Common term naturalization processing method and system based on substitute dimension large model
By building a terminology database and training a large model, we can achieve natural language to terminology parsing and interface calls, which solves the problem of inaccurate analysis of professional terms in outsourced maintenance work and improves the efficiency and accuracy of operation and maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-04-07
Smart Images
Figure CN121809656A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of artificial intelligence and operation and maintenance, in particular to a common term naturalization processing method and system based on a maintenance model. BACKGROUND
[0002] Current maintenance work relies on a large number of professional terms (such as "base station off service" and "transmission link error rate"). When trying to use a large model to call interface capabilities, the language and characters of the AI input cannot be accurately analyzed, and multiple operations and interactions are required. At the same time, the operation and maintenance personnel need to input instructions or call interfaces in the background system, rely on professional operation and maintenance personnel, and have low response efficiency.
[0003] In the prior art, the general NLP model has insufficient understanding of maintenance terms, and semantic deviation is easy to occur; the special operation and maintenance system is mostly command line or fixed form interactive, without natural language interface; and part of the maintenance knowledge base only provides term query, and cannot realize the linkage of demand analysis and interface calling. In addition, the group-level operation and maintenance guidance document is mostly in text format, and the knowledge has not been converted into structured data understandable by machines, so that the large model is difficult to directly learn the operation and maintenance specification, and cannot meet the whole-process automation demand of "natural language demand→term analysis→interface calling", which restricts the improvement of maintenance efficiency. SUMMARY
[0004] The purpose of the present application is to provide a common term naturalization processing method and system based on a maintenance large model to solve the problems raised in the background.
[0005] To achieve the above purpose, the present application provides the following technical scheme: a common term naturalization processing method based on a maintenance large model, comprising the following steps: Term library construction: collate industry maintenance terms, label term definitions, application scenarios and associated operations, form a complete term library, and provide basic term support for subsequent model training; Data acquisition and preprocessing: text data acquisition: crawling group operation and maintenance management guidance documents and historical operation and maintenance knowledge base, and extracting structured knowledge; corpus collection: recording dialog corpus of front-line maintenance personnel and non-professional personnel, and labeling "natural language demand→maintenance term→interface calling instruction" mapping relationship; data cleaning: removing duplicate and error data, and converting unstructured text into structured data in JSON format; Model construction and training: basic model selection: using a general pre-trained large model as a base; fine-tuning: inputting the preprocessed multi-source data into the model, and fine-tuning through "term prediction", "demand-term mapping" and "interface instruction generation" three tasks, so that the model has the ability to identify maintenance terms in natural language, convert natural language demand into maintenance terms, and generate background interface calling parameters based on the terms.
[0006] Preferably, in the model construction and training step, a model optimization step is further included: using small sample learning technology, learning from the collected scarce corpus to improve the model's processing ability for rare or new situations; at the same time, using knowledge distillation technology to compress the model size, so that the model adapts to edge devices, and improves the model's lightweight and generalization ability.
[0007] Preferably, after completing the model construction and training, an interaction and feedback step is further included: Interaction module building: building a natural language dialogue interface on the Web or APP side, and users input requirements through the interface; Interface adaptation: the model calls the pre-set background interface adaptation layer, and converts the instructions generated according to the user's requirements into a format recognizable by the interface to trigger the interface to perform corresponding operations; Result feedback: after the interface returns the data, the model converts the results into natural language and feeds them back to the user, realizing the complete interaction process of natural language and maintenance business operation and feedback.
[0008] Preferably, in the interaction module building step, the user's input requirements include but are not limited to device fault query, fault statistics, fault troubleshooting requirements, and various natural language requirements related to maintenance business.
[0009] Preferably, in the interface adaptation step, the format recognized by the interface is the JSON request body of RESTful API; in the result feedback step, the natural language results fed back to the user contain specific data information and related analysis and explanations, so that the user can clearly understand the maintenance business situation.
[0010] A common term naturalization processing system based on a maintenance big model, comprising: Data collection and preprocessing module: term library construction unit, used for sorting industry maintenance terms, annotating term definitions, application scenarios and related operations; text data collection unit, used for crawling group operation and maintenance management guidelines and historical operation and maintenance knowledge base, and extracting structured knowledge; corpus collection unit, used for recording dialog corpus of front-line maintenance personnel and non-professionals, and annotating "natural language requirements → maintenance terms → interface calling instructions" mapping relationship; data cleaning unit, used for removing duplicate and error data, and converting unstructured text into structured data in JSON format for subsequent model training; Model construction module: basic model selection unit, using a general pre-trained large model as a base; fine-tuning training unit, inputting preprocessed multi-source data into the model, and fine-tuning the model through three tasks of "term prediction", "demand-term mapping" and "interface instruction generation", so that the model has the ability to identify maintenance terms in natural language, convert natural language requirements into maintenance terms, and generate background interface calling parameters based on terms.
[0011] Preferably, the model construction module further comprises a model optimization unit: the unit adopts a small sample learning technique to learn from the collected scarce corpus, thereby improving the processing capability of the model for rare or newly emerging situations; and adopts a knowledge distillation technique to compress the model size, so that the model is adapted to edge devices and the lightness and generalization capability of the model are improved.
[0012] Preferably, the method further comprises an interaction and interface processing module, which comprises: an interaction unit: a Web or APP natural language dialogue interface is built, and a user inputs a demand through the interface; the demand includes but is not limited to device fault query, fault statistics, fault troubleshooting demand, and various natural language demands related to maintenance business; an interface adaptation unit: a preset background interface adaptation layer is called by the model, and an instruction generated according to the user demand is converted into a format recognizable by the interface to trigger the interface to perform a corresponding operation.
[0013] Preferably, the interaction and interface processing module further comprises a result feedback unit: when the interface returns data, the unit converts the result into a natural language and feeds it back to the user; the natural language feedback includes specific data information and related analysis and explanation, so that the user can clearly understand the maintenance business.
[0014] Preferably, in the "term prediction" task, the model has an identification accuracy of maintenance terms in natural language above a preset threshold; in the "demand-term mapping" task, the model has an accuracy of converting natural language demands into maintenance terms above a preset threshold; in the "interface instruction generation" task, the background interface calling parameters generated based on the terms accurately trigger the corresponding interface to perform a corresponding operation, and the success rate of the interface execution is above a preset threshold.
[0015] Compared with the prior art, the method and system have the following beneficial effects: The method and system based on the maintenance large model can initiate maintenance demands through daily language without learning system operation and instructions, thereby improving communication efficiency; the model automatically completes demand analysis and interface calling, thereby avoiding manual operation in the system and shortening the maintenance response time; the model is trained based on group guidance documents, thereby ensuring that the term understanding and interface calling conform to the specifications, reducing human operation errors (such as incorrect instruction input), and improving the fault processing accuracy; the model supports new maintenance scenarios (such as power maintenance and IT equipment maintenance), and only needs to supplement corresponding field data to fine-tune the model, without the need to reconstruct the system. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 The method flowchart of the application is shown in the figure. DETAILED DESCRIPTION
[0017] In order to make the purposes, technical solutions of the present application clear, complete and the advantages more clear and apparent, the embodiments of the present application are further described in detail below in combination with the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present application, rather than all the embodiments, and are only used to explain the embodiments of the present application, and do not limit the embodiments of the present application, and all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.
[0018] In an embodiment, the present application provides a technical solution: a common term naturalization processing method based on a maintenance model, through multi-source data collection and preprocessing, small model training of maintenance business, completing human-computer dialogue through natural language, realizing accurate demand analysis, interface adaptation and calling process, for example: querying RRU failure. Through natural language analysis and understanding, it is to query and analyze RRU failure, understand the demand, complete the issuance of RRU interface calling instruction, and finally complete the feedback of RRU failure state, display, etc., to quickly help maintenance personnel to understand the equipment.
[0019] For a maintenance business scenario, the following steps are mainly divided: a. Term library construction: sort out industry maintenance terms (such as communication maintenance "RRU failure" "core network element load"), mark term definition, application scenario and associated operation; b. Text data collection: crawling group operation and maintenance management guide documents (such as "Maintenance Work Specification" "Equipment Operation and Maintenance Manual"), historical operation and maintenance knowledge base (fault handling cases, operation logs), extracting structured knowledge (such as "fault phenomenon → processing steps → interface calling path"); c. Corpus collection: record the dialogue corpus of front-line maintenance personnel and non-professionals (such as customer service and operation and maintenance demand communication records), mark "natural language demand → maintenance term → interface calling instruction" mapping relationship; d. Data cleaning: remove duplicate and error data, convert unstructured text (such as case description) into JSON format structured data for model training.
[0020] e. Basic model selection: adopt a general pre-trained large model (such as BERT, LLaMA) as a base, because it has strong semantic understanding ability; f. Fine-tuning training: input pre-processed multi-source data into the model, fine-tune through the three tasks of "term prediction", "demand-term mapping", and "interface instruction generation": ① Term prediction task: let the model identify maintenance terms in natural language; ② Mapping task: train the model to convert natural language requirements (such as "What if the base station has no signal?") into maintenance terms (such as "Base station fault troubleshooting"); ③ Instruction generation task: generate background interface call parameters (such as interface address, request parameters) based on terms; g. Model optimization: use small sample learning technology (for scarce corpus) and knowledge distillation (compress model size, adapt to edge devices) to improve model lightweight and generalization ability.
[0021] h. Interaction module: build a natural language dialogue interface on the Web or APP side, and the user inputs requirements (such as "Statistics of transmission fault times this week"); i. Interface adaptation: the model calls the preset background interface adaptation layer to convert the instruction into a format recognizable by the interface (such as the JSON request body of RESTful API), and triggers the interface to execute (such as calling the fault statistics interface); j. Result feedback: after the interface returns the data, the model converts the result into natural language (such as "There were 12 transmission faults this week, mainly concentrated in the XX area"), and feeds back to the user.
[0022] In example two, based on example one, a common term naturalization processing system based on a maintenance big model is proposed, including: Data collection and preprocessing module: term library construction unit, used to organize industry maintenance terms, label term definitions, application scenarios and associated operations; text data collection unit, used to crawl group operation and maintenance management guidelines and historical operation and maintenance knowledge base, and extract structured knowledge; corpus collection unit, used to record dialogue corpus of front-line maintenance personnel and non-professionals, and label "natural language demand → maintenance term → interface call instruction" mapping relationship; data cleaning unit, used to remove duplicate and error data, and convert unstructured text into JSON format structured data for subsequent model training; Model construction module: basic model selection unit, using a general pre-trained large model as a base; fine-tuning training unit, inputting pre-processed multi-source data into the model, fine-tuning the model through the three tasks of "term prediction", "demand-term mapping", and "interface instruction generation", so that the model has the ability to identify maintenance terms in natural language, convert natural language requirements into maintenance terms, and generate background interface call parameters based on terms.
[0023] The model construction module further comprises a model optimization unit: the unit adopts a small sample learning technique, learns in view of the collected scarce corpus, improves the processing capacity of the model for rare or new situations; at the same time, a knowledge distillation technique is adopted to compress the model volume, so that the model adapts to the edge device, and the lightness and generalization ability of the model are improved.
[0024] Further comprising an interaction and interface processing module, which comprises: an interaction unit: a natural language dialogue interface of a Web or APP end is built, and a user inputs a demand through the interface; the demand includes but is not limited to device fault query, fault statistics, fault troubleshooting demand, various natural language demands related to maintenance business; an interface adaptation unit: a preset background interface adaptation layer is called by the model, instructions generated according to the user demand are converted into a format recognizable by the interface, and the interface is triggered to perform corresponding operations.
[0025] The interaction and interface processing module further comprises a result feedback unit: when the interface returns data, the unit makes the model convert the result into a natural language and feed it back to the user; the natural language result fed back includes specific data information and related analysis and explanation, so that the user can clearly understand the maintenance business.
[0026] In the "term prediction" task, the model accurately identifies the maintenance terms in the natural language to above a preset threshold; in the "demand-term mapping" task, the model accurately converts the natural language demand into the maintenance terms to above a preset threshold; in the "interface instruction generation" task, the background interface calling parameters generated based on the terms accurately trigger the corresponding interface to perform corresponding operations, and the interface execution success rate reaches above a preset threshold.
[0027] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A method for naturalizing commonly used terms based on the Daiwei large model, characterized in that: Includes the following steps: Terminology database construction: Organize industry maintenance terminology, label term definitions, application scenarios and related operations to form a complete terminology database, providing basic terminology support for subsequent model training; Data Acquisition and Preprocessing: Text Data Acquisition: Crawling the group's operation and maintenance management guidance documents and historical operation and maintenance knowledge base to extract structured knowledge; Corpus Collection: Recording the dialogue between front-line maintenance personnel and non-professionals, and marking the mapping relationship of "natural language requirements → maintenance terminology → interface call instructions"; Data Cleaning: Removing duplicate and erroneous data, and converting unstructured text into structured data in JSON format; Model Construction and Training: Basic Model Selection: A general pre-trained large model is used as the base; Fine-tuning Training: Preprocessed multi-source data is input into the model, and fine-tuning is performed through three tasks: "term prediction", "demand-term mapping" and "interface instruction generation". This enables the model to recognize substituted terms in natural language, convert natural language demands into substituted terms, and generate backend interface call parameters based on terms.
2. The method for naturalizing commonly used terms based on the Daiwei large model according to claim 1, characterized in that: The model building and training process also includes a model optimization step: employing few-shot learning techniques to learn from the collected scarce corpus, thereby improving the model's ability to handle rare or newly emerging situations; and simultaneously employing knowledge distillation techniques to compress the model size, making the model adaptable to edge devices and improving the model's lightweight and generalization capabilities.
3. The method for naturalizing commonly used terms based on the Daiwei large model according to claim 2, characterized in that: After model building and training are completed, the process also includes interaction and feedback steps: Interaction module setup: Build a natural language dialogue interface for web or mobile app, through which users input their needs; Interface adaptation: The model calls the preset backend interface adaptation layer, which converts the instructions generated according to user requirements into a format that the interface can recognize, and triggers the interface to perform the corresponding operation; Results Feedback: After the interface returns data, the model converts the results into natural language and feeds them back to the user, realizing a complete interactive process between natural language and maintenance business operations and feedback.
4. The method for naturalizing commonly used terms based on the Daiwei large model according to claim 3, characterized in that: In the process of building the interaction module, the user input requirements include, but are not limited to, equipment fault query, fault statistics, fault troubleshooting requirements, and various natural language requirements related to maintenance services.
5. The method for naturalizing commonly used terms based on the Daiwei large model according to claim 4, characterized in that: In the interface adaptation step, the interface recognition format is the JSON request body of a RESTful API; in the result feedback step, the natural language results fed back to the user include specific data information and related analysis explanations, so that the user can clearly understand the maintenance service situation.
6. A common terminology naturalization system based on the Daiwei large model, applied to the method described in claim 5, characterized in that: include: Data Acquisition and Preprocessing Module: The terminology database construction unit organizes industry maintenance terminology, annotating term definitions, application scenarios, and related operations; the text data acquisition unit crawls the group's operation and maintenance management guidance documents and historical operation and maintenance knowledge base, extracting structured knowledge; the corpus collection unit records dialogues between frontline maintenance personnel and non-professionals, annotating the mapping relationship of "natural language requirements → maintenance terminology → interface call instructions"; and the data cleaning unit removes duplicate and erroneous data, converting unstructured text into structured data in JSON format for subsequent model training. Model building module: The basic model selection unit uses a general pre-trained large model as the base; the fine-tuning training unit inputs preprocessed multi-source data into the model and fine-tunes the model through three tasks: "term prediction", "demand-term mapping" and "interface instruction generation", so that the model has the ability to recognize dimensional terms in natural language, convert natural language demands into dimensional terms, and generate backend interface call parameters based on terms.
7. The common terminology naturalization system based on the Daiwei large model according to claim 6, characterized in that: The model building module also includes a model optimization unit: this unit uses few-shot learning technology to learn from the collected scarce corpus, improving the model's ability to handle rare or new situations; at the same time, it uses knowledge distillation technology to compress the model size, making the model adaptable to edge devices and improving the model's lightweight and generalization capabilities.
8. The common terminology naturalization system based on the Daiwei large model according to claim 7, characterized in that: It also includes an interaction and interface processing module, which includes: an interaction unit: building a natural language dialogue interface on the web or APP, through which users input their needs, including but not limited to equipment fault query, fault statistics, fault troubleshooting and other natural language needs related to maintenance services; and an interface adaptation unit: the model calls a preset backend interface adaptation layer to convert the instructions generated according to user needs into a format that the interface can recognize, triggering the interface to execute the corresponding operation.
9. A common terminology naturalization processing system based on the Daiwei large model according to claim 8, characterized in that: The interaction and interface processing module also includes a result feedback unit: when the interface returns data, this unit enables the model to convert the results into natural language and feed them back to the user. The natural language results include specific data information and related analysis explanations so that the user can clearly understand the maintenance service situation.
10. A common terminology naturalization processing system based on a large-scale model according to claim 9, characterized in that: In the "Terminology Prediction" task, the model's accuracy in recognizing substituted terms in natural language reaches above the preset threshold; in the "Requirement-Terminology Mapping" task, the model's accuracy in converting natural language requirements into substituted terms reaches above the preset threshold; in the "Interface Instruction Generation" task, the backend interface call parameters based on terminology generation accurately trigger the corresponding interface to execute the corresponding operation, and the interface execution success rate reaches above the preset threshold.