Criminal judgment content extraction method based on large model

Abstract

The application discloses a criminal judgment content extraction method based on a large model, comprising the following steps: designing a nested JSON structure containing judgment elements according to the description of criminal trial supervision rules; generating training data by using a trillion-parameter model; fine-tuning and optimizing a hundred-billion-parameter model to obtain a field professional LLM; calling a VLM to perform OCR processing on a PDF version of a criminal judgment, and outputting a document format file; performing element extraction according to a grouping extraction strategy, and verifying based on a multi-layer fault tolerance mechanism; performing element standardization processing on all completed fault tolerance verified element data; identifying elements required for secondary processing, and generating business-specific elements. Based on the large model, preprocessing and format conversion are automatically performed, replacing manual input and format arrangement work; using a JSON structure to clearly indicate all elements that need to be extracted in the judgment, forming a standardized extraction framework, and avoiding element omission caused by subjective judgment deviation during manual review.

Description

Technical Field

[0001] This invention relates to a method for extracting content from criminal judgments, and more particularly to a method for extracting content from criminal judgments based on a large model. Background Technology

[0002] With the deepening of judicial reform, the procuratorate's work in supervising criminal trials faces increasingly heavy workloads. Prosecutors need to review a large number of criminal judgments for errors and omissions, identify problematic judgments, and initiate litigation supervision. Traditional methods of judgment review mainly rely on manual reading and experience-based judgment, which have significant limitations when dealing with massive amounts of unstructured judicial documents, including the following shortcomings: First, manually reviewing each judgment is time-consuming and labor-intensive, making it difficult to cope with the ever-increasing number of cases and resulting in low processing efficiency. Second, when faced with complex and lengthy judgment texts, manual review is prone to overlooking key information, posing a high risk of omission. Summary of the Invention

[0003] To address the shortcomings of the aforementioned technologies, this invention provides a method for extracting the content of criminal judgments based on a large model.

[0004] To solve the above technical problems, the technical solution adopted by this invention is: a method for extracting the content of criminal judgments based on a large model, specifically including the following steps: Step S1: Design a nested JSON structure containing elements of the judgment, based on the description of the rules for supervision of criminal trials; Step S2: Use the Deepseek-R1-0528 model to perform high-quality training data distillation to generate a labeled dataset covering criminal judgment scenarios as training data; Step S3: Based on the training data, the Qwen3-Next-80B-A3B-Thinking model is fine-tuned and optimized using the CHORD algorithm to obtain a domain-specific large model; Step S4: Call VLM to perform multimodal OCR text recognition on the PDF version of the criminal judgment and output it as a document file; Step S5: Set the grouping extraction strategy based on the JSON structure, extract elements according to the grouping extraction strategy based on the element extraction module, obtain element data, and verify it based on a multi-layer fault tolerance mechanism. Step S6: Traverse all element data that have completed fault tolerance verification, and perform element standardization processing based on the standardization processing module to obtain standardized elements; Step S7: Traverse all standardized elements, identify the elements required for secondary processing, and generate business-specific elements based on the post-processing module through a rule engine or domain-specific LLM. Step S8: Output the structured data in the PDF version of the criminal judgment.

[0005] Furthermore, in step S5, feature extraction is performed, specifically including the following steps: Step S51: Split the complete JSON structure into multiple closely related and appropriately sized groups to obtain the grouping results; Step S52: Based on the grouping results, extract specific models; Step S53: A multi-layered fault-tolerance mechanism is used for verification; if the extracted elements have abnormal JSON format, the domain-specific LLM will be automatically called again for repair; if the elements are missing, redundant, or modified, the element extraction will be triggered again; verify whether the data type meets the predefined requirements, and start the correction process if it does not.

[0006] Further, in step S52, specific model extraction is performed, which includes the following steps: Step S52-1: Design a dedicated LLM prompt template for each group; Step S52-2: Use domain-specific LLM to perform precise feature extraction by group.

[0007] Furthermore, the element standardization processing in step S6 includes date type standardization processing, sentence type standardization processing, amount type standardization processing, and date range type standardization processing.

[0008] Furthermore, step S6 involves element standardization, specifically including the following steps: Step S61: Standardize the date-type feature data; validate the date-type feature data based on regular expressions, and call the domain-specific LLM to perform format conversion, converting it to the international date standard format; Step S62: Standardize the element data of the sentence type; convert them into a unified quantitative numerical format, and identify and convert the month numbers represented by Chinese numerals into Arabic numerals; Step S63: Standardize the element data of the amount type; convert it into Arabic numeral format in yuan; Step S64: Standardize the feature data of the date range type; convert it to a uniform start time:end date format; Step S65: Standardize other types of element data; formulate specific standardization rules for elements unique to criminal judgments that are prone to format differences.

[0009] Furthermore, the specific standardization rules include converting full-width characters to half-width characters for case numbers; unifying the writing of ethnic groups to a standard format ending with "ethnicity"; removing suffixes ending with "culture" for educational level; and expressing place names according to a unified administrative division level standard.

[0010] Furthermore, the business-specific elements in step S7 include the time frame of the crime and jurisdictional compliance.

[0011] Furthermore, step S7 generates business-specific elements, specifically including the following steps: Step S71: Based on the rule engine, extract multiple crime times from standardized elements, filter out the earliest time as the start time of the crime, filter out the latest time as the end time of the crime, and calculate the duration of the crime based on the start time and end time of the crime using the standardized time difference, which is used as the crime time range. Step S72: Use the domain-specific LLM to determine whether the location of the court is in the same district or county as the place where the crime occurred, the location of the court is in the same district or county as the current residence of the perpetrator, or the location of the court is in the same district or county as the current residence of the victim, and obtain the results of the regional relationship analysis; based on the results of the regional relationship analysis and the rules of regional jurisdiction, determine whether the court in that area has legitimate jurisdiction.

[0012] This invention discloses a method for extracting content from criminal judgments based on a large model, which has the following beneficial effects: First, the scope of elements in criminal judgments is clearly defined through JSON structure, and the format of PDF criminal judgments is automatically converted based on a large model, replacing manual data entry and format sorting, thus reducing labor costs; the entire process of grouped element extraction, multi-layer fault tolerance verification, and standardization processing requires no manual intervention, significantly shortening the processing cycle of unit documents and improving processing efficiency. Second, this invention uses a JSON structure to clearly define all elements that need to be extracted from the judgment, forming a standardized extraction framework to avoid omissions of elements due to subjective judgment bias or lack of attention during manual review; it sets up a group extraction strategy to achieve refined extraction, and conducts full-dimensional verification of the extraction results through a multi-layer fault tolerance mechanism to correct missing and erroneous information in a timely manner, further avoiding the omission of key information; through domain-specific LLM coordination to generate business-specific elements, it can ensure that no key information is omitted in special business scenarios, providing accurate and complete structured data support for subsequent business processing. Attached Figure Description

[0013] Figure 1 This is a system architecture diagram of the present invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0015] like Figure 1 The method for extracting content from criminal judgments based on a large model, as shown, specifically includes the following steps: Step S1: Based on the description of the Criminal Trial Supervision Rules, design a complex nested JSON structure containing elements of the judgment. It should be noted that this JSON structure needs to consider complex situations such as multiple defendants, each with multiple charges, and is not a simple flat JSON structure, but a complex nested structure compatible with various scenarios. Simultaneously, this JSON structure incorporates appropriate data types to improve extraction accuracy and facilitate subsequent clue matching and data analysis. By setting up a nested JSON structure, and designing a deeply nested structured pattern tailored to the characteristics of the document, it comprehensively covers all elements required by the Criminal Trial Supervision Rules.

[0016] Step S2: High-quality training data distillation is performed using a model with hundreds of billions of parameters to generate a labeled dataset covering criminal judgment scenarios as training data, ensuring the diversity and representativeness of the training data, and training a large model; In this embodiment, the Deepseek-R1-0528 model is used for high-quality training data distillation. The Deepseek-R1-0528 model is a deep optimization based on the Deepseek V3 Base model and adopts a sparse hybrid expert architecture, which achieves high logical reasoning ability while maintaining a moderate parameter scale; Step S3: Based on the training data, the CHORD algorithm is used to fine-tune and optimize the model with 10 billion parameters to obtain the domain-specific large model, namely the domain-specific LLM. The core mechanism of the CHORD algorithm is to dynamically fuse expert data SFT through dual control of global weight μ and token-level weight φ during reinforcement learning, so as to achieve the optimal balance between imitation and exploration. In this embodiment, the 10 billion parameter model is Qwen3-Next-80B-A3B-Thinking. The Qwen3-Next-80B-A3B-Thinking model is the next-generation basic model architecture launched by the Qwen3 team. It has 80B total parameters, but only 3B parameters are activated in each inference, which is an extremely sparse MoE design for long context inference and inference tasks.

[0017] Step S4: Call the Visual Language Model (VLM) to perform multimodal OCR text recognition on the PDF version of the criminal judgment, and output it as a document format file. It should be noted that performing OCR text recognition on a PDF criminal judgment based on the VLM model and outputting it as a document format is a technique well-known to those skilled in the art and requires no further explanation. In this embodiment, the Qwen2.5-VL-32B-Instruct visual model is used for OCR recognition, i.e., optical character recognition, which has comprehensive capabilities in text recognition, context understanding, and format awareness. The Qwen2.5-VL-32B-Instruct visual model is a multimodal large-scale model specifically designed for processing scanned PDFs' OCR recognition. It is based on the Qwen2.5 architecture with multimodal extensions, integrating the advantages of the Transformer architecture in sequence information processing and special modules for visual information processing.

[0018] It solves the problem that some PDFs are scanned documents, which increases the difficulty of OCR recognition due to uneven lighting, dirt and blur; and it also solves the problem that the information at the end of the document is covered by red seals or ink bleeding, which affects the accuracy of OCR.

[0019] Step S5: Set the grouping extraction strategy based on the JSON structure, and extract domain features according to the grouping extraction strategy based on the feature extraction module to obtain feature data, and perform verification based on a multi-layer fault tolerance mechanism; specifically including the following steps: Step S51: Split the complete element JSON structure into multiple highly relevant and appropriately numbered groups to obtain grouping results and achieve accurate extraction; splitting the large-scale JSON structure into multiple highly relevant and appropriately numbered groups avoids the problems of attention distraction and context length limitation during single extraction; Step S52: Based on the grouping results, perform specialized model extraction, which includes the following steps: Step S52-1: Design a dedicated LLM prompt template for each group; Step S52-2: Use the trained and optimized Qwen-3-Next-80B-A3B-Thinking model to perform accurate feature extraction in groups to ensure focused attention; by setting a group extraction strategy, the problems of attention distraction and contextual constraints in large-scale structured extraction are effectively solved. Step S53: A multi-layered fault-tolerance mechanism is employed for verification, including format verification, content verification, and data type verification. Format verification involves automatically re-calling the trained and optimized Qwen3-Next-80B-A3B-Thinking model (domain-specific LLM) to repair any JSON format errors in extracted elements. Content verification triggers re-extraction if elements are missing, redundant, or modified. Data type verification checks whether the data type conforms to predefined requirements; if not, a correction process is initiated. By establishing a comprehensive fault-tolerance system from format to semantics, the robustness of the system is significantly improved.

[0020] The design incorporates multiple fault tolerance mechanisms, including JSON format error tolerance, element integrity verification, and data type validation, to ensure the accuracy and reliability of the extraction results.

[0021] Step S6: Perform element standardization processing based on the standardization processing module to obtain standardized elements and achieve format standardization; traverse all element data that has completed fault tolerance verification, perform data type verification, and formulate differentiated standardization specifications for high-frequency key element types such as date type, sentence type, amount type, and date range type. Combine regular expression verification and domain-specific LLM to achieve a unified format; specifically including the following steps: Step S61: Standardize the date-type element data; for various single date elements marked in criminal judgments, achieve format uniformity to ensure that time information can be accurately parsed and sorted by the subsequent system; for element data of the date type, perform regular expression verification and call the domain-specific LLM for format conversion to the international standard format of "YYYY-MM-DD"; Step S62: Standardize the element data of sentence type; for elements of criminal law term, uniformly convert them into quantitative numerical format to facilitate subsequent application in analysis scenarios such as sentencing comparison and sentence statistics; identify and convert the number of months represented by Chinese numerals into Arabic numerals; Step S63: Standardize the element data of the amount type; for all types of involved amount elements, convert them into Arabic numerals in yuan to eliminate the problems of differences in amount expression and confusion of units; Step S64: Standardize the date range type element data; for elements involving time intervals, uniformly convert them to the "start date:end date" format, i.e., "YYYY-MM-DD:YYYY-MM-DD", to facilitate subsequent time range queries and interval comparisons; for key element types such as time, sentence duration, and amount, implement automatic format standardization to ensure data consistency. Step S65: Standardize other types of element data; formulate special standardization rules for elements that are unique to criminal judgments and prone to format differences to further improve data standardization. Specifically, for the case number, convert full-width characters to half-width characters; for the writing of ethnic groups, uniformly convert them to the standard format ending with "ethnic group"; for the educational attainment, remove the suffix "education"; for place names, express them in accordance with the unified administrative division level norms and remove colloquial expressions or abbreviations. Based on standardization technology, the automatic standardization of judicial terms and formats is achieved, ensuring data consistency and usability.

[0022] Step S7: Traverse all standardized elements, identify the elements required for secondary processing, i.e., the derivative elements of the standardized elements, and perform business rule processing; based on the post-processing module, generate business-specific elements through a rule engine or a domain-specific LLM and perform post-processing on them. The specific steps are as follows: Step S71: Determine the time range of the crime, i.e., determine the start time and end time of the crime; based on the rule engine, extract multiple crime times from the standardized elements and select the earliest time as the start time of the criminal act, and the latest selected time as the end time of the criminal act; and based on the start time and end time of the criminal act, calculate the duration of the criminal act using the standardized time difference. Step S72: Judge the compliance of jurisdiction; call the domain-specific LLM to judge whether the location of the court, the place where the criminal act occurred, the current residence of the offender, or the current residence of the victim is in the same district or county, and obtain the regional relationship analysis result; then based on the regional relationship analysis result and the territorial jurisdiction rules, judge whether the court in this place has legal jurisdiction. That is, input information such as the name of the court, the place where the criminal act occurred, the current residence of the defendant, and the current residence of the victim, and judge whether the location of the court, the place where the criminal act occurred, the current residence of the offender, or the current residence of the victim is in the same district or county. By combining the rule engine and the domain-specific LLM, business-specific elements are derived from the standardized elements to meet the needs of specific supervision scenarios.

[0023] Step S8: Output the structured data in the PDF version of the criminal judgment to extract the content of the criminal judgment; it should be noted that the structured data includes standardized elements and business-specific elements.

[0024] By defining the scope of elements in criminal judgments through a JSON structure, and automatically preprocessing and converting PDF versions of criminal judgments based on a large model, the system replaces manual data entry and formatting, reducing labor costs. The entire process of grouped element extraction, multi-layered error correction, and standardized processing requires no manual intervention, significantly shortening the document processing cycle for each unit. Simultaneously, the JSON structure structuring of criminal judgments clarifies all elements to be extracted, forming a standardized extraction framework and preventing element omissions due to subjective judgment bias or lack of focus during manual review. A grouped extraction strategy enables refined extraction, and a multi-layered error correction mechanism conducts comprehensive checks on the extraction results, promptly correcting omissions and errors, further preventing the omission of key information. Domain-specific LLM coordination generates business-specific elements, ensuring no key information is omitted in special business scenarios, providing accurate and complete structured data support for subsequent business processing.

[0025] This invention converts unstructured criminal trial documents into structured JSON format data, facilitating subsequent automated processing and analysis. In real life, based on structured data and data governance capabilities, it automatically checks the supervisory points in criminal judgments, identifies potential problems, generates clues, and pushes them to relevant departments. At the same time, based on structured data, it enables intelligent question-and-answer of past criminal judgments and supports rapid retrieval of relevant cases by dimensions such as personnel and supervisory points.

[0026] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A method for extracting content from criminal judgments based on a large model, characterized in that, Specifically, the following steps are included: Step S1: Design a nested JSON structure containing elements of the judgment, based on the description of the rules for supervision of criminal trials; Step S2: Use the Deepseek-R1-0528 model to perform high-quality training data distillation to generate a labeled dataset covering criminal judgment scenarios as training data; Step S3: Based on the training data, the Qwen3-Next-80B-A3B-Thinking model is fine-tuned and optimized using the CHORD algorithm to obtain a domain-specific LLM. Step S4: Call VLM to perform multimodal OCR text recognition on the PDF version of the criminal judgment and output it as a document file; Step S5: Set the grouping extraction strategy based on the JSON structure, extract elements according to the grouping extraction strategy based on the element extraction module, obtain element data, and verify it based on a multi-layer fault tolerance mechanism. Step S6: Traverse all element data that have completed fault tolerance verification, and perform element standardization processing based on the standardization processing module to obtain standardized elements; Step S7: Traverse all standardized elements, identify the elements required for secondary processing, and generate business-specific elements based on the post-processing module through a rule engine or domain-specific LLM. Step S8: Output the structured data in the PDF version of the criminal judgment.

2. The method for extracting criminal judgment content based on a large model according to claim 1, characterized in that: The feature extraction in step S5 specifically includes the following steps: Step S51: Split the complete JSON structure into multiple closely related and appropriately sized groups to obtain the grouping results; Step S52: Based on the grouping results, extract specific models; Step S53: A multi-layered fault-tolerance mechanism is used for verification; if the extracted elements have abnormal JSON format, the domain-specific LLM will be automatically called again for repair; if the elements are missing, redundant, or modified, the element extraction will be triggered again; verify whether the data type meets the predefined requirements, and start the correction process if it does not.

3. The method for extracting criminal judgment content based on a large model according to claim 2, characterized in that: The specific model extraction in step S52 includes the following steps: Step S52-1: Design a specific LLM prompt template for each group; Step S52-2: Use domain-specific LLM to perform precise feature extraction by group.

4. The method for extracting criminal judgment content based on a large model according to claim 1, characterized in that: The element standardization processing in step S6 includes date type standardization processing, sentence type standardization processing, amount type standardization processing, and date range type standardization processing.

5. The method for extracting criminal judgment content based on a large model according to claim 4, characterized in that: The element standardization process in step S6 specifically includes the following steps: Step S61: Standardize the date-type feature data; validate the date-type feature data based on regular expressions, and call the domain-specific LLM to perform format conversion, converting it to the international date standard format; Step S62: Standardize the element data of the sentence type; convert them into a unified quantitative numerical format, and identify and convert the month numbers represented by Chinese numerals into Arabic numerals; Step S63: Standardize the element data of the amount type; convert it into Arabic numeral format in yuan; Step S64: Standardize the feature data of the date range type; convert it to a uniform start time:end date format; Step S65: Standardize other types of element data; formulate specific standardization rules for elements unique to criminal judgments that are prone to format differences.

6. The method for extracting criminal judgment content based on a large model according to claim 5, characterized in that: The specific standardization rules include converting full-width characters to half-width characters for case numbers; standardizing the writing of ethnic groups by converting them to a standard format ending with "ethnic group"; removing suffixes ending with "culture" for educational level; and expressing place names according to a unified administrative division hierarchy.

7. The method for extracting criminal judgment content based on a large model according to claim 1, characterized in that: The business-specific elements in step S7 include the time frame of the crime and jurisdictional compliance.

8. The method for extracting criminal judgment content based on a large model according to claim 7, characterized in that: The step S7, which generates business-specific elements, specifically includes the following steps: Step S71: Based on the rule engine, extract multiple crime times from standardized elements, filter out the earliest time as the start time of the crime, filter out the latest time as the end time of the crime, and calculate the duration of the crime based on the start time and end time of the crime using the standardized time difference, which is used as the crime time range. Step S72: Use the domain-specific LLM to determine whether the location of the court is in the same district or county as the place where the crime occurred, the location of the court is in the same district or county as the current residence of the perpetrator, or the location of the court is in the same district or county as the current residence of the victim, and obtain the results of the regional relationship analysis; based on the results of the regional relationship analysis and the rules of regional jurisdiction, determine whether the court in that area has legitimate jurisdiction.

Images