotators sort K results according to relevance, information and harmful information standards to generate sorting result data; Next, use this sorting result data to train the reward model in a pair-wise learning to rank training mode. ​ Train the InstructGPT model to generate candidate answers, and then use semantic reasoning techniques to verify and filter these answers. 5.The method of claim 1, wherein, Semantic annotation of processed power industry standard clauses: Semantic annotation of power industry standard clauses refers to using dictionaries and machine learning algorithm models to annotate power industry standard clauses; using a pre-defined label set to annotate clauses; Indexing of annotated results: Based on the annotation results of power industry standard clauses, an index of power industry standard clauses is constructed; using this index, each clause is associated with semantic labels; Specifically, the index is a data structure that sorts clauses in the database by their values and stores pointers to the corresponding clauses in the database, recording the occurrence position of the word group; when the standard clause data is updated, the index needs to be automatically updated to maintain consistency between the index and the standard clause data. 6.The method of claim 1, wherein, Semantic understanding of search conditions using deep learning models, extraction of key information and key semantic labels, search in the index based on key information and key semantic labels, and query matching power industry standard clauses: Semantic understanding of user input search conditions using deep learning models, analysis of their synonym relationships and search conditions for searching. 7.The semantic understanding based power industry standard clause search method according to claim 1, wherein, Search result display and sorting refers to sorting the searched standard clauses according to relevance and matching degree and displaying them to the user, while providing the source and publication date of the clauses; Specifically, according to the user's input search professional, the search results are intervened, the relevance of the standard is judged according to the number of keywords searched, and the sorting is performed according to relevance and matching degree; According to the user's click view of the search results, track the user's attention, feedback the accuracy of the search results, and optimize the sorting strategy and deep learning model.

8. A power industry standard clause search system based on semantic understanding, characterized by, It includes: Data collection module for collecting standard clauses of various professions in the power industry and preprocessing the collected standard clauses; Model construction module for constructing a deep learning model that can understand power industry terminology and contextual meaning; Index construction module for semantic annotation of processed power industry standard clauses, and indexing of annotated results of power industry standard clauses; Matching display module for semantic understanding of search conditions using deep learning models, extraction of key information and key semantic labels, search in the index based on key information and key semantic labels, and query matching power industry standard clauses; sorting and displaying the searched standard clauses according to relevance and matching degree.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the power industry standard clause search method based on semantic understanding according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the steps of the power industry standard clause search method based on semantic understanding according to any one of claims 1 to 7.

Images