Semantic labeling method, device and equipment for power system operation project and medium

Abstract

The application discloses a kind of semantic annotation methods, device, equipment and medium of power system operation project, the method includes the text preprocessing of the document data of power system operation project, obtains the semantic label result at Token level;According to semantic label result and text fragment set, generate triple set using preset matching rule template;Weak label assignment and logic compliance check are carried out to each keyword in triple set, and weak label candidate entity set is obtained;Each candidate entity in weak label candidate entity set is respectively labeled by using the preset fusion multi-model, and the labeling result of each candidate entity is determined according to the majority voting and confidence weighting mechanism of each model label prediction result.Relying on Token level semantic label, cross-paragraph complex semantic analysis is realized, the precision of power professional element extraction is improved, and the result is self-checked and the accuracy of the result is improved by multi-model voting and confidence weighting check.

Description

Technical Field

[0001] This invention relates to the field of semantic annotation technology, and in particular to a semantic annotation method, apparatus, equipment and medium for power system operation projects. Background Technology

[0002] In the management and evaluation of power system operation projects, a large amount of document data is characterized by unstructured format, semantic density, and a high concentration of technical terms. Common document types include feasibility study reports, project approval materials, review minutes, and acceptance summaries. These texts contain important information such as key performance indicators (KPIs), implementation phases, task milestones, and resource dependencies. If this information can be effectively extracted and structured, it will provide a foundation for subsequent knowledge graph construction and intelligent project evaluation models.

[0003] Currently, the mainstream technologies for information extraction and annotation mainly include named entity recognition (NAME) methods, template rule matching, deep learning extraction models, and automatic annotation systems for general domains. However, NAME-based methods struggle to accurately identify complex semantic relationships across sentences or paragraphs, especially in texts involving temporal and causal logic, such as those expressing project phase sequences, where they are prone to distortion. Template rule matching has poor generalization ability and easily fails when faced with slightly different document structures or expressions, limiting its application scope. Deep learning extraction models have low sensitivity to domain knowledge, typically relying on large amounts of labeled corpora, resulting in high training costs and difficulties in model transfer. Automatic annotation systems for general domains generally lack self-verification capabilities, making it difficult to backtrack and correct extraction errors or logical conflicts, leading to uncontrollable accuracy. Summary of the Invention

[0004] This invention provides a semantic annotation method, device, equipment, and medium for power system operation projects. Relying on token-level semantic tags, it achieves complex semantic parsing across paragraphs, improves the accuracy of extracting power professional elements, and overcomes the shortcomings of traditional rules in generalization by using task triples to unify structured semantics. Weak tag assignment combined with logical verification eliminates the need for large-scale manual annotation, reduces annotation costs, and filters out logical errors in advance. Multi-model voting and confidence-weighted verification compensate for the shortcomings of single models, achieves self-verification of results, and improves the accuracy of results.

[0005] To achieve the above objectives, embodiments of the present invention provide a semantic annotation method for power system operation items, including: Text preprocessing is performed on the document data of power system operation projects to obtain semantic tag results of the document data at the token level; Based on the semantic tag results and the set of text fragments in the document data, a set of triples in the document data is generated using a preset matching rule template; Weak tag assignment and logical compliance verification are performed on each keyword in the triplet set to obtain a weak tag candidate entity set of the document data. A preset fusion multi-model is used to predict the label of each candidate entity in the weakly labeled candidate entity set. Based on the label prediction results of each model, a majority voting and confidence weighting mechanism is used to determine the labeling result of each candidate entity.

[0006] As an improvement to the above solution, the text preprocessing of the document data of the power system operation project to obtain the semantic tag results of the document data at the token level includes: Obtain documentation data for power system operation projects; Text preprocessing, including word segmentation, part-of-speech tagging, and named entity recognition, is performed on the document data to obtain the word segmentation sequence, part-of-speech tagging, and named entity recognition results of the document data; Based on the word segmentation sequence, a semantic vector for each word is generated using a preset domain semantic model; Based on the part-of-speech tagging, named entity recognition results, and semantic vectors, the sentence segment to which each word belongs is determined, resulting in a set of text fragments of the document data. Based on the global semantic vector of the sentence segment to which each word belongs, the stage classification label of the sentence segment is determined, and the semantic label result of the document data at the token level is obtained.

[0007] As an improvement to the above scheme, the step of generating a set of triples for the document data based on the semantic tag results and the set of text fragments of the document data using a preset matching rule template includes: The TF-IDF method is used to perform word frequency modeling on the set of text segments of the document data to obtain a set of keyword vectors for each text segment; KMeans clustering is performed on the keyword vector set, and semantic decomposition is performed on each clustered text fragment in the clustered text fragment set. Based on the semantic decomposition results and the stage classification labels of the corresponding sentence segments in the semantic tags, the task triplet of the document data is generated; The task triples are modified using a preset matching rule template to obtain a set of triples for the document data.

[0008] As an improvement to the above scheme, the step of assigning weak tags and performing logical compliance checks on each keyword in the triplet set to obtain the weak tag candidate entity set of the document data includes: A preliminary set of weakly labeled candidate entities is obtained by assigning weak labels to each keyword in the triplet set using a pre-defined dictionary of terms in the power field. The preliminary weak tag candidate entity set is logically compliantly verified using a preset business rule base to obtain the weak tag candidate entity set of the document data.

[0009] As an improvement to the above scheme, the step of using a preset fusion multi-model to predict the label of each candidate entity in the weakly labeled candidate entity set, and determining the labeling result of each candidate entity based on the label prediction results of each model using a majority voting and confidence weighting mechanism, includes: A preset fusion multi-model is used to predict the label of each candidate entity in the weakly labeled candidate entity set, and the label prediction results of each candidate entity are obtained. The preset fusion multi-model includes a BERT-CRF sequence labeling model, a rule matching model and a graph logic reasoning model. The final label for each candidate entity is determined by majority voting based on the label prediction results of each model. The label confidence of each candidate entity is calculated using a confidence weighting mechanism based on the predicted probability of the final label. The final labels are self-verified and corrected based on the label confidence level and the preset confidence threshold to obtain the labeling results for each candidate entity.

[0010] As an improvement to the above scheme, the step of performing self-verification and correction on the final label based on the label confidence level and a preset confidence threshold to obtain the labeling result for each candidate entity includes: If the confidence level of the label is not lower than the preset confidence threshold, then the final label will be used as the labeling result of the corresponding candidate entity. If the confidence level of the label is lower than the preset confidence threshold, a preset verification strategy is adopted to perform self-verification and correction on the corresponding candidate entity, and a new final label for the corresponding candidate entity is obtained. The new final label is used as the labeling result of the corresponding candidate entity. The preset verification strategy includes a manual review mechanism, a rule secondary judgment, and a context re-encoding mechanism.

[0011] As an improvement to the above scheme, after determining the annotation result of each candidate entity, the method further includes: The annotation results are structurally transformed using a graph structure algorithm to obtain the structured semantic annotation results for each candidate entity.

[0012] To achieve the above objectives, embodiments of the present invention provide a semantic annotation device for power system operation items, comprising: The semantic tag generation module is used to perform text preprocessing on the document data of power system operation projects to obtain the semantic tag results of the document data at the Token level. The ternary combination generation module is used to generate a set of ternaries for the document data based on the semantic tag results and the set of text fragments of the document data, using a preset matching rule template. The candidate entity generation module is used to perform weak tag assignment and logical compliance verification on each keyword in the triple set to obtain a weak tag candidate entity set of the document data. The labeling result determination module is used to perform label prediction for each candidate entity in the weak label candidate entity set using a preset fusion multi-model, and to determine the labeling result of each candidate entity using a majority voting and confidence weighting mechanism based on the label prediction results of each model.

[0013] To achieve the above objectives, this invention provides a semantic annotation device for power system operation items, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the semantic annotation method for power system operation items.

[0014] To achieve the above objectives, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the semantic annotation method for the power system operation items.

[0015] Compared with existing technologies, the semantic annotation method, apparatus, equipment, and medium for power system operation projects disclosed in this invention preprocesses the document data of power system operation projects to obtain semantic tag results at the token level. Based on the semantic tag results and a set of text fragments of the document data, a set of triples for the document data is generated using a preset matching rule template. Weak tag assignment and logical compliance verification are performed on each keyword in the triple set to obtain a set of weak tag candidate entities for the document data. A preset fusion multi-model is used to predict the tag of each candidate entity in the weak tag candidate entity set. Based on the tag prediction results of each model, a majority voting and confidence-weighted mechanism is used to determine the annotation result of each candidate entity. Relying on token-level semantic tags, complex semantic parsing across paragraphs is achieved, improving the accuracy of power industry element extraction. The unified structured semantics through task triples overcomes the shortcomings of insufficient generalization ability of traditional rules. Weak tag assignment combined with logical verification eliminates the need for large-scale manual annotation, reducing annotation costs and filtering out logical errors in advance. Multi-model voting and confidence-weighted verification compensate for the shortcomings of single models, achieving self-verification of results and improving accuracy. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating a semantic annotation method for power system operation items provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a semantic annotation device for power system operation items provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of a semantic annotation device for power system operation projects provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that the terms "comprising" and "specific" in this invention, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0019] Please see Figure 1 , Figure 1 This is a flowchart illustrating a semantic annotation method for power system operation items provided in an embodiment of the present invention. The semantic annotation method for power system operation items includes: S1, perform text preprocessing on the document data of the power system operation project to obtain the semantic tag results of the document data at the Token level; S2, Based on the semantic tag results and the set of text fragments of the document data, generate a set of triples of the document data using a preset matching rule template; S3, perform weak tag assignment and logical compliance verification on each keyword in the triplet set to obtain the weak tag candidate entity set of the document data; S4. A preset fusion multi-model is used to predict the label of each candidate entity in the weak label candidate entity set. Based on the label prediction results of each model, a majority voting and confidence weighting mechanism is used to determine the labeling result of each candidate entity.

[0020] For example, the method for obtaining the preset matching rule template includes collecting historical documents of the entire process of power system operation projects, such as project initiation, implementation, grid connection, commissioning, and acceptance, and statistically analyzing high-frequency business phrases; based on the high-frequency business phrases, extracting business logic such as the sequence of stages and equipment-action combinations according to power industry standards; and encapsulating the high-frequency business phrases using regular expressions and syntactic constraints according to the business logic to form a reusable preset matching rule template. The preset fusion multi-model is obtained by fusing a BERT-CRF-based sequence labeling model, a rule matching model, and a graph logic model.

[0021] Specifically, step S1 includes: S11, Obtain document data for power system operation projects; S12, perform text preprocessing on the document data, including word segmentation, part-of-speech tagging, and named entity recognition, to obtain the word segmentation sequence, part-of-speech tagging, and named entity recognition results of the document data; S13, Generate the semantic vector of each word using a preset domain semantic model based on the word segmentation sequence; S14, determine the sentence segment to which each word belongs based on the part-of-speech tagging, named entity recognition results and semantic vector, and obtain a set of text segments of the document data; S15, determine the stage classification label of the sentence segment to which each word belongs based on the global semantic vector of the sentence segment to which it belongs, and obtain the semantic label result of the document data at the Token level.

[0022] For example, document data for power system operation projects typically contains a large number of technical terms and domain-specific sentence structures, making it difficult for general semantic models to accurately capture their contextual relationships. To improve the model's semantic understanding capability in power system professional documents, a domain-specific semantic model is constructed based on a power system professional corpus and pre-trained language models (such as BERT, RoBERTa, etc.) are fine-tuned for the domain. This model is used to understand the complex contextual semantic structure in power project documents, improving the stage-specific context awareness capability of downstream extraction tasks. For instance, a semantic modeling system for scheduling operation project documents is constructed based on a Transformer-structured pre-trained language model (using RoBERTa as the underlying architecture in this embodiment) and named the DispatchBERT model (Dispatching Bidirectional Encoder Representations from Transformers). WordPiece is used to segment the project documents to obtain the segmented word sequence. The data is then fed into the DispatchBERT model, where context modeling is performed through multiple Transformers to obtain a sequence of semantic vectors. Based on the semantic vector sequence, each word in the segmented sequence undergoes part-of-speech tagging (POS) and named entity recognition (NER) preprocessing. Combined with project lifecycle information, a softmax classification layer is added to the [CLS] bit vector (global vector) output by BERT. A stage discrimination function is constructed to classify the sentence segment to which each word belongs (e.g., project initiation, design, implementation, acceptance, etc.), resulting in the semantic tagging results of the document data at the token level. The semantic tagging results include the semantic vector of each word, part-of-speech tags (used to determine term categories, action verbs, equipment nouns, etc.), named entity tags (i.e., named entity annotation results and their entity categories, used to identify task names, key equipment, location information, etc.), the level stage tag of the paragraph to which it belongs (used to annotate the lifecycle stage, such as scheme design, commissioning preparation, grid connection implementation, acceptance review, etc.), and the original word's position information and hierarchical index in the document (for easy cross-sentence citation). The semantic vector sequence... The expression is: , The stage discrimination function is: , In the formula, The number of tokens in the word segmentation sequence of the project document; Indicates the first The semantic vector of each word element; It is the set of real numbers; For the embedded dimension; For paragraph CLS global semantic vector The predicted classification results; For activation functions; This is the weight matrix of the stage classifier; It is the bias vector; The number of project phase categories set.

[0023] Specifically, step S2 includes: S21, The TF-IDF method is used to perform word frequency modeling on the text segment set of the document data to obtain the keyword vector set for each text segment; S22, perform KMeans clustering on the keyword vector set, and perform semantic decomposition on each clustered text fragment in the clustered text fragment set; S23, Generate the task triplet of the document data based on the semantic decomposition results and the stage classification labels of the corresponding sentence segments in the semantic tags; S24, the task triples are modified using a preset matching rule template to obtain a set of triples for the document data.

[0024] For example, it's worth noting that document data typically contains multiple instructional sentences, task descriptions, or event narratives. These contents are often expressed using irregular syntax, such as "the main transformer cutover is planned for September" or "protection switching needs to be completed before the new system is put into operation." To identify the potential project behavioral logic, structural pattern aggregation and rule enhancement processing are required in the vector space. Therefore, to achieve a deep understanding of the semantic structure of project tasks, based on the semantic vector space and stage labels, logical pattern recognition methods are used to extract structural information from the document data, including elements such as action behaviors, stage anchors, resource constraints, and time nodes. For instance, core semantic units with task-expressive properties are extracted from the document, constructing structurally readable triple representations to form a semantic mapping relationship between task names, project stages, and action behaviors.

[0025] In a specific embodiment, the TF-IDF (Term Frequency-Inverse Document Frequency) method is used to perform term frequency modeling on sentence-level text segments to obtain keyword vector representations for each segment. KMeans clustering is performed on the vector set to obtain semantically similar sentence groups, thus initially filtering out text segments with the same behavioral structure features. Semantic decomposition is then performed on each clustered text segment in the resulting text segment set. For example, dependency parsing is used to identify the main verbs in the sentence (representing operational actions, such as "switching," "replacing," "removing," and "debugging"); equipment-related nouns or task objectives that form the agent structure with the verbs (such as "protection device," "main transformer," and "fiber optic link") are extracted; and the lifecycle nodes of the task (such as "design," "implementation," and "commissioning") are determined based on the stage labels output from the preprocessing stage. Task-stage-action task triples are constructed based on the semantic decomposition results and the stage classification labels of the corresponding sentence segments in the semantic labels. This triple structure achieves preliminary semantic grouping through TF-IDF feature construction and the K-Means algorithm, and introduces rule matching to enhance the recognition of specific semantic expressions. The task triples are then corrected using a preset matching rule template to obtain the triple set of the document data. For example, by using a preset matching rule template, the actions "switching" and "debugging" are identified, with the task objective being "original protection equipment." The dependency "must be completed after commissioning" is also identified based on the contextual sentence structure, thus completing the stage anchor point "commissioning." This preset matching rule template is implemented using regular expressions combined with contextual syntactic tree analysis, possessing strong flexibility and cross-sentence integration capabilities, and can handle common non-linear expression structures in power project documents. The expression for the task triple is as follows: , in, Represents the task object entity. Indicates the current stage of the task. This indicates the corresponding action behavior (such as "cut-in", "switch", "debug" etc.).

[0026] Specifically, step S3 includes: S31, Weak label values ​​are assigned to each keyword in the triplet set using a preset power field terminology dictionary to obtain a preliminary weak label candidate entity set; S32, the preliminary weak tag candidate entity set is logically compliantly verified using a preset business rule base to obtain the weak tag candidate entity set of the document data.

[0027] For example, in the processing scenario of power system operation project documents, the domain corpus is highly specialized and time-sensitive, making it difficult to rely on manually annotated corpora from general open domains for supervised training. To reduce manual costs and improve the deployability of the model in actual business, a weakly supervised annotation mechanism is introduced. This mechanism uses a pre-set power domain terminology dictionary and business rule base to initially annotate the text, and combines this with a semantic feedback strategy to form a self-correcting closed loop, constructing a high-quality training sample set. For instance, by traversing the pre-set power domain terminology dictionary, weak labels are assigned to each keyword in the triplet set; the pre-set power domain terminology dictionary matching uses a fuzzy matching method with word form restoration to ensure compatibility with surface variations of different expression styles and professional subdomain terms. Weak labels are defined as preliminary labels that do not rely on manual confirmation and can be automatically generated by the rule system, used as input for subsequent model training or label confidence evaluation. After generating weak labels, to improve label accuracy and semantic consistency and avoid annotation errors caused by rule deviations or contextual changes, a pre-set business rule base is obtained from the knowledge expression logic abstracted from power system operation management processes, dispatch operation specifications, and related institutional documents. The system performs logical compliance verification on the initial weak label candidate entity set using a preset business rule base to obtain the weak label candidate entity set of the document data. This includes: if a sentence contains keywords such as "dispatch command" or "operation ticket," the system needs to verify whether there is a clear resource description item (such as "main transformer," "line," or "protection device") in the same paragraph. If no relevant resource entity is identified, a warning flag is triggered and the current labeling is rejected; if the stage label "commissioning" is identified in the task triplet, but the timestamp of its appearance is earlier than the corresponding time of the "grid connection" stage in the document, it is considered a logical conflict, triggering time sequence verification feedback and marking the task as "time sequence abnormal"; if the task action is marked as "debugging" and appears in the "acceptance" stage paragraph, the system will revert to the original stage classification judgment according to the rules and perform a reclassification logic operation. This allows for self-learning, self-evaluation, and weak label correction without large-scale manual supervision, significantly reducing the cost of constructing initial training samples. Each keyword... Corresponding weak tags Generate according to the following rules: , Among them, weak labels This indicates that the keyword is considered by the system as a candidate tag with business semantic value; This indicates a pre-defined dictionary of terms in the power industry.

[0028] Specifically, step S4 includes: S41, a preset fusion multi-model is used to predict the label of each candidate entity in the weak label candidate entity set, and the label prediction results of each candidate entity are obtained; wherein, the preset fusion multi-model includes BERT-CRF sequence labeling model, rule matching model and graph logic reasoning model. S42, Based on the label prediction results of each model, the final label of each candidate entity is determined by majority voting; S43, the confidence level of each candidate entity is calculated using a confidence weighting mechanism based on the predicted probability of the final label; S44, perform self-verification and correction on the final label based on the label confidence level and the preset confidence threshold to obtain the labeling result for each candidate entity.

[0029] For example, to further improve the accuracy and controllability of structured extraction, a pre-defined fusion multi-model is introduced to integrate the label results from different extraction methods. The final output is determined through majority voting and confidence weighting mechanisms. Self-verification criteria are set to automatically correct low-confidence samples and review business rules to ensure the stability and error tolerance of the overall extraction chain.

[0030] In this embodiment, the preset fusion multiple model consists of a BERT-CRF-based sequence labeling model. Matching model based on rule expression And reasoning models based on graph logical dependency structures The three components generate label prediction results from the perspectives of lexical context features, explicit rule matching relationships, and semantic graph reasoning, respectively. Let a certain entity be... The label prediction results given by each model are as follows: , in, , indicating the first The label output (label prediction result) of each model. For tag spaces (e.g., "task entity", "stage identifier", "resource item").

[0031] Final Tag Determined through a majority voting mechanism, the majority voting expression is: , This involves counting the frequency of each label across the three models and selecting the label with the highest frequency as the final decision. In case of a tie (i.e., equal votes), a weighted composite score is calculated by referring to the confidence scores of each model to aid in decision-making.

[0032] The model confidence score is obtained from the predicted probability output by the softmax layer. Taking a sequence labeling model as an example, let the unnormalized score output by the model be... Then the first Confidence of each model for: , The final label confidence score It can be represented as: , in, Indicates the entity as No. The predicted probability (confidence score) of each model. This is the preset number of models for fusing multiple models.

[0033] For cases where the confidence scores differ significantly between models, the degree of divergence is recorded for subsequent dynamic weighted learning updates. For example, let's assume... The output shows a confidence score of 0.92 for a certain label, while If the output is 0.41, the final label will be determined by fusion using the following weighted decision method: , in, For the first The dynamic reliability weights of each model are adjusted in real time based on historical accuracy and current contextual features.

[0034] More specifically, step S44 includes: S441, if the confidence level of the label is not lower than the preset confidence threshold, then the final label is used as the labeling result of the corresponding candidate entity; S442, if the confidence level of the label is lower than the preset confidence threshold, a preset verification strategy is adopted to perform self-verification and correction on the corresponding candidate entity to obtain a new final label for the corresponding candidate entity, and the new final label is used as the labeling result of the corresponding candidate entity; wherein, the preset verification strategy includes a manual review mechanism, a rule secondary judgment and a context re-encoding mechanism.

[0035] For example, when the confidence level of the label is lower than a preset confidence threshold... ,Right now: , If the final label judgment of an entity is deemed to have uncertainty or ambiguity, the entity undergoes self-verification and correction to obtain a new final label for the corresponding candidate entity. The manual review mechanism involves sending the final label and its contextual information to the front-end interface for review and confirmation by business personnel. The second rule judgment involves reloading the action dependency rules and performing logical re-reasoning based on the syntactic dependency tree. The context re-encoding mechanism re-encodes and evaluates the relevant sentence segments and their context for the corresponding candidate entity, using BERT model [CLS] vectors and projection layers to re-predict the label category. This fusion mechanism and self-verification correction significantly reduce prediction conflicts between models and improve the consistency and business interpretability of the extracted output.

[0036] Furthermore, after determining the annotation result for each candidate entity, the method further includes: S5. The annotation results are structurally transformed using a graph structure algorithm to obtain the structured semantic annotation results for each candidate entity.

[0037] For example, the final confirmed annotation results are transformed into a unified structured representation format, outputting machine-parseable knowledge fragments. These structured fragments can serve as the basic units for knowledge graph construction, and can also be directly injected into the large-scale project evaluation model for semantic enhancement, improving its reasoning and context awareness capabilities. The structured representation adopts a common data modeling format, such as JSON Schema or RDF (Resource Description Framework) graph structure, to achieve unified encapsulation of three types of semantic elements: entities, attributes, and relationships, using extracted task triples. The core structure organizes labeled information according to the following semantic structure: Entity: task objectives, project stages, equipment objects, scheduling behaviors, etc.; Attribute: execution time, responsible unit, task ID, urgency level, document ID of the context, etc.; Relation: inter-task dependencies, stage inclusion relationships, operation sequence, concurrency constraints, interactive linkage, etc. This structural representation supports direct import into graph databases (such as Neo4j, GraphDB) to construct a full-process knowledge graph for power projects. The "task" serves as the central node, connected to stages, resources, actions, time, and other nodes via directed edges, supporting intelligent management tasks such as project dependency mining, risk prediction, and stage plan evaluation. In large-scale model evaluation applications, structured knowledge fragments are encoded as semantic vectors and concatenated with the context embeddings of the original documents to support the model's understanding of task logic, inter-stage causality, resource conflicts, and other issues. The embedding vectors for model input are defined as follows: , in, Textual semantic embedding for the original document; Entity representation of structured knowledge fragments after processing by a graph encoder (such as GCN or GAT); This is a concatenation operation used to merge semantic dimensions.

[0038] In this way, the project evaluation model can achieve more refined task evaluation, schedule prediction, risk warning and resource collaborative optimization with rich structured semantic support, thereby supporting intelligent management applications throughout the entire process.

[0039] This invention uses unstructured text from power system operation projects as the information source. Relying on domain semantic modeling, weak labeling mechanisms, multi-model fusion, and self-verification processes, it ultimately achieves high-precision automatic extraction and structured expression of task information, forming a stable and usable knowledge data output path. This not only reduces the cost of manual intervention and improves the quality of information extraction, but also provides strong semantic support for knowledge graph construction and project evaluation models, demonstrating significant technological innovation value and engineering application prospects.

[0040] This invention discloses a semantic annotation method for power system operation projects. The method involves preprocessing the document data of the power system operation projects to obtain semantic tags at the token level. Based on the semantic tags and a set of text fragments in the document data, a set of triples for the document data is generated using a preset matching rule template. Each keyword in the triple set is assigned a weak tag and undergoes logical compliance verification to obtain a set of weak-tag candidate entities. A preset multi-model fusion is used to predict the tags of each candidate entity in the weak-tag candidate entity set. Based on the tag prediction results of each model, a majority voting and confidence-weighted mechanism is used to determine the annotation result of each candidate entity. Relying on token-level semantic tags, this method achieves complex semantic parsing across paragraphs, improving the accuracy of power industry element extraction. By using task triples to unify structured semantics, it overcomes the shortcomings of traditional rules in terms of generalization ability. Weak tag assignment combined with logical verification eliminates the need for large-scale manual annotation, reducing annotation costs and filtering out logical errors in advance. Multi-model voting and confidence-weighted verification compensate for the shortcomings of single models, enabling self-verification of results and improving accuracy.

[0041] See Figure 2 , Figure 2 This is a schematic diagram of the structure of a semantic annotation device 10 for power system operation items provided in an embodiment of the present invention. The semantic annotation device 10 for power system operation items includes: Semantic tag generation module 11 is used to perform text preprocessing on the document data of power system operation projects to obtain semantic tag results of the document data at the Token level. The ternary combination generation module 12 is used to generate a set of ternaries for the document data based on the semantic tag results and the set of text fragments of the document data, using a preset matching rule template. The candidate entity generation module 13 is used to perform weak tag assignment and logical compliance verification on each keyword in the triple set to obtain the weak tag candidate entity set of the document data. The labeling result determination module 14 is used to perform label prediction for each candidate entity in the weak label candidate entity set using a preset fusion multi-model, and to determine the labeling result of each candidate entity using a majority voting and confidence weighting mechanism based on the label prediction results of each model.

[0042] Furthermore, the semantic annotation device 10 for power system operation items also includes: The annotation result conversion module is used to perform a structured conversion of the annotation results using a graph structure algorithm to obtain a structured semantic annotation result for each candidate entity.

[0043] The semantic annotation device 10 for power system operation items provided in this embodiment of the invention can realize all the processes of the semantic annotation method for power system operation items in the above embodiment. The functions and technical effects of each module in the device are the same as the functions and technical effects of the semantic annotation method for power system operation items in the above embodiment, and will not be repeated here.

[0044] See Figure 3 , Figure 3 This is a schematic diagram of the structure of a semantic annotation device 20 for power system operation items provided in an embodiment of the present invention. The semantic annotation device 20 for power system operation items in this embodiment includes: a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, it implements the steps in the above-described embodiment of the semantic annotation method for power system operation items. Alternatively, when the processor 21 executes the computer program, it implements the functions of each module in the above-described embodiment of the semantic annotation device for power system operation items.

[0045] For example, the computer program may be divided into one or more modules, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the semantic annotation device 20 of the power system operation project.

[0046] The semantic annotation device 20 for power system operation items can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The semantic annotation device 20 for power system operation items may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of the semantic annotation device 20 for power system operation items and does not constitute a limitation on the device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the semantic annotation device 20 for power system operation items may also include input / output devices, network access devices, buses, etc.

[0047] The processor 21 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor 21 is the control center of the semantic annotation device 20 for the power system operation project, connecting various parts of the semantic annotation device 20 for the entire power system operation project via various interfaces and lines.

[0048] The memory 22 can be used to store the computer programs and / or modules. The processor 21 implements various functions of the semantic annotation device 20 for the power system operation project by running or executing the computer programs and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0049] If the semantic annotation device 20 for the power system operation project integrates a module that is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by the processor 21, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted from the content as required by the legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium may not include electrical carrier signals and telecommunication signals.

[0050] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0051] This invention also provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the semantic annotation method for power system operation items as described in the above embodiments.

[0052] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A semantic annotation method for power system operation items, characterized in that, include: Text preprocessing is performed on the document data of power system operation projects to obtain semantic tag results of the document data at the token level; Based on the semantic tag results and the set of text fragments in the document data, a set of triples in the document data is generated using a preset matching rule template; Weak tag assignment and logical compliance verification are performed on each keyword in the triplet set to obtain a weak tag candidate entity set of the document data. A preset fusion multi-model is used to predict the label of each candidate entity in the weakly labeled candidate entity set. Based on the label prediction results of each model, a majority voting and confidence weighting mechanism is used to determine the labeling result of each candidate entity.

2. The semantic annotation method for power system operation items as described in claim 1, characterized in that, The text preprocessing of the document data for power system operation projects yields semantic tagging results at the token level, including: Obtain documentation data for power system operation projects; Text preprocessing, including word segmentation, part-of-speech tagging, and named entity recognition, is performed on the document data to obtain the word segmentation sequence, part-of-speech tagging, and named entity recognition results of the document data; Based on the word segmentation sequence, a semantic vector for each word is generated using a preset domain semantic model; Based on the part-of-speech tagging, named entity recognition results, and semantic vectors, the sentence segment to which each word belongs is determined, resulting in a set of text fragments of the document data. Based on the global semantic vector of the sentence segment to which each word belongs, the stage classification label of the sentence segment is determined, and the semantic label result of the document data at the token level is obtained.

3. The semantic annotation method for power system operation items as described in claim 1, characterized in that, The step of generating a set of triples for the document data based on the semantic tag results and the set of text fragments of the document data using a preset matching rule template includes: The TF-IDF method is used to perform word frequency modeling on the set of text segments of the document data to obtain a set of keyword vectors for each text segment; KMeans clustering is performed on the keyword vector set, and semantic decomposition is performed on each clustered text fragment in the clustered text fragment set. Based on the semantic decomposition results and the stage classification labels of the corresponding sentence segments in the semantic tags, the task triplet of the document data is generated; The task triples are modified using a preset matching rule template to obtain a set of triples for the document data.

4. The semantic annotation method for power system operation items as described in claim 1, characterized in that, The process of assigning weak tags and performing logical compliance checks on each keyword in the triplet set yields a set of weak tag candidate entities for the document data, including: A preliminary set of weakly labeled candidate entities is obtained by assigning weak labels to each keyword in the triplet set using a pre-defined dictionary of terms in the power field. The preliminary weak tag candidate entity set is logically compliantly verified using a preset business rule base to obtain the weak tag candidate entity set of the document data.

5. The semantic annotation method for power system operation items as described in claim 1, characterized in that, The process involves using a pre-defined fusion multi-model to predict the label of each candidate entity in the weakly labeled candidate entity set, and then determining the labeling result of each candidate entity based on the label prediction results of each model using a majority voting and confidence-weighted mechanism. This includes: A preset fusion multi-model is used to predict the label of each candidate entity in the weakly labeled candidate entity set, and the label prediction results of each candidate entity are obtained. The preset fusion multi-model includes a BERT-CRF sequence labeling model, a rule matching model and a graph logic reasoning model. The final label for each candidate entity is determined by majority voting based on the label prediction results of each model. The label confidence of each candidate entity is calculated using a confidence weighting mechanism based on the predicted probability of the final label. The final labels are self-verified and corrected based on the label confidence level and the preset confidence threshold to obtain the labeling results for each candidate entity.

6. The semantic annotation method for power system operation items as described in claim 5, characterized in that, The step of performing self-verification and correction on the final labels based on the label confidence level and a preset confidence threshold to obtain the labeling result for each candidate entity includes: If the confidence level of the label is not lower than the preset confidence threshold, then the final label will be used as the labeling result of the corresponding candidate entity. If the confidence level of the label is lower than the preset confidence threshold, a preset verification strategy is adopted to perform self-verification and correction on the corresponding candidate entity, and a new final label for the corresponding candidate entity is obtained. The new final label is used as the labeling result of the corresponding candidate entity. The preset verification strategy includes a manual review mechanism, a rule secondary judgment, and a context re-encoding mechanism.

7. The semantic annotation method for power system operation items as described in claim 1, characterized in that, After determining the annotation result for each candidate entity, the method further includes: The annotation results are structurally transformed using a graph structure algorithm to obtain the structured semantic annotation results for each candidate entity.

8. A semantic annotation device for power system operation items, characterized in that, include: The semantic tag generation module is used to perform text preprocessing on the document data of power system operation projects to obtain the semantic tag results of the document data at the Token level. The ternary combination generation module is used to generate a set of ternaries for the document data based on the semantic tag results and the set of text fragments of the document data, using a preset matching rule template. The candidate entity generation module is used to perform weak tag assignment and logical compliance verification on each keyword in the triple set to obtain a weak tag candidate entity set of the document data. The labeling result determination module is used to perform label prediction for each candidate entity in the weak label candidate entity set using a preset fusion multi-model, and to determine the labeling result of each candidate entity using a majority voting and confidence weighting mechanism based on the label prediction results of each model.

9. A semantic annotation device for power system operation items, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the semantic annotation method for power system operation items as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the semantic annotation method for power system operation items as described in any one of claims 1-7.

Images