A government affair text auditing method and system based on knowledge graph reasoning

Abstract

The application discloses a government affair text auditing method and system based on knowledge graph reasoning, which is used to improve the accuracy and intelligent level of government affair text auditing. The method comprises the following steps: obtaining an original policy text, and performing semantic analysis processing on the original policy text by using a double-channel semantic disambiguation engine; fusing semantic analysis results of a rule channel and a neural channel to obtain a standardized entity set subjected to semantic disambiguation and structured attribute calibration; mapping the standardized entity set, case original text data and declaration material data to a knowledge graph; learning node embedding representation in the knowledge graph by using a graph neural network model, so as to realize semantic alignment between the standardized entity set, the case original text data and the declaration material data; on the knowledge graph, performing compliance evaluation by using symbolic reasoning based on pre-defined logical rules and neural reasoning based on node embedding representation or graph path; and generating a government affair text auditing result according to a result of the compliance evaluation.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of natural language processing and intelligent government affairs, and in particular to a method and system for reviewing government documents based on knowledge graph reasoning. Background Technology

[0002] Government document review is a crucial part of the daily management and public services of government departments. It involves examining various policy documents, legal provisions, and application materials to ensure they comply with relevant laws, regulations, and policy requirements. With the continuous improvement of e-government informatization, the number of documents to be processed has increased dramatically, and traditional review methods that rely primarily on manual processes are no longer sufficient to meet the demand.

[0003] Existing technologies have attempted to leverage information technology to assist in the review of government documents, such as using Natural Language Processing (NLP). These technologies typically employ rule-based matching, keyword recognition, or shallow semantic analysis to identify information within the text or attempt to establish connections between different texts. However, due to the highly specialized and rigorous nature of government documents, especially policies and regulations, and the inevitable semantic complexity and diversity they contain, traditional NLP models struggle to accurately link legal provisions, case studies, and the logical chains of application materials. For example, they cannot dynamically parse the contextual meaning of vague qualifiers such as "shall" and "in principle" in policy clauses. In other words, existing automated government document review methods have limited capabilities in understanding complex contexts, handling ambiguous expressions, and linking different types of government documents, thus failing to meet application requirements. Summary of the Invention

[0004] This application provides a method and system for reviewing government documents based on knowledge graph reasoning, which can improve the accuracy and intelligence level of government document review.

[0005] The first aspect of this application provides a method for reviewing government documents based on knowledge graph reasoning, including:

[0006] The original policy text is obtained and semantic parsing is performed on the original policy text using a dual-channel semantic disambiguation engine. The dual-channel semantic disambiguation engine includes a rule channel built based on a government affairs domain ontology library and a neural channel built based on a pre-trained language model.

[0007] By fusing the semantic parsing results of the rule channel and the neural channel, a normalized entity set that has undergone semantic disambiguation and structured attribute labeling is obtained;

[0008] Obtain application material data and original case text data related to the original policy text, and map the standardized entity set, the original case text data, and the application material data into a knowledge graph;

[0009] learning node embedding representations in the knowledge graph by using a graph neural network model to realize semantic alignment between the standardized entity set, the case original text data and the declaration material data;

[0010] On the knowledge graph, a compliance evaluation is performed on the degree of conformity between the declaration material data and the original policy text by using symbolic reasoning based on pre-defined logical rules and neural reasoning based on the node embedding representations or graph paths;

[0011] A government text auditing result containing a compliance judgment conclusion and corresponding supporting basis is generated according to the result of the compliance evaluation.

[0012] Optionally, the rule channel is configured to perform deterministic semantic analysis on legal terms and normative expressions in the original policy text that conform to a pre-set pattern by using the government domain ontology library.

[0013] The neural channel is configured to capture dynamic semantics of ambiguous qualifiers or non-normative expressions in the original policy text by deep context modeling based on an improved ERNIE model.

[0014] Optionally, the semantic analysis results of the rule channel and the neural channel are fused to obtain a standardized entity set that has undergone semantic disambiguation and structured attribute calibration, including:

[0015] The analysis confidence of semantic analysis results of the rule channel and the neural channel that point to the same text content is calculated respectively.

[0016] The decision weights of the rule channel and the neural channel are determined according to a pre-set disambiguation weight allocation strategy.

[0017] The semantic analysis results of the rule channel and the neural channel are selected, merged or modified by integrating the analysis confidence and the decision weights, to obtain a standardized entity set that has undergone semantic disambiguation and structured attribute calibration.

[0018] Optionally, the declaration material data and the case original text data related to the original policy text are obtained, and the standardized entity set, the case original text data and the declaration material data are mapped into a knowledge graph, including:

[0019] Entities representing policy provisions and legal concepts in the standardized entity set and their relationships are converted into nodes and edges of a policy dimension in the knowledge graph.

[0020] perform information extraction on the case original text data related to the original policy text, extract target information of a case focus, a judgment reason, and cited law articles, and convert the target information into a node of a case dimension in the knowledge graph and establish an association with the policy dimension node;

[0021] parse the content of the declaration material data into a structured data item, and convert the data item into a node and an attribute of a material dimension in the knowledge graph.

[0022] Optionally, the learning of the node embedding representation in the knowledge graph by using the graph neural network model is to realize semantic alignment between the standardized entity set, the case original text data, and the declaration material data, including:

[0023] using an attention mechanism and an aggregation mechanism built in the graph neural network model, iteratively aggregating feature information of each entity node under the policy dimension, the case dimension, and the material dimension from a heterogeneous neighborhood of the entity node, to generate a node embedding representation capable of capturing cross-dimension context dependency, the heterogeneous neighborhood including entity nodes and diverse relationship edges of different dimensions;

[0024] optimizing learning of the node embedding representation based on a preset loss function until core semantic concepts between the data of each dimension reach consistent representation in an embedding space, the loss function being used to maximize the representation similarity of semantically associated entity nodes in the embedding space, while minimizing the representation similarity of irrelevant entity nodes.

[0025] Optionally, on the knowledge graph, a compliance evaluation of a degree of conformity between the declaration material data and the original policy text is performed by using symbolic reasoning based on pre-defined logical rules and neural reasoning based on the node embedding representation or a graph path, including:

[0026] extracting structured pre-defined logical rules from provisions of the original policy text or from constraint conditions defined in the government affairs domain ontology library;

[0027] on the knowledge graph, using symbolic reasoning to preliminarily check a compliance checkpoint in the declaration material data that can directly correspond to the pre-defined logical rules, and outputting a preliminary conclusion of deterministic conformity, partial conformity, non-conformity, or further analysis;

[0028] for a case where the preliminary conclusion is partial conformity, further analysis, or the compliance checkpoint involves semantic understanding complexity, using a method of neural reasoning to perform deep semantic matching by analyzing the node embedding representation or the graph path, and generating auxiliary evaluation opinions.

[0029] Optionally, the generating of the government text auditing result containing the compliance judgment conclusion and the corresponding supporting basis according to the result of the compliance assessment comprises:

[0030] integrating the preliminary conclusion of the symbolic reasoning output and the auxiliary evaluation opinion generated by the neural reasoning to obtain a comprehensive evaluation result, and determining a final compliance judgment conclusion of the declaration material data according to the comprehensive evaluation result;

[0031] obtaining supporting basis directly related to the final compliance judgment conclusion from the knowledge graph and the processes of the symbolic reasoning and the neural reasoning;

[0032] organizing and formatting the final compliance judgment conclusion and the supporting basis to generate a government text auditing result.

[0033] The second aspect of the application provides a government text auditing system based on knowledge graph reasoning, comprising:

[0034] an acquisition unit configured to acquire an original policy text and perform semantic analysis processing on the original policy text by using a double-channel semantic disambiguation engine, wherein the double-channel semantic disambiguation engine comprises a rule channel constructed based on a government domain ontology library and a neural channel constructed based on a pre-trained language model;

[0035] a fusion unit configured to fuse semantic analysis results of the rule channel and the neural channel to obtain a standardized entity set subjected to semantic disambiguation and structured attribute calibration;

[0036] a mapping unit configured to acquire declaration material data and case original text data related to the original policy text, and map the standardized entity set, the case original text data and the declaration material data to a knowledge graph;

[0037] an alignment unit configured to learn node embedding representation in the knowledge graph by using a graph neural network model, so as to realize semantic alignment between the standardized entity set, the case original text data and the declaration material data;

[0038] a reasoning unit configured to perform compliance assessment on a compliance degree between the declaration material data and the original policy text by using symbolic reasoning based on pre-defined logical rules and neural reasoning based on the node embedding representation or graph path on the knowledge graph;

[0039] a generating unit configured to generate a government text auditing result containing a compliance judgment conclusion and corresponding supporting basis according to a result of the compliance assessment.

[0040] The third aspect of the application provides a government text auditing device based on knowledge graph reasoning, comprising:

[0041] a processor, a memory, an input / output unit, and a bus;

[0042] the processor is connected with the memory, the input / output unit, and the bus;

[0043] the memory stores a program, and the processor invokes the program to execute the method for government affair text auditing based on knowledge graph reasoning in the first aspect and any one of the optional embodiments in the first aspect.

[0044] The fourth aspect of the present application provides a computer readable storage medium, which stores a program, and the program executes the method for government affair text auditing based on knowledge graph reasoning in the first aspect and any one of the optional embodiments in the first aspect when executed on a computer.

[0045] From the above technical solutions, the present application has the following advantages:

[0046] For the polysemy expression of policy provisions, the present application adopts a double-channel semantic disambiguation engine, which can more accurately understand the actual meaning of policy provisions, especially the meaning of ambiguous qualifiers, by combining the deterministic analysis of normative expressions in the rule channel and the dynamic semantic capture based on context in the neural channel. In addition, the present application further constructs a multi-dimensional knowledge graph based on the dimensions of laws, cases, and materials, maps the scattered government information to a unified knowledge framework, and realizes the effective fusion and semantic alignment of information in different dimensions through learning node embedding representation. On this unified framework, the use of symbolic reasoning for hard condition checking and neural reasoning for complex association pattern mining can accurately construct, verify, and generate complete logical chains and compliance judgment conclusions between laws, cases, and declaration materials, thereby overcoming the shortcomings of traditional NLP models in dealing with the complexity of government affair texts and constructing reliable logical chains, and significantly improving the accuracy and intelligence level of government affair text auditing. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0048] Figure 1 An embodiment flowchart of the method for government affair text auditing based on knowledge graph reasoning provided by the present application is shown in the figure.

[0049] Figure 2 Another embodiment flowchart of the method for government affair text auditing based on knowledge graph reasoning provided by the present application is shown in the figure.

[0050] Figure 3 An embodiment structure diagram of the government affair text auditing system based on knowledge graph reasoning provided in the application is provided;

[0051] Figure 4 An embodiment structure diagram of the government affair text auditing device based on knowledge graph reasoning provided in the application is provided. DETAILED DESCRIPTION

[0052] The application provides a government affair text auditing method and system based on knowledge graph reasoning, which is used to improve the accuracy and intelligent level of government affair text auditing.

[0053] It should be noted that the government affair text auditing method based on knowledge graph reasoning provided in the application can be applied to a terminal and can also be applied to a server. For example, the terminal can be a smart phone or a computer, a tablet computer, a smart television, a smart watch, a portable computer terminal, and a fixed terminal such as a desktop computer. For convenience of description, the system carried on the terminal is taken as an execution subject for example in the application.

[0054] Please refer to Figure 1 , Figure 1 An embodiment of the government affair text auditing method based on knowledge graph reasoning provided in the application includes the following steps.

[0055] 101, obtaining an original policy text and performing semantic analysis processing on the original policy text by using a double-channel semantic disambiguation engine, the double-channel semantic disambiguation engine including a rule channel constructed based on a government affair domain ontology library and a neural channel constructed based on a pre-trained language model;

[0056] Firstly, the system needs to obtain and analyze the original policy text as the basis for subsequent compliance auditing. The original policy text is the direct basis for judging whether the declaration material is compliant. However, the original policy text is usually unstructured natural language and often contains legal terms, normative expressions and ambiguous or fuzzy qualifiers such as "should" and "in principle". Traditional single natural language processing method is difficult to accurately and comprehensively understand these texts, especially to dynamically analyze the context meaning of ambiguous words. Based on this, the present application chooses to use a double-channel semantic disambiguation engine to perform semantic analysis processing on the original policy text, and to convert the natural language text in the original policy text into structured semantics understandable by a computer. The design concept of the double-channel semantic disambiguation engine is to combine the advantages of different analysis mechanisms to cope with the characteristics of government text which is both normative and complex. The double-channel engine mainly includes two parallel analysis channels: rule channel and neural channel, which are described as follows:

[0057] The implementation of the rule channel relies on a pre-constructed government domain ontology library. The government domain ontology library is a structured knowledge base containing concepts, terms, relationships, and rules in the government domain. The rule channel uses the ontology knowledge in the government domain ontology library and the preset parsing rules to perform deterministic semantic parsing on legal terms and normative expressions in the original policy text that conform to the standard mode, ensuring accurate understanding of standard language. Combined with technologies such as dependency syntax analysis, the rule channel can also assist in identifying the modifying components of ambiguous qualifiers and trigger the checking logic for related exception conditions when encountering specific ambiguous words. The advantage of the rule channel is that it has high precision and interpretability when dealing with normative and explicit language.

[0058] The neural channel is implemented based on pre-trained language model technology, and specifically can use an ERNIE model optimized and trained specifically for the government and legal domain, i.e., an improved ERNIE model. The ERNIE model is a pre-trained language model that, compared to traditional pre-trained models such as BERT, focuses on incorporating word, entity, phrase, and other knowledge unit information in the pre-training phase, thereby enhancing the model's language understanding ability, especially its grasp of semantic and conceptual relationships. The improved ERNIE model can be trained on a large amount of government and legal text to improve its language understanding ability for government and legal text. The neural channel uses deep context modeling capabilities to analyze the complex relationships of words in sentences and paragraphs, and can more accurately capture dynamic semantics in government text that rely on complex contexts, especially for ambiguous qualifiers such as "should" and "in principle" or non-standardized expressions, and can infer more accurate meanings based on the context in which they appear. The advantage of the neural channel is its flexibility and context dependence in handling language, which makes up for the shortcomings of the rule channel in understanding complex contexts.

[0059] 102、fuse the semantic parsing results of the rule channel and the neural channel to obtain a standardized entity set that has undergone semantic disambiguation and structured attribute calibration;

[0060] Since the two channels are based on different principles for semantic parsing, their outputs can differ, conflict, or complement each other. Therefore, it is necessary to fuse the semantic parsing results of the rule channel and the neural channel, comprehensively utilize the certainty of the rule channel in processing normative expressions and the flexibility of the neural channel in capturing contextual semantics, and finally generate a single and accurate semantic parsing result. Specifically, there are multiple possible ways to achieve such fusion. One way is based on preset priority rules, for example, for specific types of legal terminology, the parsing result of the rule channel can be set as the priority to adopt. Another way is to use a voting mechanism, if both channels give the same parsing for the same text segment, the confidence score is increased. It is also possible to combine the confidence scores of the outputs of the two channels based on weighting, score different results and select the highest scoring parsing, or train an independent machine learning model to learn how to best combine the outputs of the two channels. These methods can be implemented through algorithms, and the specific implementation is not limited here. The logic of fusion is to take the preliminary results of the two channels as input, and through logical judgment, calculation or model prediction to obtain a standardized entity set.

[0061] The normalized entity set contains all the key semantic information extracted from the original policy text, which has been processed by semantic disambiguation, solving potential ambiguity, and labeled with structured attributes. The normalized entity set exists in a unified, machine-friendly format, which can be directly used as input data for subsequent steps of building a multi-dimensional knowledge graph. Assuming that the original policy text fragment is: "According to Article 8 of the Certain Management Measures, the applicant unit shall be an enterprise registered in our city, and the registered capital shall not be less than 500,000 yuan." This text contains: policy reference, entity (applicant unit, enterprise, our city), attribute (registered capital), numerical value (500,000 yuan), determiner (should be, shall not be less than), and location ambiguity (our city). The rule channel may identify POLICY ("Article 8 of the Certain Management Measures"), ENTITY_TYPE ("enterprise"), LOCATION ("our city"), CONCEPT ("registered capital"), AMOUNT ("500,000 yuan"), and according to the rules, "should be" and "shall not be less than" represent "requirement" or "limitation". The neural channel models the context and may identify: ENTITY ("applicant unit") refers to ENTITY_TYPE ("enterprise"), LOCATION ("our city") is a location entity, and understands that "registered capital shall not be less than 500,000 yuan" is a numerical requirement for "enterprise", and understands that "should be" represents a qualification requirement. The final normalized entity set after fusion will decompose the information in the original text into independent entities, extract the specific content of the entity and store it as its attribute, parse the constraints in the text and store them as structured attributes of the registered capital entity, and identify the semantic relationship between entities and represent it as a relationship.

[0062] 103. Obtain the declaration material data and the case original text data related to the original policy text, and map the normalized entity set, the case original text data and the declaration material data to the knowledge graph;

[0063] The core of this step is to integrate all the information sources required for government text review: the preliminary processed policy information (the normalized entity set output in step 102), the case information that helps understand the policy (case original data), and the application information to be reviewed (application material data) into a unified data model, i.e., to build a multi-dimensional knowledge graph. Case original data provides the application and interpretation of the policy in practice, while application material data contains specific factual information of the applicant or the object to be reviewed. Since government text review involves various data sources such as policy provisions, legal cases, and application forms, they have different formats and different ways of expressing semantics, and are isolated from each other. Mapping these data to a knowledge graph can transform these scattered and heterogeneous information into a structured and interconnected unified model, obtaining a comprehensive knowledge graph containing policy specifications, case facts, and application details, providing a foundation for subsequent efficient and accurate correlation queries and logical reasoning between different information types.

[0064] 104. Learning the node embedding representation in the knowledge graph using a graph neural network model to realize semantic alignment between the normalized entity set, case original data, and application material data;

[0065] Although step 103 integrates policy, case, and material information into the same graph, simply establishing structural connections is not enough for the computer to understand the deep semantic relationship between entities that refer to the same concept or are highly related in different dimensions. For example, the "registered capital" provision node in the policy dimension and the "registered capital" attribute node of an enterprise entity in the material dimension may be far apart in structure, but they are closely related in semantics. The purpose of semantic alignment is to make these entities representing the same or related concepts in different dimensions close in the numerical representation space through learning, so that the computer can identify the implicit semantic relationship between them.

[0066] The core technology to achieve semantic alignment is the graph neural network model (GNN). GNN is a neural network model that can learn directly on graph structure data, and is used in this application to learn the embedding representation of each node in the knowledge graph. Node embedding representation is a numerical vector representation of each node in a multi-dimensional space, and this vector captures the information of the node itself and its structural and semantic context in the entire graph. The learning process enables the graph neural network model to analyze the connections between nodes and the information they carry, and map nodes with similar semantics or representing related concepts to positions close to each other in the embedding space, through this process, semantic alignment between different dimensional data can be achieved, making entities or information from different sources (policy, case, material) but related in meaning have similar numerical representations in the embedding space, improving the usability of the knowledge graph.

[0067] 105. On the knowledge graph, use symbolic reasoning based on predefined logical rules and neural reasoning based on node embedding representation or graph path to conduct compliance assessment on the degree of compliance between the declaration material data and the original policy text.

[0068] After completing the information preparation of steps 101-104, it is necessary to perform real logical judgment to determine whether the declaration material meets the policy requirements. It specifically involves comparing specific facts in the declaration material with standard conditions in the policy text, and considering relevant case explanations. Due to the complexity of government text review, simple matching is not enough to complete the task, and multiple reasoning methods need to be combined. This application chooses to use mixed reasoning methods on the constructed knowledge graph, including symbolic reasoning based on predefined logical rules and neural reasoning based on node embedding representation or graph path. The knowledge graph provides the basis for complex association and reasoning as a unified data carrier.

[0069] Symbolic reasoning is a reasoning method based on formalized logic and preset rules. Symbolic reasoning specifically makes deterministic judgments on facts in the knowledge graph through predefined logical rules. For example, if the rule states that "registered capital must be greater than or equal to 500,000 yuan", symbolic reasoning will directly check whether the node representing the registered capital value in the declaration material in the knowledge graph meets this condition. The advantage of symbolic reasoning is that its process is transparent, the result is determined, and it has strict logical validity.

[0070] Neural reasoning is a reasoning method based on machine learning models. Neural reasoning can identify and utilize complex patterns, implicit associations, or semantic similarity-based connections that cannot be distilled into explicit symbolic rules. It is suitable for handling situations with ambiguous semantics, requiring context inference, or involving complex association paths.

[0071] By comprehensively using these two reasoning mechanisms on the knowledge graph, the application can conduct comprehensive and in-depth compliance assessment on the degree of compliance between the declaration material data and the original policy text. In this application, symbolic reasoning is responsible for handling explicit hard conditions to ensure basic compliance; neural reasoning is responsible for handling complex and implicit semantics or associations to provide more detailed or auxiliary judgments.

[0072] 106. According to the results of the compliance assessment, generate a government text review result containing compliance judgment conclusions and corresponding supporting evidence.

[0073] After obtaining the results of the compliance assessment, the system first integrates all the assessment results and makes a clear compliance judgment conclusion based on this information. In order to make the final conclusion convincing and transparent, the system also needs to trace back and obtain the supporting evidence directly related to the final compliance judgment conclusion from the entire audit process. These evidence comes from multiple aspects, including but not limited to: the original declaration material data and policy clause nodes represented in the knowledge graph; specific pre-defined logical rules triggered or not triggered in symbolic reasoning process; key graph paths, semantic associations or patterns identified in neural reasoning process; and related case nodes and their association with policies or materials. The process of obtaining evidence needs to be able to trace back the reasoning chain and locate the key facts and rules that lead to the final conclusion. Finally, the system needs to organize and format the determined final compliance judgment conclusion and the obtained supporting evidence to generate government affairs text audit results that are easy for users to understand. Specifically, the conclusion and evidence can be filled into a pre-set report template to generate a structured data file (such as XML, JSON), or visualized on the user interface. The supporting evidence usually needs to be presented in a clear way, such as listing the violated clauses, related declaration data items, case explanations for auxiliary judgment, or key associated graph fragments identified in the reasoning process.

[0074] In this embodiment, for the ambiguous expression of policy clauses, the application adopts a dual-channel semantic disambiguation engine, which can more accurately understand the actual meaning of policy clauses, especially the fuzzy qualifiers, by combining the deterministic analysis of normative expressions in the rule channel and the dynamic semantic capture based on context in the neural channel. In addition, the application further constructs a multi-dimensional knowledge graph based on the dimensions of laws, cases and materials, which maps scattered government information to a unified knowledge framework, and realizes effective fusion and semantic alignment of information in different dimensions through learning node embedding representation. On this unified framework, symbolic reasoning is used for hard condition checking and neural reasoning for complex association pattern mining, which can accurately construct, verify and generate complete logical chains and compliance judgment conclusions between laws, cases and declaration materials, thereby overcoming the shortcomings of traditional NLP models in dealing with the complexity of government text and constructing reliable logical chains, significantly improving the accuracy and intelligence level of government text audit.

[0075] The knowledge graph reasoning-based government text auditing method provided by the application will be described in detail below. Please refer to Figure 2 , Figure 2 Another embodiment of the knowledge graph reasoning-based government text auditing method provided by the application, the method comprises:

[0076] 201、obtain the original policy text and perform semantic parsing on the original policy text by using a double-channel semantic disambiguation engine, the double-channel semantic disambiguation engine including a rule channel constructed based on a government domain ontology library and a neural channel constructed based on a pre-trained language model;

[0077] In the embodiment, step 201 is similar to step 101 of the foregoing embodiment, and will not be described here again.

[0078] 202、respectively calculate parsing confidence of semantic parsing results pointing to the same text content recognized by the rule channel and the neural channel;

[0079] Steps 202-204 are the process of fusing the semantic parsing results of the rule channel and the neural channel. Both channels of the double-channel engine can recognize entities, relationships or attributes for the same part of the original text (for example, the same word, phrase or sentence fragment). Therefore, the system needs to calculate a parsing confidence for the parsing results pointing to the same part of the original text recognized by the rule channel and the neural channel respectively. The parsing confidence is a numerical value for quantifying the confidence level of the channel for a specific parsing result (for example, the recognized entity type, relationship type or attribute value). The parsing confidence of the rule channel can be calculated based on the accuracy of the matched rules, the covered text range or the preset rule weight. The parsing confidence of the neural channel can be directly determined using the probability score of the underlying pre-trained language model prediction result.

[0080] 203、determine the decision weight of the rule channel and the neural channel according to a preset disambiguation weight distribution strategy;

[0081] In addition to the confidence of a specific parsing result, in the overall fusion strategy, the rule channel or the neural channel can be given higher priority, for example, the rule channel is preferred when processing standardized legal terms, and the neural channel is preferred when processing ambiguous and colloquial expressions. Specifically, a pre-designed preset disambiguation weight distribution strategy can be applied. The strategy is an algorithm or a set of rules that determines the decision weight of the rule channel and the neural channel in this fusion decision according to the current parsing context (for example, the type of the text paragraph being processed, the recognized entity type). The decision weight is also a numerical value, indicating the relative importance of the results of the two channels when considering the parsing confidence. This strategy can be configured as a fixed weight distribution strategy at system initialization, or can be dynamically adjusted during operation as needed. For example, the strategy can specify that for terms defined in the ontology library, the decision weight of the rule channel is higher, and for complex sentence patterns, the decision weight of the neural channel is higher.

[0082] 204、The integrated analysis confidence and decision weight select, merge or correct the semantic analysis results of the rule channel and the neural channel, and obtain a normalized entity set that is semantically disambiguated and structured attribute calibrated;

[0083] The system integrates the analysis confidence from step 202 and the decision weight from step 203 to perform decision on the semantic analysis results of the rule channel and the neural channel. This process can evaluate the comprehensive score of the analysis result of each channel for the same text content according to an algorithm (e.g., analysis confidence multiplied by decision weight, plus other factors). Based on this comprehensive score, the system can take one or more of the following operations:

[0084] Selection: preferentially select the analysis result of the channel with the highest comprehensive score, and discard the analysis of the same content by other channels; Merge: if the analysis results of two channels are complementary (e.g., the rule channel identifies an entity, and the neural channel identifies a certain attribute or modifier of the entity), the information is merged into the same final structured representation; Correction: if the analysis result of a channel (e.g., a channel with high decision weight and high analysis confidence) conflicts with the analysis result of another channel, the former is used to correct or overwrite the latter.

[0085] After the selection, merging or correction of the semantic analysis results of the two channels, a normalized entity set that is semantically disambiguated and structured attribute calibrated can be finally obtained. This means that each entity, relationship and attribute in the set has a definite meaning (ambiguity is eliminated), and is given a standard structure and attribute (e.g., entity type, attribute name, relationship type), and the format is unified, which can be directly used for subsequent construction of knowledge graph.

[0086] The following is an example of a policy text segment "the applicant should have no major tax law violations in the past three years." The analysis results of the rule channel and the neural channel are fused through steps 202 to 204. For the expression "in the past three years" in the text, step 202 calculates the analysis confidence of the respective analysis results of the rule channel and the neural channel (for example, the rule channel accurately matches the rule and analyzes "forward calculation for 36 months", with a confidence of 0.9; the neural channel understands the context and analyzes "the past three years", with a confidence of 0.7). Then, step 203 determines the decision weights of the two channels according to the preset disambiguation weight distribution strategy (the strategy sets the rule weight to 0.6 and the neural weight to 0.4). Finally, in step 204, the system comprehensively considers these analysis confidences and decision weights to make a decision (for example, calculate the weighted scores, rule: 0.9*0.6=0.54; neural: 0.7*0.4=0.28). Since the weighted score of the rule channel (0.54) is higher, the system finally selects the analysis result of the rule channel "forward calculation for 36 months" and includes it in the final standardized entity set, completing the final semantic determination of the expression "in the past three years".

[0087] 205. converting the entities representing policy provisions, legal concepts and their relationships in the standardized entity set into nodes and edges of the policy dimension in the knowledge graph;

[0088] Steps 205-207 are the process of mapping the standardized entity set, case original text data and declaration material data to the knowledge graph. First, the standardized entity set is traversed and processed. Each entity in the set represents a policy provision or legal concept, which is converted into a node of the policy dimension in the knowledge graph. For example, "a certain regulation", "fifteenth article", "registered capital" and the like may become independent nodes. The relationships between the entities identified in the set (such as reference, definition, application, etc.) are converted into edges connecting these corresponding nodes. For example, a "applicable" edge may be established between the node representing "fifteenth article" and the node representing "high-tech enterprise identification" according to the original text relationship.

[0089] 206. performing information extraction on the case original text data related to the original policy text, extracting target information such as case focus, judgment reason and cited articles, and converting the target information into nodes of the case dimension in the knowledge graph and establishing associations with the nodes of the policy dimension;

[0090] For the case original data, the information closely related to the policy application is extracted, such as the case focus, the judgment reason, and the law articles explicitly cited in the case. These structured target information is then converted into the nodes of the case dimension in the knowledge graph. For example, a specific case record can be a node, and its main dispute focus can be another associated node, and the points of the judgment reason can also be associated nodes. It should be noted that this step will analyze the cited law article information and establish an association (edge) between the corresponding case dimension node and the policy dimension node built in step 205. For example, if the case cites a provision of “a certain law”, a “cited” or “explained” relationship edge will be created between the node representing the case and the policy node representing the provision.

[0091] 207. The content of the declaration material data is parsed into structured data items, and the data items are converted into nodes and attributes of the material dimension in the knowledge graph;

[0092] Finally, the declaration material data is parsed and processed, especially for the structured or semi-structured content such as tables, fields, lists, etc. The content is converted into structured data items. For example, the value of the “registered capital” field, the date of the “establishment date” field, and the number of the “qualification certificate” column are extracted from the enterprise application table. These structured data items are converted into nodes and attributes of the material dimension in the knowledge graph. Generally, the subject in the declaration material (such as the applicant enterprise or individual) is represented as a node, and the specific data items extracted from the material are represented as attributes or associated attribute nodes of the node (for example, “registered capital” is an attribute of the enterprise node, and its value is “500 million yuan”). At the same time, the association between the material dimension node and the policy dimension node may also be established according to the material content (for example, the node representing the enterprise qualification is connected with the provision node about the enterprise qualification requirements in the policy).

[0093] Through steps 205-207, the system processes and maps policy information, case information, and declaration material information into different dimensions of the knowledge graph, and finally constructs a multi-dimensional knowledge graph that integrates all information sources required for government text review by establishing cross-dimensional associations.

[0094] 208. Using the attention mechanism and aggregation mechanism built into the graph neural network model, iteratively aggregate the feature information of each entity node in the policy dimension, case dimension, and material dimension from its heterogeneous neighborhood, which includes entity nodes and diverse relationship edges in different dimensions, to generate node embedding representations that can capture cross-dimensional context dependencies.

[0095] After the knowledge graph is constructed, each node has its initial feature information. However, to achieve semantic alignment, the representation of a node should not only reflect its own features, but also incorporate the context information in the entire graph, especially the information of nodes and edges associated with it in different dimensions. Specifically, the learning of node embedding representation can be completed by using the attention mechanism and aggregation mechanism built in the graph neural network model GNN. Specifically, for each entity node in the knowledge graph, the GNN model checks its heterogeneous neighborhood. The heterogeneous neighborhood refers to the set of surrounding nodes directly connected to the node, which contains entity nodes from different dimensions (policies, cases, materials) and different types of diverse relationship edges (for example, citation relationship, application relationship, attribute relationship, case association relationship, etc.) connecting them. The aggregation mechanism is responsible for collecting the feature information of each neighborhood node and its connecting edge in the heterogeneous neighborhood. The attention mechanism assigns different weights to different neighborhood nodes or different types of relationship edges when aggregating, so that the model can focus on the neighborhood information that is more important for understanding the semantics of the current node. This aggregation process is usually carried out iteratively in multiple layers, allowing nodes to gather information from more distant neighborhoods in the graph and capture broader graph structure context. In this way, the model can generate node embedding representations for each node, and these vector representations can capture cross-dimensional context dependencies and incorporate information associated with other dimensional nodes.

[0096] 209、based on the preset loss function, the node embedding representation is optimized and learned until the core semantic concepts between the data in each dimension reach consistent representation in the embedding space, and the loss function is used to maximize the similarity of the representations of the entity nodes associated with the semantics in the embedding space, while minimizing the similarity of the representations of the irrelevant entity nodes;

[0097] Step 209 is the training and optimization process of node embedding representation, specifically guiding the initial node embedding representation generated in step 208 to converge towards the goal of achieving semantic alignment in combination with the preset loss function. Optimization learning is an iterative process that continuously improves the quality of node embedding by adjusting the internal parameters of the GNN model to minimize the value of the loss function. The loss function is designed to guide the model learning to maximize the similarity of the representations of semantically related entity nodes in the embedding space, while minimizing the similarity of the representations of unrelated entity nodes. Semantically related entity nodes refer to nodes that represent the same or highly related core semantic concepts in different dimensions (for example, the node representing the definition of "registered capital" in the policy and the node corresponding to the numerical value of the enterprise's "registered capital" in the filing materials). Core semantic concepts are concepts that are important in the legal field of government affairs and are shared across text sources. The goal of learning is to achieve consistent representation of these semantically related nodes in the embedding space, that is, their vector representations become similar, reflecting their semantic association. Optimization learning continues until the loss function converges and reaches the preset stopping condition, at which point it can be considered that the core semantic concepts between the dimensions of data have achieved effective semantic alignment in the embedding space.

[0098] 210. Extract structured pre-defined logical rules from the provisions of the original policy text or from the constraints defined in the ontology library of government affairs;

[0099] The premise of symbolic reasoning is to have an executable rule set, i.e., pre-defined logical rules. Policy texts contain a large number of specifications and requirements, while the ontology library of government affairs may formally define the constraint relationships between entities. This information is the basis for human understanding of policies and making compliance judgments, and needs to be converted into a logical form that machines can understand and apply. In this embodiment, there are two ways to obtain pre-defined logical rules: one is to directly identify and extract sentences or fragments with explicit structure and judgment logic from the provisions of the original policy text; the other is to obtain the constraints defined in the ontology library of government affairs. The ontology library usually contains rules or restrictions expressed in formal language. This process is usually automated, and these pre-defined logical rules can be maintained and updated when the policy is updated or the ontology library is revised.

[0100] 211. On the knowledge graph, use symbolic reasoning to preliminarily check the compliance check points in the filing materials data that can directly correspond to the pre-defined logical rules, and output preliminary conclusions of certainty compliance, partial compliance, non-compliance, or further analysis;

[0101] In actual compliance assessment, symbolic reasoning is first used on the constructed and semantically aligned knowledge graph. Symbolic reasoning is a computational method based on explicit rules and logic for judgment. The system uses pre-extracted predefined logical rules to check specific information in the knowledge graph. These checked information points are called compliance checkpoints, which are the representation of the declaration material data in the knowledge graph, and directly match and correspond to specific predefined logical rules (for example, a rule is about "registered capital >= 500,000", and the node representing the value of the declaration material "registered capital" in the knowledge graph is the corresponding checkpoint). In symbolic reasoning, each such checkpoint is checked to determine whether it meets the corresponding rule. The result of the check is output as a preliminary conclusion, and its status is discrete, including certainty compliance (clearly meets the rules), non-compliance (clearly violates the rules), partial compliance (for example, meets part of the rule conditions), or needs further analysis (the rule cannot be directly judged, and more information or more complex reasoning is needed). The advantage of symbolic reasoning is the logical rigor and high certainty of the judgment result, which can handle the most direct and certain compliance conditions in government documents.

[0102] 212. For cases where the preliminary conclusion is partial compliance, further analysis, or the compliance checkpoint involves complex semantics, a neural reasoning method is used to analyze node embedding representations or graph paths for deep semantic matching to generate auxiliary evaluation opinions.

[0103] Step 212 is the second phase of compliance assessment, which is necessary to handle compliance issues that symbolic reasoning cannot give a definite conclusion or involve complex semantic understanding, including ambiguous situations that rules cannot fully cover, situations that require inference from multiple sources of information, or situations that involve understanding the deep meaning of the text.

[0104] This step is triggered when the preliminary conclusion output in step 211 is partially compliant, pending further analysis, or when the compliance checkpoint itself involves semantic understanding complexity (e.g., requiring judgment on whether a description meets the criteria of “innovation”). For these cases, the system utilizes the method of neural inference for deeper analysis. Neural inference relies on the previously learned knowledge graph representation, making judgments by analyzing the learned node embedding representation in the knowledge graph or exploring the graph paths in the knowledge graph. Node embedding representation captures the semantic position of nodes in the embedding space and their relevance to other nodes, while graph paths reflect the structure formed by a series of relationship connections between entities. Neural inference can perform deep semantic matching (e.g., determining whether a description in the filing material is highly related in semantics to the “core technology” concept defined in the policy), capture association patterns that symbolic rules cannot discover, or make probabilistic inferences on compliance, by calculating and pattern matching these representations and structures. The output of neural inference is supplementary evaluation opinions, which provide supplementary judgments, probability assessments, or relevant insights for complex or uncertain checkpoints, but are usually not the final deterministic conclusion.

[0105] 213. The preliminary conclusion of symbolic reasoning and the supplementary evaluation opinions generated by neural inference are integrated to obtain a comprehensive evaluation result, and the final compliance judgment conclusion of the filing material data is determined according to the comprehensive evaluation result;

[0106] Symbolic reasoning provides a deterministic or categorical preliminary judgment, while neural reasoning provides supplementary opinions that may have probability or association pattern information for complex situations. The system needs to integrate the outputs of symbolic reasoning and neural reasoning, i.e., the preliminary conclusion and the supplementary evaluation opinions, to obtain a comprehensive evaluation result. For example, for the same checkpoint, if the symbolic reasoning is “partially compliant” and the neural reasoning gives an 80% compliance probability, the comprehensive evaluation result is based on the combined score or classification of both. Subsequently, the system determines the final compliance judgment conclusion of the filing material data according to the comprehensive evaluation result. This is a decision mapping process that maps the comprehensive evaluation result to the final compliance state presented to the user. The final compliance judgment conclusion is usually one of the pre-defined discrete states, such as “compliant”, “non-compliant”, “requires manual review”, etc. The logic of the decision can be based on rules (e.g., if the comprehensive evaluation result of any key checkpoint is “non-compliant”, the final conclusion is “non-compliant”) or based on a trained classification model, which is not limited here. This decision is the final decision result given by the system's automated review.

[0107] 214. From the knowledge graph and the processes of symbolic reasoning and neural reasoning, support evidence directly related to the final compliance judgment conclusion is obtained;

[0108] After determining the final compliance judgment conclusion, the system also needs to backtrack and retrieve the supporting evidence directly related to the conclusion. These evidence come from two core parts:

[0109] 1. Knowledge graph: The knowledge graph stores all the integrated and semantically aligned original information, including the specific content of policy provisions, data facts in the reporting materials, and related case records. As supporting evidence, the system retrieves and extracts the key nodes and edges in the knowledge graph that directly participated in the final judgment, such as the specific data items in the reporting materials, the policy provisions related to them, or the case nodes that assisted in the judgment.

[0110] 2. Symbolic reasoning and neural reasoning process: When performing compliance evaluation, the reasoning engine records its working process. For symbolic reasoning, the system records which specific predefined logical rules are applied, at which compliance check points in the knowledge graph these rules are checked, and the results (initial conclusions) of the checks. For neural reasoning, the system records which node embedding representations or graph paths of the knowledge graph are analyzed, which deep semantic matching or pattern mining operations are performed, and what kind of auxiliary evaluation opinions are generated. As supporting evidence, the system extracts from these records the application of key rules that directly lead to the final conclusion, important graph path discovery, or auxiliary evaluation opinions that have a decisive impact.

[0111] By extracting original facts from the knowledge graph and tracing the logical chain from the reasoning process, the system can comprehensively obtain the supporting evidence that directly supports the final conclusion. These evidence can be used to generate the key components of the final audit report content, ensuring the traceability and credibility of the audit results.

[0112] 215. Organize and format the final compliance judgment conclusion and supporting evidence to generate the government text audit result.

[0113] Taking the final compliance judgment conclusion and the collected supporting evidence as input, and performing organization and formatting operations, the system finally generates the government text audit result. This result not only contains the final judgment of automated audit, but also contains all the evidence supporting the judgment, greatly improving the transparency, credibility, and explainability of the audit result.

[0114] The knowledge graph-based reasoning government text auditing system provided in the present application will be described in detail below. Please refer to Figure 3 , Figure 3 Another embodiment of the knowledge graph-based reasoning government text auditing system provided in the present application includes:

[0115] The acquisition unit 301 is configured to acquire an original policy text, and perform semantic analysis processing on the original policy text by using a double-channel semantic disambiguation engine, the double-channel semantic disambiguation engine including a rule channel constructed based on a government affair domain ontology library and a neural channel constructed based on a pre-trained language model;

[0116] The fusion unit 302 is configured to fuse semantic analysis results of the rule channel and the neural channel to obtain a standardized entity set subjected to semantic disambiguation and structured attribute calibration;

[0117] The mapping unit 303 is configured to acquire declaration material data and case original text data related to the original policy text, and map the standardized entity set, the case original text data and the declaration material data into a knowledge graph;

[0118] The alignment unit 304 is configured to learn node embedding representation in the knowledge graph by using a graph neural network model, so as to realize semantic alignment between the standardized entity set, the case original text data and the declaration material data;

[0119] The reasoning unit 305 is configured to perform compliance evaluation on a degree of conformity between the declaration material data and the original policy text on the knowledge graph by using symbolic reasoning based on pre-defined logic rules and neural reasoning based on node embedding representation or graph path;

[0120] The generation unit 306 is configured to generate a government affair text auditing result including a compliance judgment conclusion and corresponding supporting basis according to a result of the compliance evaluation.

[0121] In the system of the embodiment, the functions of the units correspond to the steps in the method embodiments shown in the foregoing Figure 1 or Figure 2 The foregoing will not be described here.

[0122] The application further provides a government affair text auditing device based on knowledge graph reasoning, please refer to Figure 4 , Figure 4 An embodiment of the government affair text auditing device based on knowledge graph reasoning provided by the application, the device comprises:

[0123] A processor 401, a memory 402, an input output unit 403 and a bus 404;

[0124] The processor 401 is connected with the memory 402, the input output unit 403 and the bus 404;

[0125] The memory 402 stores a program, and the processor 401 calls the program to perform any government affair text auditing method based on knowledge graph reasoning.

[0126] The application also relates to a computer readable storage medium, which stores a program, and when the program is run on a computer, the computer executes any one of the above knowledge graph reasoning-based government affair text auditing methods.

[0127] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0128] In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, and the division of units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection between units can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0129] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0130] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0131] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the various embodiments of the method of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, read-only memory), a random access memory (RAM, random access memory), a magnetic disk or an optical disk, and various media that can store program codes.

Claims

1. A method for reviewing government documents based on knowledge graph reasoning, characterized in that, The methods for reviewing government documents include: The original policy text is obtained and semantic parsing is performed on the original policy text using a dual-channel semantic disambiguation engine. The dual-channel semantic disambiguation engine includes a rule channel built based on a government affairs domain ontology library and a neural channel built based on a pre-trained language model. By fusing the semantic parsing results of the rule channel and the neural channel, a normalized entity set that has undergone semantic disambiguation and structured attribute labeling is obtained; Obtain application material data and original case text data related to the original policy text, and map the standardized entity set, the original case text data, and the application material data into a knowledge graph; A graph neural network model is used to learn the node embedding representations in the knowledge graph in order to achieve semantic alignment between the standardized entity set, the original case data, and the application material data. On the knowledge graph, symbolic reasoning based on predefined logical rules and neural reasoning based on the node embedding representation or graph path are used to conduct a compliance assessment on the degree of conformity between the application material data and the original policy text. Based on the results of the compliance assessment, a government document review result is generated, which includes the compliance judgment conclusion and the corresponding supporting evidence.

2. The method for reviewing government documents according to claim 1, characterized in that, The rule channel is used to perform deterministic semantic parsing on legal terms and normative expressions that conform to a preset pattern in the original policy text using the government affairs ontology library; The neural pathway is used to capture the dynamic semantics of ambiguous qualifiers or non-standard expressions in the original policy text through deep context modeling based on an improved ERNIE model.

3. The method for reviewing government documents according to claim 1, characterized in that, The semantic parsing results of fusing the rule channels and the neural channels yield a normalized entity set that has undergone semantic disambiguation and structured attribute labeling, including: Calculate the parsing confidence of semantic parsing results that are simultaneously identified by the rule channel and the neural channel and point to the same text content; The decision weights of the rule channel and the neural channel are determined according to a preset disambiguation weight allocation strategy; By combining the analytical confidence and the decision weight, the semantic parsing results of the rule channel and the neural channel are selected, merged, or modified to obtain a normalized entity set that has undergone semantic disambiguation and structured attribute labeling.

4. The method according to claim 1, characterized in that, The process of acquiring application material data and original case text data related to the original policy text, and mapping the standardized entity set, the original case text data, and the application material data to a knowledge graph, includes: The entities representing policy clauses and legal concepts in the standardized entity set and their relationships are converted into nodes and edges of the policy dimension in a knowledge graph; Information is extracted from the original case data related to the original policy text, extracting target information such as case focus, judgment reasoning, and cited legal provisions. The target information is then converted into nodes of the case dimension in the knowledge graph, and associations are established with the policy dimension nodes. The content of the application materials data is parsed into structured data items, and the data items are converted into nodes and attributes of the material dimension in the knowledge graph.

5. The method according to claim 4, characterized in that, The step of learning the node embedding representations in the knowledge graph using a graph neural network model to achieve semantic alignment between the standardized entity set, the original case data, and the application material data includes: By utilizing the attention and aggregation mechanisms built into the graph neural network model, the feature information of each entity node from its heterogeneous neighborhood under the policy dimension, the case dimension, and the material dimension is iteratively aggregated to generate a node embedding representation that can capture cross-dimensional contextual dependencies. The heterogeneous neighborhood contains entity nodes of different dimensions and diverse relational edges. The node embedding representation is optimized and learned based on a preset loss function until the core semantic concepts of data in each dimension achieve consistent representation in the embedding space. The loss function is used to maximize the representation similarity of semantically related entity nodes in the embedding space, while minimizing the representation similarity of unrelated entity nodes.

6. The method according to any one of claims 1 to 5, characterized in that, The compliance assessment, conducted on the knowledge graph using symbolic reasoning based on predefined logical rules and neural reasoning based on node embedding representations or graph paths, examines the degree of conformity between the submitted material data and the original policy text. This includes: Extract structured predefined logical rules from the provisions of the original policy text or from the constraints defined in the government affairs ontology library; On the knowledge graph, symbolic reasoning is used to perform preliminary verification on the compliance checkpoints in the application material data that can directly correspond to the predefined logical rules, and output preliminary conclusions of certain compliance, partial compliance, non-compliance, or need for further analysis. For cases where the preliminary conclusion is partially compliant, requires further analysis, or involves semantic complexity in the compliance checkpoints, neural reasoning methods are used to perform deep semantic matching by analyzing the node embedding representation or graph path to generate auxiliary evaluation opinions.

7. The method according to claim 6, characterized in that, The process of generating a government document review result that includes a compliance judgment conclusion and corresponding supporting evidence based on the compliance assessment results includes: The preliminary conclusions output by the symbolic reasoning and the auxiliary evaluation opinions generated by the neural reasoning are integrated to obtain a comprehensive evaluation result, and the final compliance judgment conclusion of the application material data is determined based on the comprehensive evaluation result; From the knowledge graph and the processes of symbolic reasoning and neural reasoning, obtain supporting evidence directly related to the final compliance judgment conclusion; The final compliance judgment and supporting evidence are organized and formatted to generate the government document review result.

8. A government document review system based on knowledge graph reasoning, characterized in that, The government document review system includes: The acquisition unit is used to acquire the original policy text and perform semantic parsing processing on the original policy text using a dual-channel semantic disambiguation engine. The dual-channel semantic disambiguation engine includes a rule channel built based on a government domain ontology library and a neural channel built based on a pre-trained language model. The fusion unit is used to fuse the semantic parsing results of the rule channel and the neural channel to obtain a normalized entity set after semantic disambiguation and structured attribute labeling; The mapping unit is used to acquire application material data and case text data related to the original policy text, and to map the standardized entity set, the case text data and the application material data to the knowledge graph; The alignment unit is used to learn the node embedding representation in the knowledge graph using a graph neural network model, so as to achieve semantic alignment between the normalized entity set, the original case data, and the application material data. The reasoning unit is used to perform a compliance assessment on the knowledge graph by employing symbolic reasoning based on predefined logical rules and neural reasoning based on the node embedding representation or graph path to assess the degree of conformity between the application material data and the original policy text. The generation unit is used to generate a government document review result that includes a compliance judgment conclusion and corresponding supporting evidence based on the results of the compliance assessment.

9. A government document review device based on knowledge graph reasoning, characterized in that, The government document review device includes: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method as described in any one of claims 1 to 7.

Images