Electronic file analysis method based on retrieval-enhanced generative model

Abstract

The invention relates to an electronic file analysis method based on a retrieval-enhanced generative model, and belongs to the field of target detection, character recognition, natural language processing and natural language understanding. According to the method, document data classification is performed on recognition of data titles and directories and classification and sorting results of electronic files, and a data structure for describing key information of cases is defined according to crime constitution of the cases with different crime names; performing recursive analysis on the litigation program document by using a rule-based interpretation executor; the method comprises the following steps: realizing information extraction of a record by using a generative model based on a retrieval generation enhancement technology, summarizing information extraction results of various documents according to review analysis requirements of different criminal names to form an analysis result of an electronic file of a case, and generating structured data according to a required format; and intelligent identification and structured management of key information and data of the case are realized. According to the invention, processing is convenient, and office efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the fields of target detection, character recognition, natural language processing and natural language understanding, and specifically relates to an electronic file analysis method based on a retrieval-enhanced generative model. Background Art

[0002] In today's digital age, the analysis and processing of electronic files is of great significance to improving efficiency and fairness. Traditional electronic file analysis technology and methods provide support for the work to a certain extent, but there are also some problems and limitations.

[0003] The content of electronic case files is diverse and the amount of data is large. Case handlers need to manually screen and organize the case file data and then analyze the case-related information in the documents, which is a lot of work. In actual case file data, in addition to basic case element information such as personnel and companies, more attention should be paid to characteristic elements closely related to criminal behavior. Many behaviors involve new criminal methods, which are not only an important basis for the determination of case facts, but also the core of legal application and case characterization. Therefore, the review and analysis of electronic case files needs to be able to automatically identify elements closely related to crimes and provide high-precision extraction results.

[0004] The electronic case file data is composed of scanned files of the case file, including the case file cover, file directory within the file, case report registration form, case report receipt, case filing decision, arrest approval decision, detention certificate, detention notice, request for approval of arrest, arrest warrant, arrest notice, case transfer for prosecution notice, summons for interrogation, interrogation record, questioning record, notice of litigation rights and obligations of the suspect, contract, business license, receipt and other electronic scanned data of case documents. The scanned data is stored in the server in the form of image files.

[0005] Reviewing and analyzing the electronic case file data requires character recognition, document classification, and element extraction. In practical applications, traditional character recognition technology faces data quality issues such as inconsistent clarity and resolution of scanned images. Template-based and rule-based information extraction technologies are difficult to apply templates to verbal evidence documents such as interrogation records and inquiry records, and the formulation of rules for records requires a lot of time and effort, making it difficult to cover all possible situations.

[0006] The present invention proposes an electronic case file analysis method based on a retrieval-enhanced generative model, which uses a binary image segmentation algorithm and a convolutional recurrent neural network model to identify and classify the case file content for the electronic case file data. According to the crime composition of cases with different charges, a data structure describing the key information of the case is defined. For litigation procedure documents, a rule-based interpretation executor is used to identify element data. For verbal evidence documents, a generative model based on retrieval-enhanced technology is used to identify element data. According to the correlation between the review and analysis content and the elements, the recognition results are summarized, and the summarized recognition results are analyzed using a generative model to obtain the analysis results of the review content of the case electronic case file data. Summary of the invention

[0007] 1. Technical issues to be resolved

[0008] The technical problem to be solved by the present invention is how to provide an electronic file analysis method based on a retrieval-enhanced generative model to solve the problem that traditional character recognition technology faces data quality problems such as inconsistent clarity and resolution of scanned images in practical applications, and template-based and rule-based information extraction technologies are difficult to cover all possible situations.

[0009] (II) Technical solution

[0010] In order to solve the above technical problems, the present invention proposes an electronic file analysis method based on a retrieval-enhanced generative model, which comprises the following steps:

[0011] 1. For the electronic case file data, use the binary image segmentation algorithm to detect the character position, and use the convolutional recurrent neural network model to recognize the character content, and use the recognition results to realize the recognition of data titles and directories;

[0012] 2. Classify the identification results of the electronic case file data; the electronic case file data includes: prosecution opinion, request for approval of arrest, interrogation record, inquiry record, case filing decision, process of surrender and arrest approval decision;

[0013] 3. Define the data structure that describes the key information of the case according to the crime composition of different crimes;

[0014] Fourth, use the rule-based interpretation executor to recursively parse the structured litigation procedure documents, including: prosecution opinion, request for arrest approval, case filing decision, case presentation process and arrest approval decision;

[0015] 5. Use the generative model and the search generation enhancement technology to build a prompt word project for interrogation records and inquiry records to achieve information extraction from the records;

[0016] 6. Based on the review and analysis requirements of different crimes, the information extraction results of various documents are summarized to form the analysis results of the electronic case files, and structured data are generated in the required format to achieve intelligent identification and structured management of key case information and data, and assist prosecutors in quickly focusing on key case information.

[0017] (III) Beneficial effects

[0018] The present invention proposes an electronic case file analysis method based on a retrieval-enhanced generative model, which aims to convert unstructured case electronic case file data into structured text data through image recognition technologies such as binary image segmentation algorithms and convolutional recurrent neural networks. Define a data structure that describes key case information, use a rule-based interpretation executor to extract information from litigation procedure documents, and use a generative model based on retrieval generation enhancement technology to extract information from verbal evidence documents. According to the defined relationship between the review and analysis content and the elements, summarize the recognition results, use a generative model to analyze the summarized recognition results, and obtain the analysis results of the review content. Provide basic data and review and analysis conclusions to verify the specific behavior of suspects and determine illegal facts. The above technology provides data support for the conversion of non-digital electronic case files, and at the same time provides data support for intelligent analysis of documents, special statistics and other tasks, which facilitates processing and improves office efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a recognition data flow diagram of the generative model based on the retrieval enhancement generation technology of the present invention;

[0020] Figure 2 Analyzing data flow graphs for generative model-based review. DETAILED DESCRIPTION

[0021] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0022] The present invention relates to an electronic file analysis method based on a retrieval-enhanced generative model, comprising the following steps:

[0023] 1. For the electronic case file data, use the binary image segmentation algorithm to detect the character position, and use the convolutional recurrent neural network model to recognize the character content, and use the recognition results to realize the recognition of data titles and directories;

[0024] 2. Classify the identification results of electronic case files such as prosecution opinions, arrest request, interrogation records, inquiry records, case filing decisions, process of surrender, and arrest approval decisions;

[0025] 3. Define the data structure that describes the key information of the case according to the crime composition of different crimes;

[0026] 4. Use the rule-based interpretation executor to recursively parse the litigation procedure documents with standardized structures, such as the prosecution opinion, the request for approval of arrest, the decision to file a case, and the process of surrender;

[0027] 5. Use the generative model and the search generation enhancement technology to build a prompt word project for interrogation records and inquiry records to achieve information extraction from the records;

[0028] 6. Based on the review and analysis requirements of different crimes, the information extraction results of various documents are summarized to form the analysis results of the electronic case files, and structured data are generated in the required format to achieve intelligent identification and structured management of key case information and data, and assist prosecutors in quickly focusing on key case information.

[0029] Embodiment 1:

[0030] The present invention provides an electronic file analysis method based on a retrieval-enhanced generative model, the process comprising:

[0031] 1. For the electronic case file data, use the binary image segmentation algorithm to detect the character position, and use the convolutional recurrent neural network model to recognize the character content, and use the recognition results to realize the recognition of data titles and directories;

[0032] Step S11, preprocess the case electronic file data. Perform preprocessing operations on the input image, including direction correction, grayscale, denoising and normalization. Grayscale converts color images into grayscale images, reduces the amount of data and highlights the contour features of characters. Denoising refers to removing noise interference in the image to improve the accuracy of subsequent processing. Normalization adjusts the pixel value of the image to a certain range for subsequent calculation and processing.

[0033] Step S12: Binarize the file image data. Use a binarization algorithm to convert the pre-processed image into a binary image, and divide the pixels in the image into two categories: foreground (characters) and background. Use a local adaptive threshold method to dynamically determine the threshold according to the local features of the image.

[0034] Step S13, detect the position of the characters in the file image. Use the connected region analysis method to detect the position of the characters. A connected region refers to a set of pixels in the image that have the same pixel value and are interconnected. By marking and analyzing the connected regions in the binary image, the position of the line where the text is located is detected, and the bounding box coordinates of the line where the characters are located are obtained.

[0035] Step S14: Identify the character content in the file image. Use the convolutional recurrent neural network model to identify the bounding box of the row where the character is located to obtain the character recognition result of the image. Among them, the convolution layer is used to extract the characteristics of the character, the recurrent layer is used to process the sequence data, and the fully connected layer is used to output the category probability of the character.

[0036] Step S15: According to the document title of the electronic case file data, determine the rules for identifying the title and the directory, including that the title usually appears in the first three lines, uses a larger font, is located at the top of the page or in a specific position, and the directory usually contains specific keywords or formats.

[0037] Step S16: Through the recognition rules of titles and directories, the character recognition results obtained in step S14 are pattern matched with the defined rules. If the features meet the rules of titles or directories, the corresponding character content is recognized as a title or directory. If it does not meet the specified recognition rules, the title is further confirmed based on the context information of the character content. For example, by analyzing the content and structure of the previous and next paragraphs, it is determined whether the recognized content belongs to the same document as the previous page of file data. According to the feedback and errors in actual applications, the rules and methods of title and directory recognition are optimized and adjusted to improve the accuracy and reliability of recognition.

[0038] 2. Classify the identification results of electronic case files such as prosecution opinions, arrest request, interrogation records, inquiry records, case filing decisions, process of surrender, and arrest approval decisions;

[0039] Step S21: After obtaining the title and content of the structured case electronic file data, pre-process the document data. Perform preliminary cleaning and screening on the acquired electronic file data to remove noise data, such as garbled characters, abnormal characters, etc. At the same time, standardize the data format to ensure that all data can be subsequently analyzed and classified in a standardized format.

[0040] Step S22: classify the cleaned electronic case file data into seven categories, including: litigation procedure documents, confessions and defenses of criminal suspects, witness testimonies and victim statements, physical evidence, documentary evidence, expert opinions, and investigation records.

[0041] 3. Define the data structure that describes the key information of the case according to the crime composition of different crimes;

[0042] According to the criminal composition of different crimes, define the data structure that describes the key information of the case.

[0043] S31. Clearly divide the basic information of the persons and units involved in the case. The basic information of the persons covers their identity, position, family background and suspected crimes; the basic information of the units includes but is not limited to the basic registration information of the enterprise, internal organizational structure, actual operation status and capital flow;

[0044] S32. Based on the four elements of the crime, define the data structure that describes the property, assets, and liabilities of the persons or units involved in the case, including asset status, bank account status, debt status, transfer payment status, etc.;

[0045] S33. Analyze the core behavioral characteristics of the crime and obtain the criminal behavior data structure that is broken down into specific crimes.

[0046] Taking illegal fund-raising cases as an example, the criminal behavior involves personal investment, fund-raising, fund-raising operations at the corporate level, the methods and processes of organizing and planning financing activities, the behavior patterns of repaying principal and paying interest, and public information dissemination channels. By thoroughly recording the specific facts of these key links, it provides solid data support for revealing that the nature of the criminal behavior is actually illegal fund-raising activities for the purpose of absorbing public deposits. The element data of illegal fund-raising cases are shown in Table 1.

[0047] Table 1 Data table of elements of illegal fund-raising cases

[0048]

[0049] Among them, the data attributes of personnel include basic information such as name, gender, ethnicity, and date of birth; the data attributes of units include basic information such as unit name, legal representative, actual controller, establishment time, shareholder composition, and affiliated companies; the attributes of personal investment status include personal investment information such as investment personnel, investment amount, investment process, loss amount, rebate amount, investment time, introducer, and relationship with introducer; the attributes of organizing and planning fund-raising activities include organizational planning information such as the design and organization of fund-raising activities, publicity and training, contract signing and collection, specific operations of absorbing funds, and fund-absorbing system; the repayment of principal and interest includes relevant information such as the repayment method, term, and principal repaid. Each data structure unit has independent data attributes. For different documents, the data is associated and matched from the perspective of the original data, ensuring that when the system finally uses the data, it can present the actual situation that one type of element corresponds to multiple personnel and units, or multiple personnel correspond to multiple types of elements. Taking "employment status" as an example, the specific attributes are shown in Table 2.

[0050] Table 2 Specific attributes table

[0051]

[0052]

[0053] 4. Use the rule-based interpretation executor to recursively parse the litigation procedure documents with standardized structures, such as the prosecution opinion, the request for approval of arrest, the decision to place a case, the process of surrender and the decision to approve arrest;

[0054] Extract information from litigation procedure documents. The case information contained in the electronic case file data, such as the prosecution opinion, the request for approval of arrest, the decision to file a case, the process of surrender, the decision to approve arrest, etc., is concentrated in the paragraphs such as "Basic information of the criminal suspect" and "Ascertained through investigation in accordance with the law". These paragraphs contain dense case information, and the content and structure of the context are fixed and standardized. The rule-based interpretation executor is used for recursive parsing.

[0055] Step S41, clarify the element information required to be identified in the litigation procedure document. The litigation procedure document mainly includes basic case information, suspect information, description of criminal facts, evidence enumeration, applicable law, etc. Use the relevant data structure in step S33 to define the data structure of the litigation procedure document identification information.

[0056] Step S42: Formulate extraction rules for key information required for identifying litigation procedure documents. Formulate detailed information extraction rules based on key information elements. For example, for the date of birth in the basic information of the suspect, the rules can be set as:

[0057] <[bgr].*? ((?P <result>[Date])(born|gave birth to).*? (?P <result1>[riqi]))>

[0058] By setting the format, it is used to extract dates, digital codes, etc. in specific formats from litigation procedure documents; for information on criminal suspects, the rules can be used to extract information based on common name formats, identity identifiers, etc.; for descriptions of criminal facts, the rules can focus on extracting key verbs, descriptive words for criminal behavior, etc.; for evidence enumeration, the rules can be set to identify specific evidence type keywords; for the application of law, the rules can be set to extract relevant legal article numbers and specific clause contents.

[0059] Step S43: Use the defined key information extraction rules to perform recursive parsing operations on the litigation procedure documents, and finally output complete, accurate and logical parsing results.

[0060] 5. Use the generative model and the search generation enhancement technology to build a prompt word project for interrogation records and inquiry records to achieve information extraction from the records;

[0061] Information is extracted from documents such as confessions and defenses of criminal suspects, testimonies of witnesses, and statements of victims. Interrogation records are records of interrogations of suspects and witnesses by public security organs during the investigation stage. The content often includes case background, relevant statements and testimonies. The language is colloquial and the structure is loose. Key information is often buried in a large amount of secondary information. Interrogation records are conducted on criminal suspects or defendants during the investigation or trial stage of a case. The content focuses on the facts, motives and methods of the crime. It is targeted and logical, but the authenticity of the information needs to be evaluated and verified through the model. Both are recorded in a question-and-answer format. The interrogation records and questioning records in the case contain legal terms and industry terms, and the terms have specific industry regulations and contextual dependencies in the documents. Generative language models are used for semantic analysis and logical organization.

[0062] Step S51: Build a search library for the electronic case file data. Build a search library containing key information and common expressions. The search library can search historical case data, laws and regulations, professional terms, etc. The information in the search library will serve as an important reference and basis for generating prompt words.

[0063] Step S52: Design specific prompt words instructions according to the defined case key information data structure. Combine the generative model and the search library to generate prompt words for the interrogation record and the inquiry record. When generating prompt words, the model generates relevant and targeted prompt words based on the input text content and the information in the search library. These prompt words are intended to guide the direction and focus of information extraction.

[0064] Step S53: Use the search generation enhancement technology to supplement and correct the prompt words. Set a callback mechanism to log the output results of the generative model, and use the search generation enhancement technology to further optimize the generation effect of the prompt words. By introducing the search results in the generation process, the generated prompt words are supplemented and corrected to make them more accurate and comprehensive.

[0065] Step S54: Apply the generated prompt words to the information extraction of the interrogation record and the inquiry record. By matching and analyzing the text content, key information is extracted, such as personnel positions, unit organizational structure, actual business conditions, organization and planning of fund raising, repayment of principal and interest, etc. The overall recognition process is as follows: Figure 1 shown.

[0066] 6. Based on the review and analysis requirements of different crimes, the information extraction results of various documents are summarized to form the analysis results of the electronic case files, and structured data are generated in the required format to achieve intelligent identification and structured management of key case information and data, and assist prosecutors in quickly focusing on key case information.

[0067] Summarize the element identification results of the litigation procedure documents obtained in step 4, and the element identification results of the statement documents obtained in step 5. Review and analyze the case based on the specific content of the key review and analysis in the case handling. Take the illegal fundraising case as an example to introduce the specific content of the review and analysis. The crime of illegally absorbing public deposits refers to the act of violating national financial management regulations, illegally absorbing public deposits or absorbing public deposits in disguise, and disrupting the financial order. In the review of illegal fundraising cases, it is necessary to analyze the case from the aspects of proving the fact of absorbing public deposits, subjective aspects, and the whereabouts of the funds raised.

[0068] Step S61: In handling cases of illegal absorption of public deposits, firstly, evidences related to the criminal pattern should be comprehensively collected, and on this basis, evidences should be further collected and fixed around the three characteristics of absorption of public deposits, namely publicity, inducement and sociality. According to the case elements and case characteristics proving the criminal pattern as shown in Table 3, the criminal pattern of absorption of public deposits is examined through the case elements such as personal investment, personal fund raising, unit fund raising, organization and planning of fund raising activities, public publicity, fund circulation, transfer and payment.

[0069] Table 3 Case elements related to crime patterns

[0070]

[0071] Step S62: Illegal absorption of public deposits is often carried out in the name of an organization. Therefore, in the process of reviewing evidence, attention should be paid to the review of relevant evidence such as the identity of the entity, business scope, and business conditions. The focus should be on reviewing whether the entity actually exists, whether it was established for the purpose of committing a crime, and whether the entity's main business after its establishment is to illegally absorb public deposits.

[0072] Table 4 Case elements related to subjective aspects

[0073]

[0074]

[0075] Step S63, comprehensively collect and fix evidence around the actual destination of the funds, the main causes of the losses, the business conditions of the suspects, etc., and determine whether the funds are used for production and business activities, whether the funds are squandered, and then determine whether there is an intention of illegal possession based on the actual destination of the funds and the actual business conditions.

[0076] Table 5 Case elements related to the whereabouts of raised funds

[0077]

[0078] Step S64: Summarize the element recognition results of the litigation procedure documents obtained in step 4 and the element recognition results of the statement documents obtained in step 5. According to the correlation between the review and analysis content and the elements, the recognition results are summarized, and the summarized recognition results are analyzed using a generative model to obtain the analysis results of the review content.

[0079] Step S65. Through the above review and analysis, relevant information such as the organization and planning of fund-raising, the facts of absorbing deposits, the actual whereabouts of funds, etc. that reflects the case of illegal absorption of public deposits is verified and fixed evidence is collected, so as to provide basic data and review and analysis conclusions for reviewing the criminal model of the case of illegal absorption of public deposits, analyzing the characteristics of the behavior of absorbing public deposits such as inducement, publicity, sociality, and illegality, verifying the specific behavior of the suspect, and determining the facts of the violation.

[0080] Embodiment 2:

[0081] An electronic file analysis method based on a retrieval-enhanced generative model includes the following steps:

[0082] 1. Classify and sort the electronic case file data, use the binary image segmentation algorithm to detect the character position, and use the convolutional recurrent neural network model to recognize the character content. Through the recognition results, the data title and directory can be recognized;

[0083] 2. Classify the prosecution opinions, arrest approval requests, interrogation records, inquiry records, case filing decisions, process of surrender, arrest approval decisions, and other documents in a systematic manner according to the structured data;

[0084] 3. According to the crime composition of different crimes, define the data structure describing the key information of the case. First, clearly divide the basic information of the persons and units involved in the case. The basic information of the personnel covers the identity, position, family background and suspected crimes; the basic information of the unit includes but is not limited to the basic registration information of the enterprise, the internal organizational structure, the actual operation dynamics and the flow of funds; secondly, according to the four elements of the crime composition, define the data structure describing the property, assets and liabilities of the persons or units involved in the case, including assets, bank accounts, debts, transfer payments, etc.; finally, analyze the core behavioral characteristics of the crime composition and refine the criminal behavior data structure of the specific crime. Taking illegal fund-raising cases as an example, the criminal behavior involves personal investment, fund-raising, fund-raising operations at the enterprise level, the way and process of organizing and planning financing activities, the behavior pattern of repaying principal and paying interest, and the public information dissemination channels. By recording the specific facts of these key links in detail, provide solid data support for revealing the nature of the criminal behavior.

[0085] 4. Use the rule-based interpretation executor to parse the prosecution opinion, the request for arrest approval, the decision to file a case, the process of surrendering the case, and other litigation procedure documents, and identify the case information in the documents by recursively calling the parsing rules;

[0086] 5. Use generative models and retrieval generation enhancement technology to build a prompting project for interrogation records, inquiry records, and other documents such as confessions and defenses of criminal suspects, witness testimonies, and victim statements to achieve information extraction from the records;

[0087] 6. According to the review and analysis requirements of different crimes, the information extraction results of various documents are summarized, and the summarized recognition results are analyzed using a generative model to form the analysis results of the electronic case file, and structured data is generated in the required format to achieve intelligent recognition and structured management of key case information and data.

[0088] Furthermore, the binary image segmentation algorithm and convolutional recurrent neural network model in step one are target detection algorithms and character recognition models. On the basis of character recognition, a rule template for confirming the image title is formulated according to the title and content format of the electronic file data, which includes various judgment rules corresponding to the type of book in the electronic file.

[0089] Furthermore, the classification of document data in step 2 includes classifying the electronic case file data into case data types such as litigation procedure documents, confessions and defenses of criminal suspects, witness testimonies, victim statements, physical evidence, documentary evidence, expert opinions, and records of inspection and identification experiments.

[0090] Furthermore, the data structure describing the key information of the case is defined in step three, including personnel elements, unit elements, common elements, specific crime elements, etc. The data structure of each element contains specific data attributes.

[0091] Furthermore, the rule-based interpretation executor in step 4 is a formal language based on Backus-Naur Form, which is used to express the grammar described by the context method.

[0092] Furthermore, the generative model in step five is a generative language model, and the retrieval enhanced generation technology includes retrieval (obtaining relevant document fragments from the document system) and generation (the generative language model uses the information in these fragments to generate answers) to enhance the domain knowledge of the generative model and improve the accuracy of the model recognition results by introducing external data.

[0093] Furthermore, the analysis result in step six refers to the review of the aggregated recognition results by the generative model, the aggregation of the recognition results based on the relationship between the review analysis content and the elements, and the use of the generative model to analyze the aggregated recognition results to obtain the analysis result of the review content.

[0094] The present invention is an electronic case file analysis method based on a retrieval-enhanced generative model, which aims to convert unstructured case electronic case file data into structured text data through image recognition technologies such as binary image segmentation algorithms and convolutional recurrent neural networks. Define a data structure that describes key case information, use a rule-based interpretation executor to extract information from litigation procedure documents, and use a generative model based on retrieval generation enhancement technology to extract information from verbal evidence documents. According to the defined relationship between the review and analysis content and the elements, summarize the recognition results, use a generative model to analyze the summarized recognition results, and obtain the analysis results of the review content. Provide basic data and review and analysis conclusions to verify the specific behavior of suspects and determine illegal facts. The above technology provides data support for the conversion of non-digital electronic case files, and at the same time provides data support for intelligent analysis of documents, special statistics, etc., to facilitate processing and improve office efficiency.

[0095] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention. < / result>

Claims

1. An electronic file analysis method based on a retrieval-enhanced generative model, characterized in that: The method comprises the following steps:

1. For the electronic case file data, use the binary image segmentation algorithm to detect the character position, and use the convolutional recurrent neural network model to recognize the character content, and use the recognition results to realize the recognition of data titles and directories; 2. Classify the identification results of the electronic case file data; The electronic case file data includes: prosecution opinion, request for arrest approval, interrogation record, inquiry record, case filing decision, process of surrender and arrest approval decision; 3. Define the data structure that describes the key information of the case according to the crime composition of different crimes; Fourth, use the rule-based interpretation executor to recursively parse the structured litigation procedure documents, including: prosecution opinion, request for arrest approval, case filing decision, case presentation process and arrest approval decision; 5. Use the generative model and the search generation enhancement technology to build a prompt word project for interrogation records and inquiry records to achieve information extraction from the records; 6. Based on the review and analysis requirements of different crimes, the information extraction results of various documents are summarized to form the analysis results of the electronic case files, and structured data are generated in the required format to achieve intelligent identification and structured management of key case information and data, and assist prosecutors in quickly focusing on key case information.

2. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 1, characterized in that: The step 1 specifically includes: Step S11, preprocessing the case electronic file data: preprocessing the input image; Step S12, binarization processing is performed on the file image data: the pre-processed image is converted into a binary image using a binarization algorithm, and the pixels in the image are divided into two categories: foreground characters and background; the threshold is dynamically determined according to the local features of the image using a local adaptive threshold method; Step S13, detecting the position of characters in the file image: using a connected region analysis method to detect the position of characters; by marking and analyzing the connected regions in the binary image, detecting the position of the line where the text is located, and obtaining the bounding box coordinates of the line where the characters are located; Step S14, identifying the character content in the file image: using a convolutional recurrent neural network model, identifying the bounding box of the row where the character is located, and obtaining the character recognition result of the image, wherein the convolutional recurrent neural network model includes: a convolutional layer, a recurrent layer and a fully connected layer; the convolutional layer is used to extract the features of the character, the recurrent layer is used to process the sequence data, and the fully connected layer is used to output the category probability of the character; Step S15: Determine rules for title and directory recognition based on the document title of the electronic case file data; Step S16: Through the recognition rules of titles and directories, the character recognition results obtained in step S14 are pattern matched with the defined rules; if the features meet the rules of titles or directories, the corresponding character content is recognized as a title or directory. If it does not conform to the prescribed recognition rules, the title is further confirmed based on the contextual information of the character content.

3. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 2, characterized in that: The preprocessing operation of S11 includes: direction correction, grayscale, denoising and normalization; grayscale converts a color image into a grayscale image, denoising refers to removing noise interference in the image, and normalization adjusts the pixel value of the image to a certain range.

4. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 2, characterized in that: The second step comprises: Step S21: After obtaining the title and content of the structured case electronic file data, pre-process the document data; perform preliminary cleaning and screening on the obtained electronic file data to remove noise data; at the same time, unify the data format to ensure that all data can be subsequently analyzed and classified in a standardized format; Step S22: classify the cleaned electronic case file data into seven categories, including: litigation procedure documents, confessions and defenses of criminal suspects, witness testimonies and victim statements, physical evidence, documentary evidence, expert opinions, and investigation records.

5. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 4, characterized in that: The step three specifically includes: S31. Clearly divide the basic information of the persons and units involved in the case: the basic information of the persons includes: identity, position, family background and suspected crimes; the basic information of the units includes: basic registration information of the enterprise, internal organizational structure, actual operation status and capital flow; S32. Based on the four elements of the crime, define the data structure that describes the property, assets, and liabilities of the persons or units involved in the case, including asset status, bank account status, debt status, and transfer payment status; S33. Analyze the core behavioral characteristics of the crime and obtain the criminal behavior data structure that is broken down into specific crimes.

6. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 5, characterized in that: For illegal fund-raising cases, the criminal acts involve personal investment situations, fund-raising situations, fund-raising operations at the enterprise level, the methods and processes of organizing and planning financing activities, behavioral patterns of repaying principal and paying interest, and public information dissemination channels; the data attributes of individuals include: name, gender, ethnicity and date of birth; the data attributes of units include: unit name, legal representative, actual controller, establishment time, shareholder composition and affiliated companies; the attributes of personal investment situations include: investment personnel, investment amount, investment process, loss amount, rebate amount, investment time, introducer and relationship with the introducer; the attributes of organizing and planning fund-raising activities include: design and organization of fund-raising activities, publicity and training, contract signing and collection, specific operations of absorbing funds and fund-absorbing system; the repayment of principal and interest includes: repayment method, term and principal repaid; each data structure unit has independent data attributes, and for different documents, they are associated and corresponded from the perspective of original data.

7. The electronic file analysis method based on the retrieval-enhanced generative model as claimed in claim 5, characterized in that: The fourth step comprises: Step S41, clarify the element information required to be identified in the litigation procedure document; the litigation procedure document includes: basic case information, criminal suspect information, description of criminal facts, evidence enumeration, applicable law, etc.; use the relevant data structure in step S33 to define the data structure of the litigation procedure document identification information; Step S42: Formulate extraction rules for key information required for identifying litigation procedure documents; formulate detailed information extraction rules based on key information elements; Step S43: Use the defined key information extraction rules to perform recursive parsing operations on the litigation procedure documents; and finally output complete, accurate and logical parsing results.

8. The electronic file analysis method based on the retrieval-enhanced generative model according to claim 7, characterized in that: In step S42, the date and digital codes in a specific format in the litigation procedure documents are extracted by setting the format; for the information of the suspect, the rules are used to extract the common name format and identity identification; for the description of the crime facts, the rules focus on extracting key verbs and descriptive words of the crime; for the enumeration of evidence, the rules are set to identify specific evidence type keywords; for the application of law, the rules are set to extract the relevant legal article numbers and specific clause contents.

9. The electronic file analysis method based on the retrieval-enhanced generative model according to claim 7, characterized in that: The step five comprises: Step S51, constructing a search library for the case electronic file data: establishing a search library containing key information and common expressions, the search library is used to search historical case data, laws and regulations, and professional terms; the information in the search library will serve as an important reference and basis for generating prompt words; Step S52: Design specific prompt words instructions according to the defined case key information data structure: Combine the generative model and the search library to generate prompt words for the interrogation record and the inquiry record. When generating prompt words, the model generates relevant and targeted prompt words based on the input text content and the information in the search library; Step S53, using the search generation enhancement technology to supplement and correct the prompt words: set a callback mechanism to log the output results of the generative model, and use the search generation enhancement technology to further optimize the generation effect of the prompt words; by introducing the search results in the generation process, the generated prompt words are supplemented and corrected to make them more accurate and comprehensive; Step S54, applying the generated prompt words to information extraction of interrogation records and inquiry records: extracting key information through matching and analysis with text content.

10. The electronic file analysis method based on the retrieval-enhanced generative model according to claim 9, characterized in that: The step six comprises: Step S61: In handling cases of illegal absorption of public deposits, firstly, evidences related to the criminal pattern should be comprehensively collected, and on this basis, evidences should be further collected and fixed around the three characteristics of absorption of public deposits: publicity, inducement and sociality; according to the case element table proving the criminal pattern, the criminal pattern of absorption of public deposits should be examined through personal investment, personal fund raising, unit fund raising, organization and planning of fund raising activities, public publicity, fund circulation and transfer payment; Step S62: Illegal absorption of public deposits is often carried out in the name of an organization. Therefore, during the process of reviewing the evidence, the focus is on whether the organization actually exists, whether it is established for the purpose of committing a crime, and whether the main business of the organization after its establishment is to illegally absorb public deposits. Step S63: comprehensively collect and fix evidence on the actual destination of the funds, the main causes of the losses, and the business conditions of the suspects, and judge whether the funds are used for production and business activities, whether the funds are squandered, and whether there is an illegal possession purpose according to the actual destination of the funds and the actual business conditions; Step S64: Summarize the element recognition results of the litigation procedure documents obtained in step 4 and the element recognition results of the statement documents obtained in step 5, summarize the recognition results according to the correlation between the review and analysis content and the elements, and use the generative model to analyze the summarized recognition results to obtain the analysis results of the review content; Step S65. Through the above review and analysis, the relevant information reflecting the organization and planning of the case of illegal absorption of public deposits, the facts of absorption of deposits, and the actual whereabouts of the funds is verified and fixed evidence is collected, so as to provide basic data and review and analysis conclusions for reviewing the criminal model of the case of illegal absorption of public deposits, analyzing the characteristics of the behavior of absorbing public deposits, verifying the specific behavior of the suspect, and determining the facts of the violation.

Images