Method and system for intelligent extraction and automatic processing of legal document elements

Abstract

The application discloses a kind of legal document element intelligent extraction and automatic processing method and system, the method includes: obtaining at least one unstructured legal document of the case to be processed, the unstructured legal document comes from multiple heterogeneous business systems;The unstructured legal document is analyzed in multiple modes, and the structured data containing text content and its layout information are obtained;The structured data is input into pre-trained language model, and key elements are extracted;Based on the key elements extracted, semantic retrieval is carried out in the pre-constructed processing strategy instance library, and at least one similar historical case is matched to obtain, and the processing strategy suggestion of current case is generated according to the similar historical case.The application solves the technical problems of low efficiency, easy to make mistakes and strategy response lag in traditional manual processing mode through the automation of document processing and the intelligentization of strategy generation.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for intelligent extraction and automated processing of legal document elements. Background Technology

[0002] With the deepening of digital transformation, a large amount of unstructured document data has accumulated in various business systems. For example, in the fields of judiciary, finance, and government affairs, documents such as indictments, judgments, rulings, and agreements generated during case handling are mostly stored in the form of PDF scans and images. These documents contain key elements such as the identification of relevant parties, descriptions of matters to be handled, related resource information, and time-series constraints. However, due to the lack of a unified digital representation, they are difficult for computer systems to directly recognize and process.

[0003] In existing technologies, the processing of unstructured documents mainly relies on manual reading and input, or on automated solutions using general optical character recognition (OCR) combined with keyword matching. However, the former is inefficient, taking an average of several hours to process a complex document, and manual input is prone to inconsistency and errors; while the latter can recognize text, it cannot understand the document's layout and semantic context, resulting in a high error rate in judging complex expressions such as "partially fulfilled obligations" and "no executable assets found," and it is difficult to correlate relevant information about the same task across documents.

[0004] The aforementioned deficiencies lead to the following technical problems in the process of processing multi-source heterogeneous data: First, the automated parsing capability of unstructured data is insufficient, making it impossible to convert key information in documents into structured data for subsequent programs to use; Secondly, the lack of integrated understanding of document layout structure and text content leads to low accuracy in element extraction. Third, the generation of task processing strategies relies on human experience and cannot be quantitatively matched based on historical success cases, resulting in delayed strategy response and uneven resource allocation. Summary of the Invention

[0005] This invention provides a method and system for intelligent extraction and automated processing of legal document elements, which solves the problems of low efficiency, error-proneness and delayed strategy response caused by the reliance on manual extraction in existing legal document processing.

[0006] The technical problem to be solved by this invention is to provide a method and system for intelligent extraction and automated processing of legal document elements, which can realize the automated parsing and accurate extraction of key elements of unstructured legal documents, reduce document processing time from days to hours, and intelligently generate collection strategies based on the extracted elements, thereby improving the timeliness of strategy response and the conversion rate of payment collection.

[0007] To address the aforementioned technical problems, the first aspect of this invention discloses a method for intelligent extraction and automated processing of legal document elements, the method comprising: Obtain at least one unstructured legal document for the pending case, wherein the unstructured legal document comes from multiple heterogeneous business systems; The unstructured legal document is subjected to multimodal parsing to obtain structured data containing text content and its layout information; The structured data is input into a pre-trained language model to extract key elements, which include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information. Based on the extracted key elements, semantic retrieval is performed in a pre-built processing strategy instance library to match at least one similar historical case, and processing strategy suggestions for the current case are generated based on the similar historical case.

[0008] As an optional implementation, in the first aspect of the present invention, the step of performing multimodal parsing on the unstructured legal document to obtain structured data containing text content and its layout information includes: The layout of the unstructured legal documents is analyzed to identify and retain the text blocks, table areas, image positions and their spatial coordinates. Optical character recognition is performed on the image areas in the unstructured legal documents to extract key text information, and then aligned with the layout information after page analysis. The aligned key text information and layout information are input into a multimodal large model. The semantic relationship between text content and spatial information is understood through a cross-modal attention mechanism, and the structured data is output.

[0009] As an optional implementation, in the first aspect of the present invention, the step of performing optical character recognition on the image region of the unstructured legal document to extract key text information includes: Optical character recognition is performed on the image region to generate text content and its corresponding recognition confidence level; When the recognition confidence level is lower than a preset threshold, the corresponding text content is marked as an item to be reviewed, triggering manual review or enabling the backup recognition engine; The reviewed text content and text content with a confidence level that reaches or exceeds a preset threshold are used as the key text information for alignment with the layout information after page analysis.

[0010] As an optional implementation, in the first aspect of the invention, the step of inputting the structured data into a pre-trained language model to extract key elements includes: Based on a pre-trained language model, supervised fine-tuning is performed using labeled legal document samples. Legal domain knowledge is injected by updating feature parameters through low-rank adaptation technology. At the same time, positive and negative sample pairs are introduced for preference alignment optimization, resulting in a domain-adapted language model suitable for extracting elements from legal documents. Deploy the domain-adaptive language model on a local server or in a domestic IT innovation environment to ensure that sensitive data does not leave the domain; The structured data is input into the domain-adaptive language model, and the model is guided to output standardized key elements through a preset prompt template.

[0011] As an optional implementation, in the first aspect of the invention, the step of performing semantic retrieval in a pre-built processing strategy instance library based on the extracted key elements, matching at least one similar historical case, and generating a processing strategy suggestion for the current case based on the similar historical case includes: A pre-built instance library of processing strategies is constructed. Each instance contains at least case characteristics, applicable legal basis, operation steps and historical results information. Each instance is converted into a vector through an embedding model and then stored in a vector database. The key elements of the current case are converted into feature vectors, and semantic similarity is retrieved from the vector database to obtain a preset number of historical cases that are most similar to the current case. The key elements of the current case are combined with the preset number of historical cases retrieved to form a prompt, which is then input into a lightweight large language model to generate processing strategy suggestions.

[0012] As an optional implementation, in the first aspect of the invention, the processing strategy suggestion includes strategy suggestions for collection and disposal scenarios, and the method further includes: calculating the collection priority score P of the current case based on extracted key elements, wherein the collection priority score P is calculated according to the formula... ; Determined; where W is the debtor's subjective repayment score, which is quantified based on the debtor's commitment behavior and historical performance behavior in historical communication records. The commitment behavior is identified from the transcript of the call recording based on a preset commitment statement template or small model. The historical performance behavior is determined based on at least one of the debtor's actual repayment frequency and repayment timeliness rate in past collection cycles. C represents the debtor's asset enforceability score, which is derived from a quantitative assessment based on extracted property clues. These property clues include at least one or more of the following: bank account balance, real estate, vehicles, and equity. β is a preset weighting coefficient, with a value range of [0,1].

[0013] As an optional implementation, in the first aspect of the present invention, the execution result feedback of the processing personnel on the processing strategy suggestion is obtained. When the execution result feedback is valid, the current case and its corresponding processing strategy suggestion are taken as new instances, converted into vectors and injected into the vector database in real time. Deduplication verification is performed based on semantic similarity, and the data is automatically updated to the processing strategy instance library.

[0014] As an optional implementation, in the first aspect of the present invention, the key element further includes a legal complexity dimension feature, which is a binary variable used to characterize the complexity of the case in legal procedures. The legal complexity dimension feature takes the first value when at least one of the following conditions is met: the number of defendants is greater than 1, an objection to jurisdiction has been raised, or there is a multi-layered guarantee. Otherwise, it takes the second value.

[0015] A second aspect of this invention discloses an intelligent extraction and automated processing system for legal document elements, the system comprising: The acquisition module is used to acquire at least one unstructured legal document for the case to be processed, wherein the unstructured legal document comes from multiple heterogeneous business systems; The parsing module is used to perform multimodal parsing on the unstructured legal documents to obtain structured data containing text content and its layout information; The extraction module is used to input the structured data into a pre-trained language model and extract key elements. The key elements include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information. The generation module is used to perform semantic retrieval in a pre-built processing strategy instance library based on the extracted key elements, match at least one similar historical case, and generate a processing strategy suggestion for the current case based on the similar historical case.

[0016] The third aspect of the present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the intelligent extraction and automated processing method for legal document elements as disclosed in the first aspect of the present invention.

[0017] 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 intelligent extraction and automated processing method for legal document elements disclosed in the first aspect of the present invention.

[0018] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention acquires unstructured legal documents from multiple heterogeneous business systems and performs multimodal parsing, integrating layout information with text content for a comprehensive understanding. This overcomes the limitations of traditional OCR, which only recognizes text and cannot understand table structures or paragraph relationships, achieving accurate restoration of the original document information structure and providing high-quality input data for subsequent processing. By inputting structured data into a pre-trained language model to extract key elements, the average extraction accuracy of key fields is increased to over 0.94, and the average time from document entry to usable structured data is reduced to hours, completely avoiding the risk of errors caused by manual data entry. By constructing a processing strategy instance library and performing semantic retrieval and matching of historical cases based on extracted key elements, the system can automatically generate processing strategy suggestions adapted to the current task. A feedback update mechanism dynamically optimizes the instance library, forming a closed-loop processing flow of parsing, extraction, retrieval, and feedback. Ultimately, this invention significantly improves the processing efficiency and element extraction accuracy of multi-source heterogeneous documents, significantly reduces the frequency of manual intervention and strategy response latency, and effectively improves the system's concurrent processing capacity and resource utilization balance under the same hardware resource configuration. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is a flowchart illustrating a method for intelligent extraction and automated processing of legal document elements disclosed in the first embodiment of the present invention; Figure 2 This is a detailed flowchart of S12 in the first embodiment of the present invention; Figure 3 This is a detailed flowchart of S122 in the first embodiment of the present invention; Figure 4 This is a detailed flowchart of S13 in the first embodiment of the present invention; Figure 5 This is a detailed flowchart of S14 in the first embodiment of the present invention; Figure 6 This is a structural block diagram of a legal document element intelligent extraction and automated processing system disclosed in an embodiment of the present invention.

[0021] Figure 7 This is a computer equipment hardware structure diagram of a legal document element intelligent extraction and automated processing system disclosed in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0024] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0025] This invention discloses a method and system for intelligent extraction and automated processing of legal document elements. By acquiring unstructured legal documents from multiple heterogeneous business systems and performing multimodal parsing, the invention integrates layout information with text content for understanding. This overcomes the limitations of traditional OCR, which only recognizes text and cannot understand table structures or paragraph relationships, achieving accurate restoration of the original document information structure and providing high-quality input data for subsequent processing. By inputting structured data into a pre-trained language model to extract key elements, the average extraction accuracy of key fields is increased to over 0.94, and the average time from document entry to usable structured data is reduced to hours, completely avoiding the risk of errors caused by manual data entry. By constructing a processing strategy instance library and performing semantic retrieval and matching of historical cases based on extracted key elements, the system can automatically generate processing strategy suggestions adapted to the current task. A feedback update mechanism dynamically optimizes the instance library, forming a closed-loop processing flow of parsing, extraction, retrieval, and feedback. Ultimately, this invention significantly improves the processing efficiency and element extraction accuracy of multi-source heterogeneous documents, substantially reduces the frequency of manual intervention and strategy response latency, and effectively enhances the system's concurrent processing capability and resource utilization balance under the same hardware resource configuration. These will be described in detail below.

[0026] Example 1 Please refer to Figures 1 to 5 As shown, this embodiment of the invention discloses a method for intelligent extraction and automated processing of legal document elements, including S11-S15, wherein: S11, Obtain at least one unstructured legal document for the case to be processed, wherein the unstructured legal document comes from multiple heterogeneous business systems.

[0027] This embodiment uses a disposal management system, a litigation management platform, and an enforcement investigation and control system as examples for illustration. Specifically, the multiple heterogeneous business systems may include at least one of a first business system (such as a disposal management system), a second business system (such as a litigation management platform), and a third business system (such as an enforcement investigation and control system).

[0028] Based on the above examples, the implementing entity in this embodiment can be an enterprise-level data platform deployed in the financial or judicial technology fields, or it can be an intelligent parsing engine integrated into a post-loan management system. This step aims to solve the problem of case data input, namely, pulling raw unstructured legal documents from three independent and heterogeneous business systems.

[0029] A debt collection system typically refers to a post-loan collection / debt management platform that stores call recordings, repayment commitment letters, etc.; a litigation platform refers to a court's electronic litigation platform or a law firm's case management system that stores legal documents such as complaints, judgments, and mediation agreements; an enforcement system refers to a court's enforcement case handling platform or a judicial investigation and control system that stores enforcement rulings, property investigation and control receipts, and other enforcement-related documents. Unstructured legal documents refer to text and image files without a predefined data model or not organized in rows and columns, such as scanned PDFs of complaints, images of judgments, and photos of handwritten signed mediation agreements. Their text is stored in pixel form and cannot be directly retrieved and processed by a database.

[0030] S12, perform multimodal parsing on the unstructured legal document to obtain structured data containing text content and its layout information.

[0031] The aforementioned multimodal parsing refers to a multi-dimensional understanding process that integrates text, images, and / or layout information, unlike traditional OCR solutions that only perform text recognition. Structured data refers to standardized data that organizes information in a document according to a preset format, such as containing text content, the coordinate position of the text on the page, and the table area it belongs to, and serves as direct input for subsequent element extraction.

[0032] Please refer to Figure 2 As shown, this step S12 includes S121-S123, wherein: S121. Perform layout analysis on the unstructured legal document, identify and retain the text blocks, table areas, image positions and their spatial coordinate information in the document.

[0033] Layout analysis refers to the process of region segmentation and type recognition of a document image, aiming to understand the physical structure of the document. Exemplarily, in this step, parsing tools such as PyMuPDF or pdfplumber can be used to perform lossless parsing on the input PDF document, retaining the text blocks (such as paragraphs, headings) in the original page, table areas (such as rows, columns, cells), image positions (such as seals, handwritten signatures) and their spatial coordinate systems in the page. This step avoids the problems of table misalignment and paragraph mixing caused by directly converting traditional OCR into pure text, and is especially applicable to documents with complex layouts such as court documents and bank statements. The output is a set of text elements with coordinates and a table structure tree.

[0034] S122. Perform optical character recognition on the image areas in the unstructured legal document, extract key text information, and align it with the layout information after layout analysis.

[0035] Optical Character Recognition (OCR) is a technology that converts text in an image into machine-editable text. For scanned documents, low-resolution images or pages with handwritten annotations or red seal occlusions, in this embodiment, the PaddleOCR or MiniOCR engine can be called for optical character recognition. The system automatically detects the image quality and performs preprocessing on situations such as blurring, tilting or uneven illumination, such as adaptive binarization, perspective correction, etc. The recognition result is aligned with the layout information output in step S121 to ensure that the recognized text can be accurately positioned back to its original position in the document.

[0036] As a preferred solution rather than a limitation, please refer to Figure 3 As shown, this step S122 includes S1221 - S1223, where: S1221. Perform optical character recognition on the image area to generate the text content and its corresponding recognition confidence.

[0037] During the OCR recognition process, the recognition engine not only outputs the text content, but also generates the recognition confidence for each character or each text area. Exemplarily, the confidence is a value between 0 and 1, reflecting the reliability of the recognition result. The higher the value, the more reliable the recognition. For example, for the clearly printed characters "Zhang San", the recognition confidence may reach 0.98; while for the characters "Li Si" occluded by a seal, the confidence may only be 0.65. This confidence provides a quantitative basis for subsequent quality control.

[0038] S1222, when the recognition confidence level is lower than a preset threshold, the corresponding text content is marked as an item to be reviewed, triggering manual review or enabling the backup recognition engine.

[0039] The preset threshold can be configured by business personnel according to the actual scenario. For example, a typical value can be set to 0.8 or 0.9. When the recognition confidence of a text region is lower than this threshold, the system determines that the recognition result of that region is unreliable and marks it as an item to be reviewed. Items to be reviewed can be pushed to the manual review interface for manual correction by comparing with the original image; alternatively, it can trigger the activation of a backup recognition engine, such as switching to an OCR engine optimized for handwriting for re-recognition, to improve recognition accuracy.

[0040] S1223, The reviewed text content and the text content with a confidence level that reaches or exceeds a preset threshold are used as the key text information for alignment with the layout information after page analysis.

[0041] This step is used to filter out reliable key text information, including text that meets the confidence level but has not been reviewed, as well as text that has been manually reviewed and corrected. Preferably, the filtered text information can also be aligned with the layout information in step S121 to form a text-position-confidence triplet, ensuring that each text block is accurately associated with its position on the page.

[0042] The above steps S1221-S1223 achieve quantifiable control of OCR recognition quality through a confidence mechanism. Low-confidence regions trigger manual review or a backup engine to avoid low-quality recognition results from polluting subsequent processing. At the same time, the reviewed text and the high-confidence text are used together as key information to maximize the efficiency of automated processing while ensuring recognition accuracy, thus achieving the optimal balance of human-machine collaboration.

[0043] S123, the aligned key text information and layout information are input into the multimodal large model, and the semantic relationship between text content and spatial information is understood through the cross-modal attention mechanism, and the structured data is output.

[0044] Multimodal large models refer to deep learning models capable of simultaneously processing information from multiple modalities such as text, images, and layouts. For example, multimodal large language models such as Qwen-VL or Qwen2.5-VL can be used. Cross-modal attention mechanisms are internal to the model, allowing textual, positional, and image information to be "aligned" and understood. In this embodiment, the model can use this mechanism to understand spatial semantic relationships such as "a certain amount is located in the second column of the table below the 'Execution Target' heading," thereby accurately determining the meaning of the field. For example, in a judgment, if "RMB One Hundred Thousand Yuan" appears after "The defendant shall pay within ten days from the date this judgment becomes effective," it is categorized as "Amount Payable"; if it appears in the "Fulfilled" paragraph, it is marked as "Amount Received." This step outputs structured data containing text content and its accurate semantic labels.

[0045] The above steps S121-S123, through three stages of processing—structure awareness, content recognition, and semantic alignment—solve the core pain points of judicial and financial documents being "visible but inaccurate to read, and accurately readable but unclear to understand." This embodiment uses Effective Collection Information Coverage (ECIC) as the evaluation metric. ; Where, N valid N represents the number of key fields automatically extracted and correctly verified by the system. total E represents the total number of key fields to be extracted from the document. manual For fields requiring manual intervention for correction (in hours), T process The total time (in hours) for processing a single case throughout the entire process.

[0046] This metric comprehensively measures a system's ability to parse unstructured documents from two technical dimensions: the completeness of information extraction and the degree of automation in the processing workflow. The first item... The second item reflects the system's ability to recall key fields under complex layouts. The degree of human intervention in the system's processing can be characterized, and the combination of these two aspects can quantify the system's technical effectiveness in maintaining information integrity while reducing human intervention.

[0047] Actual measurements show that the ECIC of this embodiment reaches 0.92, which is 41.5% higher than the traditional OCR + keyword matching method (ECIC≈0.65).

[0048] S13, input the structured data into the pre-trained language model and extract key elements. The key elements include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information.

[0049] As an example, the task-related party identifier may include debtor identity information, the description of the matter to be disposed of may include debt amount information, the associated resource information may include property clue information, the time constraint information may include the deadline for fulfilling obligations information, and the processing status information may include execution status information.

[0050] This step aims to accurately extract key information needed for business operations from structured data. Pre-trained language models refer to deep learning models pre-trained on large-scale general corpora, including but not limited to the Qwen series and BERT, which possess powerful semantic understanding capabilities. Through domain-specific fine-tuning, the model is able to understand legal terminology and document context, accurately identifying and extracting various key elements.

[0051] Please refer to this as a specific solution rather than a limitation. Figure 4 As shown, this step S13 includes S131-S133, wherein: S131 is based on a pre-trained language model. It uses labeled legal document samples for supervised fine-tuning and injects legal domain knowledge by updating feature parameters through low-rank adaptation technology. At the same time, it introduces positive and negative sample pairs for preference alignment optimization, resulting in a domain-adapted language model suitable for extracting legal document elements.

[0052] Supervised fine-tuning (SFT) refers to further training a pre-trained model using labeled task data to adapt it to a specific task. In this embodiment, labeled legal document samples, such as judgment-structured field pairs, are used to perform supervised fine-tuning on the pre-trained language model. Low-rank adaptation (LoRA) is an efficient fine-tuning technique that adds a low-rank matrix next to the original parameter matrix of the model for training. Only a small number of parameters need to be updated to inject domain knowledge, avoiding the computational cost of retraining the entire model. Preference alignment optimization involves introducing manually generated or rule-based positive and negative sample pairs, such as the positive sample "correctly extracted 'the person subject to enforcement, Zhang San'" and the negative sample "incorrectly extracted as 'the applicant, Zhang San'". Algorithms such as Direct Preference Optimization (DPO) are used to optimize the consistency between the model output and the business logic, reducing semantic ambiguity. Finally, a domain-adapted language model suitable for extracting elements from legal documents is obtained.

[0053] S132, Deploy the domain-adaptive language model on a local server or in a domestic IT innovation environment to ensure that sensitive data does not leave the domain.

[0054] "Information technology innovation" refers to the application of information technology innovation, requiring the use of domestically produced hardware and software environments, such as the Kylin operating system, Kunpeng chips, and DM database. This step deploys the fine-tuned model on the customer's local server or in the information technology innovation environment within the customer's intranet. All computation and storage are completed within the customer's intranet, ensuring that original documents and processing results do not leave the local area, eliminating the risk of data leakage, and meeting the requirements of the Personal Information Protection Law and financial regulations regarding the non-disclosure of sensitive data within the domain.

[0055] This embodiment implements strict data security control throughout the entire process of voice and digital data processing: all voice and digital text involving personal identity information and financial account data undergoes mandatory desensitization before entering the processing steps, and the original plaintext data is never stored on disk, in the database, or cached; at the same time, the system implements a "three-person separation" permission control mechanism, setting up three roles: system administrator, security auditor, and business operator, each with the minimum necessary permissions. Any access to the original data requires approval and is fully traceable, thus avoiding the risk of "illegally processing personal information" as stipulated in Article 51 of the Personal Information Protection Law from the architectural level.

[0056] As a preferred approach, this embodiment uses the Security Compliance Efficiency Ratio (SCER) as an evaluation metric at the deployment and security levels: ; Among them, U baseline For system availability (measured by average number of cases processed per day) without de-identification and fine-grained permissions enabled, U secure To assess actual availability after enabling end-to-end data masking, RBAC control, and domestic IT innovation deployment. Ideally, security measures should not significantly reduce efficiency. Real-world testing shows that in a bank's production environment, U... baseline =520 cases / day, U secure =490 cases / day, SCER≈0.94, which means that while meeting the compliance requirements of the Personal Information Protection Law and the Guidelines for Financial Data Security Classification, only 6% of the throughput is sacrificed, which is far better than the traditional extensive model of "security equals reduced efficiency".

[0057] S133, the structured data is input into the domain-adaptive language model, and the model is guided to output standardized key elements through a preset prompt template.

[0058] A prompt template is a pre-defined question format used to guide the model to output standardized results. For example, after inputting a document fragment, a prompt could be constructed: "Please extract from the following text: name of the person subject to enforcement, amount of enforcement, and performance period. If none are found, return 'Not mentioned'." Based on fine-tuned knowledge and contextualized generalization, the model outputs standardized key elements, such as {"Name of the person subject to enforcement": "Zhang San", "Amount of enforcement": "100,000 yuan", "Performance period": "Within ten days after the judgment takes effect"}. This method avoids the dependence of traditional rule engines on fixed templates and significantly improves the generalization ability to unseen document formats.

[0059] The above steps S131-S133 enhance semantic understanding through domain fine-tuning and prompt engineering-driven structured output, achieving accurate, fast, and stable extraction of legal document elements. This embodiment uses Collection Decision Readiness (CDR) as the evaluation metric. ; Among them, A field T represents the average extraction accuracy (F1-score) of the key fields. extract T represents the average time (in hours) from document entry to the availability of structured data. threshold The maximum allowed processing delay threshold (e.g., 24 hours) is α, where α∈[0,1] is the accuracy weighting coefficient.

[0060] This indicator integrates the accuracy parameter of feature extraction and the timeliness parameter of processing, evaluating the usability of the system's output data from two technical dimensions: data quality and response speed. Among them, the accuracy item... The timeliness item is used to measure the accuracy of the language model in identifying key elements in legal documents. The CDR value, obtained by weighting and fusing the assessment data from input to usable processing latency, quantifies the readiness of the system output data for subsequent program calls.

[0061] In this embodiment, A field =0.94, T extract =1.2 hours, let T threshold =24 hours, α=0.7, then CDR≈0.96, which is much higher than the traditional rule matching method (CDR≈0.68).

[0062] S14. Based on the extracted key elements, perform semantic retrieval in the pre-built processing strategy instance library, match at least one similar historical case, and generate a processing strategy suggestion for the current case based on the similar historical case.

[0063] As an example, the processing strategy instance library may be a collection and disposal instance library, and the processing strategy suggestion may be a collection and disposal strategy suggestion.

[0064] This step aims to transform the extracted case elements into actionable disposal strategies. The instance library is a pre-built case database covering typical debt collection scenarios. Each instance includes case characteristics, applicable legal basis, operational steps, and historical results information. Semantic retrieval refers to a retrieval method based on semantic similarity rather than simple keyword matching, capable of understanding the similarity between expressions such as "has a mortgaged property" and "owns a commercial property." For a specific solution, please refer to... Figure 5 As shown, this step S14 includes S141-S143, wherein: S141, a processing strategy instance library is pre-built. Each instance contains at least case characteristics, applicable legal basis, operation steps and historical results information. Each instance is converted into a vector through an embedding model and then stored in a vector database.

[0065] Case characteristics include structured information such as overdue days, principal amount, property type, and geographical location; applicable legal basis refers to the legal provisions upon which the case is based, such as Article 242 of the Civil Procedure Law regarding seizure; operational steps refer to the specific handling process, such as "applying to seize the salary account - initiating property appraisal and auction - participating in distribution"; historical results include quantitative indicators such as recovery rate and execution cycle. Embedding models such as bge-large-zh convert the textual descriptions of instances into vectors in a high-dimensional space, making semantically similar instances close in distance within the vector space. Vector databases such as Milvus or Chroma are used to store these vectors and support efficient similarity retrieval.

[0066] S142, convert the key elements of the current case into feature vectors, perform semantic similarity retrieval in the vector database, and obtain a preset number of historical cases that are most similar to the current case.

[0067] The key elements of the current case are also converted into feature vectors through an embedding model. Approximate nearest neighbor retrieval is then performed in the vector database to obtain a preset number of historical cases most similar to the current case, typically 3 or 5. Semantic similarity retrieval, based on cosine similarity calculation, can match the current case with the most similar historical successful cases in terms of case characteristics, property type, and ability to perform.

[0068] S143, combine the key elements of the current case with the preset number of historical cases retrieved to form a prompt, input it into a lightweight large language model, and generate processing strategy suggestions.

[0069] Lightweight large language models refer to language models with small parameter sizes and fast inference speeds, such as Qwen1.5-4B-Chat, which can respond in real time on a single GPU or high-performance CPU, meeting the latency requirements of outbound call systems. In this embodiment, the key elements of the current case are concatenated with a preset number of historical cases to form a prompt, such as: "You are a professional case handling consultant. Current client: 45 days overdue, principal of 18,000 yuan, property clues include one property. Refer to the following successful cases: Case 1: 50 days overdue, property owned, recovered through property seizure and auction; Case 2: 40 days overdue, salary account owned, repayment in installments through freezing salary account. Please generate a compliant and specific negotiation suggestion." The model outputs natural language strategy suggestions, such as "It is recommended to apply for the seizure of the property under their name, and at the same time, a repayment plan can be negotiated." The above steps S141-S143 provide experience references through an instance library, achieve accurate matching through semantic retrieval, and generate lightweight large models, forming an intelligent strategy generation mechanism that "learns from history." This embodiment uses the Effective Strategy Adoption Rate (ESAR) as the evaluation metric. ; Where, N recommended N represents the total number of handling strategies recommended by the system. executed_and_recovered The number of strategies that are adopted, implemented, and generate actual payments within 60 days.

[0070] This metric quantifies the effectiveness of the system's strategy recommendation module from the perspective of the correlation between strategy generation and business outcomes. Specifically, this metric tracks whether the strategy suggestions generated by the system are adopted and whether the adoption produces the expected results, forming a closed-loop verification of the entire process of element extraction, semantic retrieval, and strategy generation. A high ESAR value indicates that the system can accurately understand case characteristics and retrieve suitable handling solutions from historical instances, reflecting the matching accuracy of the semantic retrieval module and the content generation quality of the lightweight, large-model approach.

[0071] In this embodiment, the ESAR reaches 0.78, which is 73% higher than that of traditional human experience-based decision-making (ESAR≈0.45).

[0072] S15, output the processing strategy suggestion.

[0073] This step outputs the generated strategy recommendations in structured or natural language format to the user's interface or pushes them to an automated outbound calling system. The output can include supporting evidence, such as "This recommendation is based on Case X, which has similar characteristics to the case and a recovery rate of 85%", enhancing the interpretability and credibility of the strategy.

[0074] It should be noted that the embodiments of the present invention can also achieve end-to-end efficiency improvement at the system integration level. The End-to-End Task Efficiency Ratio (ETER) is used as the evaluation metric: ; Among them, T manual T represents the average time (in hours) to complete the entire process of similar cases under the traditional manual model. automated The actual calculation time (in hours) for this system to automatically process is T. wait ETER represents the waiting time (in hours) caused by system queuing, network latency, etc. A larger ETER indicates that the system significantly compresses the business cycle while maintaining reliability. Real-world testing shows that under a load of processing an average of 500 documents per day, T... manual =4.5 hours / case, T automated +T wait =0.6 hours / case, ETER≈7.5, which means an efficiency improvement of 650%.

[0075] It should be noted that the processing strategy suggestions include strategy suggestions for collection and disposal scenarios. In this embodiment of the invention, after S14 or before S15, the following may be included: calculating the collection priority score P of the current case based on the extracted key elements. As a preferred embodiment, the collection priority score P is calculated according to the formula... ; Confirmed. Here, W represents the debtor's subjective repayment score, quantified based on the debtor's commitments and performance in historical communication records.

[0076] The commitment behavior is identified from the transcript of the call recording based on a preset commitment statement template or a small model. The commitment statement template includes fixed phrases that indicate the willingness to repay, such as "I will pay next week", "I am raising money", and "I guarantee to repay on time". The small model is a finely tuned binary classification model used to determine whether the statement belongs to the commitment expression. The historical performance behavior is determined based on at least one of the following in the past collection cycle: the number of actual repayments, the timely repayment rate, and the promised performance rate. The timely repayment rate is the proportion of the amount repaid on time to the amount due, and the promised performance rate is the proportion of the actual repayments made in the historical promised behavior. Furthermore, the score for W is calculated using a weighted fusion method, for example, W = α·C commit +(1-α)·H perform C commit The commitment behavior is scored based on the frequency and confidence of the commitment statements; H performThe score is based on historical repayment data; α is a preset weighting coefficient.

[0077] C represents the debtor's asset enforceability score, which is derived from a quantitative assessment based on extracted property clues. These property clues include at least one or more of the following: bank account balance, real estate, vehicles, and equity. β is a preset weighting coefficient, with a value range of [0,1].

[0078] The higher the score of P, the more priority the system will allocate collection resources or initiate enforcement procedures to achieve optimal resource allocation.

[0079] As another preferred embodiment, this invention may further include a feedback update mechanism: obtaining feedback from processing personnel on the execution results of the processing strategy suggestions; when the execution result feedback is valid, the current case and its corresponding processing strategy suggestions are treated as new instances, converted into vectors, and injected into the vector database in real time. Deduplication is then performed based on semantic similarity, and the results are automatically updated to the processing strategy instance library. For example, if a processing personnel marks "valid" through a web interface, the system automatically writes the complete context of this interaction into the strategy evaluation log library, which, after deduplication, is converted into a new RAG retrieval instance. This mechanism enables the system to continuously learn, and in actual testing, it increased the strategy adoption rate by 22% and the conversion rate of high-intent customers by 15% within 3 months.

[0080] As another preferred embodiment of the invention, the key elements may further include a legal complexity dimension feature. This legal complexity dimension feature is a binary variable used to characterize the complexity of the case in legal proceedings. When at least one of the following conditions is met—the number of defendants is greater than one, a jurisdictional objection has been raised, or multiple layers of security exist—the legal complexity dimension feature takes a first value (e.g., 1); otherwise, it takes a second value (e.g., 0). This feature can be used to determine whether to match a professional lawyer or mediation institution with experience in handling complex cases during subsequent strategy generation.

[0081] It should be noted that the above description using a financial judicial scenario as an example is only for the purpose of understanding the technical solution of this invention, and is not intended to limit the application field of this invention. The technical solution of this invention is based on the structured parsing and element association of multi-source heterogeneous documents. Its core lies in solving the general technical problem of automated parsing of unstructured data and multi-source information association and fusion. Therefore, it can be widely applied to various business areas that require automated document processing, including but not limited to: 1. Government Services On the "One-Stop Government Service Platform," applications submitted by businesses or individuals involve various documents such as business licenses, qualification certificates, and application forms. The technical solution of this invention can automatically parse these documents, extract key elements such as application details, applicant information, and material lists, and generate approval suggestions based on historical successful cases, thus achieving intelligent assisted approval.

[0082] 2. Healthcare sector In regional medical collaboration platforms, patient medical information is scattered across electronic medical records, imaging reports, and laboratory test results from different medical institutions. The technical solution of this invention can be used to integrate multi-source medical documents, extract key information such as patient complaints, diagnostic conclusions, examination indicators, and medication records, and recommend treatment pathways based on similar cases to assist doctors in decision-making.

[0083] 3. Scientific research literature management field In academic research platforms, researchers need to process a large number of papers, patents, experimental reports, and other documents. The technical solution of this invention can be used to automatically parse documents, extract elements such as authors, institutions, funding projects, and experimental data, and recommend relevant literature and research methods based on research topics, thereby improving research efficiency.

[0084] 4. Enterprise Contract Management In legal management systems, enterprises need to process a large number of contract documents, such as purchase contracts, sales agreements, and confidentiality agreements. The technical solution of this invention can automatically extract key clauses from contracts, such as the subject matter, amount, performance period, and liability for breach of contract, and recommend compliant clauses and risk warnings based on historical contracts, thereby automating contract management.

[0085] The above application scenarios are merely to further illustrate the versatility of the present invention, and not to limit the scope of protection. Regardless of the specific business field in which it is applied, as long as it involves the extraction and automated processing of elements from multi-source heterogeneous documents, it falls within the protection scope of the present invention.

[0086] Example 2 Please see Figure 6 The present invention also provides a legal document element intelligent extraction and automated processing system 100, the system comprising: The acquisition module 110 is used to acquire at least one unstructured legal document of the case to be processed, wherein the unstructured legal document comes from multiple heterogeneous business systems; The parsing module 120 is used to perform multimodal parsing on the unstructured legal document to obtain structured data containing text content and its layout information; Extraction module 130 is used to input the structured data into a pre-trained language model and extract key elements, wherein the key elements include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information. The generation module 140 is used to perform semantic retrieval in a pre-built processing strategy instance library based on the extracted key elements, match at least one similar historical case, and generate a processing strategy suggestion for the current case based on the similar historical case.

[0087] The modules in this embodiment are the same as the corresponding steps in the first embodiment described above, and will not be repeated here.

[0088] All modules / units in this embodiment are the same as the corresponding steps in the above method embodiments, and their logical relationships and working principles are also the same, so they will not be repeated here. Those skilled in the art can learn the corresponding virtual modules or units from the above method embodiments to make them correspond to the steps of the above method embodiments. Virtual modules / units not disclosed in this embodiment should also be regarded as the part of the content disclosed in this invention.

[0089] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0090] This invention also provides a computer storage medium storing a computer program that, when executed by a processor, implements the intelligent extraction and automated processing method for legal document elements as described in the above embodiments.

[0091] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above-described intelligent extraction and automated processing methods for legal document elements. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0092] Alternatively, if the integrated units of the present invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, terminal, or network device, etc.) to execute all or part of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, RAM, ROM, magnetic disks, or optical disks.

[0093] Corresponding to the computer storage medium described above, one embodiment also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent extraction and automated processing method for legal document elements as described in the above embodiments.

[0094] This computer device can be a terminal, and its internal structure diagram can be as follows: Figure 7As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for intelligent extraction and automated processing of legal document elements. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for intelligent extraction and automated processing of legal document elements, characterized in that, include: Obtain at least one unstructured legal document for the pending case, wherein the unstructured legal document comes from multiple heterogeneous business systems; The unstructured legal document is subjected to multimodal parsing to obtain structured data containing text content and its layout information; The structured data is input into a pre-trained language model to extract key elements, which include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information. Based on the extracted key elements, semantic retrieval is performed in a pre-built processing strategy instance library to match at least one similar historical case, and processing strategy suggestions for the current case are generated based on the similar historical case.

2. The method according to claim 1, characterized in that, The process of performing multimodal parsing on the unstructured legal document yields structured data containing text content and its layout information, including: The layout of the unstructured legal documents is analyzed to identify and retain the text blocks, table areas, image positions and their spatial coordinates. Optical character recognition is performed on the image areas in the unstructured legal documents to extract key text information, and then aligned with the layout information after page analysis. The aligned key text information and layout information are input into a multimodal large model. The semantic relationship between text content and spatial information is understood through a cross-modal attention mechanism, and the structured data is output.

3. The method according to claim 2, characterized in that, The step of performing optical character recognition on image regions in the unstructured legal document to extract key text information includes: Optical character recognition is performed on the image region to generate text content and its corresponding recognition confidence level; When the recognition confidence level is lower than a preset threshold, the corresponding text content is marked as an item to be reviewed, triggering manual review or enabling the backup recognition engine; The reviewed text content and text content with a confidence level that reaches or exceeds a preset threshold are used as the key text information for alignment with the layout information after page analysis.

4. The method according to claim 2, characterized in that, The step of inputting the structured data into a pre-trained language model and extracting key elements includes: Based on a pre-trained language model, supervised fine-tuning is performed using labeled legal document samples. Legal domain knowledge is injected by updating feature parameters through low-rank adaptation technology. At the same time, positive and negative sample pairs are introduced for preference alignment optimization, resulting in a domain-adapted language model suitable for extracting elements from legal documents. Deploy the domain-adaptive language model on a local server or in a domestic IT innovation environment to ensure that sensitive data does not leave the domain; The structured data is input into the domain-adaptive language model, and the model is guided to output standardized key elements through a preset prompt template.

5. The method according to claim 1, characterized in that, Based on the extracted key elements, semantic retrieval is performed in a pre-built processing strategy instance library to match at least one similar historical case, and processing strategy suggestions for the current case are generated based on the similar historical cases, including: A pre-built instance library of processing strategies is constructed. Each instance contains at least case characteristics, applicable legal basis, operation steps and historical results information. Each instance is converted into a vector through an embedding model and then stored in a vector database. The key elements of the current case are converted into feature vectors, and semantic similarity is retrieved from the vector database to obtain a preset number of historical cases that are most similar to the current case. The key elements of the current case are combined with the preset number of historical cases retrieved to form a prompt, which is then input into a lightweight large language model to generate processing strategy suggestions.

6. The method according to claim 4, characterized in that, The proposed processing strategy includes strategy recommendations for debt collection and disposal scenarios. The method further includes: calculating the debt collection priority score P for the current case based on extracted key elements, wherein the debt collection priority score P is calculated according to the formula... ； Determined; where W is the debtor's subjective repayment score, which is quantified based on the debtor's commitment behavior and historical performance behavior in historical communication records. The commitment behavior is identified from the transcript of the call recording based on a preset commitment statement template or small model. The historical performance behavior is determined based on at least one of the debtor's actual repayment frequency and repayment timeliness rate in past collection cycles. C represents the debtor's asset enforceability score, which is derived from a quantitative assessment based on extracted property clues. These property clues include at least one or more of the following: bank account balance, real estate, vehicles, and equity. β is a preset weighting coefficient, with a value range of [0,1].

7. The method according to claim 5, characterized in that, The system obtains feedback on the execution results of the processing personnel regarding the processing strategy suggestions. When the execution result feedback is valid, the current case and its corresponding processing strategy suggestions are taken as new instances, converted into vectors, and injected into the vector database in real time. Deduplication is performed based on semantic similarity, and the data is automatically updated to the processing strategy instance library.

8. The method according to claim 1, characterized in that, The key elements also include legal complexity dimension features, which are binary variables used to characterize the complexity of a case in legal procedures. The legal complexity dimension feature takes the first value when at least one of the following conditions is met: the number of defendants is greater than 1, an objection to jurisdiction has been raised, or there is a multi-layered guarantee. Otherwise, it takes the second value.

9. A system for intelligent extraction and automated processing of legal document elements, characterized in that, The system includes: The acquisition module is used to acquire at least one unstructured legal document for the case to be processed, wherein the unstructured legal document comes from multiple heterogeneous business systems; The parsing module is used to perform multimodal parsing on the unstructured legal documents to obtain structured data containing text content and its layout information; The extraction module is used to input the structured data into a pre-trained language model and extract key elements. The key elements include at least one or more of the following: task stakeholder identifiers, descriptions of matters to be handled, associated resource information, time-series constraint information, and processing status information. The generation module is used to perform semantic retrieval in a pre-built processing strategy instance library based on the extracted key elements, match at least one similar historical case, and generate a processing strategy suggestion for the current case based on the similar historical case.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent extraction and automated processing method for legal document elements as described in any one of claims 1 to 8.

Images