Government affair integration intelligent service platform based on knowledge graph semantic understanding
By using a knowledge graph-based semantic understanding method, the problems of ambiguous references and cross-item role differences in joint government service processing are solved. This enables quantitative assessment and gradual disambiguation of identities, reduces the number of confirmation questions, and improves the accuracy and efficiency of government services.
Patent Information
- Application Number
- CN202611325019.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-10
AI Technical Summary
In the context of joint government service processing, existing technologies struggle to accurately extract referential information from natural language statements and to conduct quantitative assessments in cases of ambiguity and differences in roles across different matters. This results in numerous identity misjudgments and confirmation issues, and a lack of effective means to adjust for mismatched identities.
A knowledge graph-based semantic understanding approach is adopted. Through semantic decomposition and knowledge graph linking, a set of subject semantic nodes and a candidate relationship graph are generated. Role confirmation items are generated and event-level binding is performed. The same referent can be associated with multiple identities. Quantitative evaluation and conservative bias are introduced to gradually disambiguate and unbind when a mismatch is found.
It enables the quantification of the ambiguity of references and the differences in roles across matters, reduces misjudgments of identity, optimizes the number of confirmation issues, and improves the accuracy and efficiency of government services.
Smart Images

Figure CN123068564A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of e-government technology, and more specifically, to an integrated intelligent e-government service platform based on knowledge graph semantic understanding. Background Technology
[0002] With the advancement of the "one-stop service" reform for government services, government service centers across the country have integrated multiple related matters such as birth certificates, birth registration, medical insurance enrollment, and social security card applications into a unified joint service. When handling these matters, the public usually states their needs in natural language, such as, "I need to get a birth certificate for my baby, and I need to register the household registration at my maternal grandmother's house, but the child's father is out of town and can't come back."
[0003] These statements contain multiple references to people (such as "I", "baby", "grandma's house", "dad"). Since the joint service covers multiple matters, and these references correspond to different legal identities in different matters—for example, the same speaker may be the newborn's mother in the birth certificate matter, but may involve the proposed household registration recipient in the household registration matter—if the platform directly establishes these references as a definite identity relationship, it may lead to misjudgment of identity, which in turn will result in errors in the service path and material list for subsequent matters.
[0004] For the above scenarios, the common approach is to first extract personal references from user speech using entity recognition, and then map these references to specific identities based on preset rules or templates (e.g., mapping "I" to the inquirer, and "Dad" to the father). This mapping result serves as the unified identity basis for all joint-processing matters. When the platform is unsure about the identity attribution of a certain reference, it generates a batch of confirmation questions to ask the user, and confirms the identity relationship after receiving the user's response.
[0005] The above processing method has the following shortcomings:
[0006] First, the mapping from reference to identity is usually completed in one go, and the platform does not quantify the reliability of the mapping results. When the reference is ambiguous (e.g., "I" may refer to the mother, the agent, or the head of the household to be registered) and the context information is insufficient to determine a unique solution, this direct mapping is prone to misjudgment of identity, and this misjudgment will affect all subsequent matters.
[0007] Secondly, the correspondence between designations and identities is usually viewed as a whole, meaning that one designation corresponds to only one identity throughout the entire joint processing procedure. However, the same natural person may actually assume different legal roles in different matters, and the overall identity allocation cannot reflect this cross-matter role difference.
[0008] Third, when the platform generates confirmation questions for users, all questions are usually listed together without distinguishing which questions' confirmation results will have a chain reaction affecting multiple subsequent matters, or which questions will only affect a single matter. It also fails to utilize the inherent dependencies between matters to reduce the number of questions.
[0009] Furthermore, when a certain identity is found to be mismatched with the requirements of a task in subsequent tasks, the existing processing methods lack the means to adjust that identity individually, and the mismatched identity relationship may continue to affect the unfinished tasks.
[0010] In summary, in the context of joint government service processing, how to extract referential information about people from natural language statements, quantitatively assess the probability of correspondence between referential information and various legal identities without relying on preset templates for forced matching, gradually disambiguate uncertain parts in a reasonable manner, and make local adjustments when mismatches in identity relationships are found, are technical problems that have not yet been solved by existing technologies. Summary of the Invention
[0011] This invention provides an integrated intelligent e-government service platform based on knowledge graph semantic understanding, comprising:
[0012] S1: Receive the consultation message, perform semantic decomposition on the consultation message, identify the subject referential information and the intention of the matter, and generate a set of subject semantic nodes; wherein, the roles in the set of subject semantic nodes are all marked as candidate states and are not directly bound to a specific handling role;
[0013] S2: Input the set of subject semantic nodes into a preset knowledge graph to generate a subject relationship candidate graph containing multiple possible relationship paths;
[0014] S3: Based on the subject relationship candidate graph, match each candidate role with the role required by each sub-item, identify the boundary confusion of the same subject referent information between multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship;
[0015] S4: Based on the role boundary confirmation chain, bind the confirmed roles at the event level according to the event, and unbind the bindings that do not have a successor relationship, thus forming an event-role rebinding result;
[0016] S5: Based on the role rebinding result of the matter, generate and output the service result, which includes at least the processing path, role description, matters to be confirmed, and material list.
[0017] Furthermore, in S1, the set of generated main semantic nodes includes:
[0018] For the title triggers extracted from the agency consultation discourse, their title explicitness, kinship chain matching degree, event adjacency degree and information adjacency degree are weighted and calculated, and nonlinear enhancement is applied. Then, the attenuation correction of the uncertain expression suppression term is applied to obtain the candidate strength of the title trigger for each preset role.
[0019] For multiple candidate strengths of the same title trigger item, sort them and divide the difference between the highest candidate strength and the second candidate strength by the sum of the highest candidate strength and the stable parameter to get the role discrimination score. When the role discrimination score is lower than a preset threshold, retain multiple candidate roles for the title trigger item and set a temporary binding flag, and do not determine the role for the time being.
[0020] Furthermore, in S2, the generation of the candidate graph of subject relationships includes:
[0021] Each subject semantic node is attached to a standard role node in the knowledge graph. The same subject semantic node is allowed to be attached to multiple standard role nodes at the same time. The candidate role strength, semantic fit, graph relationship support and context fit are weighted and nonlinear enhancement is applied. Then, the attachment strength is obtained by role conflict degree attenuation correction and relationship path length attenuation correction.
[0022] Based on the connection results, multiple candidate roles in the same public expression are connected into a multi-path subject relationship structure. The retention strength is calculated for each path, and the paths that meet the retention conditions are summarized into the subject relationship candidate graph.
[0023] Furthermore, in S3, the organization forming a role boundary confirmation chain includes:
[0024] The carrying strength of the same subject referential information across multiple roles is compared. By combining the ratio of the carrying strength of the second candidate to the carrying strength of the highest candidate, the degree of overlap of candidate roles, and the strength of clear identity evidence, titles with confused role boundaries are identified, and targeted role confirmation items are generated.
[0025] Multiple confirmation items are organized into a role boundary confirmation chain based on the dependencies between items. Confirmation items that affect core items are arranged before those that affect secondary items, so that the role determination result of the first confirmed item directly constrains the range of optional roles for the later confirmed items.
[0026] Furthermore, in S4, the step of binding items at the item level and unbinding items that do not have a successor relationship includes:
[0027] The binding validity of candidate subjects to each task role is calculated. After solidification and judgment, task-level binding is performed on roles that meet the high confidence level condition, roles that meet the intermediate confidence level condition are kept in the candidate state to await further confirmation, and low confidence roles are not bound. The solidification and judgment introduces a conservative bias, so that the platform tends not to perform binding when the binding validity does not reach the high confidence level condition.
[0028] After the binding is completed, the bound roles are checked for any binding relationships that do not have a succession relationship. The strength of the release is determined by comprehensively considering the binding validity, the degree of role conflict, the strength of unresolved confirmation items, and the strength of the established succession basis. For binding relationships whose release strength exceeds the preset threshold, the binding is actively released and the corresponding role is downgraded to a candidate subject or reference subject. At the same time, the role is restricted from entering the processing path of subsequent key matters.
[0029] Furthermore, in S5, the service result also includes: an explanation of the reason for unbinding the role, and an indication of the key processing path that the role cannot access.
[0030] Furthermore, the preset knowledge graph is a knowledge graph of the birth of a newborn; the standard role nodes include at least: consultant, agent, newborn, father, mother, grandparents, head of household, proposed household registration recipient and guardian contact person.
[0031] This invention also provides an integrated intelligent e-government service platform based on knowledge graph semantic understanding, including:
[0032] The main semantic decomposition module is used to receive agency consultation statements, perform semantic decomposition on the agency consultation statements, identify the main referent information and the intention of the matter, and generate a set of main semantic nodes; wherein, the roles in the set of main semantic nodes are all marked as candidate states and are not directly bound to a specific handling role;
[0033] The candidate graph generation module is used to bring the set of main semantic nodes into a preset knowledge graph to generate a candidate graph of main relationships containing multiple possible relationship paths;
[0034] The role boundary confirmation module is used to match each candidate role with the role required by each sub-item based on the subject relationship candidate graph, identify the boundary confusion of the same subject referent information between multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship.
[0035] The task-role rebinding module is used to bind confirmed roles to tasks based on the role boundary confirmation chain, and to unbind bindings that do not have a successor relationship, thus forming a task-role rebinding result.
[0036] The service result output module is used to generate and output service results based on the role rebinding results of the matter. The service results include at least the processing path, role description, matters requiring confirmation, and a list of materials.
[0037] Furthermore, the main semantic decomposition module is further used to: extract the title trigger items from the consultation discourse, perform weighted calculations based on title explicitness, kinship chain matching degree, event adjacency degree, and information adjacency degree, apply nonlinear enhancement, and then apply attenuation correction with uncertain expression suppression term to obtain candidate strength; sort the multiple candidate strengths of the same title trigger item, divide the difference between the highest candidate strength and the second candidate strength by the sum of the highest candidate strength and the stable parameter as the role distinguishability, and set a temporary binding flag when the distinguishability is lower than a preset threshold;
[0038] The aforementioned role rebinding module is further used to: introduce a conservative bias in the solidification judgment, so that the binding validity does not reach the high confidence condition and the binding is not executed; for those bound roles that do not have a receiving relationship, the binding validity, the degree of role conflict, the strength of unresolved confirmation items and the strength of the received basis are comprehensively considered to determine the release strength and perform downgrading processing.
[0039] The technical effects of this invention are as follows:
[0040] This invention, after extracting person references from consultation discourse, does not immediately assign them as specific service identities. Instead, it first generates quantitative strength values for each reference corresponding to different candidate identities, and determines situations where determination needs to be postponed based on the degree of difference between the candidate strengths. Therefore, when references are ambiguous or semantic information is insufficient to support a unique judgment, the platform does not force a definitive assignment, reserving the opportunity to gradually disambiguate using more contextual information.
[0041] This invention, when associating a referential with standard identity nodes in a knowledge graph, allows the same referential to be associated with multiple possible identity nodes simultaneously, and calculates a quantified association strength for each association, forming an intermediate result containing multiple candidate relationship paths. Thus, the ambiguity of a referential is fully preserved in an explicit and measurable form, rather than being discarded at an early stage, providing ample candidate space for downstream disambiguation reasoning.
[0042] This invention, when matching candidate identities with the identities required for various joint-processing matters, automatically identifies claims with blurred role boundaries across multiple matters by comparing the identity carrying strength of the same claim across multiple matters, the degree of overlap of candidate identities, and existing clear identity evidence, and generates targeted confirmation items. Simultaneously, it sorts the confirmation items according to the dependencies between matters, placing confirmation items for core matters with broad impact at the top and confirmation items for secondary matters with narrow impact at the bottom. Thus, the results of the top-ranked confirmation items directly constrain the range of possible identities for the corresponding claims in subsequent matters (for example, once the mother of a newborn is confirmed, the legal applicant for the birth certificate and the parent-child relationship for birth registration are also determined), thereby reducing the number of confirmation items that need to be individually asked of the user.
[0043] This invention binds confirmed identity relationships at the granularity of individual matters, introducing a tiered judgment process: binding is performed for cases with sufficient confidence, candidate status is maintained for cases with insufficient confidence, and binding is not performed for cases with excessively low confidence, with an embedded conservative judgment bias. Simultaneously, a succession check is performed on the bound identity relationships; if a binding relationship is found to be incompatible with the requirements of the corresponding matter, the binding relationship is terminated, blocking its entry into the processing path of subsequent matters. Therefore, the same designation can be assigned different identities in different matters, and errors in identity relationships will not propagate indefinitely to subsequent matters. Attached Figure Description
[0044] Figure 1 A flowchart illustrating a specific implementation of the present invention is shown. Detailed Implementation
[0045] Example 1: See Figure 1 This embodiment provides an integrated intelligent e-government service platform based on knowledge graph semantic understanding, applicable to scenarios involving joint handling of e-government services (such as a single matter like a newborn's birth). It performs semantic understanding on inquiries from the public expressed in natural language, automatically identifies and gradually disambiguates the ambiguity of subject references, and outputs role binding results and service paths at the granularity of each matter. The technical details of each step are described below.
[0046] S1: Deconstruction of the main semantics
[0047] This step is used to receive consultation messages, perform semantic decomposition on the consultation messages, identify the subject referential information and the intent of the matter, and generate a set of subject semantic nodes; wherein, the roles in the set of subject semantic nodes are all marked as candidate states and are not directly bound to a specific handling role.
[0048] In existing semantic parsing schemes, after extracting referents about people from discourse, platforms typically map them directly to a pre-defined role (e.g., seeing "I" is automatically associated with "the person seeking advice"). This direct approach relies on the implicit assumption that "there is a one-to-one correspondence between referents in discourse and legal roles." However, in the context of handling inquiries on behalf of others, the same referent may actually correspond to different legal roles in different sub-items, and the user's discourse itself may contain ambiguous references (e.g., "I" in "I'm going to handle this" could refer to the mother herself or the grandmother of the person handling the matter). If the platform forcibly determines role attribution when information is insufficient, any error will propagate to the processing path of all subsequent matters, and there is a lack of corrective measures.
[0049] Therefore, this embodiment adopts a "candidate state" strategy in the semantic decomposition stage: all roles are first marked as candidates, and the reliability of each candidate affiliation is reflected by quantitative calculation. For claims with insufficient reliability, multiple candidate roles are retained and the determination is postponed, creating conditions for multi-path reasoning and step-by-step disambiguation in subsequent steps.
[0050] Specifically, S1 includes the following sub-steps:
[0051] S11: Discourse Segmentation. The consultation discourse is segmented according to pause words, punctuation marks, verbs indicating matters, kinship terms, and regional terms to obtain several basic sentence fragments. Pause words are used to identify natural breaks in the user's statement; verbs indicating matters are used to identify where the user's intention shifts; kinship terms are used to identify the boundaries of different subjects mentioned; and regional terms typically identify the boundaries of household registration-related information. Each basic sentence fragment retains its original order, source (voice input or text input), and adjacency relationship with preceding and following fragments.
[0052] S12: Title Trigger Extraction. Title triggers are extracted from basic sentence fragments. These triggers include at least first-person pronouns, kinship terms, identity terms, household registration terms, and interjections for handling matters. For compound titles (such as "child's father"), they are broken down into multiple triggers, preserving the kinship relationships between each trigger. The inventors recognize that the components of a compound title may play different roles in different matters—for example, "child's father" refers to the father in a birth certificate, but in household registration, it may involve both parent-child relationship and household registration relationship. After decomposition, the platform can assess the multiple role possibilities of each component separately, rather than forcibly binding the compound title as a whole to a single role. If the compound title is not decomposed, it will be treated as an indivisible atomic designation, and the platform can only assign it one role, thus losing the multi-matter role adaptation space of its components.
[0053] S13: Candidate Strength Calculation. For each title trigger, a weighted calculation is performed based on its title dominance, kinship chain matching degree, event adjacency, and information adjacency, and a nonlinear enhancement is applied. Then, a suppression term for uncertain expression is used for attenuation correction to obtain the candidate strength of the title trigger for each preset role.
[0054] Specifically, in this embodiment, the four positive factors (title dominance, kinship chain matching, event adjacency, and information adjacency) are weighted by preset weight parameters, and a weighted average is taken after applying nonlinear enhancement to each factor. The weighted average result is then multiplied by an exponential decay term determined by the intensity of uncertain expression. The nonlinear enhancement ensures that when multiple factors point to the same role, the candidate strength of that role receives a higher boost than linear superposition. The exponential decay term ensures that the stronger the uncertainty expressed by the user (such as phrases like "maybe," "seems," or "should be"), the greater the reduction in candidate strength.
[0055] Those skilled in the art will understand that the reason for introducing nonlinear enhancement is that, in real-world scenarios, there is a qualitative difference in reliability between situations where multiple factors jointly point to the same role and situations where a single factor points to the same role individually. For example, in a user's statement "I am the child's mother, I am here to apply for a birth certificate," "I" simultaneously possess two factors: kinship title ("mother") and proximity of matter ("applying for a birth certificate"). The candidate strength for becoming a mother should not be merely a simple average of the scores of these two factors, but should be significantly higher than in the case of a single factor. Nonlinear enhancement results in a superlinear boost when multiple pieces of evidence contribute synergistically, thereby widening the gap between "multiple evidence roles" and "single evidence roles," providing a more significant signal for subsequent role differentiation judgments. If linear weighting is used, even if multiple factors jointly point to the same role, the absolute values and relative differences in candidate strength are not significant enough, easily diluting the advantage of multiple evidence roles and triggering unnecessary postponement of binding.
[0056] Furthermore, the reason for introducing an exponential decay term to suppress uncertainty is that the wording used by users when expressing uncertainty is a global signal—it indicates that the user themselves is uncertain about the role they are referring to. This uncertainty should have a compressive effect on all factor scores, rather than affecting only one factor. Using multiplicative suppression in the form of exponential decay means that regardless of how high the scores of each factor are, as long as the user expresses sufficiently strong uncertainty, the overall candidate strength will be suppressed. If a simple subtraction correction is used (i.e., deducting a fixed penalty term from the total score), the penalty effect is relatively weak in high factor score scenarios. The platform might still make a judgment with high confidence even when the user explicitly expresses uncertainty, which contradicts reality.
[0057] S14: Role Differentiation and Delayed Binding. The strength of multiple candidate roles for the same title trigger is sorted, and the role differentiation is calculated. In this embodiment, the role differentiation is the difference between the highest candidate strength and the second candidate strength, divided by the sum of the highest candidate strength and the stability parameter. When the role differentiation is lower than a preset threshold, it is determined that there is a risk of role confusion. Multiple candidate roles are retained for the title trigger, and a delayed binding flag is set, temporarily suspending role determination.
[0058] The reason for using the ratio form instead of absolute difference to calculate role discrimination is that the absolute difference in candidate strength is limited by the numerical range of each factor, making it impossible to uniformly set a meaningful discrimination threshold. For example, when the highest candidate strength is 0.9 and the second candidate strength is 0.7, the difference of 0.2 is not significant in absolute terms, but relatively speaking, the highest candidate strength is nearly 30% higher than the second candidate strength, indicating a clear advantage. Conversely, when the highest candidate strength is 0.15 and the second candidate strength is 0.05, although the absolute difference is only 0.1, relatively speaking, the highest candidate strength is three times that of the second candidate strength, making the advantage more significant. Using the ratio form ensures that the discrimination index is comparable for both high-intensity and low-intensity scenarios, and the threshold setting has cross-scenario consistency.
[0059] The purpose of introducing a stabilizing parameter is to address the fact that when both candidate intensities are extremely low, their absolute difference is unreliable and should not be used to determine sufficient discrimination. The stabilizing parameter makes the discrimination value lower in low-intensity regions, causing the platform to be conservative in these regions—maintaining candidate status rather than risking a role when intensities are generally low.
[0060] This step does not discard low-intensity candidate roles for low-discrimination designations, but rather temporarily retains them. These temporarily retained candidate roles carry information that they "exist but are unclear." In subsequent confirmation chain stages, the platform can verify or exclude these candidate roles by asking users questions. If low-intensity candidate roles are discarded in stage S1, the platform will be unable to recover this information when the actual tasks that should be undertaken by the discarded role occur, leading to incorrect role allocation.
[0061] S15: Generation of the main semantic node set. Candidate roles are organized into a main semantic node set, with each node including: source sentence fragment, candidate role and its candidate strength, role distinguishability, and a deferred binding flag (if any). This set serves as the input to S2.
[0062] S2: Generation of candidate graphs for subject relationships
[0063] This step involves inputting the set of main semantic nodes output by S1 into a pre-defined knowledge graph to generate a candidate graph of main relationships containing multiple possible relationship paths. The knowledge graph pre-stores standard role nodes involved in the joint government affairs scenario and the legal relationships between them.
[0064] S21: Knowledge Graph Connection. Each subject semantic node is connected to a standard role node in the knowledge graph. In this embodiment, the same subject semantic node can be connected to multiple standard role nodes simultaneously. For example, the referent "I" in the discourse may be connected to three standard role nodes simultaneously: "mother," "agent," and "proposed settlement recipient."
[0065] For each connection, the candidate role strength, semantic fit, graph relationship support, and context fit are weighted and nonlinearly enhanced. Then, the connection strength is obtained after suppression and attenuation correction based on role conflict degree and relationship path length. Specifically: the candidate role strength, semantic fit, graph relationship support, and context fit are each weighted by preset weight parameters. After applying nonlinear enhancement to each factor, a weighted average is taken. The weighted average is then multiplied by an attenuation term with role conflict degree as the exponent. Finally, the result is divided by the sum of the product of the normalized relationship path length and the path attenuation coefficient plus one to obtain the connection strength.
[0066] The reason for allowing one-to-many affixing is that in joint application scenarios, the natural language references used by the public are inherently ambiguous—"I" in the joint application process could be any of the following: mother (applying for a birth certificate), agent (applying on behalf of the child's father), or head of household (registering under my own name), and may simultaneously assume multiple roles. If the platform forces a one-to-one assignment during the affixing stage, subsequent steps can only continue reasoning based on this premature and potentially erroneous choice, losing the opportunity to utilize ambiguous information for globally optimal matching. One-to-many affixing postpones decision-making from early hard judgment to later comprehensive judgment, which is a continuation of the aforementioned candidate state mechanism at the knowledge graph level.
[0067] The reason for introducing a role conflict suppression term is that, in a joint office scenario, the same standard role can usually only be assumed by one natural person. When two different subject semantic nodes are simultaneously attached to the same standard role, it indicates a competitive relationship. The attachment strength of both should be reasonably reduced to reflect the platform's uncertainty about the role's attribution. Without conflict suppression, the two nodes might attach to the same role with relatively high strength, and the platform would face a situation where two high-strength candidates compete for the same role and cannot be distinguished—which is essentially equivalent to not performing disambiguation.
[0068] The reason for introducing a path length decay term is that in knowledge graph reasoning, direct relationships (shorter path lengths) are generally more reliable than indirect relationships (longer path lengths). Without penalizing path length, the platform might overestimate the reliability of certain connections due to the extra points provided by indirect paths. The decay term naturally favors short-path reasoning, only referring to long paths when short-path information is insufficient.
[0069] S22: Generation of candidate relationship edges. Based on the attachment results, relationship edges are generated between candidate role nodes. The types of relationship edges include at least kinship relationship edges, guardianship relationship edges, household registration relationship edges, settlement acceptance relationship edges, and matter handling relationship edges.
[0070] S23: Multi-path structure and retention strength. Based on the candidate relation edge set, multiple candidate roles in the same crowd expression are connected into a multi-path subject relation structure. For example, when "I" is simultaneously associated with "mother" and "agent", and "grandmother's house" is associated with "proposed household head", the platform generates multiple possible relation paths, each path describing a complete subject relation hypothesis.
[0071] For each path, calculate the retention strength: considering the connection strength of each node on the path and the internal logical consistency of the path, retain paths that meet the preset retention conditions. The retention strength is a continuous value, which allows subsequent steps to distinguish between highly probable paths and paths that are only considered as alternatives, rather than a simple binary selection.
[0072] This embodiment transmits a multi-path structure downstream rather than a single path. If the candidate graph module directly selects and outputs an optimal path internally, the module assumes the role of disambiguation—but at this point, the platform only utilizes information at the knowledge graph level and has not yet combined it with the dependency relationships and confirmation chain information. Premature single-path decision-making will cause the subsequent role boundary confirmation module to lose the basis for multi-hypothesis reasoning, degenerating into verification based on a single hypothesis, which is not fundamentally different from the existing technology's process of determining roles before taking action.
[0073] S24: Summarize candidate graphs of principal relationships. Summarize the paths that meet the retention criteria into a candidate graph of principal relationships and output it to S3.
[0074] S3: Role Boundary Confirmation Chain
[0075] This step is used to match each candidate role with the role required by each sub-item based on the candidate relationship graph generated by S2, identify the boundary confusion of the same subject referent information between multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship.
[0076] S31: Establishing Role Boundary Templates for Items. Read the candidate graph of subject relationships and establish role boundary templates according to the roles required for each sub-item (such as birth certificate, birth registration, medical insurance enrollment, social security card application, vaccination record keeping, etc.). Each item is broken down into the required role types and the relationship requirements between roles.
[0077] S32: Calculation of Item Bearing Strength. Match each candidate entity with each item role, and calculate the item bearing strength of the candidate entity for the item role by comprehensively considering path retention strength, graph relationship consistency, item stage adaptability, and material association completeness.
[0078] S33: Boundary Confusion Identification. The carrying strength of the same subject's referential information across multiple roles is compared. In this embodiment, the boundary confusion strength is calculated as follows: the sum of the second candidate carrying strength and the stability parameter is divided by the sum of the highest candidate carrying strength and the stability parameter, multiplied by the power of the overlap enhancement of the sum of the candidate role overlap degree and one, and then divided by the power of the power of the explicit basis suppression of the sum of the explicit identity basis strength and one. Referentials with a boundary confusion strength higher than a preset threshold are identified as role boundary confusion items.
[0079] The design considerations for the above calculation method are as follows:
[0080] First, the ratio of the second candidate bearing strength to the highest candidate bearing strength (corrected by the stabilization parameter) is used instead of the ratio directly. The reason is the same as the design of the role differentiation in S1.4: when both bearing strengths are low, the instability of the ratio increases. The stabilization parameter is introduced to make the ratio in the low-strength area regress to one, avoiding misjudgment in low-strength scenarios.
[0081] Secondly, the degree of overlap in candidate roles is used as a multiplicative enhancement term rather than a linear weighting term. The inventors discovered that when the same reference is matched as three different roles in three different matters, its confusion complexity is far greater than when it is matched as two roles in two matters—the more role types there are, the less likely it is to be resolved with a single confirmation, and the user's confirmation burden grows superlinearly. Using multiplicative enhancement enables the platform to more sensitively identify and prioritize these highly confusing items in multi-role overlap scenarios, concentrating confirmation resources on the references that most need disambiguation. If linear addition is used, the platform's sensitivity to two-role overlap and three-role overlap is insufficient, which may lead to highly confusing items not being prioritized.
[0082] Third, explicitly define the strength of the identity verification criteria as an exclusionary factor. When a user has provided a unique identifier such as an ID number, the true identity of the claim can be uniquely determined. In this case, additional confirmation questions should not be generated because the claim can play different roles in multiple matters. Dividing by explicitly defining the identity verification criteria automatically excludes claims with clear evidence from the confirmation list, reducing unnecessary disturbance to the user.
[0083] S34: Confirmation Item Generation and Confirmation Chain Organization. Based on role boundary confusion items, generate targeted role confirmation items (e.g., for a scenario where "I" am a candidate for "mother" in the birth certificate and a candidate for "proposed head of household" in the household registration process, generate a confirmation item "Please confirm whether you are the newborn's mother"). Organize multiple confirmation items into a role boundary confirmation chain according to item dependencies. In the confirmation chain, confirmation items affecting core items are listed first, followed by those affecting secondary items, ensuring that the role determination result of the first confirmed item directly determines the available role range for the subsequent confirmed items.
[0084] The reason for prioritizing core confirmation items is that they confirm the most basic identity relationships, and their confirmation results will constrain the range of possible roles for multiple downstream matters. Taking the birth of a newborn as an example, once "who is the newborn's mother" is confirmed, the applicant for the birth certificate is determined, the parent-child relationship field in birth registration is determined, the family relationship field in medical insurance enrollment is determined, and the candidate range for the guardian field in social security card application is narrowed—a single confirmation generates a cascading effect of benefits in multiple places. If the core confirmation items are placed at the end, the role allocation for the preceding items will be uncertain. The platform will either have to ask questions for each item separately, increasing the number of interactions, or proceed with uncertainty, increasing the probability of errors.
[0085] The essential difference between the confirmation chain in this embodiment and the batch follow-up questioning method in the prior art is that the result of each confirmation item in the confirmation chain dynamically narrows the scope of subsequent confirmation items, and may even directly eliminate some confirmation items. For example, when a user confirms "I am the child's mother," based on the spousal relationship between the mother and father in the knowledge graph, the platform can directly infer the father's identity without having to ask "who is the child's father." Batch follow-up questioning, on the other hand, lists all uncertain items at once, with each question being independent of the others and not producing a resolution linkage effect.
[0086] S4: Event Role Rebinding
[0087] This step is used for the S3-based role boundary confirmation chain. It binds confirmed roles to items at the item level and unbinds those that do not have a successor relationship, resulting in item-role rebinding.
[0088] S41: Binding Validity Calculation. Based on the confirmation results of S3, and taking into account the confirmation status, the strength of the matter's carrying capacity, the strength of the material support, the strength of the relationship conflict, and the scope of the impact of subsequent matters, calculate the binding validity of the candidate subject to each matter role.
[0089] S42: Graded Solidification Judgment. After the binding validity is judged through solidification, it is executed according to the following three-level threshold:
[0090] Roles that meet the high confidence criteria are bound at the event level.
[0091] Roles that meet the intermediate confidence level criteria remain in candidate status and await further confirmation.
[0092] Low-confidence characters will not be bound.
[0093] The solidification judgment in this embodiment introduces a conservative bias, which is specifically achieved through conservative solidification parameters: when the binding validity is lower than the semi-solidification point, the platform solidifies with a probability lower than the linear corresponding probability, so that in the gray area where binding is optional, the platform tends not to bind rather than bind.
[0094] The inventors recognized that in government service scenarios, the cost of incorrect binding is far higher than that of delayed binding. Incorrect binding means that an unqualified entity is granted processing authority, which may lead to the collection of incorrect materials, the initiation of incorrect approval paths, or even legal defects. Delayed binding, on the other hand, only means that the platform needs to confirm one more question with the user, with a cost of only one round of interaction. Therefore, the platform's established logic should lean towards conservatism—in cases of uncertainty, it is better to confirm once more than to risk binding.
[0095] If neutral fixation is adopted (i.e., binding or not binding with symmetrical probability in the uncertainty zone), the platform's behavior in the gray area is close to random selection, failing to reflect the aforementioned asymmetric error cost. The conservative bias is essentially an algorithmic expression of the asymmetric error cost characteristic in government service scenarios, a key design feature that distinguishes it from general semantic understanding solutions.
[0096] S43: Proactive Unbinding and Downgrading. After binding is completed, check whether there are binding relationships that lack a proper connection (e.g., an entity is bound to the role of "mother" and confirmed in the birth certificate stage, but the entity's identity indicates that it does not qualify as a mother). The unbinding strength is calculated by weighting the complement of the binding validity relative to the full confidence level, the intensity of role conflicts, the intensity of unresolved confirmations, and the complement of the strength of established connection evidence relative to the full connection level. When the unbinding strength exceeds a preset threshold, the binding relationship is proactively unbound, and the corresponding role is downgraded to a candidate or reference entity. Simultaneously, the role is marked as unsuitable for subsequent critical processing steps.
[0097] The reason for using a multi-factor comprehensive assessment to determine the strength of binding removal is that binding removal triggered by a single factor carries a high risk. For example, removing a binding solely based on low binding validity might overlook situations where the binding, though lacking confidence, has clear supporting evidence. Multi-factor comprehensive assessment ensures that the platform only performs binding removal when multiple unfavorable signals converge (i.e., low confidence, significant role conflict, unresolved confirmations, and a lack of clear acceptance criteria), avoiding repeated binding-unbinding jitter caused by fluctuations in a single signal. If only a single factor is used (e.g., only checking if binding validity is below a threshold), the platform might frequently switch binding states in the boundary areas of binding validity fluctuations, leading to service path instability.
[0098] The reason for downgrading rather than deleting the unbound role is that although the unbound role can no longer fulfill the bound role for the current matter, the identity information it carries may still be valuable in subsequent matters. For example, in the vaccination registration process, if the unbound "mother" role is actually a grandmother, her kinship information with the newborn can serve as a reference for the daily guardianship contact. If directly deleted, the platform would lose this information. Downgrading to a reference subject preserves the information, while preventing it from entering critical processing paths through path isolation, thus balancing information preservation and risk control.
[0099] S44: Summary of Rebinding Results. Summarize the rebinding results of roles and tasks, including roles that have been officially bound and the scope of tasks they are bound to, roles that remain in the candidate status and the reasons for their suspension, roles that have been unbound and the reasons for their unbinding, affected tasks, and their downgraded identities, and output them to S5.
[0100] S5: Service Result Output
[0101] Based on the role rebinding results, a service result is generated and output. This service result includes at least the processing path, role description, required confirmation items, and a list of materials. Furthermore, the service result also includes an explanation of the reason for unbinding the role, and an indication of the key processing paths that the role cannot access.
[0102] The design consideration for including the reason for cancellation in the service results is that in government services, if the conclusion of "cannot be processed" lacks supporting evidence, users will be unable to take remedial measures. For example, if the output only indicates "household registration cannot be processed," users cannot determine whether it is because the head of household's identity does not meet the requirements, because specific materials are missing, or because the parent-child relationship is not confirmed. By clearly stating the reason for cancellation in the output (e.g., "This person is not the legal guardian of the newborn and cannot be the applicant for birth registration"), users can supplement information or adjust their processing path accordingly. This design makes the platform's decision-making process explainable to users, unlike a closed architecture that only outputs the conclusion while hiding the reasoning process.
[0103] Example 2: This example provides an apparatus corresponding to the method described in Example 1, comprising the following modules:
[0104] The subject semantic decomposition module receives consultation statements, performs semantic decomposition on the statements, identifies the subject referential information and the intent of the matter, and generates a subject semantic node set. The roles in the subject semantic node set are all marked as candidate states and are not directly bound to a specific handling role. This module executes all sub-steps of S1 in Embodiment 1, including statement segmentation, title trigger extraction, candidate strength calculation, role distinguishability and deferred binding determination, and node set generation.
[0105] The design consideration for embedding the candidate state marking logic within the semantic decomposition process in this module is that the candidate state marking depends on the role distinguishability calculated internally by this module, and this decision must be completed during the semantic decomposition stage. If the candidate decision is postponed to downstream modules, the semantic decomposition module will output the already determined roles, and the downstream modules will lose the basis for generating multipaths—the information has been prematurely compressed during the inter-module transmission process, and the subsequent multipath candidate graph and confirmation chain resolution mechanism will be impossible to develop.
[0106] The candidate graph generation module is used to input the set of subject semantic nodes into a preset knowledge graph to generate a candidate graph of subject relationships containing multiple possible relationship paths. This module executes all the sub-steps of S2 in Embodiment 1, including knowledge graph attachment (the same node can be attached to multiple standard role nodes at the same time), attachment strength calculation, candidate relationship edge generation, multi-path structure construction and retention strength calculation, and candidate graph aggregation.
[0107] The role boundary confirmation module is used to match each candidate role with the required role for each sub-item based on the subject relationship candidate graph, identify the boundary confusion of the same subject referential information among multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship. This module executes all the sub-steps of S3 in Embodiment 1.
[0108] The design consideration for the confirmation chain arrangement logic in this module is that the order of the confirmation chain directly determines the interaction method with the user—the user will answer confirmation questions one by one in the order of the confirmation chain. If the arrangement logic were executed in the task role rebinding module, that module would not only need to handle the binding logic, but also manage the user interaction flow and dynamically adjust the scope of confirmation items, leading to confusion of responsibilities. By placing the organization of the confirmation chain in this module, this module outputs a structured confirmation sequence ordered by dependencies. The rebinding module only needs to execute the confirmations and bindings in sequence, making the module's responsibilities clearly defined.
[0109] The task-role rebinding module is used to bind confirmed roles to tasks based on the role boundary confirmation chain, and to unbind bindings that do not have a successor relationship, thus forming a task-role rebinding result. This module executes all sub-steps of S4 in Embodiment 1, including binding validity calculation, conservative biased hierarchical solidification judgment, successor relationship check and active unbinding and downgrading, and summarizing the rebinding results.
[0110] The design consideration for treating unbinding and degradation operations as internal steps in the binding process (rather than being delayed until the output stage) in this module is that the result of the unbinding operation directly affects the processing path in subsequent service results. If mismatched bindings are only discovered and processed in the output stage, the output module either needs to determine the validity of the binding itself, exceeding its own responsibilities, or it needs to call back the rebinding module to reprocess, increasing the coupling between modules. By making unbinding an internal step of the rebinding module, a complete binding-checking-unbinding-degradation closed loop is completed before output. The output module receives the corrected results, ensuring clear responsibilities and smooth data flow.
[0111] The service result output module is used to generate and output service results based on the role rebinding results of the matter. The service results include at least the processing path, role description, matters requiring confirmation, and a list of materials. This module executes all the steps in S5 of Embodiment 1.
[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for integrated intelligent e-government services based on knowledge graph semantic understanding, characterized in that: Including the following: S1: Receive the consultation message, perform semantic decomposition on the consultation message, identify the subject referential information and the intention of the matter, and generate a set of subject semantic nodes; wherein, the roles in the set of subject semantic nodes are all marked as candidate states and are not directly bound to a specific handling role; S2: Input the set of subject semantic nodes into a preset knowledge graph to generate a subject relationship candidate graph containing multiple possible relationship paths; S3: Based on the subject relationship candidate graph, match each candidate role with the role required by each sub-item, identify the boundary confusion of the same subject referent information between multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship; S4: Based on the role boundary confirmation chain, bind the confirmed roles at the event level according to the event, and unbind the bindings that do not have a successor relationship, thus forming an event-role rebinding result; S5: Based on the role rebinding result of the matter, generate and output the service result, which includes at least the processing path, role description, matters to be confirmed, and material list.
2. The method according to claim 1, characterized in that, In S1, the set of generated main semantic nodes includes: For the title triggers extracted from the agency consultation discourse, their title explicitness, kinship chain matching degree, event adjacency degree and information adjacency degree are weighted and calculated, and nonlinear enhancement is applied. Then, the attenuation correction of the uncertain expression suppression term is applied to obtain the candidate strength of the title trigger for each preset role. For multiple candidate strengths of the same title trigger item, the difference between the highest candidate strength and the second candidate strength is divided by the sum of the highest candidate strength and the stable parameter, which is used as the role distinguishability. When the role distinguishability is lower than a preset threshold, multiple candidate roles are retained for the title trigger item and a temporary binding flag is set, and the role is not determined for the time being.
3. The method according to claim 1, characterized in that, In S2, the generation of candidate graphs for principal relationships includes: Each subject semantic node is attached to a standard role node in the knowledge graph. The same subject semantic node is allowed to be attached to multiple standard role nodes at the same time. The candidate role strength, semantic fit, graph relationship support and context fit are weighted and nonlinear enhancement is applied. Then, the attachment strength is obtained by role conflict degree attenuation correction and relationship path length attenuation correction. Based on the connection results, multiple candidate roles in the same public expression are connected into a multi-path subject relationship structure. The retention strength is calculated for each path, and the paths that meet the retention conditions are summarized into the subject relationship candidate graph.
4. The method according to claim 1, characterized in that, In S3, the organization forms a role boundary confirmation chain, including: The carrying strength of the same subject referential information across multiple roles is compared. By combining the ratio of the carrying strength of the second candidate to the carrying strength of the highest candidate, the degree of overlap of candidate roles, and the strength of clear identity evidence, titles with confused role boundaries are identified, and targeted role confirmation items are generated. Multiple confirmation items are organized into a role boundary confirmation chain based on the dependencies between items. Confirmation items that affect core items are arranged before those that affect secondary items, so that the role determination result of the first confirmed item directly constrains the range of optional roles for the later confirmed items.
5. The method according to claim 1, characterized in that, In S4, the step of binding items at the item level and unbinding items that do not have a successor relationship includes: The binding validity of candidate subjects to each task role is calculated. After solidification and judgment, task-level binding is performed on roles that meet the high confidence level condition, roles that meet the intermediate confidence level condition are kept in the candidate state to await further confirmation, and low confidence roles are not bound. The solidification and judgment introduces a conservative bias, so that the platform tends not to perform binding when the binding validity does not reach the high confidence level condition. After the binding is completed, the bound roles are checked for any binding relationships that do not have a succession relationship. The strength of the release is determined by comprehensively considering the binding validity, the degree of role conflict, the strength of unresolved confirmation items, and the strength of the established succession basis. For binding relationships whose release strength exceeds the preset threshold, the binding is actively released and the corresponding role is downgraded to a candidate subject or reference subject. At the same time, the role is restricted from entering the processing path of subsequent key matters.
6. The method according to claim 1, characterized in that, In S5, the service result also includes: an explanation of the reason for unbinding the role, and an indication of the key processing path that the role cannot access.
7. The method according to claim 3, characterized in that, The preset knowledge graph is a knowledge graph of the birth of a newborn; the standard role nodes include at least: consultant, agent, newborn, father, mother, grandparents, head of household, proposed household registration recipient and guardian contact person.
8. A government affairs integrated intelligent service platform based on knowledge graph semantic understanding, characterized in that: include: The main semantic decomposition module is used to receive agency consultation statements, perform semantic decomposition on the agency consultation statements, identify the main referent information and the intention of the matter, and generate a set of main semantic nodes; wherein, the roles in the set of main semantic nodes are all marked as candidate states and are not directly bound to a specific handling role; The candidate graph generation module is used to bring the set of main semantic nodes into a preset knowledge graph to generate a candidate graph of main relationships containing multiple possible relationship paths; The role boundary confirmation module is used to match each candidate role with the role required by each sub-item based on the subject relationship candidate graph, identify the boundary confusion of the same subject referent information between multiple item roles, generate role confirmation items, and organize them into a role boundary confirmation chain according to the item dependency relationship. The task-role rebinding module is used to bind confirmed roles to tasks based on the role boundary confirmation chain, and to unbind bindings that do not have a successor relationship, thus forming a task-role rebinding result. The service result output module is used to generate and output service results based on the role rebinding results of the matter. The service results include at least the processing path, role description, matters requiring confirmation, and a list of materials.
9. The platform according to claim 8, characterized in that: The main semantic decomposition module is further used to: calculate the title triggering items extracted from the agency consultation discourse by comprehensively considering the title explicitness, kinship chain matching degree, event adjacency degree and information adjacency degree, apply nonlinear enhancement, and then apply the attenuation correction of the uncertain expression suppression term to obtain the candidate strength; For multiple candidate strengths of the same title trigger item, sort them and divide the difference between the highest candidate strength and the second candidate strength by the sum of the highest candidate strength and the stable parameter as the role discrimination. If the discrimination is lower than the preset threshold, set a temporary binding flag. The aforementioned task role rebinding module is further used to: introduce a conservative bias in the solidification judgment, so that the binding is not executed when the binding validity does not reach the high confidence condition; For bound roles that do not have a succession relationship, the strength of the removal is determined by comprehensively considering the binding validity, the degree of role conflict, the strength of unresolved confirmation items, and the strength of the established succession basis, and a downgrade is performed.
