Electronic document refined naming method and computer equipment

Abstract

The invention relates to the technical field of artificial intelligence, and provides an electronic document refined naming method and computer equipment, and the method comprises the steps that an electronic document sequence is received, the electronic document sequence comprises a plurality of pages, and the pages correspond to various layout structures; calling a pre-trained carrier category recognition model to recognize the electronic documents in the to-be-recognized electronic document sequence to obtain a carrier category recognition result of each page of electronic document; calling a pre-trained layout category recognition model to recognize the electronic documents in the to-be-recognized electronic document sequence to obtain a layout category recognition result of each page of electronic document; querying a corresponding naming rule based on the carrier subcategory and the layout structure subcategory, wherein the naming rule indicates to-be-identified elements; the to-be-identified elements corresponding to the naming rule are extracted through the large model, and the name of the electronic document sequence is obtained based on the naming rule, so that automatic and standardized naming of the electronic document is realized, and the efficiency and accuracy of judicial archive management are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for refined naming of electronic documents. Background Technology

[0002] In the daily work of judicial institutions such as courts and procuratorates, a massive number of electronic document sequences from various litigation participants are received and generated, such as complaints, answers, lists of evidence, attorney's statements, judgments, and rulings. Currently, the standardized naming and management of such documents mainly relies on manual operation. That is, court clerks or archive managers manually read the document content and rename and classify them according to their own experience. This method is extremely inefficient, consumes a lot of manpower and time, and due to the subjectivity of human judgment, it is prone to problems such as naming errors, inconsistent standards, and inconsistencies, which bring great difficulties to the archiving, management, retrieval, and subsequent data mining and utilization of electronic case files.

[0003] To achieve in-depth application of electronic case files, higher requirements are placed on the precision of document naming. To respect the review habits of judges, lawyers, and other case handlers and improve case-handling efficiency, there is an urgent need for automated naming based on multi-dimensional characteristics such as the source of the materials (e.g., submitter), type (e.g., judgment, evidence), title, and attributes), so that reviewers can accurately find and locate the required materials simply by their file names.

[0004] However, achieving this goal presents enormous technical challenges. Case files are incredibly diverse, including not only standard internal documents but also a wide variety of evidentiary materials submitted by the parties involved, potentially exceeding ten thousand types with vastly different characteristics. Using traditional solutions based on natural language processing and named entity recognition would require customizing complex extraction rules and models for each of these massive amounts of different document types, resulting in a heavy workload and extremely high development difficulty.

[0005] Therefore, there is an urgent need in this field for an automated and refined naming method and system for electronic documents that can intelligently, accurately, and flexibly adapt to various naming rules in order to solve the aforementioned long-standing technical problems. Summary of the Invention

[0006] The purpose of this application is to provide a method and apparatus for refined naming of electronic documents, aiming to solve the urgent need in the prior art for a method that can automatically and accurately identify document types and generate standardized names accordingly.

[0007] To achieve the above objectives, this application provides a method for refined naming of electronic documents, comprising the following steps: receiving an electronic document sequence, the electronic document sequence containing multiple pages, the multiple pages corresponding to various layout structures; calling a pre-trained carrier category recognition model to identify the electronic documents in the electronic document sequence to be identified, obtaining the carrier category recognition result for each page of the electronic document, the carrier category recognition result of the electronic document including: carrier sub-category and the probability of being identified as the corresponding sub-category; calling a pre-trained layout category recognition model to identify the electronic documents in the electronic document sequence to be identified, obtaining the layout category recognition result for each page of the electronic document, the layout category recognition result of the electronic document including: layout structure sub-category and the probability of being identified as the corresponding sub-category; querying the corresponding naming rules based on the carrier sub-category and layout structure sub-category, the naming rules indicating the elements to be identified; extracting the elements to be identified corresponding to the naming rules through a large model, and obtaining the name of the electronic document sequence based on the naming rules.

[0008] Optionally, before querying the corresponding naming rules, the process further includes: calling a pre-trained primary document category recognition model to identify each page of the electronic document sequence to be identified, obtaining the document sub-category recognition result for each page of the electronic document, wherein the document sub-category recognition result includes: the document sub-category and the probability of being identified as the corresponding sub-category; statistically analyzing the document sub-category recognition results of each electronic document in the electronic document sequence, and generating a document category feature vector based on the statistical results, including: counting the number of documents identified as each document sub-category and the probability value of being identified as the corresponding document sub-category in the electronic document sequence, and using the statistical results as feature values ​​to form a document category feature vector; calling a pre-trained ultimate document category recognition model to identify the document category feature vector, obtaining the document category recognition result of the electronic document; and querying the corresponding naming rules based on the carrier sub-category, page layout sub-category, and document category recognition results.

[0009] Optionally, a pre-trained carrier category recognition model is invoked to recognize each page of the electronic document in the sequence of electronic documents to be recognized. The method includes: constructing a basic convolutional neural network model; obtaining a preset training set of electronic documents, wherein the training set includes multiple sets of data, each set of data including pages and carrier sub-category labeling information of the pages; training the basic convolutional neural network model using the training set, and iteratively optimizing the model parameters using a loss function and an optimizer to obtain the final carrier category recognition model.

[0010] Optionally, a pre-trained layout category recognition model is invoked to identify electronic documents in the sequence of electronic documents to be identified. The method includes: constructing a basic convolutional neural network model; obtaining a preset training set of electronic documents, wherein the training set includes multiple sets of data, each set of data including pages and page layout sub-category labeling information; training the basic convolutional neural network model using the training set, and iteratively optimizing the model parameters using a loss function and an optimizer to obtain the final layout category recognition model.

[0011] Optionally, the carrier subcategories include: text, charts, and images.

[0012] Optionally, the layout structure subcategories include: homepage, last page, and main text.

[0013] Optionally, the document subcategories include: judgment documents, enforcement documents, litigation procedure documents, and other documents.

[0014] Optionally, the naming rules are obtained by: inputting multiple images and the naming results corresponding to each image; performing reverse analysis on the multiple images and their corresponding naming results based on a large model to extract the implicit naming rules; performing statistical analysis on the extracted naming rules to calculate the probability of each naming rule appearing; and filtering the naming rules according to a preset probability threshold to obtain the final naming rules.

[0015] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, it implements the above-described refined electronic document naming method.

[0016] This application also provides a computer-readable storage medium storing an electronic document name generation program, which, when executed by a processor, implements the steps of the above-described refined electronic document naming method.

[0017] Compared with existing technologies, the technical solution provided in this application has the following beneficial effects: Through comprehensive analysis of pages using multiple models, a deep understanding of the multi-dimensional structure of documents is achieved; by adopting a strategy of querying naming rules based on recognition results, business logic (naming rules) and technical implementation (model recognition) are decoupled, enabling the system to easily adapt to diverse naming convention requirements of different regions and units, requiring only the configuration or updating of the corresponding naming rule library without retraining or modifying the deep learning model; the introduction of a large language model for extracting naming elements enables the processing of key information in unstructured text, improving the accuracy of name generation. Ultimately, this achieves automated and standardized naming of judicial documents, greatly improving the efficiency and accuracy of judicial file management. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings:

[0019] Figure 1 This is a flowchart of a refined naming method for electronic documents according to an embodiment of the present invention; Detailed Implementation

[0020] To better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0021] S101: Receive electronic document sequence.

[0022] Receive a sequence of electronic documents, which contains multiple pages and various layout structures. For example, the material may be a multi-page PDF file containing materials submitted by the plaintiff in a civil case, including a "Civil Complaint," a "List of Evidence," and several photocopies of evidence.

[0023] S102: Carrier category identification.

[0024] The pre-trained carrier category recognition model is invoked to identify the electronic documents in the sequence of electronic documents to be identified, and the carrier category recognition result of each page of the electronic document is obtained. The carrier category recognition result of the electronic document includes: carrier sub-category and the probability of being identified as the corresponding sub-category; the carrier sub-categories include: text, charts and images.

[0025] Optionally, the pre-trained carrier category recognition model is invoked to recognize each page of the electronic document in the sequence of electronic documents to be recognized, including: constructing a basic convolutional neural network model; obtaining a pre-set training set of electronic documents, the training set including multiple sets of data, each set of data including pages and carrier sub-category labeling information of the pages; using the training set to train the basic convolutional neural network model, using a loss function and optimizer to iteratively optimize the model parameters, and obtaining the final carrier category recognition model.

[0026] A large number of electronic document page images of judicial documents were collected to form the original training dataset. The document types should cover all types processed by the document category recognition model, including but not limited to: judgments, rulings, mediation agreements, enforcement notices, summonses, service receipts, court transcripts, and evidentiary materials, to ensure that the model can comprehensively learn the carrier features of judicial documents.

[0027] Each document page image is meticulously annotated by legal professionals or trained data labelers according to a predefined carrier subcategories system. The carrier subcategories include:

[0028] For pages containing only text, the entire page is labeled "Text".

[0029] Pages containing evidence images (such as transcripts with photos pasted on them) are labeled as "Images".

[0030] For pages containing large calculation tables (such as execution allocation tables), the entire page is labeled "Chart".

[0031] For complex pages that contain multiple carriers (such as those with both text descriptions and small illustrations), image segmentation and cropping can be performed, and each image block containing a single carrier can be individually labeled and trained to improve the model's ability to perceive local features.

[0032] Preprocessing the labeled dataset can improve the efficiency and robustness of model training.

[0033] Size normalization: Scaling all images to a uniform, fixed size to meet the input requirements of convolutional neural networks. Normalization: Normalizing image pixel values ​​from the range [0, 255] to the range [0, 1] or [-1, 1] to accelerate model convergence. Data augmentation: Expanding the data set by randomly transforming the training images to prevent overfitting and improve the model's generalization ability.

[0034] A high-performance deep convolutional neural network is selected as the backbone network of the model. Preferably, a model pre-trained on a large image dataset (such as ImageNet) is used for transfer learning to leverage its learned general feature extraction capabilities. The trained and evaluated model is saved as a deployable file format and integrated into the electronic document name generation system. During actual inference, the system calls the model to predict each page of the legal document image, outputting its category label (text, chart, or image) and the corresponding probability value.

[0035] S103: Page Layout Recognition. A pre-trained page layout category recognition model is invoked to identify the electronic documents in the sequence of electronic documents to be identified, obtaining the page layout category recognition result for each page. The page layout category recognition result includes: page layout sub-category and the probability of being identified as the corresponding sub-category; the page layout sub-category includes: first page, last page, and main text.

[0036] Optionally, the pre-trained layout category recognition model is invoked to recognize electronic documents in the sequence of electronic documents to be recognized. The method further includes: constructing a basic convolutional neural network model; obtaining a preset training set of electronic documents, the training set including multiple sets of data, each set of data including pages and page layout sub-category labeling information; training the basic convolutional neural network model using the training set, and iteratively optimizing the model parameters using a loss function and optimizer to obtain the final layout category recognition model.

[0037] Each page image is meticulously annotated by legal professionals or trained data labelers according to a predefined layout structure subcategories. The layout structure subcategories include:

[0038] First page: Refers to the first page of a document that contains distinctive elements such as a specific header, title, case number, and information about the parties involved. Its visual characteristics typically include: the court name at the top (often with a red official seal), the document title (such as "Civil Judgment"), the case number, and clearly defined information about the parties (plaintiff, defendant, and agent).

[0039] Last page: Refers to the last page of a document, including the end mark, signature area, date, and seal. Its visual features typically include: the words "No text on this page", the signature area of ​​the judge or court clerk (presiding judge, judge, court clerk), the location of the court seal, the date, and page numbering (such as "Page X of Y").

[0040] Body text: Refers to the pages between the first and last pages, which contain the core content of the document. Its visual characteristics are usually: long blocks of continuous text, possibly including item numbers (such as I, II, III), a list of evidence, legal citations, statements of reasoning for judgments, etc. The layout is relatively uniform and lacks the unique identifying elements of the first and last pages.

[0041] Preferably, for some documents (such as short rulings) that may only have one page, that page has the characteristics of both the first and last page, and should be labeled according to its most important function (usually labeled as the first page).

[0042] The training process is the same as that for the carrier category recognition model, and will not be described in detail here. The evaluated model is saved as a deployable file format and integrated into the electronic document processing system. During actual inference, the system calls the model to predict the category of each page of the court document image, outputting the category label and corresponding probability value for whether it belongs to the first page, body text, or last page.

[0043] S104: Query the corresponding naming rules based on the carrier sub-category and the layout structure sub-category, whereby the naming rules indicate the elements to be identified.

[0044] In the naming rules, the tags enclosed in curly braces {} are the identification labels, where {material type} specifies the exact material type identified, such as a judgment or a delivery receipt. The rest can be natural language descriptions; the system has some generalization ability, such as recipient's name or recipient's full name. For example, it can perform OCR recognition on a text-based homepage and then match it to the corresponding naming rules.

[0045] To improve performance, feature extraction locations were configured, divided into first page, last page, and full text, because typical features are found at the end of the first and last pages, or in the main text. This configuration aims to improve efficiency, reduce the number of pages in multimodal large-scale model recognition materials, and enhance overall cataloging efficiency.

[0046] Preferably, the naming rules are obtained by: acquiring multiple images and the naming results corresponding to each image; performing reverse analysis on the multiple images and their corresponding naming results based on a large model to extract the implicit naming rules.

[0047] The configuration tool initializes the preset naming rules by default. It can dynamically modify the built-in naming rule library on-site based on the already archived and named dossier materials to adapt to the naming characteristics of local materials.

[0048] The reverse extraction naming rules interact with prompt words and a large model, providing standard naming and extracted naming for images. Prompt words are similar to:

[0049] You are a court electronic case file preparer with strong legal expertise. The client's file name for this image is "Power of Attorney (Yin Moumou - December 9, XXXX)". Elements should be enclosed in curly braces {}. Please extract the user's naming rules based on the current file name. For example, if the sample material to be extracted is "Civil Judgment (xxx Civil First Instance No. 21).jpg", the extracted rules would be: "Civil Judgment ({Case Number})". Extracted elements should use legal terminology, such as case number, plaintiff, defendant, client, date of authorization, date of inquiry, invoice amount, and agent's name.

[0050] The extracted naming rules are statistically analyzed to calculate the probability of each rule's occurrence. These rules are then filtered according to a preset probability threshold to obtain the final naming rules. In other words, after analyzing large amounts of data, the system will obtain multiple similar rule templates. The system then performs clustering and statistical analysis on these rules to obtain the final adopted naming rules.

[0051] Example of statistical results:

[0052] Rule A: Return of Service ({Document Name}) - {Recipient} - {Date of Service} (Number of occurrences: 85, Probability: 63%)

[0053] Rule B: ({Document Name}) Return of Service - {Recipient} - {Date of Service} (Number of occurrences: 35, Probability: 26%)

[0054] Rule C: {Recipient}-{Name of Document Served} Return of Service Receipt-{Date of Service} (Number of occurrences: 15, Probability: 11%)

[0055] Calculate the probability of each rule occurring: The probability of rule A is 85 / (85+35+15) ≈ 63%.

[0056] The system administrator presets a probability threshold (e.g., 50%). Rule A has a probability (63%) higher than this threshold, therefore it is automatically adopted by the system as the final standardized naming rule for this type of document (delivery receipt) and stored in the rule base. The final naming rule is: Delivery Receipt ({Document Name}) - {Recipient} - {Delivery Date}.

[0057] S105: Extract the elements to be identified corresponding to the naming rules through the large model, and obtain the names of the electronic document sequence based on the naming rules.

[0058] Multimodal large-scale models are deep learning models jointly trained by integrating multiple types of data such as text, images, videos, and audio. Their core technologies include cross-modal encoder training, semantic alignment, and feature fusion. Typical technical paths include LLM-based task collaboration architectures (such as the HuggingGPT framework), multimodal aligned data joint training (CLIP, CoCa models), and cross-modal encoder training (LLaVA models). Multimodal large-scale models can perform question-and-answer sessions on images, videos, and audio through prompt-based dialogue. This study primarily uses image-based question-and-answer, as electronic case files are mainly composed of images.

[0059] The improved solution introduces a multimodal large model. Through interactive prompts provided by this large model, it identifies elements of the case file materials and, combined with configuration tools, assembles names. For example, the prompts might look like this:

[0060] You are an expert in electronic case file analysis, specializing in image files and capable of accurately analyzing their content and structure.

[0061] Task:

[0062] This is an image with the name {clName}. You need to extract the following relevant information from the image:

[0063] <list> Extract the information from the {field} field and represent that field using {alias} in the output JSON. Fields must be represented using {alias}.< / list> {notes}

[0064] Output format: Output in JSON format, and the output JSON must conform to the following format: {result}

[0065] Require:

[0066] 1. Extract information from a specified field in an image and output it as JSON. If the information is not available, do not return the specified field.

[0067] 2. If the extracted data is a date, the date format should be YYYY.MM.DD. If there is no date, the JSON return value will not include a date node. Do not fabricate dates.

[0068] 3. If the extracted data is a name or title, the format of the name or title is [“””].

[0069] 4. Only output JSON, do not interpret it, do not return irrelevant content, and do not output in Markdown format.

[0070] For example, the key text extracted is: "Name of document served: Payment Notice", "Recipient: Skyline Holdings Group Co., Ltd.", "Date of service: May 21, 2024".

[0071] Large model extraction output:

[0072] {Document Name} = Payment Notice

[0073] {Recipient} = Skyline Holdings Group Limited

[0074] {Delivery Date} = 2024.05.21 (The system can automatically format the date)

[0075] The system retrieved the following rules from the rule base for the delivery receipt:

[0076] Service Receipt ({Document Name}) - {Recipient} - {Date of Service} Substitute the extracted elements into the rule template. The final generated file name is: Service Receipt (Payment Notice) - Skyline Holdings Group Co., Ltd. - 2024.05.21.pdf

[0077] More preferably, it also includes a document category determination step, making the selection of naming rules more accurate. Specifically, a pre-trained primary document category recognition model is called to identify each page of electronic documents in the sequence of electronic documents to be identified, obtaining the document subcategory recognition result for each page of electronic documents. The document category recognition result of electronic documents includes: document subcategory and the probability of being identified as the corresponding subcategory; the document subcategory recognition results of each electronic document in the electronic document sequence are statistically analyzed, and a document category feature vector is generated based on the statistical results, including: counting the number of documents identified as each document subcategory in the electronic document sequence and the probability value of being identified as the corresponding document subcategory, and using the statistical results as feature values ​​to form a document category feature vector; the pre-trained ultimate document category recognition model is called to identify the document category feature vector, obtaining the document category recognition result of the electronic document; based on the carrier subcategory, page structure subcategory, and document category recognition result, the corresponding naming rules are queried. Document subcategories include four major categories: judgment documents, enforcement documents, litigation procedure documents, and other documents.

[0078] Judicial documents include judgments, rulings, mediation agreements, and decisions; enforcement documents include enforcement notices, notices of assistance in enforcement, and enforcement rulings; litigation procedure documents include notices, service receipts, and objections to jurisdiction; and other documents include state compensation documents and petitions related to litigation. In practical applications, documents can be classified into four main categories or further subcategories.

[0079] Preferably, a coarse classification of four main categories (primary classification) can be performed first, followed by a finer classification of subcategories (secondary classification). The model uses DenseNet as the base model to perform hierarchical indexing of pages. Primary labels (0, 1, 2, 3 corresponding to the four main categories) are used as secondary labels, with the subcategories within their respective primary categories indexed as secondary labels. The model outputs the primary subcategories and their probabilities. Based on the prediction results of the primary classification, the model dynamically selects the corresponding secondary classifier. For example, if the primary prediction is "judgment documents," the secondary model for judgment documents is called, outputting the classification results for judgments, rulings, mediation agreements, and decisions. The input to each secondary model is the feature vector output by the primary classifier, which can be features extracted after a global average pooling layer.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for refined naming of electronic documents, characterized in that, The method includes: Receive a sequence of electronic documents, the sequence of electronic documents containing multiple pages, the multiple pages corresponding to various page layouts; The pre-trained carrier category recognition model is called to identify the electronic documents in the sequence of electronic documents to be identified, and the carrier category recognition result of each page of the electronic document is obtained. The carrier category recognition result of the electronic document includes: carrier sub-category and the probability of being identified as the corresponding sub-category; The pre-trained layout category recognition model is called to identify the electronic documents in the sequence of electronic documents to be identified, and the layout category recognition result of each page of the electronic document is obtained. The layout category recognition result of the electronic document includes: layout structure sub-category and the probability of being identified as the corresponding sub-category; The naming rules are queried based on the carrier subcategory and the layout structure subcategory, and the naming rules indicate the elements to be identified; The model extracts the elements to be identified corresponding to the naming rules, and the electronic document sequence is named based on the naming rules.

2. The method for refined naming of electronic documents according to claim 1, characterized in that: Before querying the corresponding naming rules, the following also applies: The pre-trained primary document category recognition model is called to identify each page of the electronic document sequence to be identified, and the document sub-category recognition result of each page of the electronic document is obtained. The document sub-category recognition result of the electronic document includes: document sub-category and the probability of being identified as the corresponding sub-category; The document sub-category identification results of each electronic document in the electronic document sequence are statistically analyzed, and a document category feature vector is generated based on the statistical results. This includes: counting the number of documents in the electronic document sequence that are identified as each document sub-category and the probability value of being identified as the corresponding document sub-category, and using the statistical results as feature values ​​to form a document category feature vector. The pre-trained ultimate document category recognition model is invoked to identify the document category feature vector, thereby obtaining the document category recognition result of the electronic document; Based on the identification results of carrier subcategories, page layout subcategories, and document categories, query the corresponding naming rules.

3. The method for refined naming of electronic documents according to claim 1, characterized in that: The method involves calling a pre-trained carrier category recognition model to identify each page of an electronic document in a sequence of electronic documents to be identified. Construct a basic convolutional neural network model; Obtain a preset training set of electronic documents, the training set including multiple sets of data, each set of data including pages and the carrier sub-category labeling information of the pages; The basic convolutional neural network model is trained using the training set, and the model parameters are iteratively optimized using a loss function and an optimizer to obtain the final carrier category recognition model.

4. The method for refined naming of electronic documents according to claim 1, characterized in that: The method involves calling a pre-trained layout category recognition model to identify electronic documents in a sequence of electronic documents to be identified. Construct a basic convolutional neural network model; Obtain a preset training set of electronic documents, the training set including multiple sets of data, each set of data including pages and page layout sub-category labeling information; The basic convolutional neural network model is trained using the training set, and the model parameters are iteratively optimized using a loss function and an optimizer to obtain the final layout category recognition model.

5. The method for refined naming of electronic documents according to claim 1, characterized in that: The carrier subcategories include: text, charts, and images.

6. The method for refined naming of electronic documents according to claim 1, characterized in that: The page layout subcategories include: homepage, last page, and main content.

7. The method for refined naming of electronic documents according to claim 2, characterized in that: The document subcategories include: judgment documents, enforcement documents, litigation procedure documents, and other documents.

8. The method for refined naming of electronic documents according to claim 1, characterized in that: The naming rules are obtained in the following way: Retrieve multiple images and their corresponding names; Based on the large model, reverse analysis is performed on the multiple images and their corresponding naming results to extract the implicit naming rules; Statistical analysis was performed on the extracted naming rules to calculate the probability of each naming rule occurring. The naming rules are filtered according to a preset probability threshold to obtain the final naming rules.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, it implements the electronic document fine-grained naming method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an electronic document name generation program, which, when executed by a processor, implements the steps of the electronic document refined naming method as described in any one of claims 1-8.

Images