Policy knowledge graph construction method and system based on digital human interaction data analysis

Abstract

The present invention relates to the field of knowledge graph construction, and specifically to a policy knowledge graph construction method and system based on digital human interaction data analysis. The policy knowledge graph construction system based on digital human interaction data analysis includes: a knowledge graph preliminary construction module, a first graph update module, a second graph update module and a digital human interaction module. The present invention realizes the dynamic maintenance of the policy knowledge graph throughout its life cycle by real-time monitoring of changes in policy source web page content and deep mining of questioning information in digital human interaction data; on the one hand, it accurately captures web page policy update events and automatically completes knowledge injection in combination with a semantic unit matching algorithm; on the other hand, it uses conversation semantic analysis to extract conflicting features of user questions in interaction data, and triggers graph correction after confidence assessment and multi-level verification, ensuring that the end-to-end timeliness from policy release to user perception is controlled within a reasonable period, significantly improving the policy response accuracy of digital human services.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph construction, and specifically to a policy knowledge graph construction method and system based on digital human interaction data analysis. Background Art

[0002] In the fields of industry operations and public services, accurate, timely and easy-to-understand policy information is crucial; policy and regulatory systems are usually characterized by a wide range of sources, frequent updates, complex inter-article linkages, and significant regional differences. Traditional policy information management and service models mainly rely on manual review and updating of static document libraries or simple database query systems, which have significant bottlenecks; on the one hand, due to the diverse channels for issuing policies by management agencies at all levels and the lack of unified standards, it is difficult to achieve automatic integration and deduplication of information, resulting in scattered information and low retrieval efficiency. Users need to spend a lot of time piecing together complete rules; on the other hand, policy dynamics change frequently, and key details such as subsidy amounts, eligibility conditions, application deadlines, and compliance requirements are frequently updated. It is difficult for existing systems to automatically perceive and synchronize the latest content, which can easily cause information lags or version confusion; manual monitoring of numerous publishing websites is inefficient and costly, and it is difficult to ensure real-time performance, which cannot meet the "time-sensitive" policy requirements.

[0003] In addition, with the widespread application of intelligent interactive technologies such as virtual customer service and digital humans in the public service field, users expect to be able to obtain authoritative and accurate policy answers instantly through natural language dialogues with digital humans. However, the existing interactive systems based on static knowledge bases often rely on preset timed tasks or manual triggering by administrators to update knowledge, making it difficult to timely absorb the latest changes in policy sources, and lack effective processing mechanisms for potential information biases in user interactions (such as rule contradictions, numerical discrepancies, and policy expiration issues). The valuable real-time feedback information contained in these user interactions is difficult to be timely and automatically converted into a driving force for knowledge updates due to the lack of effective semantic analysis, credibility assessment mechanisms, and precise conflict positioning methods for policy knowledge graph nodes. This limits the self-improvement ability of the knowledge base and reduces users' trust in the authority and reliability of digital human services.

[0004] In response to the above problems, the present invention proposes a policy knowledge graph construction method and system based on digital human interaction data analysis, enabling digital humans to provide more accurate, authoritative and real-time policy responses during interaction. Summary of the Invention

[0005] The present invention realizes the dynamic maintenance of the policy knowledge graph throughout its life cycle by real-time monitoring of changes in policy source web page content and in-depth mining of questioning information in digital human interaction data. On the one hand, it accurately captures web page policy update events and automatically completes knowledge injection in combination with the semantic unit matching algorithm. On the other hand, it uses dialogue semantic analysis to extract conflicting features of user questions in interaction data, and triggers graph correction after confidence assessment and multi-level verification. The dual-path update mode breaks through the lag defect of traditional static knowledge bases, ensuring that the end-to-end timeliness from policy release to user perception is controlled within a reasonable period, significantly improving the policy response accuracy of digital human services.

[0006] The policy knowledge graph construction method based on digital human interaction data analysis includes:

[0007] Obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing;

[0008] Set a monitoring time point, and for each data source web page, calculate the hash value of its DOM tree structure at the monitoring time point. If the current hash value is inconsistent with the recorded value at the previous monitoring time point, extract the web page text data and metadata of the data source web page, and the metadata at least includes the name of the issuing agency, the policy effective date text, and the applicable geographical area code; compare the extracted web page text data with the corresponding policy nodes in the policy knowledge graph, remove duplicate text paragraphs, and extract the text fragments to be updated in the web page text data; divide the text fragments to be updated into independent policy semantic units according to the semantic boundaries of the policy terms, and use the semantic encoding model to map each semantic unit and the corresponding node text in the knowledge graph into a high-dimensional vector and calculate the similarity. When the similarity is lower than the preset similarity threshold, create a new node in the knowledge graph; when the similarity is higher than the preset similarity threshold, merge the semantic unit content into the matching node and append the version revision label, thereby completing the first graph update operation;

[0009] Collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and select a dialogue subset containing policy questions, feedback on clause contradictions, and questions about policy timeliness; parse the dialogue subset and generate a policy knowledge update request, which includes the graph node ID of the user's question, the feedback text summary, and the conflict attribute field;

[0010] The comprehensive confidence of the policy knowledge update request is calculated based on the accuracy of historical user feedback, the degree of numerical deviation, the frequency of recent changes to the target node, and the completeness of the semantic expression. The policy knowledge update request with a comprehensive confidence that reaches the preset threshold is marked as a valid policy knowledge update request. The knowledge conflict detection process is performed on the valid policy knowledge update request to verify the current version status, numerical legitimacy and real-time effectiveness of the target node in turn, and locate the node to be revised and its dependent path; for the node to be revised, the latest policy terms are automatically captured according to the conflict type or a structured task package for manual review is generated, and the node attribute correction and version rollback are executed, thereby completing the second graph update operation;

[0011] The digital human uses the updated policy knowledge graph to respond to user queries in subsequent interactions.

[0012] Preferably, the first graph updating operation specifically includes:

[0013] Based on the semantic boundary features of policy clauses, including semicolon separation, clause number prefixes, and keyword-guided sentence structures, the updated text fragments are segmented into independent policy semantic units. Each semantic unit contains a complete policy condition statement or policy rule calculation expression.

[0014] A pre-trained semantic encoding model is used to convert the semantic unit text and the node original text into high-dimensional vector representations. The spatial distance metric between the two vectors is calculated. When the metric is lower than the preset similarity threshold, a new node is created in the knowledge graph and the current semantic unit content is injected. When the metric reaches the preset threshold, the current semantic unit content is merged into the matching node and a revision identification label is added to the node.

[0015] For policy nodes whose effective date is earlier than the current date, their status mark is modified to the historical version; for newly effective policy nodes, the calculation status mark of the association rule chain is set to valid, allowing digital humans to call the association rule chain to perform calculations or reasoning of corresponding policy rules; synchronously adjust the logical relationship attributes between related entities, recalculate the conditional dependency strength between nodes based on the logical relationship type and constraint word density defined in the policy terms, and update the relationship edge weight coefficient.

[0016] Preferably, the conversation subset is parsed to generate a policy knowledge update request, and the specific operations are as follows:

[0017] The BERT multi-label classification model was used to identify intent in a subset of conversations, marking conversation segments containing the three intent categories of "rule error," "value mismatch," and "policy expired." Entity extraction was performed on these marked conversation segments to identify the policy document number, clause number, and relevant policy calculation parameter values ​​mentioned by the user.

[0018] Perform a structured comparison between the user-described policy clause text and the corresponding node attribute values ​​to extract numerical deviation fields and logical contradiction fields. Numerical deviation field processing includes identifying the user-described amount, percentage, and date values ​​and calculating the absolute difference between them and the node's stored value. Logical contradiction field processing includes detecting conflicts between the user-described rule conditions and the node's stored rules.

[0019] Combine the node ID, deviation field list, contradiction field list, conversation occurrence timestamp, and user anonymous ID to generate a policy knowledge update request.

[0020] Preferably, the confidence level of the policy knowledge update request is calculated as follows:

[0021] Four evaluation dimensions are defined, including the accuracy of historical user feedback, the degree of numerical deviation, the frequency of policy changes, and the clarity of semantic expression;

[0022] Calculation of the accuracy of user historical feedback: query the user's historical feedback requests and count the proportion of those that have been manually reviewed and confirmed to be valid;

[0023] Calculation of numerical deviation: For numerical difference fields, the percentage of the difference between the user value and the node value in the legal range width is calculated. The legal range width is the allowable value range clearly defined by policy and regulatory documents. For logical conflict fields, the logical implication conflict between the user-defined rule conditions and the node storage rules is detected.

[0024] Calculation of policy change frequency: By querying the version change record library of the policy file associated with the target node, a frequency index is generated based on the number of changes within a preset time window;

[0025] Semantic clarity uses a natural language understanding model to score the semantic completeness of the original user feedback;

[0026] Configure a preset weight distribution strategy, weight the four evaluation dimensions and combine them into a comprehensive confidence score to obtain the confidence of the policy knowledge update request.

[0027] Preferably, a knowledge conflict detection process is performed on the valid policy knowledge update request to locate the node to be corrected and its dependent path. The specific operations are as follows:

[0028] First, verify the node status of the valid policy knowledge update request and confirm that the node is marked as the current effective version rather than the historical archive version;

[0029] Then verify the legitimacy of the value and compare the value with the allowed range of the value domain defined in the policy and regulatory documents;

[0030] Finally, to verify the real-time effectiveness, call the official status query interface of the policy issuing agency, enter the policy document number and clause sequence number to obtain the real-time status code; when the interface returns a status code indicating that the clause is abolished or suspended, the node status synchronization update operation is triggered;

[0031] For valid policy knowledge update requests that have passed verification, the target node is marked as "pending correction"; the node path that has a policy calculation dependency or rule condition association relationship with the target node is retrieved in the knowledge graph; all end nodes in the path are added to the set to be tested; and a machine-readable verification report document is output, which includes the target node identifier, verification result status codes at each level, and a list of associated node identifiers.

[0032] Preferably, the second graph updating operation specifically includes:

[0033] Based on the verification report output by the knowledge conflict detection process, the nodes to be revised are classified and processed. For nodes marked with expired data, the policy issuing agency's data source is accessed to capture the latest policy text in real time. The clauses related to the node attributes in the captured text are extracted as revised data, and the node attribute values ​​and version effective timestamps are updated;

[0034] For nodes where logical contradictions are detected, a structured task package is pushed to the manual review terminal. The task package contains the current attribute data set of the target node, user feedback evidence text, visual presentation of related policy documents, and historical change records of the node; the correction data and operation instructions input by the manual review terminal are received. When it is confirmed that the node data is wrong, the version rollback operation is performed to the most recent valid historical version and the correction data confirmed by manual review is injected. When it is confirmed that the policy source update is lagging behind, the real-time monitoring task of the data source is immediately triggered.

[0035] The policy knowledge graph construction system based on digital human interaction data analysis includes:

[0036] A preliminary knowledge graph construction module is used to obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing;

[0037] The first graph update module includes a web page monitoring unit and a first graph update unit; the web page monitoring unit is used to set a monitoring time point, and for each data source web page, calculates the hash value of its DOM tree structure at the monitoring time point, and if the current hash value is inconsistent with the recorded value at the previous monitoring time point, extracts the web page text data and metadata of the data source web page; the first graph update unit is used to compare the extracted web page text data with the corresponding policy nodes in the policy knowledge graph, remove duplicate text paragraphs, and extract the text fragments to be updated in the web page text data; the text fragments to be updated are divided into independent policy semantic units according to the semantic boundaries of the policy terms, and each semantic unit is mapped to the corresponding node text in the knowledge graph into a high-dimensional vector using a semantic encoding model and the similarity is calculated. When the similarity is lower than a preset similarity threshold, a new node is created in the knowledge graph; when the similarity is higher than the preset similarity threshold, the semantic unit content is merged into the matching node and a version revision label is appended, thereby completing the first graph update operation;

[0038] The second graph update module includes a request generation unit, a request verification unit, and a second graph update unit; the request generation unit is used to collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and screen out dialogue subsets containing policy questions, feedback on clause contradictions, and questions about policy timeliness; the dialogue subset is parsed to generate a policy knowledge update request; the request verification unit is used to calculate the comprehensive confidence of the policy knowledge update request based on the accuracy of historical user feedback, the degree of numerical deviation, the recent change frequency of the target node, and the completeness of the semantic expression, and mark the policy knowledge update request whose comprehensive confidence reaches a preset threshold as a valid policy knowledge update request, and perform a knowledge conflict detection process on the valid policy knowledge update request, verifying the current version status, numerical legitimacy, and real-time effectiveness of the target node in turn, and locating the node to be corrected and its dependent path; the second graph update unit is used to automatically capture the latest policy terms or generate a structured task package for manual review for the node to be corrected according to the conflict type, execute node attribute correction and version rollback, thereby completing the second graph update operation;

[0039] The digital human interaction module is used to respond to user queries using the updated policy knowledge graph.

[0040] The present invention has the following advantages:

[0041] 1. The present invention realizes the dynamic maintenance of the policy knowledge graph throughout its life cycle by real-time monitoring of changes in policy source web page content and in-depth mining of questioning information in digital human interaction data. On the one hand, it accurately captures web page policy update events and automatically completes knowledge injection in combination with the semantic unit matching algorithm. On the other hand, it uses dialogue semantic analysis to extract conflicting features of user questions in interaction data, and triggers graph correction after confidence assessment and multi-level verification. The dual-path update mode breaks through the lag defect of traditional static knowledge bases, ensuring that the end-to-end timeliness from policy release to user perception is controlled within a reasonable period, significantly improving the policy response accuracy of digital human services.

[0042] 2. The present invention effectively solves the problem of identifying implicit policy conflicts by modeling the logical constraint relationship between clauses and integrating a real-time effectiveness verification mechanism. The system can not only detect explicit numerical deviations, but also identify the necessary and sufficient relationship conflicts between policy calculation rules and applicable conditions. When user feedback reveals rule contradictions, the scope of influence is located through associated node path analysis, and the effectiveness of the clauses is verified in real time in combination with the policy agency interface. This mechanism transforms traditional manual experience judgment into an automated and verifiable process, significantly improving the reliability of decision-making in complex policy scenarios while reducing operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a structural diagram of the policy knowledge graph construction system based on digital human interaction data analysis adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0045] Example 1, a method for constructing a policy knowledge graph based on digital human interaction data analysis, includes:

[0046] Obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing. Specific operations include:

[0047] The RoBERTa-wwm-ext-Chinese pre-trained model is used for named entity recognition, identifying five types of entities in the text: institutional entities, such as "Ministry of Finance"; policy subject entities, such as "housing subsidy"; clause entities, such as "application conditions"; amount entities, such as "5 million yuan"; and time entities, such as "60 natural days." SpaCy's Chinese dependency parsing component is also used to extract relationships in combination with manually defined rule templates. For example, when a dependency association between the core predicate structure "not exceeding" and the controlled object "subsidy amount" and the quantity "10,000 yuan" is detected, a triple structure of "subsidy rule"-[amount upper limit]→"10,000 yuan" is automatically generated.

[0048] Then it enters the knowledge fusion stage. The entity alignment process adopts the string similarity calculation based on the Jaro-Winkler algorithm and the context semantic vector matching method generated by Sentence-BERT. For entities whose comprehensive similarity exceeds the similarity threshold, such as "Ministry of Human Resources and Social Security" and "Ministry of Human Resources and Social Security", a standardized merge operation is performed; when there is a conflict in attributes, the latest national policy is automatically selected as the main version according to the rule that national policies are greater than local policies and new policies are greater than old policies; the fused structured knowledge is imported into the Neo4j graph database, and various entities are instantiated and relationship connections are constructed through node creation operations. For example, a relationship with rule association is established between the policy topic node and the clause node. Finally, a policy knowledge graph that can be used for subsequent reasoning and query is generated.

[0049] A monitoring time point is set. For each data source web page, a hash value of its DOM tree structure is calculated at the monitoring time point as a unique identification code. If the current hash value is inconsistent with the value recorded at the previous monitoring time point, the web page text data and metadata of the data source web page are extracted, and the metadata at least includes the name of the issuing agency, the policy effective date text, and the applicable geographical area code; the extracted web page text data is compared with the corresponding policy node in the policy knowledge graph, duplicate text paragraphs are removed, and the text fragments to be updated in the web page text data are extracted; the text fragments to be updated are divided into independent policy semantic units according to the semantic boundaries of the policy terms, and each semantic unit and the corresponding node text in the knowledge graph are mapped into high-dimensional vectors using a semantic encoding model and the similarity is calculated. When the similarity is lower than a preset similarity threshold, a new node is created in the knowledge graph; when the similarity is higher than the preset similarity threshold, the semantic unit content is merged into the matching node and a version revision tag is appended, thereby completing the first graph update operation; this process realizes the automatic perception and targeted capture of policy source changes, ensuring the efficient linkage between external information updates and internal knowledge systems;

[0050] Collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and screen out dialogue subsets containing policy questions, feedback on clause contradictions, and questions about policy timeliness; parse the dialogue subsets and generate policy knowledge update requests, which contain the graph node ID questioned by the user, feedback text summary, and conflict attribute fields; calculate the comprehensive confidence of the policy knowledge update request based on the accuracy of historical user feedback, the degree of numerical deviation, the recent change frequency of the target node, and the completeness of the semantic expression; mark the policy knowledge update request whose comprehensive confidence reaches the preset threshold as a valid policy knowledge update request, and perform a knowledge conflict detection process on the valid policy knowledge update request, verifying the current version status, numerical legitimacy, and real-time effectiveness of the target node in turn, and locate The nodes to be revised and their dependent paths; for the nodes to be revised, the latest policy terms are automatically captured according to the conflict type or a structured task package for manual review is generated, and node attribute revisions and version rollbacks are performed to complete the second graph update operation; this mechanism collects the interaction data between users and digital humans in real time, and uses semantic intent classification technology to screen out key dialogue fragments involving policy doubts, clause contradictions, and questions about timeliness; these dialogues are then parsed to generate structured update requests, including the questioned node ID, feedback summary, and conflict attribute fields; after the validity is verified by confidence assessment, knowledge conflict detection is performed on high-confidence requests, accurately locating the nodes to be revised and triggering graph updates; this process converts user feedback into knowledge graph optimization power, realizing a closed-loop response from interactive questioning to system correction;

[0051] The digital human uses the updated policy knowledge graph to respond to user queries in subsequent interactions.

[0052] The first graph updating operation specifically includes:

[0053] Based on the semantic boundary features of policy clauses, including semicolon separation, clause number prefixes, and keyword-guided sentence structures, the text fragments to be updated are segmented into independent policy semantic units. Each semantic unit contains a complete policy condition statement or policy rule calculation expression. Policy clauses refer to the smallest text unit in management regulations with independent legal effect or execution significance. They are usually expressed as a single regulation, an independent calculation rule, or a complete state constraint. Each clause carries an inseparable policy intention and constitutes the core content source of the knowledge graph node.

[0054] A pre-trained semantic encoding model is used to convert the semantic unit text and the original node text into high-dimensional vector representations respectively; the spatial distance metric between the two vectors is calculated; when the metric is lower than the preset similarity threshold, a new node is created in the knowledge graph and the current semantic unit content is injected; when the metric reaches the preset threshold, the current semantic unit content is merged into the matching node and a version revision identification label is added to the node; the spatial distance metric refers to the cosine similarity score between the two vectors after converting the text into a high-dimensional vector using a pre-trained semantic encoding model (such as Sentence-BERT); the value range is [-1, 1], and the closer the value is to 1, the higher the semantic similarity;

[0055] For policy nodes with an effective date earlier than the current date, their status is modified to a historical version. For newly effective policy nodes, the calculation status of the association rule chain is set to valid, allowing the digital human to call the association rule chain to perform calculations or reasoning on the corresponding policy rules. The logical relationship attributes between the associated entities are synchronously adjusted, and the conditional dependency strength between the nodes is recalculated based on the logical relationship type and constraint word density defined in the policy terms, and the relationship edge weight coefficient is updated. The constraint word density refers to the frequency ratio of keywords with mandatory constraints (such as "must", "prohibited", and "shall not exceed") in the policy terms text. The conditional dependency strength refers to the quantitative value of the constraint strength of the logical relationship between two nodes in the knowledge graph, which is determined by the basic relationship type and constraint word density. The calculation status of the rule chain refers to the logical switch attribute that controls the digital human's call for calculation or reasoning on the corresponding policy rules, which includes three states: valid: can be called by the digital human to perform real-time calculations; historical: only provides query and cannot be called; frozen: the policy effectiveness is suspended for a certain period of time. The relationship edge weight coefficient refers to the value of the logical constraint strength carried by the relationship edge connecting the nodes in the knowledge graph, which is mapped to the [0,1] interval through the conditional dependency strength.

[0056] Parse the conversation subset and generate a policy knowledge update request. The specific operations are as follows:

[0057] The BERT multi-label classification model is used to identify intent in a subset of conversations, marking conversation segments containing the three intent categories of "rule error," "value mismatch," and "policy expired." Entity extraction is performed on these marked conversation segments to identify the policy document numbers, clause numbers, and relevant policy calculation parameter values ​​mentioned by users. The BERT multi-label classification model refers to the fine-tuning of a pre-trained language model based on the Transformer architecture in the policy domain. Entity extraction refers to the technical process of extracting structured policy elements from conversation text.

[0058] Perform a structured comparison between the user-described policy terms and the corresponding node attribute values, extracting numerical deviation fields and logical contradiction fields. Numerical deviation field processing includes identifying the user-expressed amount, percentage, and date values, and calculating the absolute difference with the node storage value. Logical contradiction field processing includes detecting conflicts between the user-described rule conditions and the node storage rules (including but not limited to missing sufficient conditions and redundant necessary conditions). Structured comparison involves machine-parseable difference detection between the user-described text and the knowledge graph node attributes, including two channels: numerical deviation channel and logical contradiction channel. The logical contradiction channel includes missing sufficient conditions detection and missing necessary conditions detection.

[0059] Combine the node ID, deviation field list, contradiction field list, conversation occurrence timestamp, and user anonymous ID to generate a policy knowledge update request.

[0060] Calculate the confidence of the policy knowledge update request as follows:

[0061] Four evaluation dimensions are defined, including the accuracy of historical user feedback, the degree of numerical deviation, the frequency of policy changes, and the clarity of semantic expression;

[0062] Calculation of the accuracy of user historical feedback: query the user's historical feedback requests and count the proportion of those that have been manually reviewed and confirmed to be valid;

[0063] Calculation of numerical deviation: For numerical difference fields, the percentage of the difference between the user value and the node value in the legal range width is calculated. The legal range width is the allowable value range clearly defined in policy and regulatory documents. For logical conflict fields, the logical implications of conflicts between the user-described rule conditions and the node storage rules are detected (including but not limited to missing sufficient conditions and redundant necessary conditions).

[0064] Calculation of policy change frequency: By querying the version change record library of the policy file associated with the target node, a frequency index is generated based on the number of changes within a preset time window;

[0065] Semantic clarity uses a natural language understanding model to score the semantic completeness of the original user feedback;

[0066] Configure a preset weight allocation strategy, weight the four evaluation dimensions and combine them into a comprehensive confidence score to obtain the confidence of the policy knowledge update request. The weight coefficient of each dimension is determined by experts based on experience.

[0067] Perform a knowledge conflict detection process on valid policy knowledge update requests to locate the nodes to be corrected and their dependent paths. The specific operations are as follows:

[0068] First, verify the node status of the valid policy knowledge update request and confirm that the node is marked as the current effective version rather than the historical archive version;

[0069] Then verify the legitimacy of the value and compare the value with the allowed range of the value domain defined in the policy and regulatory documents;

[0070] Finally, to verify the real-time effectiveness, call the official status query interface of the policy issuing agency, enter the policy document number and clause sequence number to obtain the real-time status code; when the interface returns a status code indicating that the clause is abolished or suspended, the node status synchronization update operation is triggered;

[0071] For valid policy knowledge update requests that have passed verification, the target node is marked as "pending correction"; the node path that has a policy calculation dependency or rule condition association relationship with the target node is retrieved in the knowledge graph; all end nodes in the path are added to the set to be tested; and a machine-readable verification report document is output, which includes the target node identifier, verification result status codes at each level, and a list of associated node identifiers.

[0072] The second graph updating operation specifically includes:

[0073] Based on the verification report output by the knowledge conflict detection process, the nodes to be revised are classified and processed. For nodes marked with expired data, the policy issuing agency's data source is accessed to capture the latest policy text in real time. The clauses related to the node attributes in the captured text are extracted as revised data, and the node attribute values ​​and version effective timestamps are updated;

[0074] For nodes where logical contradictions are detected, a structured task package is pushed to the manual review terminal. The task package contains the current attribute data set of the target node, user feedback evidence text, visual presentation of related policy documents, and historical change records of the node; the correction data and operation instructions input by the manual review terminal are received. When it is confirmed that the node data is wrong, the version rollback operation is performed to the most recent valid historical version and the correction data confirmed by manual review is injected. When it is confirmed that the policy source update is lagging behind, the real-time monitoring task of the data source is immediately triggered.

[0075] Example 2: A policy knowledge graph construction system based on digital human interaction data analysis, such as Figure 1 As shown, including:

[0076] A preliminary knowledge graph construction module is used to obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing;

[0077] The first graph update module includes a web page monitoring unit and a first graph update unit; the web page monitoring unit is used to set a monitoring time point, and for each data source web page, calculates the hash value of its DOM tree structure at the monitoring time point, and if the current hash value is inconsistent with the recorded value at the previous monitoring time point, extracts the web page text data and metadata of the data source web page; the first graph update unit is used to compare the extracted web page text data with the corresponding policy nodes in the policy knowledge graph, remove duplicate text paragraphs, and extract the text fragments to be updated in the web page text data; the text fragments to be updated are divided into independent policy semantic units according to the semantic boundaries of the policy terms, and each semantic unit is mapped to the corresponding node text in the knowledge graph into a high-dimensional vector using a semantic encoding model and the similarity is calculated. When the similarity is lower than a preset similarity threshold, a new node is created in the knowledge graph; when the similarity is higher than the preset similarity threshold, the semantic unit content is merged into the matching node and a version revision label is appended, thereby completing the first graph update operation;

[0078] The second graph update module includes a request generation unit, a request verification unit, and a second graph update unit; the request generation unit is used to collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and screen out dialogue subsets containing policy questions, feedback on clause contradictions, and questions about policy timeliness; the dialogue subset is parsed to generate a policy knowledge update request; the request verification unit is used to calculate the comprehensive confidence of the policy knowledge update request based on the accuracy of historical user feedback, the degree of numerical deviation, the recent change frequency of the target node, and the completeness of the semantic expression, and mark the policy knowledge update request whose comprehensive confidence reaches a preset threshold as a valid policy knowledge update request, and perform a knowledge conflict detection process on the valid policy knowledge update request, verifying the current version status, numerical legitimacy, and real-time effectiveness of the target node in turn, and locating the node to be corrected and its dependent path; the second graph update unit is used to automatically capture the latest policy terms or generate a structured task package for manual review for the node to be corrected according to the conflict type, execute node attribute correction and version rollback, thereby completing the second graph update operation;

[0079] The digital human interaction module is used to respond to user queries using the updated policy knowledge graph.

[0080] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A policy knowledge graph construction method based on digital human interaction data analysis, characterized by: include: Obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing; Set a monitoring time point, and for each data source webpage, calculate the hash value of its DOM tree structure at the monitoring time point. If the current hash value is inconsistent with the value recorded at the previous monitoring time point, extract the webpage text data and metadata of the data source webpage, wherein the metadata at least includes the name of the issuing organization, the text of the policy effective date, and the applicable geographic region code; The extracted web page text data is compared with the corresponding policy nodes in the policy knowledge graph, duplicate text paragraphs are removed, and the text fragments to be updated in the web page text data are extracted; the text fragments to be updated are divided into independent policy semantic units according to the semantic boundaries of the policy terms, and each semantic unit and the corresponding node text in the knowledge graph are mapped into a high-dimensional vector using a semantic encoding model and the similarity is calculated. When the similarity is lower than a preset similarity threshold, a new node is created in the knowledge graph; when the similarity is higher than the preset similarity threshold, the semantic unit content is merged into the matching node and a version revision label is appended, thereby completing the first graph update operation; Collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and select a dialogue subset containing policy questions, feedback on clause contradictions, and questions about policy timeliness; parse the dialogue subset and generate a policy knowledge update request, which includes the graph node ID of the user's question, the feedback text summary, and the conflict attribute field; The comprehensive confidence of policy knowledge update requests is calculated based on the accuracy of historical user feedback, the degree of numerical deviation, the frequency of recent changes to the target node, and the completeness of the semantic expression. Policy knowledge update requests with a comprehensive confidence that reaches a preset threshold are marked as valid policy knowledge update requests. A knowledge conflict detection process is performed on valid policy knowledge update requests to sequentially verify the current version status, numerical legitimacy, and real-time effectiveness of the target node, and locate the node to be corrected and its dependent path. For nodes to be corrected, the latest policy terms are automatically captured according to the conflict type, or a structured task package for manual review is generated. Node attribute correction and version rollback are then performed, thereby completing the second graph update operation. The digital human uses the updated policy knowledge graph to respond to user queries in subsequent interactions.

2. The policy knowledge graph construction method based on digital human interaction data analysis according to claim 1 is characterized in that: The first graph updating operation specifically includes: Based on the semantic boundary features of policy clauses, including semicolon separation, clause number prefixes, and keyword-guided sentence structures, the updated text fragments are segmented into independent policy semantic units. Each semantic unit contains a complete policy condition statement or policy rule calculation expression. A pre-trained semantic encoding model is used to convert the semantic unit text and the node original text into high-dimensional vector representations. The spatial distance metric between the two vectors is calculated. When the metric is lower than the preset similarity threshold, a new node is created in the knowledge graph and the current semantic unit content is injected. When the metric reaches the preset threshold, the current semantic unit content is merged into the matching node and a revision identification label is added to the node. For policy nodes whose effective date is earlier than the current date, their status mark is modified to the historical version; for newly effective policy nodes, the calculation status mark of the association rule chain is set to valid, allowing digital humans to call the association rule chain to perform calculations or reasoning of corresponding policy rules; synchronously adjust the logical relationship attributes between related entities, recalculate the conditional dependency strength between nodes based on the logical relationship type and constraint word density defined in the policy terms, and update the relationship edge weight coefficient.

3. The policy knowledge graph construction method based on digital human interaction data analysis according to claim 2 is characterized in that: Parse the conversation subset and generate a policy knowledge update request. The specific operations are as follows: The BERT multi-label classification model was used to identify intent in a subset of conversations, marking conversation segments containing the following three intent categories: "rule error," "value mismatch," and "policy expired." Perform entity extraction on the marked conversation fragments to identify the policy document number, clause number, and relevant policy calculation parameter values ​​mentioned by the user; Perform a structured comparison between the user-described policy clause text and the corresponding node attribute values ​​to extract numerical deviation fields and logical contradiction fields. Numerical deviation field processing includes identifying the user-described amount, percentage, and date values ​​and calculating the absolute difference between them and the node's stored value. Logical contradiction field processing includes detecting conflicts between the user-described rule conditions and the node's stored rules. Combine the node ID, deviation field list, contradiction field list, conversation occurrence timestamp, and user anonymous ID to generate a policy knowledge update request.

4. The method for constructing a policy knowledge graph based on digital human interaction data analysis according to claim 3 is characterized in that: Calculate the confidence of the policy knowledge update request as follows: Four evaluation dimensions are defined, including the accuracy of historical user feedback, the degree of numerical deviation, the frequency of policy changes, and the clarity of semantic expression; Calculation of the accuracy of user historical feedback: query the user's historical feedback requests and count the proportion of those that have been manually reviewed and confirmed to be valid; Calculation of numerical deviation: For numerical difference fields, the percentage of the difference between the user value and the node value in the legal range width is calculated. The legal range width is the allowable value range clearly defined by policy and regulatory documents. For logical conflict fields, the logical implication conflict between the user-defined rule conditions and the node storage rules is detected. Calculation of policy change frequency: By querying the version change record library of the policy file associated with the target node, a frequency index is generated based on the number of changes within a preset time window; Semantic clarity uses a natural language understanding model to score the semantic completeness of the original user feedback; Configure a preset weight distribution strategy, weight the four evaluation dimensions and combine them into a comprehensive confidence score to obtain the confidence of the policy knowledge update request.

5. The method for constructing a policy knowledge graph based on digital human interaction data analysis according to claim 4 is characterized in that: Perform a knowledge conflict detection process on valid policy knowledge update requests to locate the nodes to be corrected and their dependent paths. The specific operations are as follows: First, verify the node status of the valid policy knowledge update request and confirm that the node is marked as the current effective version rather than the historical archive version; Then verify the legitimacy of the value and compare the value with the allowed range of the value domain defined in the policy and regulatory documents; Finally, to verify the real-time effectiveness, call the official status query interface of the policy issuing agency, enter the policy document number and clause sequence number to obtain the real-time status code; when the interface returns a status code indicating that the clause is abolished or suspended, the node status synchronization update operation is triggered; For valid policy knowledge update requests that have passed verification, the target node is marked as "pending revision"; the node path that has a policy calculation dependency or rule condition association relationship with the target node is retrieved in the knowledge graph; all terminal nodes in the path are added to the set to be checked; Output a machine-readable verification report document, which includes the target node identifier, verification result status codes at each level, and a list of associated node identifiers.

6. The method for constructing a policy knowledge graph based on digital human interaction data analysis according to claim 5 is characterized in that: The second graph updating operation specifically includes: Based on the verification report output by the knowledge conflict detection process, the nodes to be revised are classified and processed. For nodes marked with expired data, the policy issuing agency's data source is accessed to capture the latest policy text in real time. The clauses related to the node attributes in the captured text are extracted as revised data, and the node attribute values ​​and version effective timestamps are updated; For nodes where logical contradictions are detected, a structured task package is pushed to the manual review terminal. The task package contains the current attribute data set of the target node, user feedback evidence text, visual presentation of related policy documents, and historical change records of the node; the correction data and operation instructions input by the manual review terminal are received. When it is confirmed that the node data is wrong, the version rollback operation is performed to the most recent valid historical version and the correction data confirmed by manual review is injected. When it is confirmed that the policy source update is lagging behind, the real-time monitoring task of the data source is immediately triggered.

7. A policy knowledge graph construction system based on digital human interaction data analysis, characterized by: The system is applied to the policy knowledge graph construction method based on digital human interaction data analysis as described in any one of claims 1 to 6 above, comprising: A preliminary knowledge graph construction module is used to obtain policy text data from all data sources and convert all policy text data into a policy knowledge graph based on natural language processing; The first graph update module includes a web page monitoring unit and a first graph update unit; the web page monitoring unit is used to set a monitoring time point, and for each data source web page, calculates the hash value of its DOM tree structure at the monitoring time point, and if the current hash value is inconsistent with the recorded value at the previous monitoring time point, extracts the web page text data and metadata of the data source web page; the first graph update unit is used to compare the extracted web page text data with the corresponding policy nodes in the policy knowledge graph, remove duplicate text paragraphs, and extract the text fragments to be updated in the web page text data; the text fragments to be updated are divided into independent policy semantic units according to the semantic boundaries of the policy terms, and each semantic unit is mapped to the corresponding node text in the knowledge graph into a high-dimensional vector using a semantic encoding model and the similarity is calculated. When the similarity is lower than a preset similarity threshold, a new node is created in the knowledge graph; when the similarity is higher than the preset similarity threshold, the semantic unit content is merged into the matching node and a version revision label is appended, thereby completing the first graph update operation; The second graph update module includes a request generation unit, a request verification unit, and a second graph update unit; the request generation unit is used to collect interaction data between users and digital humans, perform semantic intent classification on dialogue segments involving policy queries in the interaction data, and screen out dialogue subsets containing policy questions, feedback on clause contradictions, and questions about policy timeliness; the dialogue subset is parsed to generate a policy knowledge update request; the request verification unit is used to calculate the comprehensive confidence of the policy knowledge update request based on the accuracy of historical user feedback, the degree of numerical deviation, the recent change frequency of the target node, and the completeness of the semantic expression, and mark the policy knowledge update request whose comprehensive confidence reaches a preset threshold as a valid policy knowledge update request, and perform a knowledge conflict detection process on the valid policy knowledge update request, verifying the current version status, numerical legitimacy, and real-time effectiveness of the target node in turn, and locating the node to be corrected and its dependent path; the second graph update unit is used to automatically capture the latest policy terms or generate a structured task package for manual review for the node to be corrected according to the conflict type, execute node attribute correction and version rollback, thereby completing the second graph update operation; The digital human interaction module is used to respond to user queries using the updated policy knowledge graph.

Images