An intelligent reasoning knowledge base construction method based on a knowledge graph

Abstract

The application discloses a kind of intelligent reasoning knowledge base construction method based on knowledge graph, it is related to government affair information processing and intelligent reasoning technical field, the present application realizes the accurate extraction of policy element and the formalized representation of logical relationship by constructing structured policy clause knowledge graph, provides reliable data basis for conflict detection;Adopt the method that adaptive threshold mechanism and constraint satisfiability solution are combined, both ensure the accuracy of conflict detection, and can adapt to the language characteristics and time evolution characteristics of policy text;By establishing complete conflict report generation mechanism, clear decision support is provided for policy making department;Incremental graph updating and version management function ensure the timeliness and consistency of knowledge base, and the construction of government affair responsibility sub-graph further enhances department cooperation capability;The application of multi-task learning model improves the intelligent level of appeal processing, and personalized PageRank algorithm realizes accurate department recommendation.

Description

Technical Field

[0001] This invention relates to the fields of government information processing and intelligent reasoning technology, and in particular to a method for constructing an intelligent reasoning knowledge base based on knowledge graphs. Background Technology

[0002] As the government's digital transformation deepens, the business systems built independently by various departments have gradually formed information silos, and there is a lack of effective collaboration mechanisms in the process of strategy formulation and implementation. Cross-departmental policy conflicts have become a common problem in government work, such as the contradiction between economic development strategies and environmental protection requirements, and the inconsistency of approval standards among different departments. Most of these conflicts are only discovered in the strategy implementation stage, leading to reduced administrative efficiency and waste of public resources.

[0003] Most existing intelligent reasoning knowledge base systems use rule matching and semantic similarity calculation to identify policy conflicts. Some newer solutions combine deep learning and knowledge graph technologies to build a network of connections between policy terms through entity recognition and relation extraction. Other systems introduce pre-trained language models for semantic understanding of policy texts and use graph neural networks to detect logical contradictions between terms. Such systems can identify explicit policy conflicts to a certain extent and provide a reference for decision-making.

[0004] However, existing solutions still have several limitations when dealing with complex policy conflicts. The system has limited ability to identify semantic conflicts and context-related clauses implied in policy texts, making it difficult to discover potential contradictions that require in-depth reasoning. The dynamic updating and revision management of policy clauses has not yet formed an effective real-time synchronization mechanism, resulting in insufficient timeliness of the knowledge base. In addition, the ability to make multi-hop associations between cross-department policies is weak, making it unable to effectively handle complex conflict scenarios involving multiple departments and levels. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a knowledge graph-based intelligent reasoning knowledge base construction method to solve the problems of existing conflict detection systems, such as difficulty in detecting hidden conflicts, lack of real-time performance and multi-departmental collaboration capabilities, resulting in low execution efficiency.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a method for constructing an intelligent reasoning knowledge base based on knowledge graphs, comprising:

[0009] Step S1: Parse the strategy text, use semantic role labeling and named entity recognition to extract conditional elements consisting of subject, condition, action and object;

[0010] Step S2: Construct a strategy clause knowledge graph based on the condition elements. The entities in the graph include at least strategy entities, clause entities, and department entities, and the relationships include at least subordinate, applicable, restrictive, and authorizing.

[0011] Step S3: Generate the logical relationships between clauses based on the references, constraints and scope of application between the conditional elements and write them into the graph;

[0012] Step S4: When a new clause is added, candidate comparison is performed based on graph and text features, and conflict type detection is performed on the candidate clauses.

[0013] Step S5: Write the detection results back to the map and trigger a conflict warning output.

[0014] As a preferred embodiment of the knowledge graph-based intelligent reasoning knowledge base construction method of the present invention, the pattern of the strategy clause knowledge graph includes:

[0015] The clause entity carries the effective date, applicable territory, legal level, and source of reference.

[0016] As a preferred embodiment of the knowledge graph-based intelligent reasoning knowledge base construction method of the present invention, the conflict type detection includes:

[0017] a) Calculate the semantic similarity between the new clause and the candidate clause and compare it with an adaptive threshold;

[0018] b) When the threshold is exceeded, the constraint satisfiability solver is invoked to determine the logical mutual exclusion / overwriting / repetition / condition missing of the condition elements, and the conflict type is determined in combination with the consistency of actions;

[0019] The adaptive threshold is determined as follows:

[0020] Two types of reference pairs are extracted from the knowledge graph: one type is clause pairs that already have a citation relationship or have been judged as conflicting / highly related in the past, and the other type is clause pairs that have significant differences in topic, department or scope of application; the semantic similarity sequence is uniformly calculated between the two types of reference pairs and the current newly added candidate pairs within the sliding time window;

[0021] Within the window, empirical distributions of similarity for positive and negative classes are formed, and the quantile method is used to characterize the upper tail of the negative class and the lower tail of the positive class.

[0022] In the window The trigger threshold is obtained by weighting the quantiles and performing exponential smoothing:

[0023] ,

[0024] in, Display window Similarity trigger threshold, This indicates the trigger threshold of the previous window. This represents the update coefficient, and its value is... Dimensionless The quantiles at both ends are weighted by a factor of 10, which takes the following values. Dimensionless Represents the distribution of positive class similarity within the window. Quantiles Represents the distribution of negative class similarity within the window. Quantiles This represents the tolerance for false negatives in the positive class, with a value of [value missing]. Dimensionless This represents the negative class false positive tolerance, with a value of [value missing]. Dimensionless;

[0025] use Compare the similarity of newly added candidate pairs and select those for subsequent constraint solving; if there is an overlap where the upper quantile of the negative class is higher than the lower quantile of the positive class, then temporarily adjust the quantile upwards. Alternatively, the window length can be increased to reduce jitter, and this adjustment can be recorded in the metacognitive knowledge graph.

[0026] As a preferred embodiment of the knowledge graph-based intelligent reasoning knowledge base construction method of the present invention, wherein: the conflict warning output includes generating a conflict report, and the conflict report includes at least:

[0027] The conflict clause includes location information, conflict type, affected department identification, and chain of evidence, and provides a backtracking link for manual review.

[0028] As a preferred embodiment of the intelligent reasoning knowledge base construction method based on knowledge graph described in this invention, the knowledge base supports incremental graph updates and dynamic version management. It maintains version nodes and parent-child version relationships for clause entities, records the effective / invalid timestamps and reasons for changes of versions, and achieves a consistent view based on the time dimension during querying.

[0029] As a preferred embodiment of the intelligent reasoning knowledge base construction method based on knowledge graph described in this invention, the knowledge base further includes a government affairs responsibility sub-graph, which contains department nodes, responsibility nodes, and item nodes.

[0030] For incoming requests, a personalized PageRank algorithm is used to calculate the relevance between the request and the department node and generate a recommended list of departments.

[0031] The damping factor for the personalized PageRank is set as follows:

[0032] Using departments, responsibilities, and matters as nodes, weighted edges are formed according to citation relationships, historical co-occurrence, and legal subordination; column-wise normalization is performed on the weighted adjacency matrix to obtain the transition probability matrix of the random walk;

[0033] The preference vector is synthesized from three parts:

[0034] A. Request Item Anchor Point: Map future events to item nodes, and place higher weights on this node and its first-order neighborhood.

[0035] B. Historical Preference: Extract the distribution of departments that successfully processed applications within the same subject period.

[0036] C. Rule Priority: Based on the applicable geographical area and legal level, irrelevant departments are weighted or set to zero;

[0037] The three parts are based on item anchor points / historical preferences / rule priors. The default weighted ratio is then normalized to a probability distribution. If an incoming item matches multiple items, it is then reassigned within A based on semantic similarity scores.

[0038] Under personalized settings, the relevance vector is iterated until convergence using the following formula:

[0039] ,

[0040] in, Indicates the first The departmental relevance vector obtained from the next iteration. This represents the damping factor, with values ​​ranging from 1 to 10. Dimensionless This represents the column-normalized transition probability matrix, with dimensions equal to the number of nodes multiplied by the number of nodes. Indicates the first The relevance vector of the next iteration. Let represent the preference vector, which is a probability distribution. In actual calculations, iteration stops when the norm of the difference between two adjacent vectors is lower than the convergence threshold.

[0041] As a preferred embodiment of the knowledge graph-based intelligent reasoning knowledge base construction method described in this invention, a multi-task learning model is used to simultaneously perform entity recognition, sentiment / urgency assessment, and department classification on incoming text, and historical processing cases are retrieved based on vector similarity to output recommended handling departments and reference handling solutions, and the recommendation results are used as personalized preference inputs for iterative optimization.

[0042] As a preferred embodiment of the intelligent reasoning knowledge base construction method based on knowledge graph described in this invention, the method includes: constructing an event-triggered data acquisition pipeline to access multi-source government data; utilizing message queues and streaming processing to achieve near real-time fusion; and employing time-series graph embedding with timestamps to maintain the temporal consistency of nodes and relationships.

[0043] As a preferred embodiment of the knowledge graph-based intelligent reasoning knowledge base construction method described in this invention, the method includes: constructing a strategy influence propagation network to model the influence path of strategies on departments and matters; using an independent cascading class model to simulate the influence; and establishing a multi-objective evaluation index system to optimize among compliance, processing timeliness, or resource consumption objectives, and outputting recommended strategies and trade-off explanations.

[0044] As a preferred embodiment of the intelligent reasoning knowledge base construction method based on knowledge graph described in this invention, the method involves: constructing a metacognitive knowledge graph to record the reasoning process, including: the model version called, threshold settings, the hit graph path, intermediate conclusions of conflict determination and manual review results, and periodically self-optimizing the thresholds and weights of similar cases based on the records.

[0045] The beneficial effects of this invention are as follows: By constructing a structured knowledge graph of policy terms, this invention achieves accurate extraction of policy elements and formal representation of logical relationships, providing a reliable data foundation for conflict detection; the method combining adaptive threshold mechanism and constraint satisfiability solution ensures both the accuracy of conflict detection and adaptability to the linguistic characteristics and temporal evolution of policy texts; by establishing a complete conflict report generation mechanism, it provides clear decision support for policy-making departments; incremental graph updates and version management functions ensure the timeliness and consistency of the knowledge base, while the construction of the government responsibility sub-graph further enhances departmental collaboration capabilities; the application of a multi-task learning model improves the intelligence level of appeal processing, and the personalized PageRank algorithm achieves accurate departmental recommendations; temporal graph embedding technology and streaming processing architecture ensure the real-time fusion and processing capabilities of multi-source government data, and the policy influence propagation network provides a scientific basis for policy effectiveness evaluation. The introduction of metacognitive knowledge graphs enables the system's self-optimization and continuous improvement, continuously enhancing the accuracy and reliability of conflict detection by recording the reasoning process and human feedback. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0047] Figure 1This is a flowchart of the knowledge graph-based intelligent reasoning knowledge base construction method in Example 1. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0050] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0051] This application proposes a method for constructing an intelligent reasoning knowledge base based on knowledge graphs, combining... Figure 1 As shown, the method includes:

[0052] Step S1: Parse the strategy text, use semantic role labeling and named entity recognition to extract conditional elements consisting of subject, condition, action and object;

[0053] In this embodiment, semantic role labeling is used to identify roles such as agent, patient, manner, and constraint in predicate-argument structures. Engineering implementation utilizes a domain dictionary and synonym merging through word segmentation. Named entity recognition is used to extract proper names involving organization names, legal citations, place names, time, and quantity. Specifically, text cleaning defaults to denoising, sentence segmentation, and unified number formatting. The maximum sentence length is 512 words by default, adjustable from 256 to 2048 words, set according to corpus distribution and model input limitations. The confidence threshold for roles and entities is 0.8 by default, adjustable from 0.6 to 0.95, determined by the F1 optimum of the offline validation set. For example, subject and department aliases are obtained by maintaining a synonym / abbreviation mapping, updated every 30 days by default, with temporary encrypted updates during major organizational adjustments. Optionally, when syntactic complexity leads to missing roles, nearest neighbor sentences are used to fill the gaps. If the text contains missing characters or OCR anomalies, it reverts to rule template parsing and is labeled as a low-confidence result to avoid affecting subsequent judgments.

[0054] Step S2: Construct a knowledge graph of strategy terms based on condition elements. The entities in the graph include at least strategy entities, terms entities, and department entities, and the relationships include at least subordinate, applicable, restrictive, and authorizing.

[0055] Furthermore, "subordinate" indicates the hierarchical relationship of regulations / policies to which the clauses belong; "applicable" indicates the matching relationship between the object and the geographical / time range of the clauses; "restriction" indicates the boundary imposed on the object or condition; and "authorization" indicates the source of authorization and responsibility. Each relationship carries the effective period and priority caliber through attributes for subsequent reasoning. Specifically, entity primary keys are uniquely determined by a triplet of document number + clause number + version number, with a default global uniqueness check. In case of conflict, timestamps and source credibility are used for adjudication. Relationships are deduplicated before being stored in the database, with a default duplicate threshold of 0.9, adjustable from 0.8 to 0.98, set based on historical overlap rate statistics. Similarly, graph storage can be implemented as an attribute graph. Node and edge attribute fields and indexes are whitelisted before being stored to ensure the minimum queryable scope. When a field is missing, it is placed in an unknown placeholder without triggering default value extrapolation to avoid introducing false constraints.

[0056] Step S3: Generate the logical relationships between clauses based on the references, constraints and scope of application between the conditional elements and write them into the graph;

[0057] In this embodiment, logical relationships are represented by verifiable evidence paths, including snapshots of start and end clauses, triggering relationship types, and hit conditions. Priority conflict determination follows the order of legal level over departmental regulations, specific provisions over general provisions, and the latest version over the old version. Specifically, the overlapping time and region are determined using interval intersection, with time using closed intervals and regions using administrative region coding tree matching. The path length defaults to no more than 3 hops, adjustable to 2-4 hops, and adjusted based on interpretation costs and false alarm rates. Optionally, parallel relationships whose priority cannot be determined are marked as candidate conflicts and entered into a manual review queue. When the necessary effective time or regional attributes are missing, the relationship is only recorded as a potential relationship with a reduced confidence weight to reduce interference.

[0058] Step S4: When a new clause is added, candidate comparison is performed based on graph and text features, and conflict type detection is performed on the candidate clauses.

[0059] Specifically, candidate comparison is performed jointly using semantic vector retrieval and structural element alignment, prioritizing clauses with the same theme, the same or adjacent departments, and similar scope of application in the candidate set. The default candidate set size is the first 50 entries, adjustable from 20 to 200, determined based on the recall-precision curve from offline playback. Furthermore, similarity is achieved through a weighted fusion of vector cosine and keyword coverage, with a default weight of 0.7 / 0.3, which can be optimized across different domains using development sets. When the candidate set is empty, the theme similarity threshold is relaxed or cross-level reference retrieval is enabled to ensure minimum coverage, but corresponding low confidence is indicated to prompt subsequent weight reduction in the solution.

[0060] Step S5: Write the detection results back to the map and trigger a conflict warning output;

[0061] In one embodiment, the schema of the policy terms knowledge graph includes:

[0062] The clause should include the effective date, applicable territory, legal hierarchy, and source of reference.

[0063] Relationships carry attributes such as applicable scope boundaries, exception conditions, and priority to support subsequent logical judgments and interpretations;

[0064] In one embodiment, conflict type detection includes:

[0065] a) Calculate the semantic similarity between the new clause and the candidate clause and compare it with an adaptive threshold;

[0066] b) When the threshold is exceeded, the constraint satisfiability solver is invoked to determine the logical mutual exclusion / overwriting / repetition / condition missing of the condition elements, and the conflict type is determined in combination with the consistency of actions;

[0067] The adaptive threshold is determined as follows:

[0068] Two types of reference pairs are extracted from the knowledge graph: one type is clause pairs that already have a reference relationship or have been judged as conflicting / highly related in the past (positive class), and the other type is clause pairs that have significant differences in topic, department or scope of application (negative class); semantic similarity sequence is uniformly calculated for the two types of reference pairs and the current newly added candidate pairs within the sliding time window;

[0069] Within the window, empirical distributions of similarity for positive and negative classes are formed respectively. The quantile method is used to characterize the upper tail of the negative class and the lower tail of the positive class to control the tradeoff between false alarms and false negatives.

[0070] In the window The trigger threshold is obtained by weighting the quantiles and performing exponential smoothing:

[0071] ,

[0072] in, Display window Similarity trigger threshold, This indicates the trigger threshold of the previous window. This represents the update coefficient, and its value is... Dimensionless The quantiles at both ends are weighted by a factor of 10, which takes the following values. Dimensionless Represents the distribution of positive class similarity within the window. Quantiles Represents the distribution of negative class similarity within the window. Quantiles This represents the tolerance for false negatives in the positive class, with a value of [value missing]. Dimensionless This represents the negative class false positive tolerance, with a value of [value missing]. Dimensionless;

[0073] use Compare the similarity of newly added candidate pairs and select those for subsequent constraint solving; if there is an overlap where the upper quantile of the negative class is higher than the lower quantile of the positive class, then temporarily adjust the quantile upwards. Alternatively, the window length can be increased to reduce jitter, and this adjustment can be recorded in the metacognitive knowledge graph;

[0074] In this embodiment, To control the threshold response speed, a typical range is obtained by prioritizing the trade-off between alarm jitter rate and latency cost on the playback set. Used to balance missed detections and false alarms, the value is determined jointly based on business tolerance and audit load; and The upper limit of target misclassification, derived from both positive and negative samples, is preset by the governance strategy and can be temporarily adjusted during peak periods. Specifically, the window length is 30 days by default, adjustable from 7 to 90 days, and automatically extended when data is sparse to stabilize quantile estimation; when the number of positive and negative samples is insufficient to form a stable distribution, it reverts to the static threshold and is marked as a conservative trigger mode, while the sample gap is written into the metacognitive record to drive data completion. Optionally, to suppress the impact of short-term drift on alarm frequency, the minimum threshold update interval is set to 24 hours by default; if the difference between two consecutive updates exceeds the preset range, a manual verification prompt is triggered to avoid threshold oscillation.

[0075] Specifically, the threshold acquisition method described above uses quantiles as its core, weighting the upper tail of the negative class and the lower tail of the positive class to provide a trigger boundary that is more sensitive to the current data distribution within the same time window. This approach avoids the problem of static thresholds failing when the corpus topic or writing style drifts, while suppressing short-term fluctuations through exponential smoothing, reducing the computational and communication costs caused by frequent alert switching. The steps involve first obtaining interpretable reference samples from the graph and historical labels, then characterizing the risk areas at both ends using empirical distributions, and finally implementing a traceable threshold generation chain. The threshold trade-off coefficient and the false positive / false negative tolerance can both be used as inputs for governance preferences, facilitating the switching of emphasis between different business periods and different strategy domains. If the two distributions overlap significantly in the actual data, the trigger position can be stabilized by expanding the window or dynamically adjusting the trade-off coefficient, and the adjustment process can be written into the metacognitive graph for subsequent review and reuse.

[0076] In one embodiment, the conflict warning output includes generating a conflict report, which at least includes:

[0077] The conflict clause's location information (document number, clause number, paragraph position), conflict type, affected department identifier, and evidence chain (composed of the triggering graph path and hitting conditions) are provided, along with a backtracking link for manual review.

[0078] For example, the default time limit for generating conflict reports is no more than 5 minutes, adjustable from 1 to 15 minutes, based on traffic and computing resource quotas. The severity level of the report is determined comprehensively based on the overlap of the scope of impact, legal level, and applicable time, and is categorized by default as warning / general / major, with thresholds derived from historical closed-loop statistics. Optionally, when the chain of evidence includes multi-hop paths across levels or departments, an explanatory summary field is added to compress the output; if the location information lacks a clause number, it is replaced with the document number and sentence offset, and the location completeness is indicated.

[0079] In one embodiment, the knowledge base supports incremental graph updates and dynamic version management. It maintains version nodes and parent-child version relationships for clause entities, records the effective / invalid timestamps and reasons for changes of versions, and achieves a consistent view based on the time dimension during queries.

[0080] Specifically, version relationships include three categories: additions, revisions, and repeals. The query consistency view can be switched between two modes: query time and version number, with the query time being the default. Write conflicts are resolved by the latest write time plus source credibility, with credibility defaulting to government information disclosure platform > bulletins > crawled text. Furthermore, the polling cycle for incremental updates is daily by default, but can be shortened to every 2 hours during peak release periods. When a version chain forks or a parent version is missing, it is temporarily marked as a suspended version and its participation in conflict calculations is restricted until the relationship is complete.

[0081] In one embodiment, the knowledge base further includes a government affairs responsibility subgraph, which contains department nodes, responsibility nodes, and item nodes;

[0082] For incoming requests, a personalized PageRank algorithm is used to calculate the relevance between the request and the department node and generate a recommended list of departments.

[0083] The damping factor for personalized PageRank is set as follows:

[0084] Using departments, responsibilities, and matters as nodes, weighted edges are formed according to reference relationships, historical co-occurrence, and legal subordination; the weighted adjacency matrix is ​​normalized column-wise to obtain the transition probability matrix of the random walk, which is used for subsequent calculations;

[0085] The preference vector is synthesized from three parts:

[0086] A. Request Item Anchor Point: Map future events to the request item node (including synonym merging and alias expansion), and place higher weights on this node and its first-order neighborhood.

[0087] B. Historical Preferences: Extract the distribution of departments that successfully handled similar cases within a recent period, focusing on the same subject (or region, issue).

[0088] C. Rule Priority: Based on the applicable geographical area and legal level, irrelevant departments are weighted or set to zero;

[0089] The three parts are weighted by the default ratio of item anchor point / historical preference / rule prior = 0.6 / 0.3 / 0.1 and then normalized into a probability distribution. If the incoming item matches multiple items, it is then reassigned within A according to semantic similarity score.

[0090] Under personalized settings, the relevance vector is iterated until convergence using the following formula:

[0091] ,

[0092] in, Indicates the first The departmental relevance vector obtained from the next iteration. This represents the damping factor, with values ​​ranging from 1 to 10. Dimensionless This represents the column-normalized transition probability matrix, with dimensions equal to the number of nodes multiplied by the number of nodes. Indicates the first The relevance vector of the next iteration. Let represent the preference vector, which is a probability distribution. In actual calculations, iteration stops when the norm of the difference between two adjacent vectors is lower than the convergence threshold.

[0093] The baseline is set at 0.85; if the text noise is high or the incoming documents contain multiple topics, the value is reduced to 0.75-0.80; if the cross-level propagation path is long or the relationship diffusion needs to be strengthened, the value is increased to 0.90; when there is a strong individualized historical preference, the value is set to 0.80 to highlight the preference item; the damping factor is adjusted monthly on the offline playback set, and the indicator is selected based on the combined performance of Top-k hit rate and coverage.

[0094] The converged relevance vector is sorted from high to low for departmental nodes. If there are ties, the order of ties is determined by the recent processing load, legal authority attribute and regional matching degree. The hit evidence path is output simultaneously to support the review.

[0095] In this embodiment, personalized PageRank is... The parameter value is adjusted based on the Top-k hit rate and coverage of the replay set, with a default of 0.85 and an adjustable value of 0.75-0.90. The three weights of the preference vector p are 0.6 / 0.3 / 0.1 by default, and can be reduced in scenarios with high text noise or multiple topics. To emphasize the input of preferences. Optionally, when p reaches zero on all candidate nodes, it regresses to uniform preferences and marks low-confidence recommendations; when the iteration fails to converge within the maximum number of steps (the maximum number of steps is 50 by default, adjustable from 20 to 200), it outputs the current optimal solution and records the convergence anomaly for offline investigation.

[0096] Specifically, the design approach uses random walks as a vehicle to inject business semantics into the graph structure through preference vectors, so that the scores reflect both relationship propagation and retain strong focus on the topic of the incoming document. The damping factor balances two approaches: a lower value relies more on preferences, which is suitable for scenarios with high text noise, unfocused topics, or tight deadlines; a higher value relies more on relationship diffusion, which is suitable for situations requiring cross-departmental and cross-responsibility level divergent retrieval. The preference sources adopt a three-element synthesis path of item anchors, historical preferences, and rule priors. The former provides direct semantic positioning, the middle incorporates behavioral evidence, and the latter introduces compliance and geographical constraints, making the recommendation results interpretable.

[0097] The synthesis ratio is set to a robust default value, while reserving space for fine-tuning according to data distribution and governance preferences; the overall process, from graph construction to sorting output, maintains a closed loop, and the damping range and ratio settings can be periodically checked on offline playback, and the change process is included in the metacognitive record, which is convenient for long-term maintenance and cross-domain migration;

[0098] In one embodiment, a multi-task learning model is used to simultaneously perform entity recognition, sentiment / urgency assessment, and department classification on incoming text, and retrieve historical processing cases based on vector similarity to output recommended handling departments and reference handling solutions. The recommendation results are then used as personalized preference inputs for iterative optimization.

[0099] Specifically, multi-task learning achieves parameter sharing through a shared encoder and task-specific heads. The training objective employs a weighted multi-objective loss, with task weights defaulting to a 1:1:1 ratio, adjustable to balance class imbalance. For similar case retrieval, the default Top-N is 10, adjustable from 5 to 50, and vectors are sorted by cosine similarity after normalization. Optionally, when the confidence of any subtask falls below 0.5, its influence on the final recommendation is reduced to minimize misleading information. If historical cases are insufficient, a fallback to rule priors and departmental responsibility mapping is implemented to ensure minimum usable output.

[0100] In one embodiment, an event-triggered data acquisition pipeline is constructed to access multi-source government data, near real-time fusion is achieved using message queues and streaming processing, and time-series graph embedding with timestamps is used to maintain the temporal consistency of nodes and relationships.

[0101] In this embodiment, event triggering includes policy releases, revision announcements, departmental guideline updates, and incoming requests. The message queue's retry strategy defaults to exponential backoff, with a maximum of 3 retries, adjustable from 1 to 10. The streaming fusion window defaults to 5 minutes, adjustable from 1 to 30 minutes, and the lateness tolerance defaults to 2 minutes to handle out-of-order situations. Furthermore, time-series embedding uses node / edge timestamps as query filtering conditions and writes them into time-state attributes. Updates across windows are merged based on the last written update. When resource shortages cause delays, updates to versions and clause entities are prioritized, while non-critical indicators are processed later.

[0102] In one embodiment, a strategy impact propagation network is constructed to model the impact path of the strategy on departments and matters, an independent cascading class model is used to simulate the impact, and a multi-objective evaluation index system is established to optimize among compliance, processing timeliness or resource consumption objectives, and output recommended strategies and trade-off explanations.

[0103] Specifically, the activation probability of independent cascades can be obtained by comprehensively mapping historical collaborative processing frequency, referencing strength, and responsibility adjacency, and is pruned to [0.05, 0.9] by default to avoid extreme values; the simulation steps are limited to no more than 5 steps by default, adjustable from 3 to 8 steps, and exceeding the number of steps is considered to indicate a decay in marginal propagation effect. Optionally, multi-objective optimization uses weighted summation and sensitivity analysis to provide a trade-off interpretation, with weights configured periodically by the governance side; when the constraints of any objective cannot be met, an infeasibility warning and a minimum relaxation suggestion are output for decision-making.

[0104] In one embodiment, a metacognitive knowledge graph is constructed to record the reasoning process, including: the model version called, threshold settings, the hit graph path, intermediate conclusions of conflict determination and manual review results, and periodically self-optimizing the thresholds and weights of similar cases based on the records.

[0105] In this embodiment, the smallest granularity of metacognitive recording is a single judgment, comprising three parts: input summary, key intermediate variables, and output summary. The recording period is at least one year by default. Periodic self-optimization is performed monthly by default, adjustable to weekly or quarterly, with the optimization range limited to the allowable range of thresholds and weights, without involving changes to the model structure. Optionally, to meet compliance requirements, records containing sensitive personal information are anonymized before being written to the database. When record writing fails or space is insufficient, a circular log is enabled, prioritizing the retention of data from the past three months, while simultaneously reporting maintenance alarms to arrange for expansion.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0107] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for constructing an intelligent reasoning knowledge base based on knowledge graphs, characterized in that, include, Step S1: Parse the strategy text, use semantic role labeling and named entity recognition to extract conditional elements consisting of subject, condition, action and object; Step S2: Construct a strategy clause knowledge graph based on the condition elements. The entities in the graph include at least strategy entities, clause entities, and department entities, and the relationships include at least subordinate, applicable, restrictive, and authorizing. Step S3: Generate the logical relationships between clauses based on the references, constraints and scope of application between the conditional elements and write them into the graph; Step S4: When a new clause is added, candidate comparison is performed based on graph and text features, and conflict type detection is performed on the candidate clauses. Step S5: Write the detection results back to the map and trigger a conflict warning output; The conflict type detection includes: a) Calculate the semantic similarity between the new clause and the candidate clause and compare it with an adaptive threshold; b) When the threshold is exceeded, the constraint satisfiability solver is invoked to determine the logical mutual exclusion / overwriting / repetition / condition missing of the condition elements, and the conflict type is determined in combination with the consistency of actions; The adaptive threshold is determined as follows: Two types of reference pairs are extracted from the knowledge graph: one type is clause pairs that already have a citation relationship or have been judged as conflicting / highly related in the past, and the other type is clause pairs that have significant differences in topic, department or scope of application; the semantic similarity sequence is uniformly calculated between the two types of reference pairs and the current newly added candidate pairs within the sliding time window; Within the window, empirical distributions of similarity for positive and negative classes are formed, and the quantile method is used to characterize the upper tail of the negative class and the lower tail of the positive class. In the window The trigger threshold is obtained by weighting the quantiles and performing exponential smoothing: ， in, Display window Similarity trigger threshold, This indicates the trigger threshold of the previous window. This represents the update coefficient, and its value is... Dimensionless The quantiles at both ends are weighted by a factor of 10, which takes the following values. Dimensionless Represents the distribution of positive class similarity within the window. Quantiles Represents the distribution of negative class similarity within the window. Quantiles This represents the tolerance for false negatives in the positive class, with a value of [value missing]. Dimensionless This represents the negative class false positive tolerance, with a value of [value missing]. Dimensionless; use Compare the similarity of newly added candidate pairs and select those for subsequent constraint solving; if there is an overlap where the upper quantile of the negative class is higher than the lower quantile of the positive class, then temporarily adjust the quantile upwards. Alternatively, the window length can be increased to reduce jitter, and this adjustment can be recorded in the metacognitive knowledge graph; The knowledge base further includes a government affairs responsibility sub-graph, which contains department nodes, responsibility nodes, and item nodes; For incoming requests, a personalized PageRank algorithm is used to calculate the relevance between the request and the department node and generate a recommended list of departments. The damping factor for the personalized PageRank is set as follows: Using departments, responsibilities, and matters as nodes, weighted edges are formed according to citation relationships, historical co-occurrence, and legal subordination; column-wise normalization is performed on the weighted adjacency matrix to obtain the transition probability matrix of the random walk; The preference vector is synthesized from three parts: A. Request Item Anchor Point: Map future events to item nodes, and place higher weights on this node and its first-order neighborhood. B. Historical Preference: Extract the distribution of departments that successfully processed applications within the same subject period. C. Rule Priority: Based on the applicable geographical area and legal level, irrelevant departments are weighted or set to zero; The three parts are weighted according to the default proportions of item anchor point / historical preference / rule prior and then normalized into a probability distribution. If an incoming item matches multiple items, it is then reassigned within A according to semantic similarity score. Under personalized settings, the relevance vector is iterated until convergence using the following formula: ， in, Indicates the first The departmental relevance vector obtained from the next iteration. This represents the damping factor, with values ​​ranging from 1 to 10. Dimensionless This represents the column-normalized transition probability matrix, with dimensions equal to the number of nodes multiplied by the number of nodes. Indicates the first The relevance vector of the next iteration. Let represent the preference vector, which is a probability distribution. In actual calculations, iteration stops when the norm of the difference between two adjacent vectors is lower than the convergence threshold. A multi-task learning model is used to simultaneously perform entity recognition, sentiment / urgency assessment, and department classification on incoming texts. Based on vector similarity, historical processing cases are retrieved, and recommended departments and reference handling solutions are output. The recommendation results are then used as personalized preference inputs for iterative optimization.

2. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, The patterns of the strategy terms knowledge graph include: The clause entity carries the effective date, applicable territory, legal level, and source of reference.

3. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, The conflict warning output includes generating a conflict report, which at least contains: The conflict clause includes location information, conflict type, affected department identification, and chain of evidence, and provides a backtracking link for manual review.

4. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, The knowledge base supports incremental graph updates and dynamic version management. It maintains version nodes and parent-child version relationships for clause entities, records the effective / invalid timestamps and reasons for changes of versions, and achieves a consistent view based on the time dimension during queries.

5. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, An event-triggered data acquisition pipeline is constructed to access multi-source government data. Message queues and streaming processing are used to achieve near real-time fusion, and time-series graph embedding with timestamps is adopted to maintain the time consistency of nodes and relationships.

6. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, We construct a strategy impact propagation network to model the impact path of strategies on departments and matters, use an independent cascading class model to simulate the impact, and establish a multi-objective evaluation index system to optimize among compliance, processing timeliness, or resource consumption objectives, and output recommended strategies and trade-off explanations.

7. The method for constructing an intelligent reasoning knowledge base based on knowledge graphs as described in claim 1, characterized in that, A metacognitive knowledge graph is constructed to record the reasoning process, including: the model version called, the threshold setting, the hit graph path, the intermediate conclusion of the conflict determination and the result of manual review, and the threshold and weight of similar cases are periodically self-optimized based on the record.

Images