Associated document classification method and system oriented to multi-round clearing and receiving disposal process

Abstract

The invention discloses an associated document classification method and system oriented to a multi-round clearing and receipt disposal process, and the method comprises the steps: obtaining multi-round clearing and receipt documents, and carrying out the preprocessing and text recognition, and obtaining legal documents and structured metadata corresponding to the legal documents; based on a preset classification model, outputting a category label set corresponding to the document according to the legal document; based on a preset business rule, verifying the legal document according to the category label set and the structured metadata, and generating a document classification result; and performing archiving processing on the document classification result, and performing continuous optimization on the archiving processing result based on a preset debugging and / or monitoring mechanism. Therefore, through deep fusion of intelligent classification and business rules, accurate classification of single documents and automatic association of multi-stage documents are realized, closed-loop management of multi-round clearing and receiving disposal is further realized, the clearing and receiving efficiency is remarkably improved, and the compliance risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of internet finance technology, and in particular to a method and system for classifying related documents for multi-round debt collection and disposal processes. Background Technology

[0002] In the management of non-performing assets in internet finance, the document processing process fragments the legal, litigation, and enforcement documents involved in the disposal of non-performing assets, and lacks an inherent logical connection. As a result, document processing faces challenges such as inaccurate classification, chaotic archiving, and difficulty in quantifying the disposal effect.

[0003] When facing the aforementioned challenges in non-performing asset recovery and management, the same debtor may need to undergo multiple rounds of recovery and disposal processes. Each round includes initial notification, sending a lawyer's letter, filing a lawsuit, court judgment, application for enforcement, termination of enforcement, and settlement. This process generates a large number of legal and business documents. Mixed storage of these documents further increases the complexity of classification and management, leading to inefficiency. Furthermore, existing recovery management systems lack intelligent correlation and classification mechanisms for documents across multiple recovery rounds and disposal stages, making it difficult to track the dynamic evolution of documents throughout the entire asset recovery cycle. Therefore, existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for classifying related documents in a multi-round collection and disposal process, which can realize accurate classification of single documents and automatic association of documents in multiple stages, thereby realizing closed-loop management of multi-round collection and disposal, significantly improving collection efficiency and reducing compliance risks.

[0005] To address the aforementioned technical problems, the first aspect of this invention discloses a method for classifying related documents in a multi-round collection and disposal process, the method comprising: Multiple collection documents are preprocessed and text-recognized to obtain legal documents and their corresponding structured metadata. Based on a preset classification model, the system outputs a set of category tags corresponding to the legal documents. Based on preset business rules, the legal documents are verified according to the category tag set and structured metadata to generate document classification results; The document classification results are archived, and the archiving results are continuously optimized based on preset debugging and / or monitoring mechanisms.

[0006] As an optional implementation, in the first aspect of the present invention, based on a preset classification model, a set of category tags corresponding to the legal document is output, including: Determine the modal type of the legal document, and select the corresponding modal model to construct the input representation of the legal document based on the modal type; Based on a pre-trained language model, the input representation is mapped to a dense vector representation to generate an embedding vector; The classification categories of the predefined linear classification head are connected to the embedding vector as a fully connected layer, and the corresponding classification score vector is output from the input representation. Based on a preset multi-label activation function, the independent probability of the legal document corresponding to any of the classification categories is calculated according to the classification score vector; The manually labeled category label set with the same dimension as the classification category, and the independent probability are used to construct the target loss function for classification training to obtain the label elements composed of the classification category and the independent probability, generating a structured category label set. The label elements in the category label set are filtered by a confidence threshold and weighted fusion by a multimodal supplementation mechanism.

[0007] As an optional implementation, in the first aspect of the present invention, the preset business rules include: Set a first business rule to extract the case number prefix of the document as a character identifier of the legal procedure type, construct a mapping relationship between the character identifier and the legal document, and output at least one category label for any character identifier identified in the legal document; A second business rule is set up to determine the set of keywords for category labels of the legal document based on semantic assistance according to business needs. The weights of the keyword elements in the keyword set are configured to perform confidence correction calculations for category labels, so as to adjust the confidence of the keyword association with the category label. Set a third business rule to define the target status of the disposal stage, and configure the target status for each document category in the legal documents based on its disposal stage.

[0008] As an optional implementation, in the first aspect of the present invention, based on preset business rules, the legal document is verified according to the category tag set and structured metadata to generate a document classification result, including: Load preset business rules to read case number mapping, keyword weight and stage configuration, and extract case number and keyword set from input legal documents and structured metadata; The first verification is performed based on the first business rule, and a unique category label is determined for the case number extracted from the legal document according to the case number mapping read. The second verification is performed based on the second business rule. The confidence level of the keyword is determined by the confidence level correction calculation. The category label corresponding to any keyword in the keyword set is re-determined based on the unique confidence level. The third verification is performed based on the third business rule, and the consistency of the cross-document context content in the legal document is verified and corrected according to the target status and the order of the disposal stages. Based on the triple verification results, a corrected set of category labels is output, as well as a document classification result for classifying the legal document according to the set of category labels. The document classification result includes the intervention logs corresponding to the business rules and the review flags.

[0009] A second aspect of this invention discloses a related document classification system for multi-round collection and disposal processes, the system comprising: The acquisition module is used to acquire multiple rounds of collection documents for preprocessing and text recognition to obtain legal documents and the corresponding structured metadata. The classification module is used to output a set of category tags corresponding to the legal document based on a preset classification model. The verification module is used to verify the legal document based on the preset business rules, the category tag set, and the structured metadata, and generate a document classification result. The output module is used to archive the document classification results and continuously optimize the archive processing results based on a preset debugging and / or monitoring mechanism.

[0010] A third aspect of this invention discloses another associated document classification system for multi-round collection and disposal processes, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the associated document classification method for multi-round collection and disposal processes disclosed in the first aspect of the present invention.

[0011] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the associated document classification method for multi-round collection and disposal processes disclosed in the first aspect of the present invention.

[0012] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention overcomes the bottlenecks of heterogeneous document formats and fragmented information by preprocessing documents through multiple rounds of structured metadata generation, thereby improving document compatibility and information relevance. By outputting a set of category tags based on a pre-defined classification model that covers the entire collection process, it achieves automated, high-precision classification, improving document classification accuracy and processing efficiency. Based on pre-defined business rules, it performs logical verification using category tags and structured metadata to generate consistent archiving results, eliminating logical contradictions and archiving confusion, thus improving archiving accuracy and the ability to quantify disposal effects. Furthermore, by implementing debugging and monitoring mechanisms on the archiving results, it automatically optimizes the classification model and business rules to implement disposal strategies. The continuous evolution of this approach solves the problems of fragmented document processing, inaccurate classification, mixed archiving, and difficulty in quantifying disposal effects in the management of non-performing assets in internet finance. It achieves accurate classification of single documents by stage, type, and confidence level, and automatically links documents at multiple stages based on case number and disposal process status, thereby enabling full traceability from repayment reminders to case closure. This leads to closed-loop management of multiple rounds of collection and disposal, making it particularly suitable for non-performing asset disposal scenarios of institutions such as banks, consumer finance companies, and trust companies. It can significantly improve collection efficiency while reducing compliance risks, providing a reusable technical paradigm for intelligent AI and judicial disposal. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a method for classifying related documents in a multi-round collection and disposal process disclosed in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of an associated document classification system for a multi-round collection and disposal process disclosed in an embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of another associated document classification system for a multi-round collection and disposal process disclosed in an embodiment of the present invention. Detailed Implementation

[0017] Before detailing the specific embodiments of the technical solution of this application, we will first reveal the application scenarios suitable for supporting the related document classification method for multi-round collection and disposal processes in the technical solution of this application.

[0018] The technical solution of this application is applicable to the field of Internet finance technology, especially applicable to the scenario of the recovery and disposal of non-performing assets in Internet finance. In this context, the same debtor often goes through multiple rounds of recovery and disposal processes, including the following disposal stages and corresponding document types: initial repayment reminder (phone call / sms), sending official recovery notice documents (lawyer's letter / repayment reminder letter), litigation filing (complaint, summons), court judgment (first instance / second instance judgment), application for enforcement (enforcement application, enforcement notice), enforcement termination or settlement (closing certificate, repayment agreement). In each round of disposal, a large number of legal and business documents are generated, and the documents generated at different stages of the same case are stored in a mixed manner, lacking the intelligent classification and correlation analysis capabilities based on the "integration of law, litigation, and enforcement" process of legal documents, litigation stages, and enforcement stages.

[0019] The existing disposal solutions for the recovery scenario include manual classification and filing, keyword-based rule engines, general document classification models, and stage-independent processing systems. Among them, manual classification and filing require operators to manually classify documents into corresponding folders such as "prosecution", "judgment", "enforcement", etc. according to the document title or experience, which has the defects of high labor cost, easy errors (such as confusing "ruling" and "judgment"), and inability to associate documents at multiple stages of the same case; the keyword-based rule engine uses regular expressions to match keywords such as "judgment as follows" and "application for enforcement" for label annotation, but has the defects of invalidity for scanned / OCR error texts, inability to process multi-label documents (such as "complaint + evidence list"), and complex rule maintenance and poor generalization; the general document classification model uses a public dataset to fine-tune BERT for single-label classification, with defects such as not integrating legal business logic (such as the case number prefix "execution" must belong to the enforcement stage), ignoring the sequential constraints of the law, litigation, and enforcement processes, and low recognition rate for low-frequency categories (such as "closing certificate") (such as the recognition rate < 40%); in the stage-independent processing system, independent systems are used for prosecution, trial, and enforcement in the integrated process of law, litigation, and enforcement, resulting in fragmented documents that cannot be tracked across stages, and scattered multiple recovery records of the same debtor, making it impossible to evaluate the disposal effect. It can be seen that the existing technology breaks the internal logical connection of the three stages of "law", "litigation", and "enforcement", resulting in inaccurate classification, chaotic filing, and difficult quantification of the disposal effect.

[0020] Furthermore, Table 1 corresponds to a typical multi-round recovery case to describe the defects of the existing technology in detail. The typical case is the cyclic process of the three-round recovery and disposal process of debtor A (in the table, "final termination of execution" is a professional term in the court enforcement procedure, meaning the termination of the current enforcement procedure): Table 1 - Typical Multi-round Recovery and Disposal Example

[0021] As can be seen from the above typical examples of multi-round debt collection and disposal, the existing technology lacks an intelligent association and classification mechanism that spans rounds and stages and is based on case number and process status.

[0022] To address the shortcomings of the existing technologies, this invention discloses a method and system for classifying related documents in a multi-round collection and disposal process. The method involves acquiring and preprocessing multi-round collection documents, performing text recognition to obtain legal documents and their structured metadata. Based on a preset classification model, a set of category tags is output according to the legal documents. Based on preset business rules, the legal documents are validated according to the category tag set and structured metadata to generate document classification results. The document classification results are archived, and continuous optimization is performed based on preset debugging and / or monitoring mechanisms. This enables intelligent related classification of legal documents across rounds and disposal stages, solving the problems of fragmented document processing workflows and lack of internal logical connections, and improving the accuracy of classification, archiving efficiency, and quantifiable disposal effects in non-performing asset collection management. Detailed explanations follow.

[0023] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for classifying associated documents in a multi-round collection and disposal process, as disclosed in an embodiment of the present invention. Figure 1 The described document classification method for multi-round collection and disposal processes can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 1 As shown, this document classification method for multi-round collection and disposal processes can include the following operations: 101. Obtain multiple rounds of collection documents, perform preprocessing and text recognition to obtain legal documents and corresponding structured metadata.

[0024] Optionally, the collection documents for multiple rounds include original documents, duplicate documents, and image documents generated at each stage of the collection process, such as the repayment reminder stage, mediation stage, sending lawyer's letters stage, litigation stage, judgment stage, enforcement stage, and case closure stage.

[0025] Optionally, the input format for the collected documents can be a natively searchable PDF, a scanned PDF, an image file, or a Word document. A natively searchable PDF contains embedded text, a scanned PDF consists of images without a text layer, image files can be in JPG or PNG format, and Word documents can be in .docx or .doc format. Supported formats can be configured according to the format type; this invention does not impose any limitations. Optionally, the legal documents are uniformly output structured, noise-free text and semantically complete text. The structured metadata can be case number, parties, debtors, date, disposal stage, document name and other metadata information used to assist in classification. This invention does not impose specific limitations.

[0026] 102. Based on a preset classification model, output a set of category tags corresponding to the legal document.

[0027] Optionally, the category label may be a predefined category related to the collection and disposal process, such as lawsuit, response to lawsuit, summons, judgment, ruling, mediation, enforcement application, enforcement notice, and case closure certificate. This invention does not impose any limitations on this.

[0028] 103. Based on preset business rules, the legal documents are verified according to the category tag set and structured metadata to generate document classification results.

[0029] Optionally, the business rules, based on the initial classification results of the model, can be modified or enhanced to ensure that the classification results conform to the business logic of the integration of law, litigation and enforcement (e.g., the process chain of filing a lawsuit → trial → enforcement), while improving the reliability of the judgment results of high-value categories (e.g., enforcement documents).

[0030] 104. The document classification results are archived, and the archiving results are continuously optimized based on a preset debugging and / or monitoring mechanism.

[0031] As can be seen, the above-described embodiments of the invention overcome the bottlenecks of heterogeneous document formats and fragmented information by preprocessing documents and generating structured metadata through multiple rounds, thereby improving document compatibility and information relevance. By outputting a set of category tags based on a preset classification model to cover the entire collection process, automated and high-precision classification is achieved, improving document classification accuracy and processing efficiency. Based on preset business rules, logical verification is performed using category tags and structured metadata to generate consistent archiving results, eliminating logical contradictions and archiving confusion, thus improving archiving accuracy and the ability to quantify processing effectiveness. Furthermore, by implementing debugging and monitoring mechanisms on the archiving results, the classification model and business rules are automatically optimized. The continuous evolution of current disposal strategies has solved the problems of fragmented document processing, inaccurate classification, mixed archiving, and difficulty in quantifying disposal effects in the management of non-performing assets in internet finance. It achieves accurate classification of single documents by stage, type, and confidence level, and automatically links documents at multiple stages based on case number and disposal process status, thereby achieving full traceability from repayment reminders to case closure. This enables closed-loop management of multiple rounds of collection and disposal, which is particularly suitable for non-performing asset disposal scenarios of institutions such as banks, consumer finance companies, and trust companies. It can significantly improve collection efficiency, reduce compliance risks, and provide a reusable technical paradigm for intelligent AI and judicial disposal.

[0032] As an optional embodiment, the step of obtaining and preprocessing multiple rounds of collection documents in the above steps includes the following steps: The document is identified to determine its type, and a first branch or a second branch is selected for processing based on the document type, wherein the first branch is a text extraction branch and the second branch is an image processing branch; For the collection document that selects the first branch, traverse all pages to extract a structured text sequence, which includes page number information and the text content corresponding to the page number; For the collection documents selected in the second branch, image objects are obtained through a conversion tool. Each image object is then subjected to grayscale conversion, binarization, noise reduction, tilt correction, and size normalization to obtain an image sequence.

[0033] As an example, the document type of the input collection and disposal document is determined by its file extension, such as... Then proceed to the PDF processing branch. Then the input is treated as an image for image preprocessing branch, if If Word support is already enabled, first convert it to PDF using the background tool before proceeding to the PDF processing branch.

[0034] Furthermore, for PDF files that initially enter the first branch, the system identifies them and extracts the text from the first page. If the extracted text length L > 50 characters (the character threshold is an empirical value), it is determined to be a native PDF, and the full text is extracted directly; otherwise, it is determined to be a scanned PDF, and the system enters the image processing flow of the second branch.

[0035] Furthermore, native PDF text extraction uses an extraction tool to traverse all pages, calls an extraction plugin to obtain plain text, and retains page number information to output elements as a structured text sequence: , where T represents an element of the structured text sequence. This indicates page i. This represents the text content of page i, and N represents the maximum number of pages in the original PDF text.

[0036] Furthermore, for the scanned PDF that enters the second branch, the image conversion tool (DPI set to 200) is called to convert each page of the PDF into a PIL image object and save it as an image list: the image list converted from the scanned PDF and the input image type image are preprocessed, and each image is processed in sequence as follows: (1) Grayscale conversion: convert to a single-channel grayscale image; (2) Binarization: use an adaptive threshold (such as the Sauvola or Otsu method) to enhance the contrast between text and background; (3) Denoising: use morphological operations (opening operation) to remove isolated noise points; (4) Tilt correction: estimate the rotation angle through Hough transform or text line detection, and align the rotation angle by rotating it by 180 degrees; (5) Size normalization: unify the long side to 2000 pixels and maintain the aspect ratio; finally, the preprocessed image sequence is output for subsequent OCR recognition.

[0037] Finally, a text sequence is generated from the input native PDF file, and preprocessed image sequences are generated from other non-native PDF files. These images are then fed into the recognition step for further text recognition processing.

[0038] As can be seen, through the above optional embodiments, the processing branch is dynamically selected by document type, which avoids the recognition deviation caused by uniform processing and improves document compatibility. By traversing the document pages to extract the structured text sequence containing page number markers, the context of key elements is preserved, which improves the integrity of the original document's temporal information and the accuracy of key element extraction. By performing a standardized image processing flow on non-native documents, a high-fidelity image sequence is generated, thereby breaking through the bottleneck of low-quality document recognition. In this way, the accurate conversion from isolated documents to high-fidelity structured metadata is realized, laying a data foundation for the intelligent association and classification of documents throughout the entire collection and disposal cycle.

[0039] As an optional embodiment, the step of obtaining multiple rounds of collection and disposal documents for text recognition in the above steps includes the following steps: The text line sequence of the image sequence is recognized based on a preset OCR model, and the text line sequence is concatenated into a single page of text according to a preset reading order to obtain the OCR-recognized text sequence. The text sequence and the structured text sequence are merged and unified into a text sequence to be processed; Noise characters are filtered out for each page of text in the text sequence to be processed, and the first and last lines of each page are extracted to determine the high-frequency headers and / or footers to construct a high-frequency line set. Any high-frequency line element contained in the high-frequency line set is deleted from each page of text. The text sequence to be processed is matched with business requirements using regular expressions to extract structured metadata of the legal fields corresponding to the business requirements; The processed text sequence is reorganized into paragraphs and merged for deduplication. Based on NLP enhancement, the text sequence is subjected to association processing stage and Chinese word segmentation and named entity recognition of corresponding stage documents, and legal documents with entity tags are output.

[0040] Specifically, the above steps are the operation flow of text recognition and further preprocessing layer. Its input is the original text sequence obtained from the native PDF processing of the first branch and the image sequence obtained from the non-native PDF processing of the second branch. It is used to uniformly output structured, noise-free and semantically complete legal text so as to facilitate subsequent intelligent classification.

[0041] As an example, for OCR recognition of an input image sequence, either Tesseract OCR or the high-precision PaddleOCR can be selected. Tesseract OCR is configured with the LSTM engine enabled in multilingual mixing mode, while PaddleOCR is configured with an orientation classifier and detection model to output text lines with bounding boxes. The recognition result output by the above OCR is a sequence of text lines. , of which elements This represents the identified text on page i, line j. This identified text is then concatenated into a single page of text according to the reading order. Finally, the text sequence recognized by OCR is obtained. It is understandable that using PaddleOCR for recognition can provide additional layout information (such as "title", "table", "body text" etc.) for subsequent paragraph reconstruction.

[0042] Furthermore, the original PDF text and OCR text from the first branch are merged into a unified input. Text cleaning and standardization are then performed. For each page of text after unified input, noise character filtering is first executed: garbled characters (such as consecutive non-Chinese / English / numeric characters, length > 5) are removed; meaningless symbols (such as "□□□", "...", lines accounting for > 30%) are filtered out; blank pages are deleted. Secondly, header and footer identification and removal are performed: the first and last two lines are extracted; if they appear ≥ 3 times in the entire document, they are determined to be headers or footers; a high-frequency line set is constructed. Remove from each page of the unified input Next, key legal fields are extracted to aid subsequent classification. For example, regular expressions are used to match the following fields (saved as structured metadata): Case Number: ; debtor: ; Date (e.g., case filing / judgment / enforcement) ; Output the extracted results without deleting them from the original text of the same input.

[0043] Furthermore, after extracting the execution fields, paragraph reorganization and deduplication are performed. The paragraph reorganization process involves merging lines that end with a non-punctuation mark and whose next line begins with a lowercase or Chinese character (to avoid OCR line break errors); while preserving the typical structure of legal documents (such as "After investigation:" or "This court holds:" as paragraph start markers). The deduplication process after paragraph reorganization is as follows: if the text similarity of two consecutive pages is... (Similarity can be calculated using Jaccard's algorithm or Locality Sensitive Hashing algorithm), then one of the texts is retained, and the entire text is concatenated into a single string, retaining the page number separator `<`.<PAGE i> Finally, NLP is used to enhance semantic understanding for classification (this process improves the model's sensitivity to keywords such as "XX court judgment"). The steps include using LAC or LTP to perform Chinese word segmentation and named entity recognition on the full text; labeling entity types such as `ORG` (court, company), `PER` (party), `TIME`, and `CASE-ID`; outputting text with entity labels or simply saving the entity list for classifier feature fusion. The final output is the cleaned full legal text. and structured metadata To provide intelligent classification down to the next level, the structured metadata, as an example, includes a unique identifier associated with the case number, parties or participants related to the case, date information related to the case, and entities related to the case, where entities can be organizations associated with the collection case (such as XX court), events (such as the enforcement phase), etc., as a collection that connects the entire collection process through identifiers, participants, and date information.

[0044] As can be seen, through the above optional embodiments, by recognizing text lines in an image sequence based on a preset OCR model and splicing them into a single page of text in the reading order to generate a coherent OCR text sequence, the coherence of the text is improved while reducing the omission rate of key elements. By merging the OCR text sequence with the structured text sequence into a unified sequence to be processed, information consistency is ensured, multi-source information conflicts are eliminated, the quality of the structured foundation is improved, and the information redundancy rate is reduced. By noise filtering and high-frequency line processing, the text semantic environment is evolved, improving the purity of the text and the proportion of key content. By using regular expressions to accurately extract structured metadata, the field capture is automated. By using NLP to enhance semantic association and label entity tags such as "lawyer's letter" and "mediation" at the disposal stage, multi-stage dynamic documents are constructed, significantly improving the depth of document semantic understanding and the accuracy of the association at the disposal stage. Through multi-round collection document text recognition and deep semantic enhancement mechanisms, the problems of fragmented document information, severe noise interference, and coarse extraction of key fields in the management of non-performing assets in Internet finance are solved, realizing the intelligent conversion from raw text to high-precision structured metadata, providing high-quality input for full-cycle intelligent association archiving.

[0045] As an optional embodiment, the step above, outputting a set of category tags corresponding to the legal document based on a preset classification model, includes: Determine the modal type of the legal document, and select the corresponding modal model to construct the input representation of the legal document based on the modal type; Based on a pre-trained language model, the input representation is mapped to a dense vector representation to generate an embedding vector; The classification categories of the predefined linear classification head are connected to the embedding vector as a fully connected layer, and the corresponding classification score vector is output from the input representation. Based on a preset multi-label activation function, the independent probability of the legal document corresponding to any of the classification categories is calculated according to the classification score vector; The manually labeled category label set with the same dimension as the classification category, and the independent probability are used to construct the target loss function for classification training to obtain the label elements composed of the classification category and the independent probability, generating a structured category label set. The label elements in the category label set are filtered by a confidence threshold and weighted fusion by a multimodal supplementation mechanism.

[0046] It should be noted that this embodiment describes the operation flow of the intelligent classification core layer in the classification system. This layer aims to automatically classify cleaned legal documents (which may include images or metadata) and output their respective integrated legal, litigation, and enforcement category tag sets, such as {"complaint", "judgment", "enforcement notice", "list of evidence"}, etc. It also supports multiple tags, high accuracy, and local deployment.

[0047] As an example, for plain text input types, legal documents... Cut off or padded to maximum length To adapt to most BERT models, special tags (i.e., input representations) are added. ,in The word segmenter output of WordPiece or SentencePiece. For multimodal image-text input types (such as complex documents containing tables, seals, and handwritten annotations), if the original input is a scanned document and the image is retained, the first page image is taken and the legal document and the first page image are fed into the language model as joint input. The image and text semantics are fused through cross-modal attention. It is worth noting that the multimodal fusion strategy is only enabled when the text confidence is low (such as high OCR error rate) or the categories are easily confused (such as "mediation document" vs "ruling document").

[0048] Furthermore, a pre-trained language model is selected to represent the input. The mapping is used to generate embedding vectors from dense vector representations. The pre-trained language model can be either the basic choice RoBERTa-wwm-ext or the high-retrieval-demand-adapted BGE-large-zh. It's worth noting that due to the existence of few-shot scenarios in sample classification, choosing the BGE-large-zh pre-trained language model provides better generalization in few-shot classification. The [CLS] vector in the input representation is optimized by comparing it to alcoholism, making it more suitable for tasks with clear classification boundaries. The generated embedding vector is represented as follows: Optional, dimension d=768 to fit BERT-like models, where This represents the corresponding model parameters.

[0049] Furthermore, the structure of the classification model in the classification step is embedded in the embedding vector. A fully connected layer (linear classification head) above it can be represented as: ,in Let W represent the classification score vector, b represent the weight matrix of the fully connected layer, b represent the bias vector of the fully connected layer, and K represent the dimension of the predefined categories. For example, if K=9, the predefined categories cover "lawsuit, response to lawsuit, summons, judgment, ruling, mediation, enforcement application, enforcement notice, case closure certificate," etc. Multi-label output is achieved through embedding vectors and sigmoid activation. The calculation formula is as follows: ; In the formula, This represents the independent probability that a document belongs to the Kth predefined category. This represents the embedding vector of the k-th class; the training objective is set using the binary cross-entropy loss function for multi-label outputs, expressed as: ; In the formula This indicates manually labeled categories, such as the combination of "judgment and enforcement notice," which corresponds to two categories. .

[0050] Optionally, for the deployment optimization of classification models, the model can be exported to ONNX format and loaded using ONNX Runtime to improve inference speed. At the same time, multiple documents input to the model can be grouped into mini-batc (e.g., batch_size=8) for batch processing to improve throughput.

[0051] Optionally, the label elements in the category label set can be controlled by a confidence threshold, such as setting a category-level threshold. (High-risk categories such as "Execution Notification" have high thresholds), and the final predicted set of category labels is output based on the confidence threshold: It is worth noting that if That is, category classification results with no labels that meet the criteria or with overly vague labels, such as If the result is not satisfactory, it will be transferred to the manual review queue for re-evaluation.

[0052] Optionally, when the input document is a scanned copy with a complex layout, and the maximum probability of the main classification model is less than a preset threshold, the tag elements in the category tag set may be omitted. If this happens, a multimodal supplementation mechanism is activated. The label elements are output using a language model such as Qwen-VL. Simultaneously, a weighted fusion is applied to the outputs of the classification model and the language model. For example, the weight of the classification model output is 0.7, and the weight of the language model output is 0.3. It's understandable that the language model output still needs to be processed by a Sigmoid function. Multi-label output of activation functions and determination of confidence threshold.

[0053] As an example, the final structured output includes: a list of predicted category labels; the predicted probability for each label; a marker indicating whether manual review is required; and the name of the model that generated the result. The results can be formatted using LaTeX mathematical typesetting, presenting the key-value pairs in a clear, aligned visual format for easy reading and understanding. This structured result can then be input into the subsequent process engine layer and business rule layer to drive the automatic archiving and task assignment of legal and enforcement documents.

[0054] It can be seen that through the above optional embodiments, by dynamically selecting the corresponding model based on the document modality type, semantic distortion caused by unified processing is avoided, the model adaptation rate is increased, and the allocation error rate is reduced. By mapping the input representation into a dense vector based on the pre-trained language model, deep semantic associations of legal elements such as "person subject to execution" and "subject matter of execution" are captured, the document semantics are deeply restored, and the relevance and classification quality of semantic representation are improved. By accessing the linear classification head through predefined classification categories to output the classification score vector, multi-label parallel computing is supported, precise mapping and efficient output of multiple categories are achieved, the Sigmoid activation function is used to calculate the independent probability of each category, the mutually exclusive assumption is avoided, the category overlap rate is reduced, and the probability calculation accuracy is improved. Based on manual annotation, a target loss function is constructed, the labels are screened through a confidence threshold, and a multi-modal supplementary mechanism is introduced for weighted fusion to generate high-confidence structured labels, solving the problems of inaccurate document classification, category conflicts, and messy filing in the management of non-performing assets in Internet finance, realizing the precise generation of high-confidence structured labels from fuzzy labels, and providing high-quality classification input for full-cycle intelligent association filing.

[0055] As an optional embodiment, in the above steps, the preset business rules include: Set the first business rule, extract the case number prefix of the document as the character identifier of the legal procedure type, construct a mapping relationship between the character identifier and the legal document, and the mapping relationship outputs at least one category label for any character identifier recognized in the legal document; Set the second business rule, determine the keyword set for the category label based on the semantics of the legal document according to business requirements, configure weights for the keyword elements in the keyword set for confidence correction calculation of the category label, so as to adjust the confidence of the keyword associated with the category label; Set the third business rule, define the target state of the disposal stage, and configure the target state for the category of each document in the legal document based on its belonging disposal stage.

[0056] Specifically, the preset business rules can be configured through rule library design. The rules can be maintained in the form of YAML or JSON files and are divided into three categories. The first business rule is the case number prefix rule of strong business signals The case number format example: `(2024) Jing0105 Zhixing 1234 Hao`, and the extracted case number prefix is expressed as: , where the last Chinese characters in the case number format example (such as "Zhi", "MinChu", and "MinZhong", etc.) are program type identifiers. The next relationship configuration example is: case_prefix_map: "Zhi": ["Execution Notice", "Execution Ruling", "Execution Application"]; "MinChu": ["Complaint", "Acceptance Notice"]; "MinZhong": ["Second Instance Judgment", "Second Instance Ruling"]; "XingChu": ["Criminal Indictment", "Criminal Judgment"]. If the rule action then it is forced to ensure that at least one "execution class" label is included in the output label; otherwise, it is marked as "rule conflict" and a review is triggered.

[0057] Furthermore, the second business rule is the key word weight rule for semantic assistance, which assigns weights to keyword phrases to adjust the confidence of each category, as shown in Table 2: Table 2 - Exemplary Keyword Confidence Assignment

[0058] Then, the confidence correction calculation is performed through the configured weights, and the correction calculation formula is expressed as: ; In the formula, represents the independent probability of the category label after the correction calculation, represents the keyword set corresponding to the k-th category, represents the just function, with a value of 1 or 0, represents the mediation factor, with a value range of 0.2 - 0.3. The category label after the correction calculation still needs to be rejudged as the final label through the confidence threshold.

[0059] Furthermore, the third business rule is the litigation execution state machine rule for process consistency verification. Exemplarily, three target states are defined as: , and the target state of the preset belonging stage is configured for each document category. Exemplarily, the configuration is: the complaint document is in the prosecution state stage, the acceptance notice document is in the prosecution state stage, the first instance judgment document is in the trial state stage, the second instance ruling document is in the trial state stage, and both the execution application document and the execution notice document are in the execution state stage.

[0060] Finally, business constraint rules are constructed. If the system already knows that the case has entered the "execution" stage previously (such as the historical document contains an "Execution Notice"), then the "Complaint" should not appear in the new document; if the document contains an "execution" class label but the case number is "MinChu", it is regarded as an anomaly and requires manual review. This business rule constraint depends on the case-level context that associates historical documents through the case number to achieve cross-document logical consistency.

[0061] As can be seen, through the above optional embodiments, by extracting the document case number prefix to construct a character identifier-category mapping table, it is ensured that each identifier strictly outputs the corresponding category label, thereby establishing a standardized classification benchmark to eliminate semantic ambiguity. By defining keyword sets and configuring weights based on business needs, and dynamically adjusting the reliability of labels through a confidence correction formula, the confidence level is dynamically optimized and the label credibility is accurately quantified. By defining the target state of the disposal stage, the target state of the document category is configured based on the stage to ensure the consistency of stage logic, support full-cycle association, and solve the problems of inaccurate document classification, category conflict, and chaotic archiving logic in Internet finance non-performing asset management. It realizes the accurate generation of high-confidence structured labels from fuzzy labels, providing highly consistent and highly quantifiable classification input for full-cycle intelligent association archiving.

[0062] As an optional embodiment, the above steps, including verifying the legal document based on the preset business rules, the category tag set, and structured metadata, and generating a document classification result, include: Load preset business rules to read case number mapping, keyword weight and stage configuration, and extract case number and keyword set from input legal documents and structured metadata; The first verification is performed based on the first business rule, and a unique category label is determined for the case number extracted from the legal document according to the case number mapping read. The second verification is performed based on the second business rule. The confidence level of the keyword is determined by the confidence level correction calculation. The category label corresponding to any keyword in the keyword set is re-determined based on the unique confidence level. The third verification is performed based on the third business rule, and the consistency of the cross-document context content in the legal document is verified and corrected according to the target status and the order of the disposal stages. Based on the triple verification results, a corrected set of category labels is output, as well as a document classification result for classifying the legal document according to the set of category labels. The document classification result includes the intervention logs corresponding to the business rules and the review flags.

[0063] As an example, the rule base is loaded as follows: Case number mapping, keyword weights, and stage configurations are read from `rules.yaml`; case number C and the matching keyword set are obtained from metadata and text; a triple validation is performed, where the first validation is a mandatory case number validation: if the case number specifies "execution" but the model does not output an execution class label, the highest weight execution label is automatically added; the second validation is keyword confidence correction, adjusted according to the correction calculation formula. The category labels are re-determined; the third verification is a state machine consistency check, which marks the classification results that violate the order of the handling stages of the target state as `needs_review = True`; the final output is the enhanced classification result, which is used for archiving and task distribution. Its structured classification result includes the final classification labels (final_labels); the adjusted classification scores (adjusted_scores); the rule actions used (rule_actions); and a flag indicating whether review is required. The legal business rule mechanism integrated in this embodiment ensures that the model classification results are both intelligent and compliant, closely matching real business scenarios such as courts, arbitration, and enforcement, significantly reducing the misclassification rate, and improving robustness, especially in the recognition of high-value enforcement documents.

[0064] As can be seen, through the above optional embodiments, by loading the case number mapping table, the case numbers of legal documents are uniquely matched to ensure that the category label output is unambiguous, reducing the classification error rate and improving the consistency of the archiving benchmark; by correcting the confidence based on keyword weights, the confidence of the labels is dynamically optimized by re-determining the keyword-related labels; by verifying the cross-document context consistency based on the target status of the disposal stage, the consistency of the document time sequence throughout the entire cycle is ensured; and by generating a highly reliable classification result through triple verification and fusion output, the problems of high document classification error rate, stage logic breakage, and lack of archiving consistency in Internet finance non-performing asset management are solved, realizing the accurate output of document classification from fuzzy labels to strongly consistent structured documents.

[0065] As an optional embodiment, the document classification results are archived in the above steps. The system integration and service layer operation process is executed to perform batch-running, archived and debuggable complete document processing on the classification content output in the above embodiment. At the same time, it can achieve localized, high-throughput and structured output, avoid dependence on external message queues or distributed frameworks, and conform to thread pool concurrency preferences and local path operation habits.

[0066] Specifically, the archiving process includes input configuration, concurrency and batch processing mechanisms, output generation, automatic archiving and file organization, API services, and running mode switching. The input configuration steps are as follows: The input source specifies a local directory path, supports automatic traversal of subdirectories, and collects files with extensions `.pdf`, `.jpg`, `.png`, and `.docx`; a file list is generated to create a processing queue. , where M is the total number of files.

[0067] Furthermore, the implementation of concurrency and batch processing mechanisms includes task sharding and thread pool concurrency, where task sharding sets the batch size. Preferred To balance memory and throughput, the file list is then divided into blocks; this process is represented as follows: , This represents the total number of batches, where thread pool concurrency refers to the number of worker threads in the startup thread pool. Each batch is submitted as a task to a `ThreadPoolExecutor`, and the process within the thread processes each file in the batch. The system sequentially executes steps such as document acquisition, OCR / NLP, model classification, and rule processing, without any global locks other than I / O blocking. The concurrency and batch processing mechanisms are released during I / O operations (such as disk read / write and OCR calls) through the Python GIL, and the thread pool can effectively improve the throughput of multiple files.

[0068] Furthermore, after each file is processed, a structured result is generated and uniformly written to the output directory, such as `... / law and litigation_output / `. The output generation implementation includes the result data structure and batch output format, where the result data is formatted for each file. The output is a JSON object, which includes the file storage path (identifying the file's location in the system), a unique hash value for the file (used for deduplication, ensuring consistent hashes for files with the same content to avoid duplicate processing, and for tracing results by associating hash values ​​with historical processing records), a list of predicted file tags (representing the "document types" identified by the AI ​​model / rule module; for example, ["Judgment", "Execution Notice"] indicates the presence of both "Judgment" and "Execution Notice" legal documents), a tag confidence score (corresponding to the "prediction confidence level" of each tag in the predicted file tag list; the value range is usually [0, 1], with higher values ​​indicating greater confidence in the model's prediction of the tag; for example, "Judgment": 0.92 indicates a 92% confidence level for the "Judgment" tag), a case number (a unique case identifier extracted from the file, used to associate the case's legal process and information), and a tag-page range mapping (the key is the document type, such as "Judgment" or "Execution Notice," and the value is the corresponding page number range for that type of content in the file, in array form, such as [1, 3]). The structured results indicate whether manual review is required (false indicates the result is reliable and no manual intervention is needed, while true indicates manual verification is required, such as when the confidence level is low or there are rule conflicts), and the file processing time (used to monitor system performance, such as 2.4 seconds, which means that it took 2.4 seconds from reading the file to generating the classification result). These structured results serve the automated processing of legal documents, realizing the "automatic classification → content location → quality verification → performance tracking" of documents, thereby improving the efficiency and accuracy of legal document processing.

[0069] Furthermore, the implementation of automatic archiving and file organization involves automatically moving the original files to the archive directory according to the classification results after the classification process is completed. An exemplary structure is shown as follows: D_archive / ├── Complaint / ├── Judgment / ├── Execution Notice / ├── Evidence / ├── Files pending review / ← Files with needs_review = true └── multi_label / ← Contains multiple main categories (such as "prosecution + evidence") and has no dominant label; The archiving strategy in this process is: if the final classification result If the final classification result is... If the case is filed in the order of the legal proceedings stage (execution > trial > prosecution), the main category should be selected first, or the case should be placed in the `multi_label / ` directory. If `needs_review = true`, it should be forcibly moved to `pending review / `. For multiple PDFs archived in the same category, they can be automatically merged into a single file by case number.

[0070] Furthermore, the API service is implemented by launching a lightweight REST service (such as based on FastAPI), listening on a local port (such as `http: / / localhost:8000`); providing two interfaces, one of which is a single file classification interface that returns JSON classification results for uploaded files; and the other of which is a batch status query interface that enables an asynchronous task queue to query the progress of tasks being processed in the background.

[0071] Furthermore, the implementation of the running mode switching supports either the default batch processing mode or the debug mode. In batch processing mode, the input directory is specified in the command line, all data is processed, and the output is archived as JSON or CSV. In debug mode, enabling `debug=true` records the intermediate results (OCR text, input, rule logs) of each file to the `debug / ` subdirectory, with each file corresponding to a `.txt` log file. The input configuration, concurrency and batch processing mechanisms, output generation, automatic archiving and file organization, API services, and running mode switching in the above embodiments ensure that the system is ready to use, locally efficient, and results are traceable, fully adapting to the closed-loop requirements of batch classification, archiving, and quality inspection of legal documents.

[0072] As an optional embodiment, the above steps, based on a preset debugging mechanism, continuously optimize the archiving process results, including: For each input collection document, a corresponding debugging log is generated. The content of the debugging log is sorted in the processing order as follows: document meta information, OCR output original text, extracted case number, extracted associated roles, extracted keywords, classification model input, original category labels and confidence of corresponding category labels, business rule intervention records, and document classification results. The debugging log specifically records samples below the confidence threshold and samples with classification anomalies, and uses these specific records as negative samples to iteratively train and optimize the document classification model and business rules.

[0073] Specifically, the debugging mechanism ensures that the classification process is observable, traceable, and optimizable, and particularly supports local debugging and performance evaluation, perfectly meeting the explicit requirements for log callbacks, error sample retention, and efficiency quantification.

[0074] As can be seen, through the above optional embodiments, by generating structured debugging logs for each collection document, the complete link from metadata to classification results is recorded in the processing order, achieving problem traceability and accurate location; by automatically labeling samples with confidence levels and classification anomalies, special records are generated as a negative sample library to focus on optimizing key problem points and improve the targeting of optimization; by driving iterative optimization through negative samples, the model and rules achieve self-learning evolution, thereby forming a closed-loop optimization mechanism driven by the full-link debugging log, solving the problems of difficulty in locating problems in archived results, low optimization efficiency, and weak system self-learning ability in the management of non-performing assets in Internet finance, and thus achieving an intelligent upgrade from static archiving to dynamic self-evolution.

[0075] As an optional embodiment, the above steps, based on a preset monitoring mechanism, continuously optimize the archiving process results, including: Evaluation metrics are set for the generated document classification results to output a classification performance report. The evaluation metrics include processing time, throughput, error rate, and proportion of manual review. Based on the classification performance report, the classification errors of the document classification results are identified, and the root causes of the classification errors are located. The classification errors and corresponding root causes are then fed back to the model and rules of the classification process for iteration.

[0076] Specifically, the monitoring mechanism ensures that the classification process is observable, traceable, and optimizable, and particularly supports local debugging and performance evaluation, perfectly meeting the explicit requirements for log callbacks, error sample retention, and efficiency quantification.

[0077] Optionally, the expression for processing time (in seconds) in the classification performance report is as follows: ; in, For initial input The time required for end-to-end processing This indicates the average time taken for a single file, and can also be used to calculate the total time taken. .

[0078] Optionally, throughput (in documents per second) in the classification performance report is used as the core efficiency metric, and its expression is: ; In the formula, This indicates the number of files that were correctly classified and archived.

[0079] Optionally, the expression for the error rate in the classification performance report is: ; In the formula, This indicates the number of files that failed to be categorized due to factors such as unsupported format or corrupted files.

[0080] Optionally, the expression for the proportion of manual review in the classification performance report is: ; In the formula, This indicates the number of documents manually reviewed. The percentage of documents manually reviewed reflects the degree of automation in the classification process and is typically less than 5%.

[0081] As can be seen, through the above optional embodiments, by setting four core indicators—processing time (single document processing time), throughput (average number of documents processed per day), error rate (percentage of classification errors), and manual review ratio (percentage of cases requiring manual intervention)—a structured classification performance report is generated. This multi-dimensional evaluation index quantitatively assesses and precisely defines the archiving quality baseline, improving the objectivity of classification quality assessment and the efficiency of problem identification. By deeply analyzing the root causes of errors based on the performance report to generate a structured root cause report, precise root cause identification and classification error tracing are achieved to overcome optimization bottlenecks. Furthermore, by directly feeding the root causes of errors back to the classification model and business rules through a feedback-driven dynamic iteration mechanism, a closed loop of "monitoring, analysis, and optimization" is realized. This addresses the problems of difficulty in locating archiving results issues, long optimization cycles, and weak system self-learning capabilities in internet finance non-performing asset management. It achieves an intelligent leap from passive response to proactive predictive optimization, resulting in rapid identification of classification error root causes, quantification of efficiency and stability, and support for continuous iteration of models and rules.

[0082] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an associated document classification system for a multi-round collection and disposal process disclosed in an embodiment of the present invention. Figure 2The described document classification system for multi-round collection and disposal processes can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 2 As shown, this document classification system for multi-round collection and disposal processes may include: The acquisition module 201 is used to acquire multiple rounds of collection documents for preprocessing and text recognition to obtain legal documents and the corresponding structured metadata of the legal documents; The classification module 202 is used to output a set of category tags corresponding to the document based on a preset classification model and the legal document. Verification module 203 is used to verify the legal document based on the preset business rules, the category tag set, and the structured metadata, and generate a document classification result. The output module 204 is used to archive the document classification results and continuously optimize the archive processing results based on a preset debugging and / or monitoring mechanism.

[0083] As an optional implementation, preprocessing the collection documents from multiple rounds includes the following steps: The document is identified to determine its type, and a first branch or a second branch is selected for processing based on the document type, wherein the first branch is a text extraction branch and the second branch is an image processing branch; For the collection document that selects the first branch, traverse all pages to extract a structured text sequence, which includes page number information and the text content corresponding to the page number; For the collection documents selected in the second branch, image objects are obtained through a conversion tool. Each image object is then subjected to grayscale conversion, binarization, noise reduction, tilt correction, and size normalization to obtain an image sequence.

[0084] As an optional implementation, text recognition is performed on multiple rounds of collection and disposal documents, including the following steps: The text line sequence of the image sequence is recognized based on a preset OCR model, and the text line sequence is concatenated into a single page of text according to a preset reading order to obtain the OCR-recognized text sequence. The text sequence and the structured text sequence are merged and unified into a text sequence to be processed; Noise characters are filtered out for each page of text in the text sequence to be processed, and the first and last lines of each page are extracted to determine the high-frequency headers and / or footers to construct a high-frequency line set. Any high-frequency line element contained in the high-frequency line set is deleted from each page of text. The text sequence to be processed is matched with business requirements using regular expressions to extract structured metadata of the legal fields corresponding to the business requirements; The processed text sequence is reorganized into paragraphs and merged for deduplication. Based on NLP enhancement, the text sequence is subjected to association processing stage and Chinese word segmentation and named entity recognition of corresponding stage documents, and legal documents with entity tags are output.

[0085] As an optional implementation, based on a preset classification model, a set of category tags corresponding to the legal document is output, including: Determine the modal type of the legal document, and select the corresponding modal model to construct the input representation of the legal document based on the modal type; Based on a pre-trained language model, the input representation is mapped to a dense vector representation to generate an embedding vector; The classification categories of the predefined linear classification head are connected to the embedding vector as a fully connected layer, and the corresponding classification score vector is output from the input representation. Based on a preset multi-label activation function, the independent probability of the legal document corresponding to any of the classification categories is calculated according to the classification score vector; The manually labeled category label set with the same dimension as the classification category, and the independent probability are used to construct the target loss function for classification training to obtain the label elements composed of the classification category and the independent probability, generating a structured category label set. The label elements in the category label set are filtered by a confidence threshold and weighted fusion by a multimodal supplementation mechanism.

[0086] As an optional implementation, the preset business rules include: Set a first business rule to extract the case number prefix of the document as a character identifier of the legal procedure type, construct a mapping relationship between the character identifier and the legal document, and output at least one category label for any character identifier identified in the legal document; A second business rule is set up to determine the set of keywords for category labels of the legal document based on semantic assistance according to business needs. The weights of the keyword elements in the keyword set are configured to perform confidence correction calculations for category labels, so as to adjust the confidence of the keyword association with the category label. Set a third business rule to define the target status of the disposal stage, and configure the target status for each document category in the legal documents based on its disposal stage.

[0087] As an optional implementation, based on preset business rules, the legal documents are validated according to the category tag set and structured metadata to generate document classification results, including: Load preset business rules to read case number mapping, keyword weight and stage configuration, and extract case number and keyword set from input legal documents and structured metadata; The first verification is performed based on the first business rule, and a unique category label is determined for the case number extracted from the legal document according to the case number mapping read. The second verification is performed based on the second business rule. The confidence level of the keyword is determined by the confidence level correction calculation. The category label corresponding to any keyword in the keyword set is re-determined based on the unique confidence level. The third verification is performed based on the third business rule, and the consistency of the cross-document context content in the legal document is verified and corrected according to the target status and the order of the disposal stages. Based on the triple verification results, a corrected set of category labels is output, as well as a document classification result for classifying the legal document according to the set of category labels. The document classification result includes the intervention logs corresponding to the business rules and the review flags.

[0088] As an optional implementation, based on a preset debugging mechanism, the archiving process results are continuously optimized, including: For each input collection document, a corresponding debugging log is generated. The content of the debugging log is sorted in the processing order as follows: document meta information, OCR output original text, extracted case number, extracted associated roles, extracted keywords, classification model input, original category labels and confidence of corresponding category labels, business rule intervention records, and document classification results. The debugging log specifically records samples below the confidence threshold and samples with classification anomalies, and uses these specific records as negative samples to iteratively train and optimize the document classification model and business rules.

[0089] As an optional implementation, based on a preset monitoring mechanism, the archiving process results are continuously optimized, including: Evaluation metrics are set for the generated document classification results to output a classification performance report. The evaluation metrics include processing time, throughput, error rate, and proportion of manual review. Based on the classification performance report, the classification errors of the document classification results are identified, and the root causes of the classification errors are located. The classification errors and corresponding root causes are then fed back to the model and rules of the classification process for iteration.

[0090] Example 3 Please see Figure 3, Figure 3 This is another document classification system for multi-round collection and disposal processes disclosed in the embodiments of the present invention. Figure 3 The described document classification system for multi-round collection and disposal processes is applied in data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 3 As shown, this document classification system for multi-round collection and disposal processes may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the associated document classification method for a multi-round collection and disposal process described in Embodiment 1.

[0091] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps of the associated document classification method for a multi-round collection and disposal process described in Embodiment 1.

[0092] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the associated document classification method for a multi-round collection and disposal process described in Embodiment 1.

[0093] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0094] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0095] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0096] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0101] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0102] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0105] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0106] Finally, it should be noted that the document classification method and system for multi-round collection and disposal processes disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; 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 classifying related documents in a multi-round collection and disposal process, characterized in that, The method includes: Multiple collection documents are preprocessed and text-recognized to obtain legal documents and their corresponding structured metadata. Based on a preset classification model, the system outputs a set of category tags corresponding to the legal documents. Based on preset business rules, the legal documents are verified according to the category tag set and structured metadata to generate document classification results; The document classification results are archived, and the archiving results are continuously optimized based on preset debugging and / or monitoring mechanisms.

2. The document classification method for multi-round collection and disposal processes according to claim 1, characterized in that, Obtain and preprocess collection documents from multiple rounds, including the following steps: The document is identified to determine its type, and a first branch or a second branch is selected for processing based on the document type, wherein the first branch is a text extraction branch and the second branch is an image processing branch; For the collection document that selects the first branch, traverse all pages to extract a structured text sequence, which includes page number information and the text content corresponding to the page number; For the collection documents selected in the second branch, image objects are obtained through a conversion tool. Each image object is then subjected to grayscale conversion, binarization, noise reduction, tilt correction, and size normalization to obtain an image sequence.

3. The document classification method for multi-round collection and disposal processes according to claim 2, characterized in that, The process involves obtaining multiple rounds of collection and disposal documents for text recognition, including the following steps: The text line sequence of the image sequence is recognized based on a preset OCR model, and the text line sequence is concatenated into a single page of text according to a preset reading order to obtain the OCR-recognized text sequence. The text sequence and the structured text sequence are merged and unified into a text sequence to be processed; Noise characters are filtered out for each page of text in the text sequence to be processed, and the first and last lines of each page are extracted to determine the high-frequency headers and / or footers to construct a high-frequency line set. Any high-frequency line element contained in the high-frequency line set is deleted from each page of text. The text sequence to be processed is matched with business requirements using regular expressions to extract structured metadata of the legal fields corresponding to the business requirements; The processed text sequence is reorganized into paragraphs and merged for deduplication. Based on NLP enhancement, the text sequence is subjected to association processing stage and Chinese word segmentation and named entity recognition of corresponding stage documents, and legal documents with entity tags are output.

4. The method for classifying related documents for multi-round collection and disposal processes according to claim 1, characterized in that, Based on a preset classification model, the system outputs a set of category tags corresponding to the legal document, including: Determine the modal type of the legal document, and select the corresponding modal model to construct the input representation of the legal document based on the modal type; Based on a pre-trained language model, the input representation is mapped to a dense vector representation to generate an embedding vector; The classification categories of the predefined linear classification head are connected to the embedding vector as a fully connected layer, and the corresponding classification score vector is output from the input representation. Based on a preset multi-label activation function, the independent probability of the legal document corresponding to any of the classification categories is calculated according to the classification score vector; The manually labeled category label set with the same dimension as the classification category, and the independent probability are used to construct the target loss function for classification training to obtain the label elements composed of the classification category and the independent probability, generating a structured category label set. The label elements in the category label set are filtered by a confidence threshold and weighted fusion by a multimodal supplementation mechanism.

5. The method for classifying related documents for multi-round collection and disposal processes according to claim 1, characterized in that, The preset business rules include: Set a first business rule to extract the case number prefix of the document as a character identifier of the legal procedure type, construct a mapping relationship between the character identifier and the legal document, and output at least one category label for any character identifier identified in the legal document; A second business rule is set up to determine the set of keywords for category labels of the legal document based on semantic assistance according to business needs. The weights of the keyword elements in the keyword set are configured to perform confidence correction calculations for category labels, so as to adjust the confidence of the keyword association with the category label. Set a third business rule to define the target status of the disposal stage, and configure the target status for each document category in the legal documents based on its disposal stage.

6. The method for classifying related documents for multi-round collection and disposal processes according to claim 5, characterized in that, Based on preset business rules, the legal documents are validated according to the category tag set and structured metadata to generate document classification results, including: Load preset business rules to read case number mapping, keyword weight and stage configuration, and extract case number and keyword set from input legal documents and structured metadata; The first verification is performed based on the first business rule, and a unique category label is determined for the case number extracted from the legal document according to the case number mapping read. The second verification is performed based on the second business rule. The confidence level of the keyword is determined by the confidence level correction calculation. The category label corresponding to any keyword in the keyword set is re-determined based on the unique confidence level. The third verification is performed based on the third business rule, and the consistency of the cross-document context content in the legal document is verified and corrected according to the target status and the order of the disposal stages. Based on the triple verification results, a corrected set of category labels is output, as well as a document classification result for classifying the legal document according to the set of category labels. The document classification result includes the intervention logs corresponding to the business rules and the review flags.

7. The method for classifying related documents for multi-round collection and disposal processes according to claim 1, characterized in that, Based on a pre-defined debugging mechanism, the archiving process results are continuously optimized, including: For each input collection document, a corresponding debugging log is generated. The content of the debugging log is sorted in the processing order as follows: document meta information, OCR output original text, extracted case number, extracted associated roles, extracted keywords, classification model input, original category labels and confidence of corresponding category labels, business rule intervention records, and document classification results. The debugging log specifically records samples below the confidence threshold and samples with classification anomalies, and uses these specific records as negative samples to iteratively train and optimize the document classification model and business rules.

8. The method for classifying related documents for a multi-round collection and disposal process according to claim 7, characterized in that, Based on a pre-defined monitoring mechanism, the archiving process results are continuously optimized, including: Evaluation metrics are set for the generated document classification results to output a classification performance report. The evaluation metrics include processing time, throughput, error rate, and proportion of manual review. Based on the classification performance report, the classification errors of the document classification results are identified, and the root causes of the classification errors are located. The classification errors and corresponding root causes are then fed back to the model and rules of the classification process for iteration.

9. A document classification system for multi-round collection and disposal processes, characterized in that, The system includes: The acquisition module is used to acquire multiple rounds of collection documents for preprocessing and text recognition to obtain legal documents and the corresponding structured metadata. The classification module is used to output a set of category tags corresponding to the legal document based on a preset classification model. The verification module is used to verify the legal document based on the preset business rules, the category tag set, and the structured metadata, and generate a document classification result. The output module is used to archive the document classification results and continuously optimize the archive processing results based on a preset debugging and / or monitoring mechanism.

10. A document classification system for multi-round collection and disposal processes, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the associated document classification method for multi-round collection and disposal processes as described in any one of claims 1-8.

Images