Power information extraction method and device and computer equipment

Abstract

The application discloses an electric power information extraction method, device and computer equipment. The method comprises the following steps: recognizing key characters in an industry expansion document by using a multi-modal large model, and performing annotation processing on the key characters; incrementally training an expert model based on the annotated key characters; in response to an information extraction request of a target document uploaded by a user, inputting the information extraction request into the multi-modal large model to obtain an extraction task; calling an expert model corresponding to the extraction task to process the target document to obtain a text extraction result of the target document. The method can improve the intention understanding ability of complex and fuzzy requirements, help improve the accuracy, robustness and response speed of document information extraction for complex and variable business requirements, and greatly improve the automation degree and efficiency of business processing while ensuring high accuracy.

Description

Technical Field

[0001] This application relates to the field of information extraction technology, and in particular to a method, apparatus and computer equipment for extracting power information. Background Technology

[0002] In physical business operations, processes such as business expansion applications, change requests, and property rights verification often involve a large number of certificates and licenses, including ID cards, business licenses, real estate certificates, and powers of attorney. These documents come from diverse sources and have different formats, containing both image information and complex text content. Furthermore, some documents are prone to blurring, tilting, and noise interference due to factors such as shooting angle, lighting, and resolution, which increases the workload and risk of errors for manual review.

[0003] Traditional data entry methods can only achieve single-field recognition or simple OCR extraction. For certificate documents with complex document types and large differences in field rules, there is a lack of end-to-end intelligent processing capabilities that integrate data collection, preprocessing, recognition, semantic understanding, task scheduling, and result management, resulting in problems such as insufficient data accuracy and insufficient depth of business semantic understanding. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, and computer device for extracting power information to solve the problems mentioned in the background art.

[0005] According to a first aspect of this application, a method for extracting electricity information is provided, the method comprising: The key characters in the business expansion document are identified using a multimodal large model, and the key characters are annotated. The annotation process includes field type annotation, position association annotation, and business attribute annotation. Incremental training of the expert model is performed based on the labeled key characters, wherein the expert model includes a document type classification model, a field extraction model, a relation recognition model, and a calibration model; In response to an information extraction request for a target document uploaded by a user, the information extraction request is input into the multimodal large model to obtain an extraction task; The expert model corresponding to the extraction task is invoked to process the target document, and the text extraction result of the target document is obtained.

[0006] According to a second aspect of this application, an apparatus for extracting power information is provided, the apparatus comprising: The preprocessing module is used to perform image enhancement, layout structure analysis, and preprocessing quality assessment on diverse business expansion document samples. The sample training module is used to identify key characters in the business expansion document using a multimodal large model, and to perform annotation processing on the key characters, wherein the annotation processing includes field type annotation, position association annotation and business attribute annotation; and to incrementally train the expert model based on the annotated key characters, wherein the expert model includes a document type classification model, a field extraction model, a relationship recognition model and a calibration model. The information extraction module is used to respond to an information extraction request for a target document uploaded by a user, input the information extraction request into the multimodal large model to obtain an extraction task; and call the expert model corresponding to the extraction task to process the target document to obtain the text extraction result of the target document.

[0007] According to a third aspect of this application, a readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of the above-described method for extracting power information.

[0008] According to a fourth aspect of this application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for extracting power information.

[0009] By employing the aforementioned technical solution, the powerful semantic understanding capabilities of the multimodal large-scale model are first utilized to identify and deeply annotate key characters in the business expansion documents. Then, the high-quality data preprocessed and annotated by the multimodal large-scale model is used for continuous incremental training of the domain expert model. This enables the expert model to adaptively learn and iterate its performance in response to the business logic of the business expansion documents, solving the problems of weak generalization ability and slow adaptation to new business scenarios in traditional models. Furthermore, the multimodal large-scale model and the functionally decoupled expert model are collaboratively scheduled. When a user needs to extract the content of the target document, the multimodal large-scale model is invoked again to automatically parse the user's natural language or business request into a structured extraction task, and the task-driven trained expert model completes the target document processing. This improves the system's ability to understand complex and ambiguous requests, helps improve the accuracy, robustness, and response speed to complex and changing business needs of document information extraction, and significantly improves the automation and efficiency of business processing while ensuring high accuracy.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating the method for extracting power information provided in an embodiment of this application is shown. Figure 2 A structural block diagram of the power information extraction device provided in an embodiment of this application is shown; Figure 3 A schematic diagram of the electronic structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0014] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “attached” to another element, it can be directly connected or attached to the other element, or there may be intermediate elements. Furthermore, “connected” or “attached” as used herein can include wireless connections or wireless interconnections. The term “and / or” as used herein includes all or any unit and all combinations of one or more associated listed items.

[0015] Exemplary embodiments according to this application will now be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments may be implemented in many different forms and should not be construed as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided so that the disclosure of this application is thorough and complete, and that the concept of these exemplary embodiments is fully conveyed to those skilled in the art.

[0016] This embodiment provides a method for extracting power information, such as Figure 1As shown, the method includes: Step 101: Perform image enhancement, layout structure analysis, and preprocessing quality assessment on diverse business expansion document samples.

[0017] Among them, business expansion documents refer to power business expansion application-related documents with a fixed format. For example, customer application documents (such as identity certificate, business license, power of attorney, etc.), power supply plan documents (such as power supply plan response form, power supply plan approval form, etc.), engineering design and construction documents (power supply engineering design drawings, construction contract, equipment inspection report, etc.), and acceptance documents (such as defect rectification notice, acceptance confirmation form, etc.).

[0018] It should be noted that the format of the business expansion document sample can be an editable electronic document, a PDF converted or scanned from an editable document, an image / video file, etc., and this application embodiment does not make specific limitations.

[0019] In this embodiment, the accuracy and efficiency of document digitization are improved from multiple dimensions through image enhancement, layout structure analysis and preprocessing quality assessment, and high-quality data support is provided for subsequent automated information recognition, entry and intelligent analysis.

[0020] In practical application scenarios, step 101, which involves image enhancement, layout structure analysis, and preprocessing quality assessment of diverse business expansion document samples, specifically includes the following steps: Step 101-1: Based on the document source, format, and image quality, classify and organize the business expansion document samples to form multiple sample sets.

[0021] For example, data collection is performed on business expansion documents of various content categories, such as ID cards, business licenses, and real estate certificates. During the collection process, documents are categorized and organized according to indicators such as document source (e.g., scanned copies, mobile phone photos, system-transferred documents), file format (PDF, JPG, PNG, etc.), and image clarity to construct a basic corpus covering the differences in real business scenarios, providing diverse samples for model training. Data collection is performed on samples of business expansion documents such as ID cards, business licenses, and real estate certificates, and the documents are graded and organized according to document source, format, and image quality to establish a basic corpus covering multiple types of scenarios. This allows the model to learn document features from multiple scenarios, improving its applicability in real business situations.

[0022] Step 101-2: Perform image enhancement and structured preprocessing on multiple sample sets.

[0023] The structured preprocessing step is used to ensure that the preprocessed results have a uniform format, which facilitates field recognition by the OCR model. This can include layout analysis and field region segmentation. Image enhancement is used to improve the quality of image / video documents and can include conventional image processing methods such as smoothing filtering, tilt correction, edge sharpening, noise reduction, white balance correction, and abnormal image removal. These will not be elaborated upon in the embodiments of this application. It is understood that if the sample set is not an image / video dataset, the image enhancement step can be omitted.

[0024] Further, in one embodiment, step 101-2 specifically includes: matching the features of the business expansion document sample with the document features of different document content categories in the document template library; if the similarity between the document features in the document template library and the features of the business expansion document sample is greater than the similarity threshold, the document content category to which the document features belong is taken as the document content category of the business expansion document sample, and a key field region box is generated on the business expansion document sample based on the document content category of the business expansion document sample; if the document features in the document template library do not match the features of the business expansion document sample, the main structural blocks of the business expansion document sample are extracted by image segmentation, and the key field region boxes are generated on the business expansion document sample by using a field clustering algorithm based on the title recognition results of the main structural blocks.

[0025] Document features include layout features, content / topic features, visual features, and metadata features. Image segmentation can be semantic segmentation or instance segmentation; image segmentation can be achieved through Mask R-CNN, U... 2 - Implementation of the visual branch of Net or LayoutLMv3. Key field bounding boxes include, but are not limited to, the photo area, the code area, and the issuing authority area.

[0026] In this embodiment, a pre-built document template library is used to match document features, enabling rapid and accurate classification of documents of known types. After determining the document content category, the business expansion document sample is divided into multiple key field bounding boxes based on the standard layout features of the document content category in the document template library. This allows for the rapid and direct location of the bounding boxes of key fields using pre-defined knowledge. In cases where a template library match cannot be found, the system performs adaptive layout inference. Specifically, semantic segmentation or instance segmentation is performed on the business expansion document sample to extract the main structural blocks. For each structural block, an OCR engine is invoked to recognize the title. Using each recognized title structural block as an anchor point, a field clustering algorithm is used to locally cluster the fields within their neighborhood. The minimum bounding rectangle of each cluster is generated as the bounding box of the key field. This achieves automatic adaptation to new document formats, giving the system strong adaptability and scalability. Even when faced with new formats, variants, or completely unseen documents, it can attempt to understand and structure them without pre-creating templates, significantly reducing the cost and time of accessing new document types and ensuring the timeliness and accuracy of key information extraction.

[0027] Step 101-3: Based on business scenarios, construct a preprocessing quality assessment system for different sample sets. Through fuzzy matching rules, content integrity judgment, and layout stability scoring, review and redistribute the business expansion document samples output by the structured preprocessing of different sample sets.

[0028] Among them, fuzzy matching rules are used to detect whether the extracted content of specific fields has a reasonable association with known dictionaries, preset values, or context, and can be identified even with minor errors. For example, intelligent validation of fields such as name and address. Content integrity judgment is used to check whether all necessary business fields in the document have been correctly extracted and whether there are any omissions. Layout stability score is used to evaluate whether the document layout analysis is accurate and whether the field positioning is stable and reliable, especially suitable for template changes or non-standard documents.

[0029] Step 101-4: Output a sample of business expansion documents that meet the review requirements.

[0030] In this embodiment, documents with similar characteristics are grouped into the same set, making the processing complexity and expected results of documents within each sample set consistent. Then, through image enhancement and structured preprocessing, disordered image pixels or unstructured fields are transformed into ordered correspondences, providing a structured data carrier for subsequent intelligent review and data entry. After structured preprocessing, different preprocessing quality evaluation systems are constructed for different sample sets to evaluate their preprocessing quality. Fuzzy matching rules are used to determine whether the document content is consistent with the preset field structure, and field completeness and layout stability scores are used to determine whether the document can be used for subsequent tasks. When the preprocessing result does not meet the quality requirements, the system will automatically reassign or return the data for supplementation, improving the overall data flow quality. This not only makes the evaluation system align with specific document evaluation standards, improving the accuracy of quality evaluation, but also guides the iterative optimization of image enhancement, layout analysis, and field extraction algorithms in the previous stage, thereby improving the overall quality and usability of structured data, providing a reliable foundation for downstream intelligent analysis, and reducing operational errors and potential risks caused by data errors.

[0031] Step 102: Use a multimodal large model to identify key characters in the business expansion document and perform annotation processing on the key characters.

[0032] The annotation process includes field type annotation, location association annotation, and business attribute annotation.

[0033] In this embodiment, the multimodal large model can simultaneously process and fuse visual information such as images, text, and semantic relationships in a document. By integrating information from these different modalities, the multimodal large model can perform more comprehensive reasoning and understanding, thereby accurately identifying key characters in the document. This not only avoids the limitations of traditional OCR, which can only process image information, and can quickly adapt to new document types, enabling multi-source document processing, but also accurately identifies text in complex, ambiguous, or low-quality documents through linguistic context and structural logical relationships, significantly improving recognition accuracy and robustness.

[0034] In practical applications, step 102, which involves using a multimodal large model to identify key characters in the business expansion document and then performing annotation processing on these key characters, specifically includes the following steps: Step 102-1: Based on the key field region box, identify key characters in the business expansion document through a multimodal large model.

[0035] Step 102-2: Annotate the key characters based on their field type, the business attributes corresponding to the key field area box to which they belong, and their positional associations.

[0036] The location association includes the contextual relationship of characters in the key field area box and the page coordinates.

[0037] In this embodiment, key field bounding boxes are used as the localization basis. A multimodal large model is used to identify specified areas segment by segment, achieving efficient recognition of characters within key areas such as ID card numbers, unified social credit codes, and property certificate numbers. This reduces redundant interference in model recognition and improves character extraction accuracy. Simultaneously, the system performs comprehensive annotation of key fields based on field type, business attributes, and positional associations including contextual relationships and page coordinates. This multi-dimensional annotation endows the recognition results with business semantics and spatial attributes, enabling the system to retain the original layout features of the document, improve the accuracy of subsequent field extraction and relationship parsing, significantly reduce manual annotation and data entry costs, and provide downstream business systems with a high-quality, highly reliable, and seamlessly integrated data source. This enhances the semantic understanding capabilities of the expert model and reduces misidentification and missing information.

[0038] For example, taking an image document as an example, combining the previously generated field region bounding boxes and layout structure, the OCR algorithm in the model scans the specified areas of the document image segment by segment. Based on character shape, text arrangement, and the inherent structure of the certificate layout, it automatically locates the positions of key fields such as ID card number, unified social credit code, and property certificate number, thus obtaining preliminary key character recognition results. After character recognition, the system performs layout-level binding processing on the text content according to the correspondence between the recognized text and the layout structure. The binding process records the boundary coordinates of text paragraphs, their respective field regions, and contextual structure, giving each text paragraph clear spatial positioning information. With the help of this spatial positioning link, the system can accurately reflect the layout relationships between different fields, such as the adjacency between the unified social credit code and the company name, and the logical order of the real estate number and the property owner field, thus preserving the original structural semantics of the document. Subsequently, the system performs semantic association processing on the text content based on the field's region, the identified key phrases, and the overall format of the document. During this process, the system automatically infers the business relationships between fields based on the field's region and its fixed layout position. For example, in the business license format, the system can identify the enterprise name field based on the field paragraph structure and automatically infer the region where the unified social credit code is located based on its fixed page position; in the real estate certificate scenario, the system will extract the property owner information based on the title style and field paragraph position and establish a corresponding relationship with the real estate number, so that the text content and document structure maintain consistent semantic expression.

[0039] Location association annotation specifically includes: binding text to page coordinates to give each field a unique spatial index for subsequent field conflict detection and cross-certificate comparison; generating semantic links of "enterprise name - unified social credit code" and "property owner - real estate number" based on the contextual relationship between fields; and performing formatted label binding on fields with special formats, including dates, segment codes, and geographical names, to ensure consistency in subsequent parsing.

[0040] Understandably, when faced with documents of low resolution or containing stains, reflections, creases, etc., the system will automatically trigger a quality compensation process based on character confidence and layout stability indicators. This compensation process includes enhancing clarity in localized areas, repairing character deformations, and performing secondary recognition in specific areas to repair or re-recognize blurry characters.

[0041] For example, for character regions exhibiting obvious recognition anomalies, the system performs local correction operations based on contextual semantics and the structure to which the field belongs, thereby improving the stability and reliability of recognition. Through these processing actions, the system can maintain high recognition accuracy even under complex document conditions.

[0042] Step 103: Incrementally train the expert model based on the labeled keywords.

[0043] The expert model comprises a document type classification model, a field extraction model, a relationship recognition model, and a calibration model. These models are configured hierarchically according to their functions. During runtime, each model is processed independently by the scheduler based on task requirements. At the output stage, results from multiple models are merged to allow different tasks to perform reasoning within independent modules and complete the overall information extraction process through a collaborative mechanism. Specifically: the document type classification model identifies the category of uploaded documents, providing an entry point for the subsequent field extraction process. The field extraction model locates and extracts key field content within the identified field areas. The relationship recognition model infers the logical relationships between fields to restore the inherent structure and semantic links of the document. The calibration model compares the model's prediction results with the business knowledge base, performing a secondary check on field format, field logic, and relationships between documents.

[0044] Specifically, the training content includes field-level extraction, feature consistency verification, cross-document field association, and secondary calibration.

[0045] In this embodiment, annotated keywords and original business expansion document samples are used as training data to analyze the structural features, field classification methods, and business rules of different document types, and accordingly establish corresponding document type classification models, field extraction models, and relationship recognition models. During incremental training, the system inputs samples batch by batch according to the differences in real-world document scenarios, enabling the model to identify patterns of change between fields, format fluctuations, and consistency characteristics across documents. As training continues, the model gradually masters the field patterns of different certificate types, the relationship structure between fields, and various business expressions that may appear in documents. To further improve the stability of the model's output, the system introduces a knowledge rule-based verification mechanism. This mechanism compares the prediction results of the document type classification model, field extraction model, and relationship recognition model with the business knowledge base, achieving a secondary check of field format, field logic, and relationships between documents. This enhances the model's generalization ability in complex scenarios.

[0046] For example, when identifying certificate numbers, the system automatically verifies their format according to the specific numbering rules for various types of certificates; when processing enterprise address information, the system makes consistency judgments based on the administrative division structure of the address; and when analyzing property documents, the system performs logical verification of field relationships based on the mapping rules between property owners and real estate numbers. Through this process, the system can promptly identify and automatically correct potential errors in model reasoning.

[0047] Step 104: In response to the information extraction request of the target document uploaded by the user, the information extraction request is input into the multimodal large model to obtain the extraction task.

[0048] In this embodiment, a multimodal large model is used to semantically decompose the user's input natural language information extraction requirements. Combined with the visual and textual information of the target document, the document structure, content intent, and semantic requirements are converted into a unified vector representation. This allows for the real-time, dynamic generation of extraction tasks tailored to the specific needs. This ensures that the extraction tasks focus on core requirements, avoids redundant data interference, and improves the system's adaptability and generalization capabilities when facing new document types or customized extraction requirements. It also helps to accurately extract information even in complex or ambiguous contexts.

[0049] For example, when a user uploads multiple documents and needs to confirm their consistency, the system automatically breaks them down into tasks such as "identifying certificate types," "extracting core fields," and "performing cross-certificate comparison." When a user only needs to extract fields from a single document, the system only triggers the field extraction-related tasks. The task parsing module parses the user input and document content, transforming the processing requirements into a set of executable tasks, which are then assigned to the corresponding expert models for execution by the scheduling module.

[0050] In one embodiment, prior to step 104, the method for extracting electricity information further includes: displaying a certificate information extraction interface; in response to an upload operation on the certificate information extraction interface, obtaining a target document and assigning a number to the target document; and calling a format conversion engine to convert the target document into a specified format.

[0051] The certificate information extraction interface includes a local file import entry, an external system integration entry, and a scanning entry.

[0052] In this embodiment, the system interface integrates three document import methods: local file import, external system integration, and scanning. It covers the upload needs of all types of business expansion certificates, including electronic documents, system-shared documents, and scanned paper documents. This adapts to different document delivery scenarios for customers and power companies, improving operational flexibility and significantly reducing the time costs associated with manual transfer and repetitive data entry. Upon receiving a document, the system assigns a unique number to the uploaded target document, establishing a unique association of "document number - certificate content - processing record." This not only meets the compliance archiving and traceability requirements for business expansion, facilitating subsequent searching, verification, and auditing, but also allows for sequential processing in batch upload scenarios, ensuring uninterrupted access to subsequent recognition pipelines. Simultaneously, the system calls a format conversion engine to convert documents to a specified format, eliminating format differences between different source documents (such as JPG, PDF, and Word). This provides standardized input for subsequent image enhancement, multimodal large-scale model recognition, and other processes, preventing processing failures or accuracy degradation due to format incompatibility.

[0053] Step 105: Call the expert model corresponding to the extraction task to process the target document and obtain the text extraction result of the target document.

[0054] The power information extraction method provided in this application first utilizes the powerful semantic understanding capabilities of a multimodal large-scale model to identify and deeply annotate key characters in business expansion documents. Then, it uses the high-quality data preprocessed and annotated by the multimodal large-scale model to continuously and incrementally train a domain expert model. This enables the expert model to adaptively learn and iterate on the business logic of the business expansion documents, solving the problems of weak generalization ability and slow adaptation to new business scenarios in traditional models. Furthermore, the multimodal large-scale model and the functionally decoupled expert model are collaboratively scheduled. When a user needs to extract the content of a target document, the multimodal large-scale model is invoked again to automatically parse the user's natural language or business request into a structured extraction task. The system then applies a priority strategy to schedule the task using the expert model based on the actual document situation and business requirements. This improves the system's ability to understand complex and ambiguous needs, helping to improve the accuracy, robustness, and response speed of document information extraction to complex and changing business requirements. While ensuring high accuracy, it significantly improves the automation and efficiency of business processing, enabling the system to maintain stable and accurate business processing capabilities in complex and diverse document scenarios.

[0055] In one embodiment, step 105, which is to call the expert model corresponding to the extraction task to process the target document, specifically includes: matching the priority of the extraction task corresponding to the target document based on the complexity of the target document, the number of missing fields, and the urgency of the business; and calling the expert model corresponding to the extraction task to process the target document based on the priority order.

[0056] In this embodiment, the expert model scheduling priority can be dynamically adjusted based on the complexity of the document type, the number of missing fields, and the urgency of the business. This enables a flexible and efficient processing flow according to business needs, ensuring smooth task execution and accurate results.

[0057] For example, when processing complex real estate certificates, the system will prioritize calling models with strong format parsing capabilities to confirm field structure; when key fields are missing in the document, the system will prioritize initiating the completion reasoning for the relevant fields; in business scenarios that require rapid response, the system will adjust computing resources to prioritize the execution of critical tasks.

[0058] In practical application scenarios, for the relationship recognition model, step 105 specifically includes: establishing a field mapping table for text fields in multiple target documents uploaded by the same user; traversing the field mapping table; if the text fields for the same field type in the field mapping table are different, it is determined that there is a cross-document conflict in the field type, and different text fields are filtered sequentially based on recognition confidence, preset logical rules, and document version weights until the text fields for the same field type are the same; based on the field mapping table and the text fields that have cross-document conflicts, a cross-document consistency score is generated to indicate the degree of matching between documents to the user.

[0059] Field types are used to distinguish different content within the same document. The field mapping table, indexed by user identifiers, records text fields of the same type in different target documents and their identification confidence levels. This allows the system to accurately locate related fields in each document by recording the field's source, version information, page location, and semantic adjacency structure. For example, a field mapping table can be created for multiple documents uploaded by the same user, establishing a correspondence between elements such as name, address, and ID number across different documents.

[0060] In this embodiment, by traversing the field mapping table, inconsistencies in the same field type across multiple documents from the same user can be automatically located. When inconsistencies in the same field type across multiple documents are detected, i.e., cross-document conflicts exist, the system will activate a conflict resolution strategy based on factors such as field confidence, format compliance, document version time, and business logic priority to determine the field that best meets business requirements as the final output. This resolves conflicts without manual intervention, ensuring accurate data matching between multiple documents and avoiding information bias caused by field inconsistencies. Simultaneously, the generated cross-document consistency score provides users with intuitive feedback on the information matching status of multiple documents. This assists staff in quickly assessing document compliance and provides targeted alerts for conflicting fields, offering clear evidence for subsequent manual review or business approval, significantly reducing data risks in business expansion processes.

[0061] Furthermore, after step 105, the method for extracting electricity information further includes: in response to obtaining the text extraction results of the target document, displaying the extraction source, confidence score and / or cross-document consistency score of each field in the text extraction results on the certificate information extraction interface based on the regional position of the fields in the document.

[0062] In this embodiment, the system generates field-level information displays for each field based on its location within the document and the model's recognition results. This information includes the field's extraction source, confidence score, and cross-document consistency score. This allows users to intuitively understand business expansion document data from different sources, in different formats, and of varying quality. Users gain a comprehensive understanding of the system's recognition process, improving review efficiency and the accuracy of manual review. Simultaneously, displaying the quantified confidence score and cross-document consistency score assists manual review, thereby strengthening data compliance and traceability, and reducing subsequent approval risks.

[0063] In one embodiment, the method for extracting power information further includes: generating a unique tracking number for each extraction task, and associating and storing the tracking number with the target document and its text extraction results; and labeling the text extraction results with a model version identifier.

[0064] In this embodiment, the system automatically generates a unique tracking number for each document processing step and associates this number with the user-uploaded document, the recognized output content, and subsequent manual correction information, forming a complete processing chain record. Simultaneously, to ensure stable backtracking capabilities after model upgrades or rule adjustments, the system retains a processing version for each recognition task in the background, recording the text extraction results, recognition algorithm version, rule base version, and model parameters used. This allows users to recover historical results based on the tracking number. This provides a basis for comparative verification and quality analysis before and after model upgrades. The audit log's complete record of operational behavior and data flow ensures compliance and traceability of the process, making it particularly suitable for the stringent data security requirements in power marketing operations. Through a transparent display mechanism and a comprehensive log system, users can clearly see the source and processing process of each field, thereby enhancing their acceptance of the system's automatic recognition results.

[0065] Furthermore, the system can also record all user operations, model calls, field change records, and data flow, making the entire process traceable at any time and meeting the data security and compliance management requirements in the power marketing field.

[0066] The method for extracting power information provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0067] It should be noted that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0068] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0069] Furthermore, such as Figure 2 As shown, as a specific implementation of the above-mentioned method for extracting power information, this application provides a power information extraction device 200, which includes a preprocessing module 201, a sample training module 202, and an information extraction module 203.

[0070] Among them, the preprocessing module 201 is used to perform image enhancement, layout structure analysis and preprocessing quality assessment on diverse business expansion document samples; The sample training module 202 is used to identify key characters in business expansion documents using a multimodal large model and to perform annotation processing on the key characters, including field type annotation, position association annotation and business attribute annotation; and to incrementally train the expert model based on the annotated key characters, including a document type classification model, a field extraction model, a relationship recognition model and a calibration model. The information extraction module 203 is used to respond to the information extraction request of the target document uploaded by the user, input the information extraction request into the multimodal large model to obtain the extraction task; and call the expert model corresponding to the extraction task to process the target document to obtain the text extraction result of the target document.

[0071] Furthermore, the preprocessing module 201 is specifically used to classify and organize the business expansion document samples according to document source, format, and image quality to form multiple sample sets; perform image enhancement and structured preprocessing on multiple sample sets, wherein the structured preprocessing includes layout analysis and field region division; construct a preprocessing quality evaluation system for different sample sets based on business scenarios, and review and redistribute the business expansion document samples output by the structured preprocessing of different sample sets through fuzzy matching rules, content integrity judgment, and layout stability scoring; and output business expansion document samples that meet the review requirements.

[0072] Further, the preprocessing module 201 is specifically used to perform structured preprocessing on multiple sample sets, including: matching the features of the business expansion document samples with the document features of different document content categories in the document template library; if the similarity between the document features in the document template library and the features of the business expansion document samples is greater than the similarity threshold, the document content category to which the document features belong is taken as the document content category of the business expansion document sample, and a key field region box is generated on the business expansion document sample based on the document content category of the business expansion document sample; if the document features in the document template library do not match the features of the business expansion document samples, the main structural blocks of the business expansion document samples are extracted through image segmentation, and the key field region boxes are generated on the business expansion document samples based on the title recognition results of the main structural blocks using a field clustering algorithm; wherein, the key field region boxes include the photo area, the encoding area, and the issuing agency area.

[0073] Furthermore, the sample training module 202 is specifically used to identify key characters in business expansion documents based on key field region boxes using a multimodal large model; and to annotate key characters based on the field type of the key characters, the business attributes corresponding to the key field region boxes to which the key characters belong, and the positional association; wherein, the positional association includes the contextual relationship of the characters in the key field region boxes and the page coordinates.

[0074] Furthermore, the expert model includes a relationship recognition model and an information extraction module 203, which is specifically used to establish a field mapping table for text fields in multiple target documents uploaded by the same user. The field mapping table is indexed by the user identifier and records text fields of the same field type in different target documents and their recognition confidence scores. The field mapping table is traversed. If the text fields of the same field type in the field mapping table are different, it is determined that there is a cross-document conflict in the field type. Based on the recognition confidence score, preset logical rules, and document version weight, different text fields are filtered sequentially until the text fields of the same field type are the same. Based on the field mapping table and the text fields that have cross-document conflicts, a cross-document consistency score is generated to indicate the degree of matching between documents to the user.

[0075] Furthermore, the power information extraction device 200 includes: The tracking module generates a unique tracking number for each extraction task and associates and stores the tracking number with the target document and its text extraction results; it also labels the text extraction results with model version identifiers.

[0076] Furthermore, the power information extraction device 200 includes: The display module is used to display the certificate information extraction interface, which includes a local file import entry, an external system integration entry, and a scanning entry. The document acquisition module is used to respond to the upload operation of the certificate information extraction interface, acquire the target document, assign a number to the target document, and call the format conversion engine to convert the target document into a specified format.

[0077] Furthermore, the display module is also used to respond to the text extraction results of the target document, and based on the regional position of the fields in the document in the text extraction results, display the extraction source, confidence score and / or cross-document consistency score of each field in the text extraction results on the certificate information extraction interface.

[0078] Furthermore, the power information extraction device 200 includes: The scheduling module is used to match the priority of the extraction task corresponding to the target document based on the complexity of the target document, the number of missing fields, and the urgency of the business. The information extraction module 203 is specifically used to process the target document by calling the expert model corresponding to the extraction task based on priority order.

[0079] Specific limitations regarding the power information extraction device can be found in the limitations of the power information extraction method described above, and will not be repeated here. Each module in the aforementioned power information extraction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0080] Based on the above, Figure 1 Accordingly, embodiments of this application also provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method for extracting power information is shown.

[0081] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0082] Based on the above, Figure 1 The method shown, and Figure 2 The virtual device embodiment shown is designed to achieve the above objectives, such as... Figure 3As shown in the figure, this application embodiment also provides a computer device 300, which includes a processor 301 and a memory 302. The memory 302 stores a program or instructions that can run on the processor 301. When the program or instructions are executed by the processor 301, they implement the above-mentioned... Figure 1 The method for extracting power information is shown.

[0083] The memory 302 can be used to store software programs and various data. The memory 302 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 302 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 302 in the embodiments of this application includes, but is not limited to, these and any other suitable types of memory.

[0084] Processor 301 may include one or more processing units; optionally, processor 301 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 301.

[0085] Computer equipment can specifically include personal computers, servers, network devices, etc.

[0086] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB ports, card reader ports, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0087] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0089] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0090] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for extracting electricity information, characterized in that, The method includes: Image enhancement, layout structure analysis, and preprocessing quality assessment are performed on diverse business expansion document samples. The key characters in the business expansion document are identified using a multimodal large model, and the key characters are annotated. The annotation process includes field type annotation, position association annotation, and business attribute annotation. Incremental training of the expert model is performed based on the labeled key characters, wherein the expert model includes a document type classification model, a field extraction model, a relation recognition model, and a calibration model; In response to an information extraction request for a target document uploaded by a user, the information extraction request is input into the multimodal large model to obtain an extraction task; The expert model corresponding to the extraction task is invoked to process the target document, and the text extraction result of the target document is obtained.

2. The method for extracting power information according to claim 1, characterized in that, The process of image enhancement, layout structure analysis, and preprocessing quality assessment of diverse business expansion document samples includes: The business expansion document samples are classified and organized according to document source, format, and image quality to form multiple sample sets; Image enhancement and structured preprocessing are performed on the multiple sample sets, wherein the structured preprocessing includes layout analysis and field region segmentation; Based on business scenarios, a preprocessing quality assessment system for different sample sets is constructed. Through fuzzy matching rules, content integrity judgment, and layout stability scoring, the business expansion document samples output by the structured preprocessing of different sample sets are reviewed and redistributed. Output a sample of business expansion documents that meets the review requirements.

3. The method for extracting power information according to claim 2, characterized in that, The structured preprocessing performed on the plurality of sample sets includes: The features of the business expansion document samples are matched with the features of documents of different document content categories in the document template library; If the similarity between the document features in the document template library and the features of the business expansion document sample is greater than the similarity threshold, the document content category to which the document features belong is taken as the document content category of the business expansion document sample, and a key field region box is generated on the business expansion document sample based on the document content category of the business expansion document sample. If the document features in the document template library do not match the features of the business expansion document sample, the main structural blocks of the business expansion document sample are extracted by image segmentation. Based on the title recognition results of the main structural blocks, a field clustering algorithm is used to determine the key field region bounding boxes on the business expansion document sample. The key field area box includes the photo area, the code area, and the issuing authority area.

4. The method for extracting power information according to claim 3, characterized in that, The process of identifying key characters in the business expansion document using a multimodal large model and performing annotation processing on the key characters includes: Based on the key field region box, the key characters in the business expansion document are identified through the multimodal large model; The key characters are labeled based on their field type, the business attributes corresponding to the key field area box to which they belong, and their location associations. The location association includes the contextual relationship of characters in the key field area box and the page coordinates.

5. The method for extracting power information according to claim 1, characterized in that, The expert model includes a relationship recognition model. The step of calling the expert model corresponding to the extraction task to process the target document includes: A field mapping table is established for text fields in multiple target documents uploaded by the same user. The field mapping table is indexed by the user identifier and records the text fields of the same field type in different target documents and their identification confidence scores. Traverse the field mapping table; If the text fields for the same field type in the field mapping table are different, it is determined that there is a cross-document conflict in the field type. Based on the identification confidence, preset logical rules and document version weight, the different text fields are filtered in turn until the text fields for the same field type are the same. Based on the field mapping table and the text fields that conflict across documents, a cross-document consistency score is generated to indicate to the user the degree of matching between documents.

6. The method for extracting power information according to claim 1, characterized in that, The method further includes: A unique tracking number is generated for each extraction task, and the tracking number is associated with and stored in relation to the target document and its text extraction results; The extracted text results are labeled with a model version identifier.

7. The method for extracting power information according to claim 1, characterized in that, The method further includes: The document information extraction interface is displayed, which includes a local file import entry, an external system integration entry, and a scanning entry. In response to the upload operation of the certificate information extraction interface, the target document is obtained and numbered; The format conversion engine is invoked to convert the target document into the specified format. In response to obtaining the text extraction results of the target document, based on the regional position of the fields in the document in the text extraction results, the extraction source, confidence score and / or cross-document consistency score of each field in the text extraction results are displayed on the certificate information extraction interface.

8. The method for extracting power information according to claim 1, characterized in that, The step of calling the expert model corresponding to the extraction task to process the target document includes: Based on the complexity of the target document, the number of missing fields, and the urgency of the business, the priority of the extraction task corresponding to the target document is matched; The expert model corresponding to the extraction task is invoked according to the priority order to process the target document.

9. A device for extracting power information, characterized in that, The device includes: The preprocessing module is used to perform image enhancement, layout structure analysis, and preprocessing quality assessment on diverse business expansion document samples. The sample training module is used to identify key characters in the business expansion document using a multimodal large model, and to perform annotation processing on the key characters, wherein the annotation processing includes field type annotation, position association annotation, and business attribute annotation; and, Incremental training of the expert model is performed based on the labeled key characters, wherein the expert model includes a document type classification model, a field extraction model, a relation recognition model, and a calibration model; The information extraction module is configured to respond to an information extraction request from a user-uploaded target document by inputting the information extraction request into the multimodal large model to obtain an extraction task; and, The expert model corresponding to the extraction task is invoked to process the target document, and the text extraction result of the target document is obtained.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for extracting power information as described in any one of claims 1 to 8.

Images