Radio frequency domain knowledge graph construction method and device, and computer program product

Abstract

The invention discloses a radio frequency domain knowledge graph construction method and device and a computer program product, and relates to the technical field of knowledge graph creation, and the method comprises the steps: carrying out the semantic segmentation of a pre-obtained radio frequency text data stream, and generating a plurality of semantic blocks; performing sub-field identification on the plurality of semantic blocks by utilizing a pre-constructed field dictionary, and performing semantic matching on a cue word template from a pre-constructed multi-level cue word library based on a sub-field identification result; calling a pre-constructed large language model, and performing entity relationship extraction on the semantic blocks based on the cue word template to obtain an entity relationship extraction result; performing structure verification and fault-tolerant mechanism processing on an entity relationship extraction result to obtain a candidate knowledge triple; and constructing the radio frequency domain knowledge graph based on the candidate knowledge triples. High-consistency and high-precision radio frequency domain knowledge graph construction is realized through semantic segmentation, sub-domain identification, multi-level prompt word bank, structure verification and fault-tolerant mechanism processing.

Description

Technical Field

[0001] This application relates to the field of knowledge graph creation technology, and in particular to methods, devices and computer program products for constructing knowledge graphs in the radio frequency field. Background Technology

[0002] Due to the high complexity and multi-subdomain differences in RF technology documents, existing knowledge extraction methods driven by unified or static Prompt templates cannot adaptively adjust to the semantic feature differences among the multiple subdomains of RF. This leads to problems such as entity omissions, relationship errors, and incomplete structures, resulting in the inability to construct an accurate RF domain knowledge graph. In addition, existing knowledge extraction methods lack semantic segmentation and context preservation mechanisms, which interrupts parameter links, structural relationships, or process relationships in the RF domain during the extraction process. This damages the semantic integrity of complex RF documents, causing knowledge fragmentation and making it difficult to unify the knowledge.

[0003] Therefore, how to construct a highly consistent and accurate knowledge graph in the radio frequency field has become a technical problem that this application urgently needs to solve. Summary of the Invention

[0004] The main purpose of this application is to provide a method, device and computer program product for constructing a knowledge graph in the radio frequency (RF) field, aiming to solve the technical problem of how to construct a highly consistent and accurate RF field knowledge graph.

[0005] To achieve the above objectives, this application proposes a method for constructing a knowledge graph in the radio frequency (RF) field, the method comprising: Semantic segmentation is performed on the pre-acquired radio frequency text data stream to generate multiple semantic blocks; Using a pre-built domain dictionary, sub-domain identification is performed on the multiple semantic blocks respectively, and based on the results of the sub-domain identification, semantic matching of prompt word templates is performed from a pre-built multi-level prompt word library; A pre-built large language model is invoked, and entity relations are extracted from the semantic block based on the prompt word template to obtain the entity relation extraction result; The entity relation extraction results are subjected to structural verification and fault tolerance mechanism processing to obtain candidate knowledge triples; A knowledge graph for the radio frequency domain is constructed based on the candidate knowledge triples.

[0006] In one embodiment, the step of semantically segmenting the pre-acquired radio frequency text data stream to generate multiple semantic blocks further includes: Acquire technical documents in the field of radio frequency; The technical documents in the radio frequency field are encoded, converted, and their layout parsed to obtain the original radio frequency text data stream; The original radio frequency text data stream is filtered to obtain the radio frequency text data stream.

[0007] In one embodiment, the step of performing sub-domain identification on the plurality of semantic blocks using a pre-built domain dictionary, and semantically matching prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification, further includes: Constructing a domain dictionary based on the knowledge system of the radio frequency field; Based on the task objective of knowledge extraction in the radio frequency domain, a multi-level prompt word is designed. The multi-level prompt word includes task-driven prompt words, domain-guided prompt words, and format constraint prompt words. Each level of prompt word contains multiple prompt word templates. The multi-level prompt words are categorized and stored to construct a multi-level prompt word library.

[0008] In one embodiment, the step of performing sub-domain identification on the plurality of semantic blocks using a pre-built domain dictionary, and semantically matching prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification includes: The domain dictionary is used to identify sub-domains for each of the multiple semantic blocks to obtain the core sub-domain corresponding to each semantic block. The key features of the semantic block are analyzed, and semantic matching is performed in the multi-level prompt word library based on the core sub-domain and the key features to obtain a set of candidate prompt word templates. The optimal prompt word template is selected from the candidate prompt word template set by combining the key features, historical extraction performance, and current task type.

[0009] In one embodiment, the step of invoking a pre-built large language model to extract entity relations from the semantic block based on the prompt word template, and obtaining the entity relation extraction result, includes: The contextual relationships of the prompt word template are supplemented, and example triples and domain explanations are added to the prompt word template to obtain the enhanced prompt word template: The pre-built large language model is invoked, and the enhanced prompt word template and the semantic block are input into the pre-built large language model for entity relation extraction to obtain the entity relation extraction result.

[0010] In one embodiment, the step of performing structural verification and fault-tolerance mechanism processing on the entity relation extraction results to obtain candidate knowledge triples includes: From the entity relation extraction results, knowledge triples that conform to both the predefined triple structure and the predefined semantic specification are selected as the first candidate knowledge triples. The knowledge triples in the entity relation extraction results that do not conform to the predefined triple structure or the predefined semantic specification are regarded as the fourth candidate knowledge triples. Verify whether the first candidate knowledge triple belongs to a pre-set set of allowed types; If the first candidate knowledge triple belongs to the allowed type set, then the numerical parameters in the first candidate knowledge triple are formatted and the units are standardized, and the entity writing of the first candidate knowledge triple is unified to obtain the second candidate knowledge triple. Delete conflicting or incomplete relations in the second candidate knowledge triplet to obtain the third candidate knowledge triplet; Based on a pre-set fault tolerance mechanism, the fourth candidate knowledge triplet is repaired to obtain the fifth candidate knowledge triplet. The candidate knowledge triplet is obtained based on the third candidate knowledge triplet and the fifth candidate knowledge triplet.

[0011] In one embodiment, the step of constructing a radio frequency domain knowledge graph based on the candidate knowledge triples includes: The candidate knowledge triples and the semantic blocks corresponding to the candidate knowledge triples are written into the radio frequency domain knowledge base to form knowledge graph nodes and knowledge graph relationships. A radio frequency domain knowledge graph is constructed based on the knowledge graph nodes and the knowledge graph relationships.

[0012] In one embodiment, the step of constructing a radio frequency domain knowledge graph based on the candidate knowledge triples further includes: Obtain the error logs and extract statistical results generated by the fault tolerance mechanism; Analyze the error logs and the extracted statistical results to identify high-frequency error types and weak sub-regions; Adjust the domain dictionary and the multi-level prompt word library according to the high-frequency error types and the weak sub-regions; The execution steps are returned based on the adjusted domain dictionary and multi-level prompt word library: sub-domain identification is performed on the multiple semantic blocks using the pre-built domain dictionary, and prompt word templates are semantically matched from the pre-built multi-level prompt word library based on the results of the sub-domain identification.

[0013] In addition, to achieve the above objectives, this application also proposes a knowledge graph construction device in the radio frequency field, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the knowledge graph construction method in the radio frequency field as described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the radio frequency domain knowledge graph construction method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: Semantic blocks that maintain complete semantic meaning are generated through semantic segmentation, avoiding knowledge breaks across sentences and paragraphs, thus laying the foundation for high-precision extraction. The domain dictionary identifies subdomains and matches them with suitable multi-level prompt word templates, making the prompt strategy fit the characteristics of the radio frequency subdomain, reducing semantic understanding bias, and improving extraction consistency. The large language model extracts entity relations based on the adapted templates, giving full play to the model's capabilities while using domain guidance to improve extraction accuracy. Structural verification and fault tolerance mechanisms verify and correct the extraction results, further ensuring the consistency of the candidate knowledge triples. Finally, a graph is constructed based on the verified candidate knowledge triples, achieving a highly consistent and high-precision radio frequency domain knowledge graph construction. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

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

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the radio frequency field knowledge graph construction method of this application; Figure 2 This is a flowchart illustrating the second embodiment of the radio frequency field knowledge graph construction method of this application; Figure 3 This is a flowchart illustrating the third embodiment of the radio frequency field knowledge graph construction method of this application; Figure 4 This is a flowchart illustrating the fifth embodiment of the radio frequency field knowledge graph construction method of this application; Figure 5 The flowchart for constructing a multi-level prompting knowledge graph in the radio frequency field provided in this application; Figure 6 A schematic diagram illustrating the entity recognition and relation extraction results provided in this application; Figure 7 Schematic diagrams illustrating the effects of the radio frequency domain knowledge graph construction method provided in this application on different models; Figure 8 This is a schematic diagram of the module structure of the radio frequency knowledge graph construction device according to an embodiment of this application; Figure 9 This is a schematic diagram of the device structure of the hardware operating environment involved in the radio frequency domain knowledge graph construction method in this application embodiment.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: semantically segment the pre-acquired radio frequency text data stream to generate multiple semantic blocks; perform sub-domain identification on the multiple semantic blocks using a pre-built domain dictionary, and semantically match prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification; call a pre-built large language model to extract entity relations from the semantic blocks based on the prompt word templates to obtain entity relation extraction results; perform structural verification and fault tolerance mechanism processing on the entity relation extraction results to obtain candidate knowledge triples; and construct a radio frequency domain knowledge graph based on the candidate knowledge triples.

[0023] In this embodiment, for ease of description, the following description will focus on a radio frequency domain knowledge graph construction system.

[0024] This application's embodiments take into account that the radio frequency (RF) field covers multiple areas such as antenna design, RF circuits, interconnect structures, manufacturing processes, thermal management, testing and simulation. Its engineering documentation highly relies on technical terminology, physical constraints, and cross-sentence structured descriptions, typically presented in the form of design documents, system specifications, process flows, simulation records, test reports, technical manuals, journal articles, and authoritative books. These documents contain a large amount of key parameters, structural combination relationships, signal link and process link knowledge, but are generally fragmented, diverse in expression, have large context spans, and complex cross-domain semantic relationships, severely restricting the reuse and structured accumulation of engineering knowledge. RF domain knowledge graphs (RF-KG), as an important technology supporting intelligent search, automatic question answering, design assistance, and process traceability, urgently require high-quality entity and relationship extraction capabilities. However, the high complexity and multi-subdomain differences of RF technical documents make manual knowledge graph construction costly, time-consuming, and difficult to maintain. Existing automated methods are insufficient when facing RF professional semantics, cross-segment semantic links, and domain characteristics, specifically manifested in: Existing large language model knowledge extraction methods typically employ unified or static Prompt templates to drive extraction. While this reduces the cost of manual rule design to some extent, the fixed prompt structure and singular strategies fail to adapt to the semantic differences across multiple subdomains in RF documents. This leads to incomplete extraction or semantic misunderstandings when handling different types of content, such as antennas, RF circuits, interconnect structures, process links, and test documents. Furthermore, existing methods lack mechanisms for handling chain-like semantic structures across sentences and paragraphs in technical documents, and lack semantic segmentation and context preservation techniques, causing parameter links, structural relationships, or process relationships to be interrupted during extraction. In addition, existing extraction methods have limited capabilities in structured constraints, format validation, and abnormal output correction. If the model output format is invalid or fields are missing, the system lacks effective fault tolerance and correction mechanisms, failing to meet the stability and consistency requirements of engineering-level knowledge extraction in the RF field. Ultimately, this results in the inability to construct an accurate RF domain knowledge graph, the destruction of the semantic integrity of complex RF documents, knowledge fragmentation, and difficulty in achieving consistency.

[0025] Therefore, this application provides a solution that generates semantic blocks that maintain complete semantics through semantic segmentation, avoiding knowledge breaks across sentences and paragraphs, thus laying the foundation for high-precision extraction. A domain dictionary identifies sub-domains and matches them with suitable multi-level prompt word templates, ensuring that the prompting strategy aligns with the characteristics of the radio frequency (RF) sub-domain, reducing semantic understanding bias and improving extraction consistency. A large language model extracts entity relationships based on the adapted templates, fully leveraging the model's capabilities while utilizing domain guidance to improve extraction accuracy. Structural verification and fault-tolerance mechanisms validate and correct the extraction results, further ensuring the consistency of candidate knowledge triples. Finally, a graph is constructed based on the verified candidate knowledge triples, achieving highly consistent and high-precision RF domain knowledge graph construction.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a radio frequency domain knowledge graph construction system. The following description uses a radio frequency domain knowledge graph construction system as an example to illustrate this embodiment and the subsequent embodiments.

[0027] Based on this, embodiments of this application provide a method for constructing a knowledge graph in the radio frequency field, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the radio frequency field knowledge graph construction method of this application.

[0028] In this embodiment, the radio frequency domain knowledge graph construction method includes steps S10~S50: Step S10: Semantically segment the pre-acquired radio frequency text data stream to generate multiple semantic blocks; Radio frequency (RF) text data stream refers to a collection of preprocessed RF technical document texts; semantic segmentation refers to dividing text into segments based on semantic integrity; semantic block refers to a text segment corresponding to a complete technical description.

[0029] By generating semantic blocks that maintain complete semantics, the problem of semantic link breaks across sentences / paragraphs in long documents in the radio frequency domain is avoided, ensuring that each semantic block contains complete technical logic units, thereby improving the accuracy of subsequent subdomain recognition and prompt word matching.

[0030] Specifically, the system acquires the preprocessed radio frequency text data stream, uses syntactic boundaries, punctuation marks, heading levels, and key radio frequency terms as the basis for segmentation, and analyzes and divides each segment to generate semantic blocks; at the same time, it attaches metadata such as document ID, page number, and paragraph number to each semantic block for traceability.

[0031] In one possible implementation, a Transformer-based semantic segmentation model can be used to segment a pre-acquired radio frequency (RF) text data stream to generate multiple semantic blocks. Specifically, the Transformer-based semantic segmentation model first learns complete boundary features described by techniques using a pre-trained RF domain corpus; then, it receives the pre-processed RF text data stream and uses the learned RF domain semantic features to identify syntactic boundaries, punctuation marks, heading levels, and key RF terms as segmentation criteria; next, it analyzes the text segment by segment, dividing it into semantic blocks that maintain semantic continuity (each block corresponds to a complete technical description); finally, it adds metadata such as document ID, page number, and paragraph number to each semantic block to complete the semantic block generation.

[0032] Step S20: Using a pre-built domain dictionary, perform sub-domain identification on the multiple semantic blocks respectively, and based on the results of the sub-domain identification, semantically match prompt word templates from a pre-built multi-level prompt word library; Domain dictionary refers to a dictionary set containing terms from radio frequency (RF) subdomains; subdomain identification refers to determining the RF subdirection to which a semantic block belongs; multi-level prompt lexicon refers to a database containing task-driven / domain-guided / format-constrained prompt word templates.

[0033] Specifically, the system calls the domain dictionary to match key terms of semantic blocks and identify subdomains; combines the subdomain identification results to match the semantic block information with a multi-level prompt word library to filter candidate template sets; and determines the appropriate template based on the task type and semantic block features.

[0034] Additionally, it should be noted that the sub-domain identification results serve as constraints for multi-level cue scheduling. Since different sub-domains correspond to different sets of cue templates, and each type of cue template contains cue templates from multiple sub-domains, it is necessary to combine the sub-domain identification results to limit the candidate range of cue templates. There are significant differences in the legal entity types and relation types across different radio frequency sub-domains. The sub-domain identification results are used to limit the entity categories and relation sets that the large language model can use during extraction, thereby reducing ambiguous extraction and illegal relation generation.

[0035] In addition, it is understandable that since subdomain identification occurs before entity relation extraction, it directly participates in prompt scheduling and extraction strategy selection, directly affecting model behavior, rather than just being used for result labeling. It belongs to a control-type intermediate processing mechanism, rather than a simple data classification operation.

[0036] Step S30: Invoke the pre-built large language model, extract entity relations from the semantic block based on the prompt word template, and obtain the entity relation extraction result; Entity-relation extraction results refer to the raw data containing entities and relations output by the large language model.

[0037] The pre-built large language model is invoked to extract entity relations from the semantic block based on the prompt word template. By leveraging the semantic understanding capabilities of the large language model and combining it with the adapted prompt word template, radio frequency professional knowledge can be accurately extracted, thereby reducing output deviation and improving the accuracy and completeness of entity relation extraction.

[0038] Specifically, the system combines semantic blocks with adapted prompt word templates to form input text; it calls a pre-built large language model to process the input text; the large language model identifies entities and relationships in the semantic blocks based on the prompt word templates and outputs the results.

[0039] Step S40: Perform structural verification and fault tolerance mechanism processing on the entity relation extraction results to obtain candidate knowledge triples; Structural verification refers to checking whether the entity relation extraction results conform to the predefined specifications; fault tolerance mechanism processing refers to repairing abnormal results in the entity relation extraction results; candidate knowledge triples refer to the standardized knowledge units that have undergone verification processing.

[0040] Specifically, the system performs structural verification on the extracted results, which includes JSON format and field integrity verification; it triggers fault tolerance processing for abnormal results that fail structural verification, which includes automatic correction of format errors and enhanced retry prompts; and it converts the results that pass structural verification into candidate knowledge triples.

[0041] Step S50: Construct a radio frequency domain knowledge graph based on the candidate knowledge triples.

[0042] Radio frequency (RF) knowledge graphs refer to structured knowledge networks containing RF entities and relationships.

[0043] Specifically, entities in candidate triples are mapped to nodes in the knowledge graph, and relationships between entities are mapped to edges between nodes, forming a semantically related network structure. Source metadata is attached to each node and edge, including the corresponding semantic block content, document ID, page number, paragraph number, etc., to achieve knowledge traceability. The organized nodes, edges, and metadata are written into the radio frequency domain knowledge base, and a storage method that supports graph structure query and reasoning is adopted to form a queryable and reasonable knowledge graph.

[0044] In one feasible implementation, the entity distribution, relation types, knowledge coverage of each subdomain, and extraction quality indicators in the knowledge graph can be statistically analyzed and displayed in a visual manner to help evaluate the completeness and accuracy of the knowledge graph and provide a basis for subsequent iterative optimization.

[0045] Furthermore, it should be noted that the legitimate entity types and relationship types differ significantly across different RF subdomains. For example, in the process subdomain, the key relationships are "includes," "depends on," and "precedes," while in the antenna subdomain, the key relationships are "operates on," "has," and "applies to." Therefore, the RF domain knowledge graph construction stage also incorporates the results of subdomain identification. Specifically, the domain identification results are used to group and classify the extracted triples, forming different subgraphs or schema fragments in the RF domain knowledge graph, providing a structural foundation for subsequent retrieval, reasoning, and applications.

[0046] This embodiment provides a method for constructing a knowledge graph in the radio frequency (RF) domain. Semantic segmentation generates semantic blocks that maintain complete semantics, avoiding knowledge fragmentation across sentences and paragraphs, thus laying the foundation for high-precision extraction. A domain dictionary identifies sub-domains and matches them with suitable multi-level prompt word templates, ensuring that the prompting strategy aligns with the characteristics of the RF sub-domain, reducing semantic understanding bias and improving extraction consistency. A large language model extracts entity relationships based on the adapted templates, fully leveraging the model's capabilities while utilizing domain guidance to improve extraction accuracy. Structural verification and fault tolerance mechanisms validate and correct the extraction results, further ensuring the consistency of candidate knowledge triples. Finally, a graph is constructed based on the verified candidate knowledge triples, achieving highly consistent and high-precision RF domain knowledge graph construction.

[0047] In one feasible implementation, step S30 may include steps S31-S32: Step S31: Supplement the contextual association of the prompt word template, and add example triples and domain explanations to the prompt word template to obtain the enhanced prompt word template; Contextual association refers to the related textual information before and after a semantic block; example triples refer to high-quality entity relation triples that have been verified in the radio frequency domain; domain explanation refers to the definition or description of radio frequency professional terms; and enhanced prompt word templates refer to optimized templates after contextual supplementation, example addition, and domain explanation injection.

[0048] The purpose of this step is to address the issues of insufficient adaptability of the original prompt word template and semantic understanding bias. By supplementing the context to maintain cross-sentence semantic links, adding examples to clarify the extraction target, and injecting domain explanations to help the model understand professional concepts, the accuracy and consistency of entity relationship extraction can be improved.

[0049] Specifically, by utilizing the metadata of semantic blocks, such as document ID, page number, and content of adjacent semantic blocks, a sliding window mechanism is used to extract N related sentences before and after the current semantic block and add them as context to the prompt word template to ensure that the model understands the semantic links across sentences.

[0050] In addition, high-quality triples consistent with the current semantic block subdomain are matched from a pre-built radio frequency domain example library, and prompt word templates are inserted as reference examples to clearly extract the target format and relation type; combined with the professional term definitions in the domain dictionary, domain explanations of key terms in the semantic block are automatically inserted to help the model accurately understand radio frequency domain concepts.

[0051] Step S32: Invoke the pre-built large language model, input the enhanced prompt word template and the semantic block into the pre-built large language model for entity relation extraction, and obtain the entity relation extraction result.

[0052] Enhanced prompt word templates are used to guide large models to accurately extract entities and relationships in the radio frequency (RF) domain, improving the extraction precision and recall rate and ensuring that the results meet the professional standards of the RF domain.

[0053] Specifically, the enhanced prompt word template and semantic block content are concatenated in a fixed format. The pre-built large language model is called through the API interface, and the corresponding temperature parameters and output format parameters are set to ensure the stability of the results. The input text concatenated in a fixed format is input into the large language model, which then extracts entity relations from the input text to obtain the entity relation extraction results.

[0054] In one feasible implementation, step S50 may include steps S51-S52: Step S51: Write the candidate knowledge triples and the semantic blocks corresponding to the candidate knowledge triples into the radio frequency domain knowledge base to form knowledge graph nodes and knowledge graph relationships. A knowledge graph node refers to an element in a knowledge graph that represents an entity; a knowledge graph relation refers to an element in a knowledge graph that represents the connection between entities.

[0055] Specifically, the candidate knowledge triples are first preprocessed by normalization to unify the representation of the same entity node, and the relations are normalized by predicate: for example, antenna elements and elements are unified as the same entity node; parameter association and attribute correspondence are unified as attribute relations.

[0056] Furthermore, the subject and object in the triple are mapped to entity nodes in the knowledge graph, and the predicate is mapped to relation edges between nodes. Entity type and key attributes from the source semantic block are added to each node, and metadata such as source semantic block content, document ID, page number, and paragraph number are added to the relation edges.

[0057] Step S52: Construct a radio frequency domain knowledge graph based on the knowledge graph nodes and the knowledge graph relationships.

[0058] Specifically, the first step is to construct the graph topology: nodes and relationships are categorized and organized according to entity and relationship types, and connections within and across sub-domains are established based on sub-domain identification results, forming a hierarchical knowledge network. Next, graph database storage and indexing are performed: the graph structure is stored using graph databases such as Neo4j, leveraging the native graph query capabilities of the graph database to support fast node and relationship retrieval and path analysis; simultaneously, indexes are created for frequently queried entities to improve query efficiency.

[0059] Furthermore, based on ontology rules and physical constraints in the radio frequency (RF) domain, implicit knowledge relationships are automatically derived and added to the graph database to improve the accuracy of the RF domain knowledge graph.

[0060] Based on the first embodiment of this application, a second embodiment of this application is proposed. In the second embodiment of this application, content that is the same as or similar to that in the first embodiment described above can be referred to the above description and will not be repeated hereafter.

[0061] Based on this, please refer to Figure 2 , Figure 2 This is a schematic diagram of the process for applying for the second embodiment, such as... Figure 2 As shown, before step S10, the radio frequency domain knowledge graph construction method further includes steps S01~S03: Step S01: Collect technical documents in the field of radio frequency; Radio frequency (RF) technical documents refer to various technical documents from the RF field, including but not limited to antenna design reports, RF circuit specifications, interconnect structure design documents, manufacturing and processing records, test and verification reports, etc.

[0062] Specifically, the system provides a unified document collection portal, which supports integration with the enterprise's internal document management system as well as manual uploading of external publicly available resources.

[0063] Step S02: Encode and convert the RF technical document and parse its layout to obtain the original RF text data stream; We perform unified encoding conversion on RF technical documents to avoid garbled characters; then we parse the layout of documents of different formats. For example, we use OCR technology to extract text and table structure from PDF scans, use the DOM model to parse paragraphs, headings and embedded tables from Word documents, and extract cell content from Excel documents while preserving row and column relationships. At the same time, we extract non-text elements such as LaTeX descriptions of formulas and text descriptions of charts, and finally form a structured raw RF text data stream containing the main text, table data and metadata (document type, source).

[0064] Step S03: Filter the original radio frequency text data stream to obtain the radio frequency text data stream.

[0065] Filter irrelevant content from the raw radio frequency (RF) text data stream: remove headers, footers, watermarks, advertisements, and duplicate paragraphs; filter non-RF domain text; retain parameter-related rows / columns for table data and delete blank rows or invalid data; in addition, correct OCR errors and remove special symbol noise using regular expressions, ultimately obtaining a clean text data stream containing only the core RF domain content.

[0066] In this embodiment, various technical documents in the radio frequency (RF) field are collected through multiple channels and categorized and labeled according to sub-fields, providing rich and accurate data sources for subsequent processing. Heterogeneous documents are uniformly encoded and parsed into structured raw data streams containing text, tables, and metadata, achieving document format standardization and information integrity preservation. Irrelevant content is filtered, noise is corrected, and core RF text is purified, effectively improving the efficiency and accuracy of subsequent semantic segmentation and prompt scheduling, and jointly laying a high-quality data foundation for the construction of a knowledge graph in the RF field.

[0067] Based on the first and / or second embodiments of this application, a third embodiment of this application is proposed. In the third embodiment of this application, content that is the same as or similar to the first and / or second embodiments described above can be referred to the above description and will not be repeated hereafter.

[0068] Based on this, please refer to Figure 3 , Figure 3 This is a schematic diagram of the process for applying for the third embodiment, such as... Figure 3 As shown, before step S20, the radio frequency domain knowledge graph construction method further includes steps A01~A03: Step A01: Construct a domain dictionary based on a pre-defined knowledge system in the radio frequency field; The knowledge system in the radio frequency field refers to a professional knowledge framework covering sub-fields such as antenna design, radio frequency circuits, interconnect structures, manufacturing and processing, and testing and verification. It includes the core terminology, physical parameters, component relationships and industry standards of each sub-field.

[0069] Specifically, based on the sub-fields covered by the radio frequency (RF) field, and in conjunction with industry standards, authoritative technical manuals, and core terms, physical parameters, and component relationships in engineering practice, the entity types, relationship patterns, and professional constraints of each sub-field can be clarified to form a structured knowledge framework, thereby determining the knowledge system of the RF field.

[0070] Furthermore, based on the knowledge system of the radio frequency field, core terms, professional vocabulary, physical parameters and component names of each subfield are collected, and term definitions, synonyms and hierarchical relationships are supplemented by industry standards and authoritative technical manuals; then, the subfield affiliation and common related entity / relationship types of each term are marked to form a structured domain dictionary; finally, the dictionary is updated regularly to include newly emerging radio frequency technology terms and industry standards.

[0071] Step A02: Design multi-level prompts based on the predetermined radio frequency domain knowledge extraction task objectives. The multi-level prompts include task-driven prompts, domain-guided prompts, and format-constrained prompts. Each level of prompts contains multiple prompt templates. Specifically, in response to the practical needs of intelligent search, design assistance, and process traceability in RF engineering, and in combination with the common structured knowledge types in the documents, the types of entities and relationships to be extracted are identified, forming the RF domain knowledge extraction task objectives, which provide direction for subsequent multi-level prompt design and scheduling.

[0072] Furthermore, by clearly defining the task objectives of knowledge extraction in the radio frequency (RF) domain, the types of entities and relational patterns to be extracted are determined. Next, task-driven prompt templates are designed to address the task objectives, clearly defining the types of entities and relations to be extracted. Then, combining the RF domain knowledge system, domain-guided prompt templates are designed, embedding professional expressions and physical rules from each subdomain. Finally, format constraint prompt templates are designed, specifying the output structure and field requirements. Each level of prompts contains multiple templates to adapt to different subdomains and task scenarios.

[0073] Step A03: Classify and store the multi-level prompt words to construct a multi-level prompt word library.

[0074] The system categorizes multi-level prompts according to their type (task-driven, domain-guided, format-constrained) and subdomain. The categorized prompt templates are then stored in corresponding storage units. Each storage unit contains metadata such as the prompt template's content, applicable scenarios (e.g., subdomain, task type), priority, and historical usage effects. Simultaneously, an indexing mechanism is established, with index dimensions including subdomain, task type, and template type, to enable rapid retrieval and retrieval of corresponding prompt templates based on the characteristics of the current semantic block.

[0075] In this embodiment, a radio frequency (RF) domain dictionary is constructed to support the subdomain identification of semantic blocks and the supplementation of domain information for prompt words. Multi-level prompt words can provide refined guidance for different tasks and professional characteristics in the RF domain, effectively improving the adaptability of the large language model to complex documents in the RF domain, reducing entity omissions, relational errors and structural inconsistencies, and enhancing the accuracy and stability of knowledge extraction.

[0076] Based on the above embodiments of this application, a fourth embodiment of this application is proposed. In this fourth embodiment, content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0077] In this embodiment, the step S20, which involves using a pre-built domain dictionary to perform sub-domain identification on the multiple semantic blocks and semantically matching the prompt word template from the pre-built multi-level prompt word library based on the results of the sub-domain identification, may include steps S21 to S23: Step S21: Use the domain dictionary to perform sub-domain identification on the multiple semantic blocks respectively to obtain the core sub-domain corresponding to each semantic block; The core subdomain refers to the radio frequency (RF) subdomain type to which the semantic block primarily belongs. The purpose of subdomain identification is to address the problem of poor compatibility of prompt words caused by RF documents spanning multiple subdomains, avoid semantic misunderstandings of unified prompt words in different subdomains, and improve the targeting and accuracy of knowledge extraction.

[0078] Specifically, the system first loads a pre-built domain dictionary, then performs term matching on each semantic block, counts the frequency and weight of terms belonging to each subdomain in the semantic block, determines multiple core subdomains corresponding to the semantic block based on the frequency and weight, and identifies the core subdomain. The core subdomain identifier is not the final output result, but is used as a control variable for subsequent semantic matching.

[0079] Step S22: Analyze the key features of the semantic block, and perform semantic matching in the multi-level prompt word library based on the core sub-domain and the key features to obtain a set of candidate prompt word templates; Key features refer to the features in the semantic block that reflect the type of technical content; the candidate prompt word template set refers to the set of prompt word templates that meet the conditions after matching.

[0080] Specifically, the system first parses the key features of the semantic block, such as identifying whether the semantic block is a structural description, parameter relationship, or process step; then, based on the core subdomain, it locates the corresponding subdomain's prompt word template group in a multi-level prompt word library; finally, it performs semantic matching by combining the key features with the applicable scenarios of the templates in the template group, and includes templates with a matching degree higher than a threshold into the candidate set. In one possible implementation, semantic matching can use cosine similarity calculation, comparing the key feature vector of the semantic block with the applicable feature vector of the prompt word template, and selecting templates with high similarity.

[0081] The multi-level prompt word library contains different types of prompt templates. Each type of prompt template contains multiple prompt templates from different sub-domains. The sub-domain identification results are used to limit the candidate range of prompt word templates. For example, if the core sub-domain corresponding to a semantic block is an antenna-related semantic block, then a prompt template containing antenna parameters, radiation characteristics, and frequency band relationships is selected; if the core sub-domain corresponding to a semantic block is a process-related semantic block, then a prompt template containing process sequence and process dependencies is selected, thus avoiding the problem of semantic offset caused by a unified Prompt in different RF sub-domains.

[0082] Step S23: Select the optimal prompt word template from the candidate prompt word template set by combining the key features, historical extraction performance and current task type.

[0083] Historical extraction performance refers to the performance metrics such as accuracy and recall of candidate prompt word templates in the same or similar past scenarios; current task type refers to the type of knowledge extraction task that needs to be performed now; optimal prompt word template refers to the prompt word template that is most suitable for the current semantic block after comprehensively considering key features, historical performance and task type.

[0084] Specifically, firstly, historical extraction performance data of candidate templates are obtained; then, templates that match the current task type are selected; and finally, based on the matching degree between key features and templates and the weighted score of historical performance, the template with the highest score is selected as the optimal template.

[0085] In this embodiment, the core sub-domain of the semantic block is identified by the domain dictionary, providing accurate domain basis for prompt word matching; candidate templates are matched from the multi-level prompt word library by combining the core sub-domain and key features to reduce scheduling blindness; the optimal template is selected by comprehensively considering key features, historical extraction performance and task type, thereby improving the pertinence, accuracy and stability of the large model in radio frequency domain knowledge extraction.

[0086] Based on the above embodiments of this application, a fifth embodiment of this application is proposed. In this fifth embodiment, content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0087] In this embodiment, step S40, which involves performing structural verification and fault-tolerance processing on the entity relation extraction results to obtain candidate knowledge triples, may include steps S41 to S47: Step S41: Select knowledge triples from the entity relation extraction results that conform to both the predefined triple structure and the predefined semantic specification as the first candidate knowledge triples. The predefined triple structure refers to the pre-defined triple format that conforms to the requirements of the knowledge graph; the predefined semantic specification refers to the pre-defined JSON rules that conform to the semantic logic of the radio frequency domain; the first candidate knowledge triple refers to the triple selected from the entity relation extraction results that conforms to both the predefined triple structure and the predefined semantic specification.

[0088] Specifically, each triple in the entity relation extraction result is subjected to two checks: the first check is whether it conforms to the predefined triple structure, and the second check is whether it conforms to the predefined semantic specification. Only triples that pass both checks are marked as the first candidate knowledge triples.

[0089] Step S42: The knowledge triples in the entity relation extraction results that do not conform to the predefined triple structure or do not conform to the predefined semantic specification are taken as the fourth candidate knowledge triples. The fourth candidate knowledge triple refers to the triple selected from the entity relation extraction results that does not conform to the predefined triple structure or the predefined semantic specification.

[0090] Specifically, each triple in the entity relation extraction result is traversed, and a predefined triple structure check and a predefined semantic specification check are performed on each triple; if either check fails, the triple is marked as the fourth candidate knowledge triple.

[0091] Step S43: Verify whether the first candidate knowledge triple belongs to the pre-set set of allowed types; The allowed type set refers to the predefined set of entity types and relation types that are allowed to exist in the radio frequency domain knowledge graph. For example, entity types include antennas and radio frequency circuits, and relation types include inclusion and connection.

[0092] Specifically, for each first candidate knowledge triple, check whether it belongs to a type in the allowed type set; if the candidate knowledge triple belongs to the allowed type set, it passes the check; otherwise, it fails the check.

[0093] Step S44: If the first candidate knowledge triplet belongs to the allowed type set, then the numerical parameters in the first candidate knowledge triplet are formatted and the units are standardized, and the entity writing of the first candidate knowledge triplet is unified to obtain the second candidate knowledge triplet. Numerical parameters refer to the parameters in a triplet that represent physical quantities or numerical values; format standardization refers to unifying the format of numerical parameters; unit standardization refers to unifying the units of numerical parameters; entity notation standardization refers to unifying different ways of representing the same entity into a standard notation.

[0094] The second candidate knowledge triple refers to the triple after numerical normalization and entity writing unification.

[0095] Specifically, for the first candidate knowledge triplet that passes the type validation, the system first performs format normalization and unit normalization, and then replaces different ways of writing the same entity with the standard way of writing by querying the thesaurus in the domain dictionary; after processing, the second candidate knowledge triplet is obtained.

[0096] Step S45: Delete conflicting or incomplete relations in the second candidate knowledge triple to obtain the third candidate knowledge triple; Conflicting relations refer to contradictory relationships between the same entity pairs; incomplete relations refer to relations in which necessary information or fields are missing in the triples; third candidate knowledge triples refer to triples after deleting conflicting and incomplete relations.

[0097] The system iterates through the second candidate knowledge triples, establishes a relation index for entity pairs, searches for contradictory relations between the same entity pairs, and checks whether the triples are missing necessary fields to detect incomplete relations. Triples in the second candidate knowledge triples that have either conflicting or incomplete relations are deleted to obtain the third candidate knowledge triples.

[0098] Step S46: Based on the pre-set fault tolerance mechanism, the fourth candidate knowledge triplet is repaired to obtain the fifth candidate knowledge triplet. Fault tolerance mechanisms refer to a set of pre-defined strategies for repairing triples that do not conform to predefined triple structures or predefined semantic specifications. These include mechanisms such as automatic formatting correction, enhanced retries with prompts, and constraint completion.

[0099] Specifically, the system first identifies the error type of the fourth candidate knowledge triple and selects the corresponding fault tolerance strategy according to the error type: for format errors, the automatic format repair module is used to correct them; for semantic errors, the prompt enhancement retry module is used to re-invoke the large language model for extraction; for missing fields, the constraint completion module is used to supplement the missing information; triples that conform to the predefined triple structure and predefined semantic specifications after repair are marked as the fifth candidate knowledge triples.

[0100] However, for cases where automatic repair is not possible, a policy-based retry process will be triggered, including adjusting the Prompt combination, strengthening structural constraints, narrowing the extraction range, and adding example triples to retry the large model. The exception handling module also records the semantic block where the error occurred, the Prompt used, and the error type, providing a basis for subsequent parameter tuning.

[0101] Step S47: Obtain the candidate knowledge triplet based on the third candidate knowledge triplet and the fifth candidate knowledge triplet.

[0102] The third and fifth candidate knowledge triples are merged into a temporary set; then the temporary set is deduplicated: by calculating the hash value of each triple, duplicate triples are removed, and the final set is the candidate knowledge triples.

[0103] In this embodiment, the extraction results are divided into compliant first candidate triples and non-compliant fourth candidate triples. After verifying the legality of the type of the first candidate, the parameter format, unit and entity writing are unified, and conflicting or incomplete relationships are deleted. At the same time, the fourth candidate triples are repaired through a fault tolerance mechanism. Finally, the two types of results are integrated to obtain high-quality candidate triples, which effectively improves the standardization, accuracy and completeness of the knowledge extraction results in the radio frequency field, reduces knowledge loss and errors, and meets the requirements of stability and consistency for engineering-level knowledge graph construction.

[0104] Based on the above embodiments of this application, a fifth embodiment of this application is proposed. In this fifth embodiment, content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter.

[0105] Based on this, please refer to Figure 4 , Figure 4 This is a schematic diagram of the process of applying for the fifth embodiment, as shown below. Figure 4 As shown, after step S50, the radio frequency domain knowledge graph construction method further includes steps S60~S90: Step S60: Obtain the error logs and extracted statistical results generated by the fault tolerance mechanism. Error logs refer to log files that record information such as the type, content, and location of errors that occur during the fault tolerance mechanism process; extraction statistics refer to the results obtained after statistical analysis of various indicators (such as extraction success rate, error rate, extraction quality of each subdomain, etc.) in the knowledge extraction process.

[0106] By collecting error information and quality data during the knowledge extraction process, we can provide a basis for subsequent analysis of high-frequency error types and weak sub-regions, thereby supporting the optimization of the domain dictionary and multi-level prompting lexicon, and improving the long-term adaptability and extraction accuracy of the system.

[0107] Specifically, the system periodically exports error logs from the fault tolerance mechanism module, including error type (such as formatting error, missing field, semantic error), the semantic block ID corresponding to the error, the prompt word template ID used, and the time the error occurred. Simultaneously, it obtains extraction statistics from the extraction results module, including metrics such as the number of extractions, correct extractions, incorrect extractions, recall rate, and precision rate for each subdomain. This data is then integrated and stored in a data warehouse for subsequent analysis. In one possible implementation, the system can be configured with a scheduled task to automatically export the previous day's error logs and extraction statistics every morning at midnight, and perform deduplication and formatting on the data to ensure accuracy and consistency.

[0108] Step S70: Analyze the error log and the extracted statistical results to identify high-frequency error types and weak sub-regions; High-frequency error types refer to error types that appear frequently in error logs; weak sub-regions refer to sub-domains or semantic block types that perform poorly in the extracted statistical results.

[0109] Specifically, firstly, error logs are categorized and statistically analyzed to calculate the frequency of each error type and identify high-frequency error types. Then, the extraction statistics are analyzed at the sub-domain level to compare the extraction indicators of each sub-domain and identify weak sub-regions with indicators below the threshold. At the same time, the correlation between high-frequency error types and weak sub-regions is analyzed by combining error logs and extraction statistics, such as a certain error type occurring frequently in a certain weak sub-region.

[0110] In one possible implementation, the system can use data visualization tools, such as bar charts and heatmaps, to display the distribution of error types and extraction indicators of sub-domains, and intuitively identify high-frequency error types and weak sub-regions.

[0111] Step S80: Adjust the domain dictionary and the multi-level prompt word library according to the high-frequency error type and the weak sub-region; For high-frequency error types, the system adjusts the multi-level prompt dictionary based on the error cause. For example, for format errors, the structural constraint prompt template is optimized and descriptions of format validation rules are added; for semantic errors, more domain-specific knowledge is added to the domain guidance prompt template.

[0112] For weak sub-regions, the system adjusts the domain dictionary, adding professional terms and sub-domain classification rules for that sub-region, while optimizing the multi-level prompt word templates corresponding to that sub-region to better suit the semantic features of that sub-region.

[0113] Step S90: Return to the execution steps based on the adjusted domain dictionary and multi-level prompt word library: perform sub-domain identification on the multiple semantic blocks using the pre-built domain dictionary, and semantically match prompt word templates from the pre-built multi-level prompt word library based on the results of the sub-domain identification.

[0114] The system loads the adjusted domain dictionary and multi-level prompt word library, uses the terms and classification rules in the adjusted domain dictionary to identify sub-domains for new or unprocessed semantic blocks, and then, based on the features of the identified sub-domains and semantic blocks, semantically matches appropriate prompt word templates from the adjusted multi-level prompt word library to drive the large language model to extract knowledge; at the same time, the system records the results of this extraction for subsequent analysis and optimization.

[0115] In this embodiment, by collecting error logs and extraction statistics of fault-tolerant processing, analyzing high-frequency error types and weak sub-regions, and specifically adjusting the domain dictionary and multi-level prompt word library and reapplying them to the sub-domain identification and prompt matching of semantic blocks, a closed-loop feedback mechanism is formed. This effectively improves the system's ability to adapt to the semantic differences of multiple sub-domains in the radio frequency domain, reduces extraction errors, and continuously enhances the accuracy, robustness, and engineering usability of knowledge extraction, providing dynamic optimization support for the construction of a high-quality radio frequency domain knowledge graph.

[0116] For example, to aid in understanding the radio frequency domain knowledge graph construction provided in the embodiments of this application, the following is combined with Figure 5 This application provides a detailed explanation of the knowledge graph construction process in the radio frequency field. For example... Figure 5 As shown, Figure 5 Please refer to the flowchart for constructing the multi-level prompting knowledge graph in the radio frequency field provided for this application. Figure 5 The specific implementation steps of this application are as follows: Step 1: Document Acquisition and Preprocessing. The system receives various technical documents from the RF field, including but not limited to antenna design reports, RF circuit specifications, interconnect structure design documents, manufacturing and processing records, test and verification reports, etc. It performs unified encoding conversion and layout parsing on documents of different formats (PDF, Word, etc.), extracts the main text, tables, and necessary annotation information, forming a raw text data stream suitable for subsequent processing, and filters out obviously irrelevant content.

[0117] Step 2: Document Semantic Segmentation and Semantic Block Generation. Based on the preprocessed text, the document is semantically segmented using syntactic boundaries, punctuation, heading levels, and key RF terminology, dividing the long document into several semantic blocks. Each semantic block typically corresponds to a relatively complete technical description, such as an antenna structure description, a process flow description, or a test scenario description. During segmentation, metadata such as document ID, page number, and paragraph number are appended to each semantic block for subsequent traceability. Figure 5 The multiple texts arranged horizontally in the middle are a semantic block diagram.

[0118] Step 3: Domain Dictionary and Multi-level Prompts. Based on the knowledge system of the RF domain, a domain dictionary is constructed, containing terms such as antennas, RF circuits, interconnect structures, process steps, and test indicators. This dictionary is used to identify the subdomain to which the semantic blocks belong and to supplement the domain information in the prompts. Simultaneously, multi-level prompts are pre-designed for different task objectives, including a task-driven prompt (defining the entity types and relational schemas to be extracted), a domain-guided prompt (embedding RF professional expressions and physical constraint information), and a format and structure constraint prompt (specifying output JSON fields, triple structures, etc.). Figure 5 The “Domain Dictionary” and “Prompt1…Prompt9” modules correspond to the preparation results of this step.

[0119] Step 4: Semantic Matching Extraction and Prompt Scheduling Driven Large Model Call. For each semantic block, the system first uses a domain dictionary for sub-domain identification, and then uses the "Semantic Matching Extraction" module to match key terms, parameters, and structural information in the semantic block with descriptions in a multi-level Prompt library to select a candidate Prompt set. Subsequently, the "Prompt Scheduling Mechanism" selects or combines the most suitable set of Prompts from the candidate set based on the current task type (e.g., structural relation extraction, parameter relation extraction, process chain extraction, etc.), semantic block features, and historical extraction performance. The "Prompts Enhancement" module can automatically supplement contextual information, example triples, or domain explanations to enhance the Prompts. After scheduling is complete, the system sends the "semantic block + enhanced Prompt" as input to the large language model to obtain preliminary entity and relation extraction results. This process... Figure 5 The middle section is represented by a loop arrow and the bottom section by the "Prompts Enhancement / Prompt Scheduling Mechanism / Semantic Matching Extraction" module.

[0120] Step 5: Structure Verification and Normalization. After parsing, the initial output of the large language model enters the "Structure Verification" module. The system checks whether the output conforms to the predefined triple structure and JSON specification, verifies whether entities and relations belong to the allowed type set, and normalizes the units and formats of numerical parameters (such as frequency, gain, and power). Simultaneously, different ways of writing the same entity are standardized, and obviously conflicting or incomplete relations are marked and filtered out. Results that pass structure verification are considered candidate knowledge triples.

[0121] Step 6: Exception Handling and Fault Tolerance Mechanism. For outputs deemed abnormal during structural validation or semantic checking, this invention automatically handles them through the "Exception Handling / Fault Tolerance Mechanism" module. On one hand, the system automatically corrects common format errors (such as missing parentheses or incorrect field order); on the other hand, for cases that cannot be automatically corrected, a policy-based retry process is triggered, including adjusting the Prompt combination, strengthening structural constraints, narrowing the extraction scope, and adding example triples to re-invoke the large model. The exception handling module also records the semantic block where the error occurred, the Prompt used, and the error type, providing a basis for subsequent parameter tuning. Figure 5 The “Exception Handling / Fault Tolerance Mechanism” on the right corresponds to this step.

[0122] Step 7: Knowledge Input, Result Output, and Statistical Analysis. The triples retained after structural verification and fault tolerance mechanisms will be written into the radio frequency domain knowledge base along with their source semantic blocks, page numbers, and document information, forming queryable and reasonable knowledge graph nodes and relationships. Simultaneously, the system can statistically analyze the relationships, entity distribution, and extraction quality indicators in the knowledge graph according to business needs, and... Figure 5 The top chart area displays information in the form of bar charts, etc., to help evaluate the knowledge coverage and model performance in different sub-domains.

[0123] Step 8: Optimization and Iterative Updates Based on Error Logs. The system periodically analyzes the error logs and extracted statistical results recorded by the exception handling module to identify high-frequency error types and weak subdomains. Correspondingly, it adjusts domain dictionary entries, modifies or adds Prompt templates, optimizes Prompt scheduling strategies, and enhances Prompt rules, thus forming a closed-loop optimization process around "semantic matching extraction—Prompt scheduling—exception handling—structural verification". Figure 5 The central circular arrow embodies this iterative update mechanism, enabling the invention to continuously improve its adaptability and extraction accuracy to documents in the radio frequency field during continuous operation.

[0124] To verify the effectiveness and reliability of this application in knowledge extraction tasks in the radio frequency field, systematic experimental verification was conducted on multiple real radio frequency technology document datasets. The experimental documents cover radio frequency circuit descriptions, antenna design documents, test and measurement reports, and manufacturing process technical data, which have typical characteristics such as dense technical terms, strong cross-segment semantic links, and complex structures.

[0125] The experiment first generated 864 semantic blocks from the documents according to the semantic segmentation strategy proposed in this application. These included 54% manufacturing-related documents and 46% antenna and design-related documents. A gold-standard triplet set was manually constructed by domain experts as the evaluation benchmark. Subsequently, based on the multi-level prompting scheduling mechanism, semantic matching extraction, anomaly handling, and structural verification process of this application, knowledge extraction was performed on each semantic block. The results were then compared item by item using expert-annotated entities, relations, and triples to calculate metrics such as precision, recall, and F1 score. Based on this, please refer to... Figure 6 , Figure 6 This is a schematic diagram of the entity recognition and relation extraction results provided in this application, as shown below. Figure 6 As shown, this application achieved a precision of 0.95, a recall of 0.86, and an F1 score of 0.90 in the entity recognition task, and a precision of 0.90, a recall of 0.82, and an F1 score of 0.86 in the relation extraction task. These results are significantly better than baseline methods such as unified Prompt or multiple Prompt (without context enhancement), and many indicators show an improvement of 5%–10%. This demonstrates that the semantic segmentation strategy, domain guidance prompts, and structural verification mechanism in this invention can effectively reduce semantic omissions, relation errors, and structural inconsistencies.

[0126] It should be noted that the F1 score refers to the harmonic mean of precision and recall, which is a commonly used model performance evaluation metric in the fields of machine learning and information retrieval.

[0127] In addition, to evaluate the adaptability of different large language models within the framework of this application, the Qwen-plus model, the DeepSeek model (DeepSeek-R1), and the ChatGPT model were selected for comparative testing. Figure 7 The schematic diagrams illustrating the effects of the radio frequency domain knowledge graph construction method provided in this application on different models are shown in the reference. Figure 7 The results show that the present application maintains stable performance under different models, with all precision, recall and F1 scores above 0.80, and the performance difference between large and medium-sized models is further reduced, verifying the robustness of the present application under the condition of model size variation.

[0128] The knowledge graph constructed in this application for the radio frequency (RF) field is also used in downstream RF concept question answering and short answer tasks. Experimental results show that, compared with the model without a knowledge graph, using the knowledge graph constructed in this application can improve the concept question answering score from 8.78 to 9.06, and the short answer score also shows a significant increase, indicating that structured knowledge has a significant enhancing effect on RF professional question answering tasks. The comprehensive experimental results show that the knowledge extraction quality, output stability, and engineering reliability of this application in complex documents in the RF field have been fully verified, demonstrating its practical application value.

[0129] This application also provides a knowledge graph construction device in the field of radio frequency. Please refer to... Figure 8 The radio frequency domain knowledge graph construction device includes: The semantic segmentation module 10 is used to perform semantic segmentation on the pre-acquired radio frequency text data stream to generate multiple semantic blocks; The semantic matching module 20 is used to perform sub-domain identification on the multiple semantic blocks using a pre-built domain dictionary, and semantically match prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification. The model extraction module 30 is used to call a pre-built large language model, extract entity relations from the semantic block based on the prompt word template, and obtain entity relation extraction results. Verification processing module 40 is used to perform structural verification and fault tolerance mechanism processing on the entity relation extraction results to obtain candidate knowledge triples; The graph construction module 50 is used to construct a radio frequency domain knowledge graph based on the candidate knowledge triples.

[0130] The radio frequency (RF) domain knowledge graph construction apparatus provided in this application, employing the RF domain knowledge graph construction method described in the above embodiments, can solve the technical problems of RF domain knowledge graph construction. Compared with the prior art, the beneficial effects of the RF domain knowledge graph construction apparatus provided in this application are the same as those of the RF domain knowledge graph construction method provided in the above embodiments, and other technical features in the RF domain knowledge graph construction apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0131] This application provides a radio frequency (RF) domain knowledge graph construction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the RF domain knowledge graph construction method in the first embodiment described above.

[0132] The following is for reference. Figure 9This document illustrates a structural schematic diagram of a radio frequency (RF) domain knowledge graph construction device suitable for implementing embodiments of this application. The RF domain knowledge graph construction device in this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 9 The radio frequency domain knowledge graph construction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0133] like Figure 9 As shown, the radio frequency (RF) domain knowledge graph construction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the RF domain knowledge graph construction device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the radio frequency domain knowledge graph construction device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows radio frequency domain knowledge graph construction devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0134] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0135] The radio frequency (RF) domain knowledge graph construction device provided in this application, employing the RF domain knowledge graph construction method described in the above embodiments, can solve the technical problem of RF domain knowledge graph construction. Compared with the prior art, the beneficial effects of the RF domain knowledge graph construction device provided in this application are the same as those of the RF domain knowledge graph construction method provided in the above embodiments, and other technical features in this RF domain knowledge graph construction device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0136] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0137] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0138] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the radio frequency domain knowledge graph construction method in the above embodiments.

[0139] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0140] The aforementioned computer-readable storage medium may be included in a radio frequency domain knowledge graph construction device; or it may exist independently and not be assembled into a radio frequency domain knowledge graph construction device.

[0141] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a radio frequency (RF) domain knowledge graph construction device, the RF domain knowledge graph construction device performs the following: semantic segmentation on a pre-acquired RF text data stream to generate multiple semantic blocks; performs sub-domain identification on each of the multiple semantic blocks using a pre-built domain dictionary; semantically matches prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification; invokes a pre-built large language model to extract entity relations from the semantic blocks based on the prompt word templates, obtaining entity relation extraction results; performs structural verification and fault tolerance mechanism processing on the entity relation extraction results to obtain candidate knowledge triples; and constructs an RF domain knowledge graph based on the candidate knowledge triples.

[0142] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0144] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0145] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described radio frequency domain knowledge graph construction method, thereby solving the technical problem of radio frequency domain knowledge graph construction. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the radio frequency domain knowledge graph construction method provided in the above embodiments, and will not be repeated here.

[0146] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the radio frequency domain knowledge graph construction method described above.

[0147] The computer program product provided in this application can solve the technical problem of constructing knowledge graphs in the radio frequency (RF) field. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the RF field knowledge graph construction method provided in the above embodiments, and will not be repeated here.

[0148] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for constructing a knowledge graph in the radio frequency field, characterized in that, The method for constructing a knowledge graph in the radio frequency field includes: Semantic segmentation is performed on the pre-acquired radio frequency text data stream to generate multiple semantic blocks; Using a pre-built domain dictionary, sub-domain identification is performed on the multiple semantic blocks respectively, and based on the results of the sub-domain identification, semantic matching of prompt word templates is performed from a pre-built multi-level prompt word library; A pre-built large language model is invoked, and entity relations are extracted from the semantic block based on the prompt word template to obtain the entity relation extraction result; The entity relation extraction results are subjected to structural verification and fault tolerance mechanism processing to obtain candidate knowledge triples; A knowledge graph for the radio frequency domain is constructed based on the candidate knowledge triples.

2. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The step of semantically segmenting the pre-acquired radio frequency text data stream to generate multiple semantic blocks also includes the following before: Acquire technical documents in the field of radio frequency; The technical documents in the radio frequency field are encoded, converted, and their layout parsed to obtain the original radio frequency text data stream; The original radio frequency text data stream is filtered to obtain the radio frequency text data stream.

3. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, Before the step of performing sub-domain identification on the multiple semantic blocks using a pre-built domain dictionary, and semantically matching prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification, the following steps are also included: Construct a domain dictionary based on a pre-defined knowledge system in the radio frequency field; Based on the predetermined task objective of extracting radio frequency domain knowledge, a multi-level prompt word is designed. The multi-level prompt word includes task-driven prompt words, domain-guided prompt words, and format-constrained prompt words. Each level of prompt word contains multiple prompt word templates. The multi-level prompt words are categorized and stored to construct a multi-level prompt word library.

4. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The step of performing sub-domain identification on the multiple semantic blocks using a pre-built domain dictionary, and semantically matching prompt word templates from a pre-built multi-level prompt word library based on the results of the sub-domain identification includes: The domain dictionary is used to identify sub-domains for each of the multiple semantic blocks to obtain the core sub-domain corresponding to each semantic block. The key features of the semantic block are analyzed, and semantic matching is performed in the multi-level prompt word library based on the core sub-domain and the key features to obtain a set of candidate prompt word templates. The optimal prompt word template is selected from the candidate prompt word template set by combining the key features, historical extraction performance, and current task type.

5. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The step of calling a pre-built large language model, extracting entity relations from the semantic block based on the prompt word template, and obtaining the entity relation extraction result includes: The contextual relationships of the prompt word template are supplemented, and example triples and domain explanations are added to the prompt word template to obtain the enhanced prompt word template: The pre-built large language model is invoked, and the enhanced prompt word template and the semantic block are input into the pre-built large language model for entity relation extraction to obtain the entity relation extraction result.

6. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The step of performing structural verification and fault-tolerance processing on the entity relation extraction results to obtain candidate knowledge triples includes: From the entity relation extraction results, knowledge triples that conform to both the predefined triple structure and the predefined semantic specification are selected as the first candidate knowledge triples. The knowledge triples in the entity relation extraction results that do not conform to the predefined triple structure or the predefined semantic specification are regarded as the fourth candidate knowledge triples. Verify whether the first candidate knowledge triple belongs to a pre-set set of allowed types; If the first candidate knowledge triple belongs to the allowed type set, then the numerical parameters in the first candidate knowledge triple are formatted and the units are standardized, and the entity writing of the first candidate knowledge triple is unified to obtain the second candidate knowledge triple. Delete conflicting or incomplete relations in the second candidate knowledge triplet to obtain the third candidate knowledge triplet; Based on a pre-set fault tolerance mechanism, the fourth candidate knowledge triplet is repaired to obtain the fifth candidate knowledge triplet. The candidate knowledge triplet is obtained based on the third candidate knowledge triplet and the fifth candidate knowledge triplet.

7. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The steps for constructing a radio frequency domain knowledge graph based on the candidate knowledge triples include: The candidate knowledge triples and the semantic blocks corresponding to the candidate knowledge triples are written into the radio frequency domain knowledge base to form knowledge graph nodes and knowledge graph relationships. A radio frequency domain knowledge graph is constructed based on the knowledge graph nodes and the knowledge graph relationships.

8. The method for constructing a knowledge graph in the radio frequency field as described in claim 1, characterized in that, The step of constructing a radio frequency domain knowledge graph based on the candidate knowledge triples further includes: Obtain the error logs and extract statistical results generated by the fault tolerance mechanism; Analyze the error logs and the extracted statistical results to identify high-frequency error types and weak sub-regions; Adjust the domain dictionary and the multi-level prompt word library according to the high-frequency error types and the weak sub-regions; The execution steps are returned based on the adjusted domain dictionary and multi-level prompt word library: sub-domain identification is performed on the multiple semantic blocks using the pre-built domain dictionary, and prompt word templates are semantically matched from the pre-built multi-level prompt word library based on the results of the sub-domain identification.

9. A knowledge graph construction device in the field of radio frequency, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the radio frequency domain knowledge graph construction method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the radio frequency domain knowledge graph construction method as described in any one of claims 1 to 8.

Images