Intelligent processing method and system for community correction investigation and evaluation based on LLM large model

Abstract

The invention discloses a community correction investigation and evaluation intelligent processing method and system based on an LLM large model, and relates to the technical field of intelligent information processing, and the method comprises the specific steps of element extraction, information fusion, intelligent research and judgment, compliance verification and auditing and archiving. According to the method, element extraction and information fusion operation is executed through an LLM large model trained by corpora in the community correction field, core information is extracted from survey and evaluation entrusting materials, logic association is established, then a standardized questionnaire is generated based on the core information, meanwhile, semantic coding and association integration are performed on multi-source heterogeneous data, and therefore, a standardized questionnaire is generated. Then constructing a Prompt template library and carrying out intelligent research and judgment and compliance verification in combination with an LLM large model, determining unified evaluation dimensions and rules according to community correction work requirements, completing multi-dimensional quantitative scoring through the LLM large model, and splitting an evaluation conclusion into feature items and standard texts to carry out comparison verification; and standardization of the evaluation process and automation of compliance verification are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, specifically to an intelligent processing method and system for community corrections investigation and assessment based on an LLM large model. Background Technology

[0002] Community corrections investigation and assessment is a crucial preliminary step in community corrections work, and its conclusions directly affect the formulation and implementation of subsequent correction measures. As the standardization and refinement of community corrections work continue to improve, higher requirements are placed on the comprehensiveness, accuracy, and efficiency of investigation and assessment. Currently, the scope of community corrections work is gradually expanding, the circumstances of those being assessed are becoming increasingly complex, and the sources of information required for investigation and assessment are becoming more diverse. Traditional working models are no longer suitable for the actual needs of community corrections work under the new circumstances.

[0003] However, current community corrections investigation and assessment work largely relies on manual information collection, analysis, and judgment, which has many obvious shortcomings. The formats of information collected manually from multiple channels vary greatly, and there is a lack of efficient integration methods, making it difficult to achieve semantic association between data from different sources. The construction of assessment templates lacks unified standards and is easily affected by subjective human factors, resulting in inconsistent assessment dimensions and rules. The generated questionnaires are not targeted enough and cannot accurately match the specific circumstances of the assessed individuals. The compliance verification process is mainly based on manual review, which is not only inefficient but also prone to omissions and makes it difficult to quickly and comprehensively compare assessment conclusions with relevant regulations.

[0004] Therefore, we developed an intelligent processing method and system for community corrections investigation and assessment based on an LLM large model. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent processing method and system for community corrections investigation and assessment based on an LLM large-scale model. This invention performs element extraction and information fusion operations through an LLM large-scale model trained on corpora in the community corrections field. It utilizes named entity recognition and relation extraction functions to extract core information from the investigation and assessment commission materials and establish logical connections. Then, it generates a standardized questionnaire based on the core information. At the same time, it performs semantic encoding and association integration on multi-source heterogeneous data, transforming scattered data of different formats into a unified summary text, providing a comprehensive and structured basis for subsequent assessment work, and avoiding assessment loopholes caused by information dispersion or format differences.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a smart processing method for community corrections investigation and assessment based on an LLM large model, the specific steps of which are as follows:

[0007] Element extraction: Upon receiving the commission letter materials for investigation and evaluation, based on the legal element dictionary and through the LLM model, we first perform named entity recognition to extract core entities, then perform relationship extraction to establish logical connections between core entities, integrate them into core information, and standardize them into structured data to generate a list of core elements for investigation and evaluation.

[0008] Information fusion: Basic data is collected through multi-source information systems. Based on the list of core elements of the survey and evaluation, a standardized questionnaire is generated using an LLM model and distributed and collected. Semantic encoding and association mapping are performed on multi-source heterogeneous data to generate a summary text of multi-source survey information.

[0009] Intelligent assessment: The Prompt template library is retrieved, and the summary text of multi-source survey information is input into the LLM large model for multi-dimensional quantitative scoring to determine the preliminary assessment conclusion;

[0010] Compliance verification: Input the preliminary assessment conclusions into the LLM model for compliance verification, automatically correct any issues found during the verification, and generate a draft of the investigation and assessment opinions;

[0011] Review and Archiving: The initial draft of the investigation and evaluation opinions is pushed to the terminal for manual review. After the review is approved, a formal evaluation opinion is generated, which is then sent to relevant parties and the entire process data is encrypted and archived.

[0012] Furthermore, in the element extraction process, the LLM model is trained on a corpus from the community corrections domain, which includes community corrections case texts, assessment standard texts, and typical document texts. The LLM model includes named entity recognition, relation extraction, text generation, semantic encoding, logical reasoning, and compliance verification functions. Named entity recognition is performed by segmenting the investigation and assessment commission letter text into sentences and words, matching the segmentation results with entity entries in the legal element dictionary, and marking the successfully matched core entities. Relation extraction is performed by establishing logical connections between core entities and integrating these connections into core information, including basic information of the assessed person, relevant case details, commissioning party information, execution location information, and investigation period requirements. After the core information integration is complete, the LLM model's text generation function outputs a list of core elements for the investigation and assessment.

[0013] Furthermore, in the information fusion, the multi-source information system includes a community corrections agency database, a village and residents' committee information system, and a household registration system; the collected basic data includes residence information, family and social relationships, and community security status; based on the core information in the core element list of the investigation and assessment, the text generation function of the LLM model is called to generate standardized questionnaires, including victim opinion collection questionnaires and community public opinion collection questionnaires, supporting both online QR code and offline manual distribution and collection channels; the multi-source heterogeneous data consists of basic data and collected questionnaire texts; the semantic encoding and association mapping operation specifically involves: converting the format of the multi-source heterogeneous data, mapping various types of data into semantic vectors through the LLM model, and analyzing the cosine similarity between semantic vectors before association and integration, thereby generating a summary text of multi-source investigation information.

[0014] Furthermore, in the aforementioned intelligent assessment, the steps for constructing the Prompt template library are as follows: Combining the core needs of community correction investigation and assessment with the requirements of regional assessment work, six core assessment dimensions are determined, including residence stability, family support strength, consequences and impacts of the involved behavior, social danger, community acceptance, and suitability of the proposed prohibited items; assessment indicators and scoring rules are formulated for each dimension, and the assessment indicators and scoring rules of each dimension are transformed into prompt text that can be recognized by the LLM large model; the prompt texts of all dimensions are integrated to form the Prompt template library; the specific steps for multi-dimensional quantitative scoring are as follows: the summary text of multi-source investigation information is split into the six core assessment dimensions corresponding to the Prompt template library; the LLM large model extracts information, analyzes causal relationships, and determines the risk level of the split text according to the assessment indicators and scoring rules of each dimension, giving a quantitative score from 0 to 10 points; and the comprehensive assessment score is calculated using the comprehensive assessment score calculation formula, and the preliminary assessment conclusion is determined based on the comprehensive assessment score.

[0015] Furthermore, in the aforementioned intelligent assessment, the comprehensive evaluation score is calculated using the following formula: ,in, To comprehensively evaluate the score, The core evaluation dimensions are set to 6. For the first The weighting coefficients for each assessment dimension are determined through typical community correction cases, taking into account regional assessment requirements. For the first Quantitative score values ​​for each evaluation dimension;

[0016] Based on the comprehensive evaluation score, the preliminary evaluation conclusion is determined as follows: When At that time, the preliminary assessment conclusion was that community corrections were appropriate; when At that time, the preliminary assessment concluded that further investigation was needed; when At that time, the preliminary assessment concluded that community corrections were not applicable.

[0017] Furthermore, the specific steps for compliance verification are as follows: the preliminary assessment conclusion is broken down into four feature items: assessment basis, assessment logic, expression content, and privacy information. These are then input into the LLM model and compared with community correction case texts, assessment standard texts, and typical document texts. The compliance matching degree of each feature item is calculated using the compliance verification matching degree calculation formula to determine whether the feature item is compliant. Non-compliant feature items are automatically corrected. After all four feature items meet the standards, a draft of the investigation and assessment opinion is generated.

[0018] Furthermore, in the compliance verification, the formula for calculating the compliance verification matching degree is: ,in, For the compliance matching degree of feature items, This represents the number of feature terms, with a value of 4. For the first Evaluation content vector in each feature item With standard text vectors cosine similarity, Using the LLM large model to analyze the first The evaluation content of each feature term is semantically encoded and generated. Semantic encoding of corresponding feature items in the corpus of the community correction domain is generated using an LLM large model;

[0019] The compliance matching score determines whether a feature is compliant: when When it is deemed compliant, When it is determined that correction is needed, If so, it is determined that the corresponding feature item content needs to be regenerated.

[0020] Furthermore, during the review and archiving process, the initial draft of the investigation and assessment opinions is modified and annotated during the manual review stage. The LLM model extracts the change points for the modifications, compares the change points with the assessment standard text to generate verification prompts, and proceeds to the formal assessment opinion generation stage after the verification is passed. After the formal assessment opinion is generated, it is affixed with an electronic seal and sent to relevant parties, including the client and the supervisory agency. The entire process data is encrypted and then archived in the community correction agency database.

[0021] On the other hand, a community corrections investigation and assessment intelligent processing system based on an LLM large model includes:

[0022] Element Extraction Module: Receives the investigation and assessment commission letter and accompanying materials. Based on the preset legal element dictionary, it uses the LLM model to first perform named entity recognition to extract core entities, then performs relationship extraction to establish logical connections between core entities, integrates them into core information, and standardizes them into structured data to generate a list of core elements for investigation and assessment.

[0023] Information fusion module: Collects basic data through multi-source information systems, generates standardized questionnaires based on the list of core elements of the survey and evaluation using the LLM model and distributes and collects them, performs semantic encoding and association mapping on multi-source heterogeneous data, and generates a summary text of multi-source survey information;

[0024] Intelligent assessment module: retrieves the Prompt template library, inputs the summary text of multi-source survey information into the LLM large model for multi-dimensional quantitative scoring, and determines the preliminary assessment conclusion;

[0025] Compliance verification module: Input the preliminary assessment conclusions into the LLM large model for compliance verification, automatically correct the problems found in the verification, and generate a draft of the investigation and assessment opinions;

[0026] Review and archiving module: The initial draft of the investigation and evaluation opinions is pushed to the terminal for manual review. After the review is approved, a formal evaluation opinion is generated, which is pushed to relevant parties and the entire process data is encrypted and archived.

[0027] Compared with existing technologies, this intelligent processing method and system for community corrections investigation and assessment based on an LLM large model has the following beneficial effects:

[0028] I. This invention utilizes an LLM model trained on corpora from the community corrections domain to perform element extraction and information fusion operations. By employing named entity recognition and relation extraction functions, it extracts core information from the survey and assessment commission materials and establishes logical connections. Based on the core information, it generates a standardized questionnaire. Simultaneously, it performs semantic encoding and association integration on multi-source heterogeneous data, transforming scattered data of different formats into a unified summary text. This provides a comprehensive and structured basis for subsequent assessment work, avoiding assessment loopholes caused by information dispersion or format differences.

[0029] Second, this invention constructs a standardized Prompt template library and combines it with an LLM (Limited Ledger Model) for intelligent assessment and compliance verification. Based on the needs of community correction work, it determines unified assessment dimensions and rules, transforms them into prompt text that can be recognized by the LLM model, and then completes multi-dimensional quantitative scoring through the LLM model. At the same time, it breaks down the assessment conclusions into feature items and compares and verifies them with the standardized text, thereby achieving standardization of the assessment process and automation of compliance verification. This frees the assessment process from the interference of human subjective factors and ensures the objectivity and accuracy of the assessment conclusions.

[0030] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0032] Figure 1 A flowchart of an intelligent processing method for community corrections investigation and assessment based on an LLM large model;

[0033] Figure 2 This is a framework diagram of an intelligent processing system for community corrections investigation and assessment based on an LLM large model.

[0034] Figure 3 This is a flowchart of information fusion in an intelligent processing method for community corrections investigation and assessment based on an LLM large model. Detailed Implementation

[0035] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0036] Example 1:

[0037] In investigation and assessment scenarios applicable to community corrections, the community corrections agency receives an investigation and assessment request letter, which includes relevant explanations of the case and personal circumstances of the person being assessed. The agency operates based on a legal element dictionary and utilizes a large-scale LLM model trained with community corrections case texts, assessment standard texts, and typical document texts. This LLM model possesses six functions: named entity recognition, relation extraction, text generation, semantic encoding, logical reasoning, and compliance verification. The agency first performs named entity recognition on the investigation and assessment request letter, such as... Figure 1As shown, the material text is processed by sentence segmentation and word segmentation. The word segmentation results are matched with entity entries in the legal element dictionary to mark core entities such as the basic information of the person being evaluated, relevant case information, client information, execution location information, and investigation period requirements. Then, logical connections between the core entities are established through relation extraction operations. The connected core entities are integrated into core information. At the same time, the core information is standardized to form structured data, and finally, a list of core elements for investigation and evaluation is generated.

[0038] Based on the core element list for investigation and assessment, a multi-source information system comprised of a community corrections institution database, a village and resident committee information system, and a household registration system was used to collect basic data on the assessed person's residence, family and social relationships, and community security. The text generation function of the LLM (Local Level Management) model was then utilized to generate two standardized questionnaires based on the core information: a victim opinion collection questionnaire and a community public opinion collection questionnaire. These were distributed and collected through both online QR codes and offline manual channels. After the questionnaires were collected, the basic data and the collected questionnaire texts were integrated into multi-source heterogeneous data. This data underwent format conversion, and the LLM model was used to map various data types into semantic vectors. The cosine similarity between semantic vectors was analyzed and correlation integration was performed to form a summary text of multi-source investigation information, such as... Figure 2 As shown.

[0039] A pre-built Prompt template library was retrieved. This library, designed to meet the core needs of community corrections investigation and assessment, as well as regional assessment requirements, identified six core assessment dimensions: residential stability, family support, consequences and impacts of the offending behavior, social risk, community acceptance, and suitability of the proposed prohibitions. Corresponding assessment indicators and scoring rules were developed for each dimension, and these were transformed into prompt text recognizable by the LLM (Local Management Model). The resulting template library was then integrated. The aggregated multi-source survey information was broken down into the six core assessment dimensions corresponding to the template library. Based on the assessment indicators and scoring rules for each dimension, the LLM model extracted information, analyzed causal relationships, and determined risk levels from the broken-down text, assigning a quantitative score of 0 to 10 to each dimension. The overall assessment score was calculated using the following formula: ,in, To comprehensively evaluate the score, The core evaluation dimensions are set to 6. For the first The weighting coefficients for each assessment dimension are determined through typical community correction cases, taking into account regional assessment requirements. For the first The quantitative score values ​​for each evaluation dimension; the preliminary evaluation conclusion is determined based on the comprehensive evaluation score: when At that time, the preliminary assessment conclusion was that community corrections were appropriate; when At that time, the preliminary assessment concluded that further investigation was needed; when At that time, the preliminary assessment conclusion was that community correction was not applicable; in this assessment, the quantitative scores of the six dimensions of the assessed person were 8, 9, 7, 8, 9, and 8, respectively, and the final comprehensive assessment score was 8 points, and the preliminary assessment conclusion was that community correction was applicable.

[0040] The preliminary assessment conclusions are broken down into four feature items: assessment basis, assessment logic, content of expression, and privacy information. These are then input into the LLM model and compared with community correction case texts, assessment standard texts, and typical document texts. The compliance matching degree of each feature item is calculated using the compliance verification matching degree calculation formula, which is as follows: ,in, For the compliance matching degree of feature items, This represents the number of feature terms, with a value of 4. For the first Evaluation content vector in each feature item With standard text vectors cosine similarity, Using the LLM large model to analyze the first The evaluation content of each feature term is semantically encoded and generated. Semantic encoding of corresponding feature items in the corpus of community corrections is performed using an LLM large-scale model; compliance matching is used to determine whether the feature items are compliant: when When it is deemed compliant, When it is determined that correction is needed, If the content of the corresponding feature items needs to be regenerated, it is determined that the compliance matching degree of the four feature items in this assessment is 0.85, 0.88, 0.90 and 0.87 respectively. According to the compliance matching degree judgment rule, all four feature items are judged to be fully compliant and do not need to be corrected. The initial draft of the investigation and assessment opinion can be directly generated.

[0041] The initial draft of the investigation and assessment opinions is pushed to the staff's terminal for manual review. The staff modifies and annotates the draft. The LLM model extracts the change points for the modifications and compares the change points with the assessment standard text to generate verification prompts. After the verification is passed, the formal assessment opinion is generated. After the formal assessment opinion is generated, it is affixed with an electronic seal and pushed to the client and the supervisory agency. At the same time, the data of the entire process is encrypted and archived in the community correction agency database.

[0042] In summary, the community corrections investigation and evaluation work was carried out by relying on an LLM model trained on corpus data in the community corrections field to complete relevant information processing, generating standardized questionnaires based on data collected from multiple information systems, integrating multi-source survey information into a summary text, using a Prompt template library to complete the evaluation and judgment, ensuring that all feature items meet the standards after compliance verification, and then undergoing manual review to finally generate a formal evaluation report, which was then pushed out and archived, leading to an evaluation conclusion that community corrections are applicable.

[0043] Example 2:

[0044] In a scenario requiring supplementary investigation and assessment, a community corrections agency receives a letter of entrustment for investigation and assessment. The letter concerns an application filed by the person being assessed due to a dispute. Based on a pre-set legal element dictionary, the agency utilizes a large-scale LLM model trained with community corrections case texts, assessment standard texts, and typical document texts, and equipped with six functions: named entity recognition, relation extraction, text generation, semantic encoding, logical reasoning, and compliance verification. This model performs sentence and word segmentation on the letter of entrustment. The word segmentation results are matched with entity entries in the legal element dictionary to identify core entities such as the person being assessed's basic information, relevant details of the case, information of the entrusting party, information on the place of execution, and the required investigation period. Then, through relation extraction, logical connections between these core entities are established. After being integrated into core information, the data is standardized and processed into structured data, generating a list of core elements for investigation and assessment.

[0045] Based on the list of core elements for investigation and assessment, a multi-source information system comprising the community corrections agency database, village and resident committee information system, and household registration system is used to collect basic data on the assessed person's residence, family and social relationships, and community security status, such as... Figure 3 As shown, the text generation function of the LLM model is invoked to generate two types of standardized survey questionnaires based on core information: a victim opinion collection questionnaire and a community public opinion collection questionnaire. The questionnaires are distributed and collected using both online QR codes and offline manual methods. After the questionnaires are collected, the basic data and questionnaire texts are integrated into multi-source heterogeneous data. The data are then formatted and mapped into semantic vectors using the LLM model. The cosine similarity between semantic vectors is analyzed and correlation integration is carried out to generate a summary text of multi-source survey information.

[0046] A pre-built Prompt template library is retrieved, which includes six core assessment dimensions: residential stability, family support, consequences and impacts of the alleged conduct, social danger, community acceptance, and suitability of the proposed prohibitions. Each dimension has corresponding assessment indicators and scoring rules. The aggregated text of multi-source survey information is broken down into the six core assessment dimensions. The LLM model extracts information, performs causal relationship analysis, and determines the risk level of the broken text based on the assessment indicators and scoring rules of each dimension, and gives quantitative scores for each dimension. The quantitative scores for the six dimensions are 6, 5, 4, 6, 5, and 5, respectively. The comprehensive assessment score is calculated using the comprehensive assessment score calculation formula, which is: Based on the comprehensive evaluation score, the preliminary evaluation conclusion is determined as follows: When At that time, the preliminary assessment conclusion was that community corrections were appropriate; when At that time, the preliminary assessment concluded that further investigation was needed; when At that time, the preliminary assessment concluded that community correction was not applicable; the comprehensive assessment score was 5 points, and the preliminary assessment concluded that supplementary investigation was required.

[0047] The preliminary assessment conclusions are broken down into four feature items: assessment basis, assessment logic, content of expression, and privacy information. These are then input into the LLM model and compared with community correction case texts, assessment standard texts, and typical document texts. The compliance matching degree of each feature item is calculated using the compliance verification matching degree calculation formula, which is as follows: ; Determine whether a feature is compliant by using compliance matching degree: when When it is deemed compliant, When it is determined that correction is needed, When the time is right, it is determined that the corresponding feature items need to be regenerated. In this assessment, the compliance matching degree of the three feature items of assessment basis, assessment logic and expression content are 0.82, 0.85 and 0.83 respectively, all of which meet the compliance standard. The compliance matching degree of privacy information feature item is 0.65, which is determined to be corrected. Then, the privacy information feature item content is automatically corrected by the LLM big model. After the correction, all four feature items meet the standard, and the initial draft of the survey assessment opinion is generated.

[0048] The initial draft of the investigation and assessment opinion is sent to the staff terminal for manual review. After review, the staff will suggest that relevant materials regarding the recent social activities of the person being assessed need to be supplemented. The LLM model will extract the change points based on the modification suggestions and generate verification prompts. After the verification is passed, the staff will notify the relevant units to supplement the materials. After the materials are supplemented, the initial draft of the investigation and assessment opinion is revised and a formal assessment opinion is generated. After being stamped with an electronic seal, it is sent to the client and the supervisory agency. The entire process of data is encrypted and archived in the community correction agency database.

[0049] In summary, community correction investigations and assessments were conducted for relevant applications. The LLM (Limited Ledger Model) was used to process the authorization letter materials and combine them with data collected from multiple information systems. A questionnaire was generated and integrated into a summary text. Based on the Prompt template library, preliminary conclusions were drawn that supplementary investigations were needed. One feature item in the compliance verification required correction. After automatic correction by the LLM model, a draft opinion was generated. The manual review stage required supplementary materials. After the materials were supplemented and improved, the draft was revised, a formal assessment opinion was generated, and sent to relevant parties. Simultaneously, encrypted archiving of all data throughout the process was completed.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart processing method for community corrections investigation and assessment based on an LLM large model, characterized in that, The specific steps of this method are as follows: Element extraction: Upon receiving the commission letter materials for investigation and evaluation, based on the legal element dictionary and through the LLM model, we first perform named entity recognition to extract core entities, then perform relationship extraction to establish logical connections between core entities, integrate them into core information, and standardize them into structured data to generate a list of core elements for investigation and evaluation. Information fusion: Basic data is collected through multi-source information systems. Based on the list of core elements of the survey and evaluation, a standardized questionnaire is generated using an LLM model and distributed and collected. Semantic encoding and association mapping are performed on multi-source heterogeneous data to generate a summary text of multi-source survey information. Intelligent assessment: The Prompt template library is retrieved, and the summary text of multi-source survey information is input into the LLM large model for multi-dimensional quantitative scoring to determine the preliminary assessment conclusion; Compliance verification: Input the preliminary assessment conclusions into the LLM model for compliance verification, automatically correct any issues found during the verification, and generate a draft of the investigation and assessment opinions; Review and Archiving: The initial draft of the investigation and evaluation opinions is pushed to the terminal for manual review. After the review is approved, a formal evaluation opinion is generated, which is then sent to relevant parties and the entire process data is encrypted and archived.

2. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 1, characterized in that, In the element extraction process, the LLM model is trained on a corpus from the community corrections domain, which includes community corrections case texts, assessment standard texts, and typical document texts. The LLM model includes named entity recognition, relation extraction, text generation, semantic encoding, logical reasoning, and compliance verification functions. Named entity recognition is performed by segmenting the investigation and assessment entrustment letter text into sentences and words, matching the segmentation results with entity entries in the legal element dictionary, and marking the successfully matched core entities. Relation extraction is performed by establishing logical connections between core entities and integrating these connections into core information, including basic information of the person being assessed, relevant case details, entrusting party information, execution location information, and investigation period requirements. After the core information is integrated, the LLM model's text generation function outputs a list of core elements for the investigation and assessment.

3. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 1, characterized in that, In the information fusion, the multi-source information system includes a community corrections agency database, a village and residents' committee information system, and a household registration system; the basic data collected includes residence information, family and social relationships, and community security status. Based on the core information in the core element list of the investigation and assessment, the text generation function of the LLM model is used to generate standardized questionnaires, including questionnaires for collecting victim opinions and questionnaires for collecting community public opinions. It supports two distribution and collection channels: online QR code and offline manual distribution. The multi-source heterogeneous data consists of basic data and collected questionnaire texts. The semantic encoding and association mapping operation is as follows: the format of the multi-source heterogeneous data is converted, and the various types of data are mapped into semantic vectors through the LLM model. After analyzing the cosine similarity between semantic vectors, they are associated and integrated to generate a summary text of multi-source investigation information.

4. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 1, characterized in that, In the intelligent assessment, the steps for constructing the Prompt template library are as follows: Combining the core needs of community correction investigation and assessment with the requirements of regional assessment work, six core assessment dimensions are determined, including residence stability, family support, consequences and impacts of the involved behavior, social danger, community acceptance, and suitability of the proposed prohibited items; assessment indicators and scoring rules are formulated for each dimension, and the assessment indicators and scoring rules of each dimension are transformed into prompt text that can be recognized by the LLM large model; the prompt texts of all dimensions are integrated to form the Prompt template library; the specific steps for multi-dimensional quantitative scoring are as follows: the summary text of multi-source investigation information is split into the six core assessment dimensions corresponding to the Prompt template library; the LLM large model extracts information, analyzes causal relationships, and determines the risk level of the split text according to the assessment indicators and scoring rules of each dimension, giving a quantitative score from 0 to 10, and calculates the comprehensive assessment score through the comprehensive assessment score calculation formula, and determines the preliminary assessment conclusion based on the comprehensive assessment score.

5. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 4, characterized in that, In the intelligent analysis, the comprehensive evaluation score is calculated using the following formula: ,in, To comprehensively evaluate the score, The core evaluation dimensions are set to 6. For the first The weighting coefficients of each evaluation dimension, For the first Quantitative score values ​​for each evaluation dimension; Based on the comprehensive evaluation score, the preliminary evaluation conclusion is determined as follows: When At that time, the preliminary assessment conclusion was that community corrections were appropriate; when At that time, the preliminary assessment concluded that further investigation was needed; when At that time, the preliminary assessment concluded that community corrections were not applicable.

6. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 1, characterized in that, The specific steps for compliance verification are as follows: the preliminary assessment conclusion is broken down into four feature items: assessment basis, assessment logic, expression content, and privacy information. These are then input into the LLM model and compared with community correction case texts, assessment standard texts, and typical document texts. The compliance matching degree of each feature item is calculated using the compliance verification matching degree calculation formula to determine whether the feature item is compliant. Non-compliant feature items are automatically corrected. After all four feature items meet the standards, a draft of the investigation and assessment opinion is generated.

7. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 6, characterized in that, In the compliance verification, the formula for calculating the compliance verification matching degree is: ,in, For the compliance matching degree of feature items, This represents the number of feature terms, with a value of 4. For the first Evaluation content vector in each feature item With standard text vectors cosine similarity, Using the LLM large model to analyze the first The evaluation content of each feature term is semantically encoded and generated. Semantic encoding of corresponding feature items in the corpus of the community correction domain is generated using an LLM large model; The compliance matching score determines whether a feature is compliant: when When it is deemed compliant, When it is determined that correction is needed, If so, it is determined that the corresponding feature item content needs to be regenerated.

8. The intelligent processing method for community corrections investigation and assessment based on an LLM large model according to claim 1, characterized in that, During the review and archiving process, the initial draft of the investigation and assessment opinions is modified and annotated during the manual review stage. The LLM model extracts the change points for the modifications, compares the change points with the assessment standard text to generate verification prompts, and proceeds to the formal assessment opinion generation stage after the verification is passed. After the formal assessment opinion is generated, it is affixed with an electronic seal and sent to relevant parties, including the client and the supervisory agency. All data in the process is encrypted and archived to the community correction agency database.

9. A community corrections investigation and assessment intelligent processing system based on an LLM large model, the system being applicable to the community corrections investigation and assessment intelligent processing method based on an LLM large model as described in any one of claims 1-8, characterized in that, The system includes: Element Extraction Module: Receives the investigation and assessment commission letter and accompanying materials. Based on the preset legal element dictionary, it uses the LLM model to first perform named entity recognition to extract core entities, then performs relationship extraction to establish logical connections between core entities, integrates them into core information, and standardizes them into structured data to generate a list of core elements for investigation and assessment. Information fusion module: Collects basic data through multi-source information systems, generates standardized questionnaires based on the list of core elements of the survey and evaluation using the LLM model and distributes and collects them, performs semantic encoding and association mapping on multi-source heterogeneous data, and generates a summary text of multi-source survey information; Intelligent assessment module: retrieves the Prompt template library, inputs the summary text of multi-source survey information into the LLM large model for multi-dimensional quantitative scoring, and determines the preliminary assessment conclusion; Compliance verification module: Input the preliminary assessment conclusions into the LLM large model for compliance verification, automatically correct the problems found in the verification, and generate a draft of the investigation and assessment opinions; Review and archiving module: The initial draft of the investigation and evaluation opinions is pushed to the terminal for manual review. After the review is approved, a formal evaluation opinion is generated, which is pushed to relevant parties and the entire process data is encrypted and archived.

Images