A disciplinary inspection case data reasoning method and device, electronic equipment and storage medium

Abstract

This invention discloses a reasoning method, apparatus, electronic device, and storage medium for disciplinary inspection and supervision case data, belonging to the field of natural language processing technology. The method includes: acquiring multimodal case data and query requests; performing structured parsing on the data and combining it with semantic matching from a knowledge base to extract candidate information; determining the structure type through dynamic routing, reconstructing structured knowledge using a generative model, and generating descriptive information; semantically decomposing the query to obtain sub-problems and performing coverage verification; and extracting precise knowledge fragments based on the verification results to generate reasoning documents. Therefore, by implementing this invention, the problem of existing technologies where the underlying structural logic of multimodal evidence is easily fragmented during parsing and retrieval, leading to an incomplete chain of documentary evidence and the potential for factual illusions can be solved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a reasoning method, apparatus, electronic device, and storage medium for data in disciplinary inspection and supervision cases. Background Technology

[0002] Disciplinary inspection and supervision work involves massive amounts of case files from diverse sources. These materials have complex data modalities and are often interwoven with strict legal and regulatory application logic. Intelligent reasoning based on case data can effectively assist investigators in quickly sorting through evidence, thereby automatically constructing evidence chains and generating accurate qualitative review documents. This not only improves the efficiency and accuracy of case clue sorting and reduces the risk of oversights in manual review, but also serves as a crucial support for promoting the standardization and intelligent transformation of investigation and review work.

[0003] However, current intelligent reasoning technologies generally face challenges when dealing with complex evidentiary logic, including the tendency for the reasoning process to deviate from objective facts and the lack of rigorous document generation logic. The main reason for this is that existing retrieval-enhanced reasoning solutions often employ a fixed, one-way generation model. On the one hand, they cannot dynamically predict the structure and type of fragmented, heterogeneous information retrieved based on different query requests, making it difficult for underlying evidence to form a clear and coherent knowledge system. On the other hand, existing solutions lack a robust error-prevention loop when generating the final document, failing to break down macro-level query requests downwards and quantitatively verify their consistency and coverage with the underlying reconstructed knowledge. Once the model experiences misunderstandings or omissions of evidence during processing, the system cannot automatically detect and intercept them, ultimately resulting in logical gaps or false statements in the generated reasoning document, failing to meet the high requirements of the rigor of the evidence chain in disciplinary inspection and supervision operations. Summary of the Invention

[0004] This invention provides a reasoning method, apparatus, electronic device, and storage medium for data in discipline inspection and supervision cases. It can solve the problems in the prior art where the underlying structural logic of multimodal evidence is easily fragmented during parsing and retrieval, and large models lack a rigorous verification and error prevention closed loop during reasoning generation, resulting in an unreliable chain of documentary evidence and the potential for factual illusions.

[0005] An embodiment of the present invention provides a reasoning method for data in disciplinary inspection and supervision cases, comprising: Acquire multimodal case data and natural language query requests for target disciplinary inspection and supervision cases; The multimodal case data is processed through structured parsing to generate an initial structured data file; Based on natural language query requests, semantic matching is performed on the initial structured data file and the preset legal knowledge base to extract a set of candidate information fragments; The natural language query request and the candidate information fragment set are input into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set; Based on the target structure type, the candidate information fragment set is reconstructed using a pre-defined generative language model combined with a pre-defined first prompt template, and the target structured knowledge is output. Based on the natural language query request and the target structured knowledge, the features are summarized using a generative language model combined with a pre-defined second prompt template, and the descriptive information of the target structured knowledge is output. Based on the natural language query request and description information, semantic decomposition is performed using a generative language model combined with a preset third prompt template to output multiple sub-questions; the multiple sub-questions are then subjected to coverage verification with the target structured knowledge, and a quantitative verification score is calculated. Based on the quantitative verification score, multiple sub-questions are mapped and matched with the target structured knowledge, the corresponding precise knowledge fragments are extracted, and the precise knowledge fragments are combined to generate reasoning documents for natural language query requests.

[0006] Furthermore, the multimodal case data undergoes structured parsing processing to generate an initial structured data file, including: Based on the pre-defined intermediate representation structure of case documents, multimodal case data is parsed to extract document metadata, the smallest semantic unit, and the original layout relationships. The intermediate representation structure of case documents includes a document metadata layer, a content block layer, a structural information layer, and a vectorized input layer. The document metadata, the smallest semantic unit, and the original layout relationship are written into the document metadata layer, content block layer, and structural information layer of the intermediate representation structure of the case document, respectively. For each smallest semantic unit in the content block layer, a standardized text for vectorized retrieval is generated based on the current smallest semantic unit, and the standardized text is written into the vectorized input layer of the intermediate representation structure of the case document. The intermediate representation structure of the case document after data writing is completed is used as the initial structured data file.

[0007] Furthermore, based on natural language query requests, semantic matching is performed on the initial structured data file and the pre-defined legal and regulatory knowledge base to extract a set of candidate information fragments, including: Feature extraction is performed on natural language query requests to obtain query features; Input the query features into a preset embedding model so that the embedding model generates a query semantic vector based on the query features; Extract standardized text from the vectorized input layer of the initial structured data file; Extract the contents of each article and case from the pre-set legal knowledge base; Each standardized text is input into the embedding model so that the embedding model generates a corresponding first feature vector based on each standardized text. Each clause and case content is input into the embedding model so that the embedding model generates a corresponding second feature vector based on each clause and case content; Calculate the similarity between the query semantic vector and each first feature vector, and generate the corresponding first semantic similarity; Calculate the similarity between the query semantic vector and each second feature vector, and generate the corresponding second semantic similarity. The standardized texts are sorted based on the first semantic similarity to obtain the first ranking sequence, and the standardized texts that meet the preset first condition are extracted from the first ranking sequence as the first candidate information fragments. The content of each article and case is sorted based on the second semantic similarity to obtain the second ranking sequence, and the content of articles and cases that meet the preset second conditions are extracted from the second ranking sequence as the second candidate information fragments. The first candidate information fragment and the second candidate information fragment are fused together to generate a set of candidate information fragments.

[0008] Furthermore, the natural language query request and the candidate information fragment set are input into a preset dynamic routing model to generate the target structure type of the candidate information fragment set, including: The natural language query request and the candidate information fragment set are fused to generate joint input features; The joint input features are input into a preset dynamic routing model, so that the dynamic routing model calculates and generates the preference probabilities corresponding to the joint input features under multiple preset candidate structure types based on the joint input features; wherein, the multiple candidate structure types include tables, graphs, algorithm flows, directories, and text blocks; Compare the preference probabilities corresponding to each candidate structure type, and determine the candidate structure type corresponding to the highest preference probability as the target structure type.

[0009] Furthermore, based on the target structure type, the candidate information fragment set is reconstructed using a pre-defined generative language model combined with a pre-defined first prompt template, outputting the target structured knowledge; based on the natural language query request and the target structured knowledge, features are summarized using a generative language model combined with a pre-defined second prompt template, outputting descriptive information of the target structured knowledge, including: The natural language query request, the target structure type, and the set of candidate information fragments are concatenated into a preset first prompt template to generate a first prompt instruction; The first prompt instruction is input into the preset generative language model so that the generative language model extracts text content that matches the natural language query request from the candidate information fragment set and reconstructs it according to the target structure type to generate target structured knowledge. The natural language query request and the target structured knowledge are concatenated into a preset second prompt template to generate a second prompt instruction; The second prompt instruction is input into the generative language model so that the generative language model can extract the target structured knowledge, generate a summary text containing the topic, key fields and the relationship with the natural language query request, and use the summary text as the descriptive information of the target structured knowledge.

[0010] Furthermore, based on the natural language query request and description information, semantic decomposition is performed using a generative language model combined with a pre-defined third-party prompt template, outputting multiple sub-questions. These sub-questions are then subjected to coverage verification against the target structured knowledge, and a quantitative verification score is calculated, including: The natural language query request and description information are concatenated into a preset third prompt template to generate a third prompt instruction; The third prompt instruction is input into the generative language model so that the generative language model can decompose the natural language query request into multiple logically independent sub-problems; Extract the first entity information from the target structured knowledge; For each subproblem, extract the second entity information from the current subproblem; Calculate the proportion of the second entity information corresponding to the current sub-problem in the first entity information to obtain the semantic coverage value of the current sub-problem. After calculating the semantic coverage values ​​for all sub-problems, the semantic coverage values ​​for each sub-problem are comprehensively calculated to obtain the quantitative verification score.

[0011] Furthermore, based on the quantitative verification score, multiple sub-questions are mapped and matched with the target structured knowledge, the corresponding precise knowledge fragments are extracted, and the precise knowledge fragments are combined to generate reasoning documents for natural language query requests, including: For each sub-problem, the current sub-problem is mapped and matched with the target structured knowledge, and the precise knowledge fragment corresponding to the current sub-problem is extracted; The quantitative verification score, natural language query request, and precise knowledge fragments corresponding to each sub-question are concatenated into the preset fourth prompt template to generate the fourth prompt instruction; The fourth prompt instruction is input into the generative language model, so that the generative language model can perform logical reasoning based on the quantized verification score, natural language query request, and the precise knowledge fragments corresponding to each sub-question, and generate a reasoning document.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] One embodiment of the present invention provides a reasoning device for disciplinary inspection and supervision case data, comprising: a data parsing module, a semantic matching module, a dynamic routing module, a knowledge reconstruction module, and a reasoning generation module; The data parsing module is used to acquire multimodal case data and natural language query requests for target discipline inspection and supervision cases; perform structured parsing processing on the multimodal case data to generate an initial structured data file; The semantic matching module is used to perform semantic matching on the initial structured data file and the preset legal knowledge base based on natural language query requests, and extract a set of candidate information fragments. The dynamic routing module is used to input natural language query requests and candidate information fragment sets into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set. The knowledge reconstruction module is used to reconstruct the candidate information fragment set according to the target structure type, based on a preset generative language model and a preset first prompt template, and output the target structured knowledge; based on the natural language query request and the target structured knowledge, it summarizes the features based on a generative language model and a preset second prompt template, and outputs the descriptive information of the target structured knowledge. The reasoning generation module is used to perform semantic decomposition processing based on a generative language model and a preset third prompt template according to the natural language query request and description information, and output multiple sub-questions; perform coverage verification processing on the multiple sub-questions and target structured knowledge, and calculate the quantitative verification score; based on the quantitative verification score, map and match the multiple sub-questions with the target structured knowledge respectively, extract the corresponding precise knowledge fragments, and generate a reasoning document for the natural language query request in combination with the precise knowledge fragments.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the reasoning method for disciplinary inspection and supervision case data as described in any of the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the reasoning method for disciplinary inspection and supervision case data as described in any of the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a reasoning method, apparatus, electronic device, and storage medium for disciplinary inspection and supervision case data. The method acquires multimodal case data and natural language query requests for target disciplinary inspection and supervision cases; performs structured parsing on the multimodal case data to generate an initial structured data file; based on the natural language query request, semantically matches the initial structured data file with a preset legal and regulatory knowledge base to extract a set of candidate information fragments; inputs the natural language query request and the set of candidate information fragments into a dynamic routing selection model to determine the target structure type of the candidate information fragments; based on the target structure type, reconstructs the candidate information fragments using a generative language model combined with a first prompt template to generate target structured knowledge, and summarizes the features of the target structured knowledge using a second prompt template to obtain descriptive information; based on the natural language query request and descriptive information, semantically decomposes the data using a third prompt template to generate multiple sub-questions; performs coverage verification on the multiple sub-questions and the target structured knowledge to obtain a quantitative verification score; based on the quantitative verification score, maps and matches the sub-questions with the target structured knowledge to extract precise knowledge fragments, and generates a reasoning document for the natural language query request accordingly.

[0019] This invention introduces a dynamic routing model to generate target structure types for candidate information fragments, and reconstructs the target structured knowledge by combining a generative language model with a first prompt template, thereby transforming underlying heterogeneous evidence into a well-organized knowledge system. Furthermore, addressing the lack of error-proofing loops and the susceptibility to false outputs in the reasoning generation process, this invention further decomposes the macro-query request semantics into multiple sub-problems, and performs consistency and coverage checks on these sub-problems with the reconstructed target structured knowledge. Finally, based on the quantitative verification score, precise knowledge fragments are extracted to generate reasoning documents, effectively intercepting logical gaps and evidence omissions, and improving the rigor and reliability of the evidence chain in disciplinary inspection and supervision documents. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a reasoning method for disciplinary inspection and supervision case data provided in an embodiment of the present invention.

[0021] Figure 2This is a schematic diagram of the structure of a reasoning device for disciplinary inspection and supervision case data provided in an embodiment of the present invention. Detailed Implementation

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

[0023] like Figure 1 As shown, to address the problems in existing technologies where the underlying structural logic of multimodal evidence is easily fragmented during parsing and retrieval, and where large models lack a rigorous verification and error-proofing loop during reasoning generation, leading to loose evidence chains and the potential for factual illusions, an embodiment of this invention provides a reasoning method for disciplinary inspection and supervision case data, comprising at least the following steps: Step S1: Obtain multimodal case data and natural language query requests for the target discipline inspection and supervision cases; perform structured parsing processing on the multimodal case data to generate an initial structured data file; In a preferred embodiment, the multimodal case data undergoes structured parsing processing to generate an initial structured data file, including: Based on the pre-defined intermediate representation structure of case documents, multimodal case data is parsed to extract document metadata, the smallest semantic unit, and the original layout relationships. The intermediate representation structure of case documents includes a document metadata layer, a content block layer, a structural information layer, and a vectorized input layer. The document metadata, the smallest semantic unit, and the original layout relationship are written into the document metadata layer, content block layer, and structural information layer of the intermediate representation structure of the case document, respectively. For each smallest semantic unit in the content block layer, a standardized text for vectorized retrieval is generated based on the current smallest semantic unit, and the standardized text is written into the vectorized input layer of the intermediate representation structure of the case document. The intermediate representation structure of the case document after data writing is completed is used as the initial structured data file.

[0024] Specifically, the steps involve acquiring multimodal case data and natural language query requests for the target disciplinary inspection and supervision cases. The multimodal case data covers various heterogeneous file formats within disciplinary inspection and supervision business scenarios, specifically including text documents, tabular documents, and multimedia audiovisual files. The text documents include portable document format files and word processing software documents; the tabular documents include financial statements and spreadsheet files; and the multimedia audiovisual files include audio recordings and surveillance videos. The natural language query requests are the case search intent or evidence collection needs for clues of violations input by the case handlers.

[0025] To overcome the feature barriers caused by heterogeneous evidence, the multimodal case data needs to undergo structured parsing to generate an initial structured data file. Specifically, the multimodal case data is parsed according to a pre-defined intermediate representation structure for case documents to extract document metadata, the smallest semantic units, and the original layout relationships. This intermediate representation structure is a hierarchical logical model constructed to unify heterogeneous document formats. The hierarchical logical model is strictly divided from top to bottom into a document metadata layer, a content block layer, a structural information layer, and a vectorized input layer.

[0026] In the specific parsing stage, the extracted document metadata, the smallest semantic unit, and the original layout relationship are respectively written into the document metadata layer, content block layer, and structural information layer of the intermediate representation structure of the case document. Specifically, the case number, source format, upload time, and page number information are used as the document metadata and uniformly aggregated into the document metadata layer. For text documents, the smallest semantic unit is extracted according to the logical hierarchy of title, paragraph, and attachment description; for multimedia audiovisual files, a speech recognition algorithm is used to extract transcribed text fragments as the smallest semantic unit, and then all semantic segments are uniformly aggregated into the content block layer. Simultaneously, the original hierarchical nesting order and physical location arrangement features of the multimodal case data are preserved as the original layout relationship and solidified into the structural information layer.

[0027] Specifically, when the multimodal case data includes the tabular document, to avoid losing semantic logic by directly flattening the two-dimensional table into long text, the unique row and column structure relationships of the tabular document must be fully preserved in the structural information layer. This means strictly mapping the worksheet name, header fields, row indexes, and the hierarchical relationships of cell content. Subsequently, for each smallest semantic unit in the content block layer, standardized text is specifically generated for subsequent vectorized retrieval based on the currently traversed smallest semantic unit, and this standardized text is synchronously written into the vectorized input layer of the intermediate representation structure of the case document. If the traversed object is the row data of the tabular document, then each row field of the tabular document is extracted, and multiple discrete cell features are fused and concatenated into a coherent declarative sentence as row-level semantic expanded text. This row-level semantic expanded text is then used as the standardized text, thus balancing the structural fidelity requirements of the two-dimensional table with the embedding effect of high-dimensional feature vector transformation.

[0028] Finally, the intermediate representation structure of the case document, composed of all levels of data writing, is used as the initial structured data file. By performing the above multimodal data parsing and hierarchical aggregation steps, the semantic silos caused by heterogeneous evidence in disciplinary inspection and supervision case files can be effectively overcome. The disorganized original case materials are transformed into a set of basic elements with clear hierarchy and rich semantics, laying a reliable data foundation for subsequent high-precision feature vector matching and generative language model logical reasoning.

[0029] Step S2: Based on the natural language query request, perform semantic matching on the initial structured data file and the preset legal knowledge base to extract a set of candidate information fragments; In a preferred embodiment, based on a natural language query request, semantic matching is performed on an initial structured data file and a preset legal and regulatory knowledge base to extract a set of candidate information fragments, including: Feature extraction is performed on natural language query requests to obtain query features; Input the query features into a preset embedding model so that the embedding model generates a query semantic vector based on the query features; Extract standardized text from the vectorized input layer of the initial structured data file; Extract the contents of each article and case from the pre-set legal knowledge base; Each standardized text is input into the embedding model so that the embedding model generates a corresponding first feature vector based on each standardized text. Each clause and case content is input into the embedding model so that the embedding model generates a corresponding second feature vector based on each clause and case content; Calculate the similarity between the query semantic vector and each first feature vector, and generate the corresponding first semantic similarity; Calculate the similarity between the query semantic vector and each second feature vector, and generate the corresponding second semantic similarity. The standardized texts are sorted based on the first semantic similarity to obtain the first ranking sequence, and the standardized texts that meet the preset first condition are extracted from the first ranking sequence as the first candidate information fragments. The content of each article and case is sorted based on the second semantic similarity to obtain the second ranking sequence, and the content of articles and cases that meet the preset second conditions are extracted from the second ranking sequence as the second candidate information fragments. The first candidate information fragment and the second candidate information fragment are fused together to generate a set of candidate information fragments.

[0030] Specifically, after obtaining the initial structured data file, the process can proceed to the core step of semantic matching between the initial structured data file and a preset legal knowledge base based on the natural language query request, thereby extracting a set of candidate information fragments.

[0031] In the specific feature transformation operation, deep feature extraction is first performed on the natural language query request to obtain the corresponding query features. Then, the query features are input into a preset embedding model, enabling the embedding model to generate a high-dimensional query semantic vector based on the query features. The embedding model employs a bidirectional encoder representation model or a sentence-level bidirectional encoder representation model, which can accurately capture the complex contextual semantic relationships within the natural language query request.

[0032] While constructing query-side features, the matching objects from the evidence and knowledge sides are simultaneously acquired. Specifically, the previously structured standardized texts are extracted from the vectorized input layer of the initial structured data file; simultaneously, relevant articles and case content are extracted from the preset legal and regulatory knowledge base. This preset legal and regulatory knowledge base is a structured data center pre-built using an automated web crawling mechanism and a timed incremental indexing mechanism. It is frequently updated and stores massive amounts of real case data, compliance policy documents, and legal articles related to discipline inspection and supervision.

[0033] To project information from different sources into a unified metric space, the extracted standardized texts are batch-input into the embedding model, enabling the embedding model to generate corresponding high-dimensional first feature vectors based on each standardized text. Using the same feature space mapping logic, each clause and case content is input into the embedding model together, enabling the embedding model to generate corresponding second feature vectors based on each clause and case content.

[0034] After completing the vectorization representation of all text modal data, the similarity comparison process is immediately initiated. The spatial vector distance between the query semantic vector and each of the first feature vectors is calculated using a cosine similarity algorithm or an inner product search algorithm, and a corresponding first semantic similarity is generated based on the calculated distance. Following the same metric, the vector similarity between the query semantic vector and each of the second feature vectors is calculated, generating a corresponding second semantic similarity.

[0035] To accurately extract core evidence and regulations that best match the investigators' search intent, the standardized texts are sorted in descending order based on the numerical value of the first semantic similarity to obtain a first ranking sequence. Standardized texts meeting a preset first condition are then extracted from this first ranking sequence as first candidate information fragments. The preset first condition is configured to either have a similarity value exceeding a preset similarity threshold or a ranking order falling within a preset recall range. Similarly, the provisions and case content are sorted based on the numerical value of the second semantic similarity to obtain a second ranking sequence. Provisions and case content meeting a preset second condition are then extracted from this second ranking sequence as second candidate information fragments. The determination logic for the preset second condition is strictly aligned with the preset first condition.

[0036] Finally, a hybrid retrieval strategy is employed to deeply fuse the extracted first and second candidate information fragments, generating a final set of candidate information fragments. The specific deep fusion process includes cross-referencing and redundancy removal of the first and second candidate information fragments, eliminating overlapping fragments with duplicate content, and globally reordering the deduplicated information fragments in descending order based on their respective first and second semantic similarity scores. This ensures that the final set of candidate information fragments possesses both high information density and factual comprehensiveness when input into subsequent models.

[0037] By performing the aforementioned semantic matching and dual-path high-dimensional vector retrieval steps, the data barrier between case file factual evidence and macro-level legal provisions is effectively broken down, realizing the dynamic integration of fragmented case clues and standard legal guidelines. This provides the large language model with an absolutely accurate and context-enhanced information source that takes into account both legal principles and facts for subsequent rigorous and compliant intelligent reasoning.

[0038] Step S3: Input the natural language query request and the candidate information fragment set into the preset dynamic routing selection model to generate the target structure type of the candidate information fragment set; In a preferred embodiment, a natural language query request and a set of candidate information fragments are input into a preset dynamic routing model to generate a target structure type for the set of candidate information fragments, including: The natural language query request and the candidate information fragment set are fused to generate joint input features; The joint input features are input into a preset dynamic routing model, so that the dynamic routing model calculates and generates the preference probabilities corresponding to the joint input features under multiple preset candidate structure types based on the joint input features; wherein, the multiple candidate structure types include tables, graphs, algorithm flows, directories, and text blocks; Compare the preference probabilities corresponding to each candidate structure type, and determine the candidate structure type corresponding to the highest preference probability as the target structure type.

[0039] Specifically, after obtaining evidence fragments retrieved from multiple sources, the core reasoning control process is executed to input the natural language query request and the candidate information fragment set into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set.

[0040] First, deep semantic fusion processing is performed on the natural language query request and the candidate information fragment set. Natural language representation features and heterogeneous evidence representation features are aligned and concatenated in the same high-dimensional computational space to generate joint input features that characterize query intent and contextual association. Then, the joint input features are input into a pre-set dynamic routing model. This dynamic routing model is pre-trained using the direct preference optimization algorithm. The direct preference optimization algorithm aims to bypass the traditional explicit reward modeling steps and directly transform the human judgment preference learning task in discipline inspection and supervision operations into a single-policy network optimization objective.

[0041] When constructing the dynamic routing model, the forward propagation logic for structure type selection follows the following mathematical expression: And it satisfies the feature set constraint: In the formula, Indicates the target structure type of the final output. This represents a dynamic routing model. This represents the input natural language query request. This represents the set of candidate information fragments extracted through semantic matching. Represents the first element in the set. One candidate information fragment, This indicates the total number of candidate information fragments.

[0042] The core mechanism of the direct preference optimization algorithm lies in using a mathematical model to represent and fit the degree of preference for the optimal structured form judged by business experts. The preference probability calculation formula based on the Bradley model is configured as follows: In the formula, This represents the probability of a structure type that tends to win rather than lose, given the joint input features. This indicates the winning structural types that human experts believe are more conducive to logical expression and subsequent reasoning during the annotation process. This indicates a failed structure type that does not meet business standards. This represents the joint input features obtained after feature fusion processing. This represents the ideal human reward function used to accurately evaluate the matching degree between input features and candidate structure types. This represents a non-linear activation function.

[0043] The nonlinear activation function is responsible for mapping the reward difference to a standard probability interval between zero and one. Its specific function expression is defined as follows: In the formula, This represents the difference in reward values ​​output by the winning and losing structure types under the ideal human reward function.

[0044] Based on the aforementioned mathematical foundations of preference representation, the core loss function used by the dynamic routing model during the parameter training phase is constructed as follows: In the formula, This represents the overall loss function of the direct preference optimization algorithm. This represents the network weight parameters of the dynamic routing model to be optimized and updated. This represents the mathematical expectation operation based on probability distribution calculation. This indicates a high-quality preference dataset pre-constructed based on real case review standards. The temperature hyperparameter represents the degree to which the optimization strategy deviates from the initial baseline model. This indicates the dynamic routing network architecture currently undergoing iterative training. This represents an initial reference selection model with all parameters frozen. By minimizing the loss function, the dynamic routing model can accurately learn and output a probability distribution that conforms to compliance verification logic.

[0045] In the actual reasoning phase, after acquiring the joint input features, the dynamic routing model strictly calculates and generates the preference probability values ​​corresponding to the joint input features under multiple preset candidate structure types based on the learned parameter distribution. These multiple candidate structure types are explicitly limited to five forms: tables, graphs, algorithm flows, directories, and text blocks. The table type specifically corresponds to the layout presentation when performing tasks such as amount verification, abnormal fund comparison, and factual element statistics. The graph type corresponds to the topological network presentation when sorting out complex personnel relationships and multi-level fund flow chains. The algorithm flow type corresponds to the forward logic presentation when reconstructing engineering approval nodes and institutional compliance decision-making tasks. The directory type corresponds to the hierarchical framework presentation when handling long case summaries and legal basis hierarchical division tasks. The text block type corresponds to the concise paragraph presentation when answering single factual questions and extracting simple regulations.

[0046] After completing the probability calculations for all candidate forms, the system comprehensively compares the magnitude of the preference probabilities output by each candidate structural type. The system selects the candidate structural type with the highest preference probability and determines it as the target structural type for subsequent knowledge reconstruction. By executing joint feature fusion and dynamic preference routing extraction steps, the system completely breaks through the rigid unidirectional text output limitation of traditional retrieval enhancement models when facing multimodal evidence. This allows fragmented case information to be adaptively matched to the organizational form with the highest explanatory power based strictly on the query dimensions of the investigators, thus building a solid cognitive framework foundation for subsequently improving the rigor of legal logic deduction and the intuitiveness of the evidence chain.

[0047] Step S4: Based on the target structure type, reconstruct the candidate information fragment set using a preset generative language model and a preset first prompt template to output the target structured knowledge; based on the natural language query request and the target structured knowledge, summarize the features using a generative language model and a preset second prompt template to output the descriptive information of the target structured knowledge. In a preferred embodiment, based on the target structure type, the candidate information fragment set is reconstructed using a preset generative language model combined with a preset first prompt template to output target structured knowledge; based on the natural language query request and the target structured knowledge, features are summarized using a generative language model combined with a preset second prompt template to output descriptive information of the target structured knowledge, including: The natural language query request, the target structure type, and the set of candidate information fragments are concatenated into a preset first prompt template to generate a first prompt instruction; The first prompt instruction is input into the preset generative language model so that the generative language model extracts text content that matches the natural language query request from the candidate information fragment set and reconstructs it according to the target structure type to generate target structured knowledge. The natural language query request and the target structured knowledge are concatenated into a preset second prompt template to generate a second prompt instruction; The second prompt instruction is input into the generative language model so that the generative language model can extract the target structured knowledge, generate a summary text containing the topic, key fields and the relationship with the natural language query request, and use the summary text as the descriptive information of the target structured knowledge.

[0048] Specifically, after determining the target structure type, the process proceeds to the stage of fusing and reconstructing the scattered set of candidate information fragments. Specifically, the natural language query request, the target structure type, and the set of candidate information fragments are simultaneously concatenated into a preset first prompt template to generate a first prompt instruction used to control the output format of the generative language model. The preset first prompt template is a pre-configured set of structured instructions, embedding mandatory text conversion rules that match various candidate structure forms.

[0049] The first prompt instruction is input into the preset generative language model. Upon receiving the first prompt instruction, the generative language model, leveraging its deep semantic understanding capabilities, accurately extracts text content matching the natural language query request from the candidate information fragment set and reconstructs features strictly according to the target structure type. If the target structure type is a table, the generative language model extracts information suitable for statistics, comparison, or time sorting, and reconstructs it into a structured table containing explicit column names and corresponding fact records. If the target structure type is a graph, the generative language model identifies key entities and the associated actions between entities, and reconstructs it into a graph data set containing entities and their relationships. If the target structure type is an algorithm flow, the generative language model extracts sequential, conditional, or decision-making processing actions, and reconstructs it into a step-by-step logical flow with input conditions and output results. If the target structure type is a table of contents, the generative language model hierarchically organizes long documents, and reconstructs them into a framework containing multi-level headings and corresponding summary points. If the target structure type is a text block, the generative language model extracts the most relevant factual content and reconstructs it into concise paragraphs centered around the core facts.

[0050] Through the aforementioned multi-dimensional reconstruction processes, the generative language model output can intuitively present the target structured knowledge of the core evidence logic. The mathematical process of disparate knowledge reconstruction is expressed by a mapping equation: In the formula, Representing target structured knowledge, The mapping function representing the generative language model, This indicates the default first prompt template.

[0051] To facilitate subsequent logical decomposition and precise verification, the generated structured content needs to be condensed and summarized. The natural language query request and the target structured knowledge are concatenated together into a preset second prompt template to generate a second prompt instruction. This preset second prompt template is also a pre-configured set of structured instructions, internally containing content feature constraints for element summarization and abstract generation.

[0052] The second prompt instruction is input into the generative language model, enabling it to perform in-depth feature summarization of the target structured knowledge. The generative language model extracts core topics, key fields, and deep relationships between the target structured knowledge and the natural language query request from the target structured knowledge, and integrates the extracted core elements to generate a summary text. Subsequently, the generated summary text is directly output as descriptive information for the target structured knowledge.

[0053] The mathematical process of generating descriptive information follows the logical expression below: In the above logical expression, Descriptive information representing the structured knowledge of the target. This indicates the preset second prompt template.

[0054] By performing the above reconstruction and feature summarization steps, the fragmented and unstructured heterogeneous evidence in the original case file is completely transformed into a structured knowledge asset that highly conforms to the layout logic of specific discipline inspection and supervision tasks. At the same time, a descriptive summary for summarizing the overall picture of the evidence is constructed, laying a highly interpretable foundation for subsequent accurate semantic decomposition and knowledge tracing of complex cases.

[0055] Step S5: Based on the natural language query request and description information, semantic decomposition processing is performed using a generative language model combined with a preset third prompt template to output multiple sub-questions; the multiple sub-questions are subjected to coverage verification processing with the target structured knowledge, and a quantitative verification score is calculated; based on the quantitative verification score, the multiple sub-questions are mapped and matched with the target structured knowledge respectively, the corresponding precise knowledge fragments are extracted, and the precise knowledge fragments are combined to generate a reasoning document for the natural language query request.

[0056] In a preferred embodiment, based on the natural language query request and description information, semantic decomposition is performed using a generative language model combined with a preset third prompt template to output multiple sub-questions; the multiple sub-questions are then subjected to coverage verification with the target structured knowledge, and a quantitative verification score is calculated, including: The natural language query request and description information are concatenated into a preset third prompt template to generate a third prompt instruction; The third prompt instruction is input into the generative language model so that the generative language model can decompose the natural language query request into multiple logically independent sub-problems; Extract the first entity information from the target structured knowledge; For each subproblem, extract the second entity information from the current subproblem; Calculate the proportion of the second entity information corresponding to the current sub-problem in the first entity information to obtain the semantic coverage value of the current sub-problem. After calculating the semantic coverage values ​​for all sub-problems, the semantic coverage values ​​for each sub-problem are comprehensively calculated to obtain the quantitative verification score.

[0057] In a preferred embodiment, based on the quantized verification score, multiple sub-questions are mapped and matched with target structured knowledge to extract corresponding precise knowledge fragments, and the precise knowledge fragments are combined to generate reasoning documents for natural language query requests, including: For each sub-problem, the current sub-problem is mapped and matched with the target structured knowledge, and the precise knowledge fragment corresponding to the current sub-problem is extracted; The quantitative verification score, natural language query request, and precise knowledge fragments corresponding to each sub-question are concatenated into the preset fourth prompt template to generate the fourth prompt instruction; The fourth prompt instruction is input into the generative language model, so that the generative language model can perform logical reasoning based on the quantized verification score, natural language query request, and the precise knowledge fragments corresponding to each sub-question, and generate a reasoning document.

[0058] Specifically, after obtaining the descriptive information of the target structured knowledge, the process moves to the core stage of decomposing the macro-intention and performing logical verification. In practice, the natural language query request and the descriptive information are concatenated into a pre-defined third prompt template to generate a third prompt instruction. This pre-defined third prompt template sets specific task decomposition constraints, guiding subsequent models to focus on the overall evidence outlined by the descriptive information. Subsequently, the third prompt instruction is input into the generative language model, enabling it to decompose the macro-level natural language query request into multiple logically independent sub-problems. The mathematical process of semantic decomposition follows the following logical expression: In the formula, The first one obtained from disassembly The issue of height. This represents the total number of subproblems obtained from the decomposition. This represents a semantic decomposition mapping function constructed based on a preset third prompt template and a generative language model.

[0059] To prevent generative language models from developing logical illusions detached from reality during decomposition, rigorous coverage verification of multiple sub-problems with the target structured knowledge is necessary to calculate a quantitative verification score. In the specific verification process, the first entity information embedded within the target structured knowledge is extracted. This first entity information encompasses the names of the individuals involved in the case, the amount of funds transferred, and key time points. Simultaneously, for each sub-problem, the second entity information relevant to the currently being traversed sub-problem is extracted. Next, the coverage ratio of the second entity information corresponding to the current sub-problem within the first entity information is calculated, and this calculated coverage ratio is directly used as the semantic coverage rate for the current sub-problem. After calculating the semantic coverage rates for all sub-problems, a comprehensive weighted calculation is performed on the semantic coverage rates for each sub-problem to obtain the final quantitative verification score.

[0060] Furthermore, if the calculated quantization verification score does not meet the preset score condition (i.e., it is lower than the preset verification threshold), it proves that there is a factual omission or logical deviation in the current semantic decomposition or the description information. At this time, the system triggers the feedback correction process, sends back the feedback signal of verification failure, and drives the generative language model to reconstruct or supplement the target structured knowledge until the recalculated quantization verification score meets the preset score condition, thereby forming a rigorous error prevention closed-loop verification mechanism.

[0061] If the quantitative verification score meets the preset scoring conditions, proving that the semantic decomposition has not deviated from the underlying evidentiary facts, then subsequent precise tracing and generation actions can be triggered based on the quantitative verification score. For each sub-problem, the current sub-problem is mapped and matched with the target structured knowledge to accurately extract the precise knowledge fragment corresponding to the current sub-problem. The mathematical mapping relationship for extracting precise knowledge fragments is configured as follows: In the formula, Indicates that for the first Precise knowledge fragments extracted from each sub-question This represents the mapping function for precise knowledge extraction. Represents the target structured knowledge.

[0062] After obtaining fine-grained supporting evidence, the final stage involves generating a reasoning document tailored to the natural language query request by combining precise knowledge fragments. Specifically, the natural language query request, multiple sub-questions, and the precise knowledge fragments corresponding to each sub-question are simultaneously concatenated into a pre-defined fourth prompt template to generate a fourth prompt instruction to guide the final argument. The pre-defined fourth prompt template incorporates standard writing conventions for compliance review reports and evidence chain deduction paradigms. Finally, the fourth prompt instruction is input into a generative language model, enabling the model to perform objective logical reasoning based strictly on the original intent of the natural language query request and the precise knowledge fragments corresponding to each sub-question, ultimately generating and outputting the reasoning document.

[0063] The reasoning document is specifically divided into a fact-finding module, an evidence chain module, and a violation determination module. The fact-finding module describes the objective actions within the reviewed case file; the evidence chain module visually demonstrates the mapping and supporting relationships between each sub-issue and precise knowledge fragments; and the violation determination module outputs the final review conclusion derived from a pre-set legal and regulatory knowledge base. This modular output ensures that the final document fully complies with the standardized review requirements of disciplinary inspection and supervision operations in terms of structure and logic.

[0064] The mathematical process for generating reasoning documents is constructed as follows: In the mathematical process of generating reasoning documents, This refers to the final output of the reasoning document. This represents a logical reasoning generation function constructed based on a preset fourth prompt template and a generative language model.

[0065] By performing the aforementioned semantic decomposition, coverage verification, and precise mapping generation steps, the macro-level requirements for discipline inspection and supervision review are effectively broken down into verifiable micro-level investigation tasks. Furthermore, relying on a strict entity coverage verification mechanism, illusory outputs that deviate from the evidence benchmark are intercepted, ultimately ensuring that the output reasoning documents have a flawless and rigorous connection between the logic of applying Party discipline and national laws and the underlying evidence facts.

[0066] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0067] like Figure 2 As shown, an embodiment of the present invention provides a reasoning device for disciplinary inspection and supervision case data, including: a data parsing module, a semantic matching module, a dynamic routing module, a knowledge reconstruction module, and a reasoning generation module; The data parsing module is used to acquire multimodal case data and natural language query requests for target discipline inspection and supervision cases; perform structured parsing processing on the multimodal case data to generate an initial structured data file; The semantic matching module is used to perform semantic matching on the initial structured data file and the preset legal knowledge base based on natural language query requests, and extract a set of candidate information fragments. The dynamic routing module is used to input natural language query requests and candidate information fragment sets into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set. The knowledge reconstruction module is used to reconstruct the candidate information fragment set according to the target structure type, based on a preset generative language model and a preset first prompt template, and output the target structured knowledge; based on the natural language query request and the target structured knowledge, it summarizes the features based on a generative language model and a preset second prompt template, and outputs the descriptive information of the target structured knowledge. The reasoning generation module is used to perform semantic decomposition processing based on a generative language model and a preset third prompt template according to the natural language query request and description information, and output multiple sub-questions; perform coverage verification processing on the multiple sub-questions and target structured knowledge, and calculate the quantitative verification score; based on the quantitative verification score, map and match the multiple sub-questions with the target structured knowledge respectively, extract the corresponding precise knowledge fragments, and generate a reasoning document for the natural language query request in combination with the precise knowledge fragments.

[0068] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the reasoning method for disciplinary inspection and supervision case data as described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0069] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0070] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the reasoning method for disciplinary inspection and supervision case data as described in any one of the present invention, or the processor executes the computer program to implement the functions of each module in the above-described device embodiments.

[0071] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0072] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0073] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0074] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0075] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the reasoning method for any of the disciplinary inspection and supervision case data described above.

[0076] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0077] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0078] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An inference method for discipline inspection and supervision case data, characterized in that, include: Acquire multimodal case data and natural language query requests for target disciplinary inspection and supervision cases; The multimodal case data is processed through structured parsing to generate an initial structured data file; Based on natural language query requests, semantic matching is performed on the initial structured data file and the preset legal knowledge base to extract a set of candidate information fragments; The natural language query request and the candidate information fragment set are input into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set; Based on the target structure type, the candidate information fragment set is reconstructed using a preset generative language model and a preset first prompt template, and the target structured knowledge is output. Based on the natural language query request and the target structured knowledge, feature summarization is performed using a generative language model combined with a preset second prompt template, and descriptive information of the target structured knowledge is output. Based on the natural language query request and description information, semantic decomposition is performed using a generative language model combined with a preset third prompt template, and multiple sub-questions are output. Multiple sub-problems are subjected to coverage verification with the target structured knowledge, and a quantitative verification score is calculated. Based on the quantitative verification score, multiple sub-questions are mapped and matched with the target structured knowledge, the corresponding precise knowledge fragments are extracted, and the precise knowledge fragments are combined to generate reasoning documents for natural language query requests.

2. The inference method of discipline inspection case data according to claim 1, wherein, The multimodal case data is processed using structured parsing to generate an initial structured data file, including: Based on the pre-defined intermediate representation structure of case documents, multimodal case data is parsed to extract document metadata, the smallest semantic unit, and the original layout relationships. The intermediate representation structure of case documents includes a document metadata layer, a content block layer, a structural information layer, and a vectorized input layer. The document metadata, the smallest semantic unit, and the original layout relationship are written into the document metadata layer, content block layer, and structural information layer of the intermediate representation structure of the case document, respectively. For each smallest semantic unit in the content block layer, a standardized text for vectorized retrieval is generated based on the current smallest semantic unit, and the standardized text is written into the vectorized input layer of the intermediate representation structure of the case document. The intermediate representation structure of the case document after data writing is completed is used as the initial structured data file.

3. The inference method of disciplinary inspection case data according to claim 2, wherein, Based on natural language query requests, semantic matching is performed on the initial structured data file and the pre-set legal and regulatory knowledge base to extract a set of candidate information fragments, including: Feature extraction is performed on natural language query requests to obtain query features; Input the query features into a preset embedding model so that the embedding model generates a query semantic vector based on the query features; Extract standardized text from the vectorized input layer of the initial structured data file; Extract the contents of each article and case from the pre-set legal knowledge base; Each standardized text is input into the embedding model so that the embedding model generates a corresponding first feature vector based on each standardized text. Each clause and case content is input into the embedding model so that the embedding model generates a corresponding second feature vector based on each clause and case content; Calculate the similarity between the query semantic vector and each first feature vector, and generate the corresponding first semantic similarity; Calculate the similarity between the query semantic vector and each second feature vector, and generate the corresponding second semantic similarity. The standardized texts are sorted based on the first semantic similarity to obtain the first ranking sequence, and the standardized texts that meet the preset first condition are extracted from the first ranking sequence as the first candidate information fragments. The content of each article and case is sorted based on the second semantic similarity to obtain the second ranking sequence, and the content of articles and cases that meet the preset second conditions are extracted from the second ranking sequence as the second candidate information fragments. The first candidate information fragment and the second candidate information fragment are fused together to generate a set of candidate information fragments.

4. The inference method of disciplinary inspection case data according to claim 3, wherein, The natural language query request and the candidate information fragment set are input into a pre-defined dynamic routing model to generate the target structure type of the candidate information fragment set, including: The natural language query request and the candidate information fragment set are fused to generate joint input features; The joint input features are input into a preset dynamic routing model, so that the dynamic routing model calculates and generates the preference probabilities corresponding to the joint input features under multiple preset candidate structure types based on the joint input features; wherein, the multiple candidate structure types include tables, graphs, algorithm flows, directories, and text blocks; Compare the preference probabilities corresponding to each candidate structure type, and determine the candidate structure type corresponding to the highest preference probability as the target structure type.

5. The inference method of discipline inspection case data according to claim 4, characterized in that, Based on the target structure type, the candidate information fragment set is reconstructed using a preset generative language model and a preset first prompt template, and the target structured knowledge is output. Based on the natural language query request and the target structured knowledge, a feature summary is performed using a generative language model combined with a pre-set second prompt template. The output is a description of the target structured knowledge, including: The natural language query request, the target structure type, and the set of candidate information fragments are concatenated into a preset first prompt template to generate a first prompt instruction; The first prompt instruction is input into the preset generative language model so that the generative language model extracts text content that matches the natural language query request from the candidate information fragment set and reconstructs it according to the target structure type to generate target structured knowledge. The natural language query request and the target structured knowledge are concatenated into a preset second prompt template to generate a second prompt instruction; The second prompt instruction is input into the generative language model so that the generative language model can extract the target structured knowledge, generate a summary text containing the topic, key fields and the relationship with the natural language query request, and use the summary text as descriptive information of the target structured knowledge.

6. The inference method of discipline inspection case data according to claim 5, characterized in that, Based on the natural language query request and description information, semantic decomposition is performed using a generative language model combined with a preset third prompt template, and multiple sub-questions are output. Multiple sub-problems are subjected to coverage verification against the target structured knowledge, and a quantitative verification score is calculated, including: The natural language query request and description information are concatenated into a preset third prompt template to generate a third prompt instruction; The third prompt instruction is input into the generative language model so that the generative language model can decompose the natural language query request into multiple logically independent sub-problems; Extract the first entity information from the target structured knowledge; For each subproblem, extract the second entity information from the current subproblem; Calculate the proportion of the second entity information corresponding to the current sub-problem in the first entity information to obtain the semantic coverage value of the current sub-problem. After calculating the semantic coverage values ​​for all sub-problems, the semantic coverage values ​​for each sub-problem are comprehensively calculated to obtain the quantitative verification score.

7. The inference method of disciplinary inspection case data according to claim 6, wherein, Based on the quantitative verification score, multiple sub-questions are mapped and matched with the target structured knowledge, and corresponding precise knowledge fragments are extracted. These precise knowledge fragments are then combined to generate reasoning documents for natural language query requests, including: For each sub-problem, the current sub-problem is mapped and matched with the target structured knowledge, and the precise knowledge fragment corresponding to the current sub-problem is extracted; The quantitative verification score, natural language query request, and precise knowledge fragments corresponding to each sub-question are concatenated into the preset fourth prompt template to generate the fourth prompt instruction; The fourth prompt instruction is input into the generative language model, so that the generative language model can perform logical reasoning based on the quantized verification score, natural language query request, and the precise knowledge fragments corresponding to each sub-question, and generate a reasoning document.

8. An inference device for discipline inspection and supervision case data, characterized in that, include: The module includes a data parsing module, a semantic matching module, a dynamic routing module, a knowledge reconstruction module, and a reasoning generation module. The data parsing module is used to acquire multimodal case data and natural language query requests for target discipline inspection and supervision cases; perform structured parsing processing on the multimodal case data to generate an initial structured data file; The semantic matching module is used to perform semantic matching on the initial structured data file and the preset legal knowledge base based on natural language query requests, and extract a set of candidate information fragments. The dynamic routing module is used to input natural language query requests and candidate information fragment sets into a preset dynamic routing selection model to generate the target structure type of the candidate information fragment set. The knowledge reconstruction module is used to reconstruct the set of candidate information fragments based on the target structure type, a preset generative language model, and a preset first prompt template, and output the target structured knowledge. Based on the natural language query request and the target structured knowledge, feature summarization is performed using a generative language model combined with a preset second prompt template, and descriptive information of the target structured knowledge is output. The reasoning generation module is used to perform semantic decomposition processing based on the natural language query request and description information, combined with a preset third prompt template and a generative language model, and output multiple sub-questions. Multiple sub-problems are subjected to coverage verification with the target structured knowledge, and a quantitative verification score is calculated. Based on the quantitative verification score, multiple sub-questions are mapped and matched with the target structured knowledge, the corresponding precise knowledge fragments are extracted, and the precise knowledge fragments are combined to generate reasoning documents for natural language query requests.

9. An electronic device, comprising: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the reasoning method for disciplinary inspection and supervision case data as described in any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium includes a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the reasoning method for disciplinary inspection and supervision case data as described in any one of claims 1 to 7.

Images