Archive management system and method based on artificial intelligence

Abstract

The invention discloses an archive management system and method based on artificial intelligence, relates to the technical field of archive management systems, realizes seamless self-adaptive triggering and content generation of three differentiated services (precise pushing, association expansion and risk early warning) through a decision function, maps a complex multi-factor decision into a continuous pedigree value, and improves the efficiency of archive management. According to the method, the service content and the decision logic are deeply coupled, so that a user does not need to switch a function module, the system can provide the most adaptive composite service based on the current context, risk prevention and control, knowledge association and accurate retrieval capabilities are fused into unified intelligent interaction experience, and the user experience is improved. The system can be converted into active perception from passive retrieval, and the usability and the intelligent level of the system are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of records management technology, and in particular to an artificial intelligence-based records management system and method. Background Technology

[0002] Records management is a fundamental task for organizations to preserve core memories, support compliant operations, and facilitate knowledge innovation. Its development has gone through two main stages: from traditional physical management to modern digital management. However, the existing technology system has fundamental limitations in terms of intelligence, especially in terms of management and service models, which are significantly passive and cannot meet the needs of today's data-driven era for proactive knowledge empowerment.

[0003] While traditional digital record management systems have solved the problems of physical storage and retrieval efficiency, their core remains a "storage-query" architecture based on manually preset metadata, which has significant limitations: First, the information correlation is weak, and it is difficult to establish semantic connections between records, forming knowledge silos; second, the service model is passive, and it cannot proactively push relevant knowledge according to user roles and task scenarios, resulting in a large number of records becoming unused "dark data". Summary of the Invention

[0004] In response to the above situation, the present invention adopts a proactive service triggering mechanism driven by both real-time user intent and profile characteristics. It makes real-time decisions and proactively provides three differentiated services: accurate push, related expansion, or risk warning, which can transform from passive retrieval to proactive perception.

[0005] The technical solution is an artificial intelligence-based record management system, comprising:

[0006] The document feature module is used to process a single document and output a set of quantified document feature parameters.

[0007] The associated scheduling module is used to perform the following operations when a user operates on a focused file D_a:

[0008] (a) Based on the archive feature parameters, calculate the comprehensive correlation degree R(D_a, D_x) between the focus archive D_a and any target archive D_x in the knowledge base, where R(D_a, D_x)∈[0, 1];

[0009] (b) Calculate the user intent entropy value H_U, which represents the degree of user intent focus or dispersion, based on the user's file operation behavior sequence within the current session window, where H_U ∈ [0, 1].

[0010] (c) A strategy decision function is used to map R(D_a, D_x), H_U, and the file category identifier C_a of the focus file D_a together into a strategy selection signal; the strategy decision function is defined as:

[0011] Decision =σ( W * R(D_a, D_x)-(1- W) * H_U);

[0012] Where σ is the Sigmoid activation function, and W is the dynamic decision weight, which is calculated as follows:

[0013] W = max( 0, min( 1, Base_W(C_a)+η*(R(D_a, D_x) - H_U)));

[0014] Where Base_W(C_a) is the basic weight value corresponding to the file category C_a obtained by querying a preset category-basic weight mapping table, and Base_W(C_a) ∈ (0, 1); η is a preset sensitivity coefficient, and 0 <η≤ 0.5; the max and min functions are used to constrain the calculation results within the interval [0, 1];

[0015] Based on the output of the Decision function, one of the following three proactive service strategies will be dynamically triggered:

[0016] When Decision > λ_1, the precise push strategy is triggered;

[0017] When λ_2 < Decision ≤ λ_1, the association expansion strategy is triggered;

[0018] When Decision ≤ λ_2, the risk warning strategy is triggered;

[0019] Where λ_1 and λ_2 are preset policy decision thresholds, and 1 > λ_1 > λ_2 > 0;

[0020] The control execution module is used to generate and output the service content corresponding to the specific policy that is triggered.

[0021] Furthermore, the quantified archival feature parameters output by the archival feature module include:

[0022] a) Issuing agency level value L: obtained by querying a pre-stored organizational tree structure table, where L is the depth of the issuing agency node in the tree structure, the depth of the root node is 1, and the depth of its child nodes increases layer by layer.

[0023] b) Timeliness Status Coefficient τ: Based on the file's effective timestamp t_s, expiration timestamp t_e, current system timestamp t_now, and file status flag, it is assigned a value according to the following rules: If the file is marked as "revoked", then τ = -1; if the file is marked as "valid indefinitely" and not marked as "revoked", then τ = 1; if t_now < t_s, then τ = 0; if t_s ≤ t_now ≤ t_e, then τ = 1; if t_now > t_e, then τ = -1.

[0024] c) Thematic feature vector F: It is generated by using a predefined vocabulary containing N core terms in the field of archives; segmenting the full text of the archives into words, counting the word frequency of each term and calculating its TF-IDF value to form an N-dimensional vector; and performing L2 norm normalization on the vector to obtain the thematic feature vector F.

[0025] d) Archive Category Identifier C: Obtained by classifying the archive title and abstract using a pre-trained text classification model. The categories include: "Regulations and Policies", "Operational Guidelines", "Historical Records", "Contracts and Agreements", and "Notices and Announcements".

[0026] Furthermore, the formula for calculating the comprehensive correlation degree R(D_a, D_x) is as follows:

[0027] R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c * J(C_a, C_x);

[0028] in:

[0029] F_a and F_x are the topic feature vectors of the focus file D_a and the target file D_x, respectively;

[0030] S(F_a, F_x) is the cosine similarity between vectors F_a and F_x;

[0031] P(L_a, L_x, τ_a, τ_x) is the effectiveness correlation function;

[0032] C_a and C_x are the file category identifiers for the focus file D_a and the target file D_x, respectively;

[0033] J(C_a, C_x) represents the category association degree. When C_a = C_x, J = 1.0; when C_a and C_x belong to a preset association category group, J = 0.5; otherwise, J = 0.

[0034] ω_s, ω_p, and ω_c are preset weight coefficients, and satisfy ω_s + ω_p + ω_c = 1.

[0035] Furthermore, the formula for calculating the effectiveness correlation function P(L_a, L_x, τ_a, τ_x) is as follows:

[0036] P(L_a, L_x,τ_a,τ_x) = [A(L_a, L_x)+T(τ_a, τ_x)] / 2;

[0037] in:

[0038] A(L_a, L_x) is a hierarchical association factor, which is obtained by querying a pre-defined authority relationship mapping table based on an organizational tree structure. This table defines the degree of influence of the validity of documents issued between different levels of organizations.

[0039] T(τ_a, τ_x) is a time-dependent synergy factor, and its value takes the following rules: when τ_a = 1 and τ_x = 1, T = 1; when τ_a = 1 and τ_x = 0, T = δ, where δ is a preset constant between 0 and 1; in other cases, T = -1.

[0040] Furthermore, the user intent entropy value H_U is calculated as follows: Let the set of files operated on by the user within the current session window be {D_1, D_2, ..., D_M}, and its corresponding set of topic feature vectors be {F_1, F_2, ..., F_M}, then...

[0041] H_U = [ 2 / (M*(M-1)) ] * Σ_{i=1}^{M-1} Σ_{j=i+1}^{M} (1 - S(F_i,F_j))

[0042] Where S(F_i, F_j) is the cosine similarity between vectors F_i and F_j.

[0043] Furthermore, the preset model parameters are configured as follows: the category-base weight mapping table is as follows: for the "laws and regulations" or "contracts and agreements" category, Base_W is 0.7; for the "operation guide" category, Base_W is 0.5; for the "historical records" or "notices and announcements" category, Base_W is 0.3; and the sensitivity coefficient η is 0.2.

[0044] The strategy decision threshold λ_1 is set to 0.7, and λ_2 is set to 0.4.

[0045] Furthermore, the service content generated by the control execution module includes:

[0046] When the precise push strategy is triggered, the service content is a recommendation card containing the target file D_x identifier and the comprehensive relevance R(D_a, D_x) value;

[0047] When the association expansion strategy is triggered, the service content is a visualization graph consisting of the K most frequently accessed file nodes and their association relationships, which are constructed with files as nodes and the comprehensive association degree R between files as edge weights, starting from the focus file D_a and performing several random walks.

[0048] When the risk warning strategy is triggered, the service content is a warning report that clearly points out that there is a validity conflict between files D_a and D_x or that file D_x has expired;

[0049] The risk warning strategy is triggered when any of the following conditions are met: the timeliness coordination factor T(τ_a, τ_x) = -1, or the comprehensive correlation degree R(D_a, D_x) is lower than the preset risk threshold R_alert; wherein the risk threshold R_alert is 0.3.

[0050] An artificial intelligence-based record management method, applied to the system described in any one of claims 1-7, includes:

[0051] S1: Extract the quantitative feature parameters of the archives, including the issuing agency level value L, the timeliness status coefficient τ, the subject feature vector F, and the archive category identifier C;

[0052] S2: When a user interacts with the focused file D_a, the comprehensive correlation degree R(D_a, D_x) between the focused file D_a and the target file D_x is calculated according to the formula R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c * J(C_a, C_x), where R(D_a, D_x) ∈ [0, 1];

[0053] S3: Calculate the user intent entropy value H_U based on the user's operation sequence within the current session window, where H_U ∈ [0, 1];

[0054] S4: Based on the category identifier C_a of the focus file D_a, query the preset category-basic weight mapping table to obtain its basic weight value Base_W(C_a), where Base_W(C_a) ∈ (0, 1);

[0055] S5: Calculate the dynamic decision weight W = max(0, min(1, Base_W(C_a)+η* (R(D_a, D_x) -H_U) ) ), where η is the preset sensitivity coefficient, and 0 < η ≤ 0.5;

[0056] S6: Substitute W into the decision function Decision = σ( W * R(D_a, D_x) - (1 - W) * H_U ), and trigger precise push, association expansion or risk warning strategies based on the comparison results of the output value with the thresholds λ_1 and λ_2, where 1 > λ_1 > λ_2 > 0;

[0057] S7: Generate and output the corresponding service content based on the triggered strategy.

[0058] Furthermore, in step S3, the user's current session window is determined by an adaptive time mechanism: starting from the current operation time and going back, if the time interval between two consecutive file operation behaviors exceeds a preset silence threshold T_idle, then the session window is interrupted with the silence threshold as the boundary; otherwise, the window continues to accumulate; the value range of the silence threshold T_idle is 5 to 30 minutes.

[0059] Furthermore, in step S7, the generated risk warning report classifies and labels the types of validity conflicts: when the validity status coefficient τ_x of the target file D_x is -1, it is labeled as "referring to an expired file" risk; when the validity status coefficient τ_a of the focus file D_a is 1 and the validity status coefficient τ_x of the target file D_x is -1, and L_a < L_x, it is labeled as "lower-level file referencing an expired upper-level file" high risk.

[0060] Due to the adoption of the above technical solutions, the present invention has the following advantages compared with the prior art;

[0061] 1. By quantifying the user's current focus in real time through "user intent entropy (H_U)," and using the difference between relevance and intent entropy (R-H_U) as a single feedback signal, scattered files are connected into a knowledge network with business logic, solving the problem of knowledge silos. Bounded and smooth real-time fine-tuning of decision weights is performed, significantly improving service matching.

[0062] 2. Through decision functions, seamless adaptive triggering and content generation of three differentiated services (precise push, related expansion, and risk warning) are achieved. Complex multi-factor decisions are mapped to a continuous spectrum value, and the system is clearly divided into three strategy intervals through thresholds. The service content and decision logic are deeply coupled, so that users do not need to switch function modules. The system can provide the most suitable composite service based on the current context. Risk prevention and control, knowledge association and precise retrieval capabilities are integrated into a unified intelligent interactive experience, enabling the system to transform from passive retrieval to active perception, which greatly improves the system's usability and intelligence level. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the modules of an artificial intelligence-based archive management system and method according to the present invention.

[0064] Figure 2 This is a flowchart of an artificial intelligence-based archive management system and method according to the present invention. Detailed Implementation

[0065] The foregoing and other technical contents, features and effects of the present invention are described in conjunction with the appendix below. Figures 1 to 2 As will be clearly shown in the detailed description of the embodiments, the structural content mentioned in the following embodiments is with reference to the accompanying drawings.

[0066] Example 1: Based on existing technology, an artificial intelligence-based archive management system includes:

[0067] The document feature module is used to process a single document and output a set of quantified document feature parameters.

[0068] The associated scheduling module is used to perform the following operations when a user operates on a focused file D_a:

[0069] (a) Based on the archive feature parameters, calculate the comprehensive correlation degree R(D_a, D_x) between the focus archive D_a and any target archive D_x in the knowledge base, where R(D_a, D_x)∈[0, 1];

[0070] (b) Calculate the user intent entropy value H_U, which represents the degree of user intent focus or dispersion, based on the user's file operation behavior sequence within the current session window, where H_U ∈ [0, 1].

[0071] (c) A strategy decision function is used to map R(D_a, D_x), H_U, and the file category identifier C_a of the focus file D_a together into a strategy selection signal; the strategy decision function is defined as:

[0072] Decision =σ( W * R(D_a, D_x)-(1- W) * H_U);

[0073] Where σ is the Sigmoid activation function, and W is the dynamic decision weight, which is calculated as follows:

[0074] W = max( 0, min( 1, Base_W(C_a)+η*(R(D_a, D_x) - H_U)));

[0075] Where Base_W(C_a) is the basic weight value corresponding to the file category C_a obtained by querying a preset category-basic weight mapping table, and Base_W(C_a) ∈ (0, 1); η is a preset sensitivity coefficient, and 0 <η≤ 0.5; the max and min functions are used to constrain the calculation results within the interval [0, 1];

[0076] Based on the output of the Decision function, one of the following three proactive service strategies will be dynamically triggered:

[0077] When Decision > λ_1, the precise push strategy is triggered;

[0078] When λ_2 < Decision ≤ λ_1, the association expansion strategy is triggered;

[0079] When Decision ≤ λ_2, the risk warning strategy is triggered;

[0080] Where λ_1 and λ_2 are preset policy decision thresholds, and 1 > λ_1 > λ_2 > 0;

[0081] The control execution module is used to generate and output the service content corresponding to the specific policy that is triggered.

[0082] Furthermore, the quantified archival feature parameters output by the archival feature module include:

[0083] a) Issuing agency level value L: obtained by querying a pre-stored organizational tree structure table, where L is the depth of the issuing agency node in the tree structure, the depth of the root node is 1, and the depth of its child nodes increases layer by layer.

[0084] b) Timeliness Status Coefficient τ: Based on the file's effective timestamp t_s, expiration timestamp t_e, current system timestamp t_now, and file status flag, it is assigned a value according to the following rules: If the file is marked as "revoked", then τ = -1; if the file is marked as "valid indefinitely" and not marked as "revoked", then τ = 1; if t_now < t_s, then τ = 0; if t_s ≤ t_now ≤ t_e, then τ = 1; if t_now > t_e, then τ = -1.

[0085] c) Thematic feature vector F: It is generated by using a predefined vocabulary containing N core terms in the field of archives; segmenting the full text of the archives into words, counting the word frequency of each term and calculating its TF-IDF value to form an N-dimensional vector; and performing L2 norm normalization on the vector to obtain the thematic feature vector F.

[0086] d) Archive Category Identifier C: Obtained by classifying the archive title and abstract using a pre-trained text classification model. The categories include: "Regulations and Policies", "Operational Guidelines", "Historical Records", "Contracts and Agreements", and "Notices and Announcements".

[0087] Furthermore, the formula for calculating the comprehensive correlation degree R(D_a, D_x) is as follows:

[0088] R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c * J(C_a, C_x);

[0089] in:

[0090] F_a and F_x are the topic feature vectors of the focus file D_a and the target file D_x, respectively;

[0091] S(F_a, F_x) is the cosine similarity between vectors F_a and F_x;

[0092] P(L_a, L_x, τ_a, τ_x) is the effectiveness correlation function;

[0093] C_a and C_x are the file category identifiers for the focus file D_a and the target file D_x, respectively;

[0094] J(C_a, C_x) represents the category association degree. When C_a = C_x, J = 1.0; when C_a and C_x belong to a preset association category group, J = 0.5; otherwise, J = 0.

[0095] ω_s, ω_p, and ω_c are preset weight coefficients, and satisfy ω_s + ω_p + ω_c = 1.

[0096] Where the topic similarity S(F_a, F_x):

[0097] Implementation method: Cosine similarity is used. First, a predefined vocabulary containing core terms in the archival field is constructed, such as: [“punishment”, “approval”, “filing”, “effectiveness”, “implementation”, “repeal”, “revise”, “authorization”, “supervision”, “compliance”, “term”, “institution”, “level”, “effective”, “expiration”]. The full text of each archive is segmented and stop words are removed. The term frequency (TF) of each term in the vocabulary is calculated, and its inverse document frequency (IDF) is calculated based on a corpus containing tens of thousands of archives. The TF-IDF value of each term is then obtained, forming an N-dimensional vector. Finally, the vector is normalized to L2 norm to obtain the topic feature vector F.

[0098] Example: The similarity S is obtained by calculating the dot product of the topic vector F_a of file D_a and the topic vector F_x of file D_x and dividing it by their respective magnitudes. The value range is [0, 1]. The calculated value is S(F_a, F_x) = 0.85, indicating that the two topics are highly related.

[0099] Category association degree J(C_a, C_x):

[0100] Implementation method: Pre-set associated category groups based on business logic. Details are as follows:

[0101] Group 1 (Regulatory): {"Laws and Policies", "Operational Guidelines"};

[0102] Group 2 (Transaction Classes): {"Contracts and Agreements", "Notices and Announcements"};

[0103] "Historical records" are a category of their own.

[0104] Value selection rules:

[0105] If C_a = C_x, then J = 1.0.

[0106] If C_a and C_x belong to the same association category group, then J = 0.5.

[0107] Otherwise, J = 0.

[0108] Example: D_a is "laws and policies", and D_x is "operational guidelines". Both belong to the "regulatory category", so J = 0.5.

[0109] Weighting coefficients (ω_s, ω_p, ω_c):

[0110] Example configuration 1: ω_s = 0.4, ω_p = 0.4, ω_c = 0.2. This configuration takes into account content, effectiveness and category, and is applicable to general scenarios.

[0111] Example configuration 2: ω_s = 0.3, ω_p = 0.5, ω_c = 0.2. This configuration strengthens the weight of the effectiveness association and is suitable for risk management scenarios.

[0112] Furthermore, the formula for calculating the effectiveness correlation function P(L_a, L_x, τ_a, τ_x) is as follows:

[0113] P(L_a, L_x,τ_a,τ_x) = [A(L_a, L_x)+T(τ_a, τ_x)] / 2;

[0114] in:

[0115] A(L_a, L_x) is a hierarchical association factor, which is obtained by querying a pre-defined authority relationship mapping table based on an organizational tree structure. This table defines the degree of influence of the validity of documents issued between different levels of organizations.

[0116] T(τ_a, τ_x) is a time-dependent synergy factor, and its value takes the following rules: when τ_a = 1 and τ_x = 1, T = 1; when τ_a = 1 and τ_x = 0, T = δ, where δ is a preset constant between 0 and 1; in other cases, T = -1.

[0117] The effectiveness correlation function P(L_a, L_x, τ_a, τ_x):

[0118] Implementation: P = [ A(L_a, L_x) + T(τ_a, τ_x) ] / 2.

[0119] Hierarchical association factor A(L_a, L_x): obtained by querying the preset "authority relationship mapping table".

[0120] This table is based on the fundamental principles of administrative organization law and internal management practices, as follows:

[0121] Hierarchical relationship Values ​​of A(L_a, L_x) illustrate L_a == L_x (same level organization) 1.0 Equal effect L_a == L_x - 1 (direct superior to subordinate) 0.8 Superiors have the authority to guide subordinates. L_a == L_x + 1 (direct subordinate to superior) 0.6 Subordinates must obey superiors, but the effect is slightly weaker. abs(L_a - L_x) >= 2 (stepwise) 0.4 The effectiveness decreases with increasing hierarchical distance. Unable to determine the affiliation 0.1 Default weak association

[0122] The time-dependent synergy factor T(τ_a, τ_x) has the following value rules:

[0123] When τ_a = 1 and τ_x = 1 (both are valid), T = 1.

[0124] When τ_a = 1 and τ_x = 0 (focus is effective, target is not effective), T = δ; δ is a preset constant that reflects the reference value for ineffective documents, usually taken as δ = 0.5.

[0125] In other cases (such as when any file has been decommissioned or expired), T = -1. This negative value will significantly lower the P value, thereby triggering a risk warning.

[0126] Furthermore, the user intent entropy value H_U is calculated as follows: Let the set of files operated on by the user within the current session window be {D_1, D_2, ..., D_M}, and its corresponding set of topic feature vectors be {F_1, F_2, ..., F_M}, then...

[0127] H_U = [ 2 / (M*(M-1)) ] * Σ_{i=1}^{M-1} Σ_{j=i+1}^{M} (1 - S(F_i,F_j))

[0128] Where S(F_i, F_j) is the cosine similarity between vectors F_i and F_j.

[0129] The calculation of the user intent entropy value H_U depends on the definition of user actions within the "current session window". To achieve dynamic and accurate window division, an adaptive determination mechanism based on the time interval of the action is adopted.

[0130] The specific implementation of the mechanism is as follows: The system monitors the user's opening and viewing operations on files in real time, and records the file identifier and precise timestamp corresponding to each operation. When it is necessary to calculate the user intent entropy H_U at the current moment, the system uses the time point of the most recent operation as the benchmark and traces back the user's operation history sequence.

[0131] During the backtracking process, the system introduces a preset "silent time threshold" (denoted as T_idle, with a typical value range of 5 to 30 minutes, such as 15 minutes). This threshold represents the time limit for the system to determine whether two operations belong to the same continuous work session. The system calculates the time interval between adjacent operations in turn. If a certain time interval is greater than the silent time threshold T_idle, the system determines that the user's work focus has been interrupted at this point. Using this time point as the boundary, the subsequent (closer to the present in time) continuous operation sequence is defined as the "current session window". If all the checked time intervals do not exceed T_idle, the backtracking will continue until the maximum session length or the starting point of the operation sequence set by the system is reached.

[0132] In this way, the "current session window" is no longer a static interval with a fixed duration or a fixed number of operations, but an interval that is dynamically adjusted according to the actual density of user operations. This design ensures that the user intent entropy H_U can accurately reflect the degree of user intent focus or dispersion during a recent period of continuous and focused work, thereby providing a more reliable basis for subsequent dynamic decision-making.

[0133] Furthermore, the preset model parameters are configured as follows: the category-base weight mapping table is as follows: for the "laws and regulations" or "contracts and agreements" category, Base_W is 0.7; for the "operation guide" category, Base_W is 0.5; for the "historical records" or "notices and announcements" category, Base_W is 0.3; and the sensitivity coefficient η is 0.2.

[0134] Category-Base Weight Mapping Table (Base_W(C_a));

[0135] File Category C Base_W value Reasons for setting Laws and policies 0.7 Citations must be precise, and objective relevance (R) should be given high priority. Contract Agreement 0.7 The terms are closely related and require precise matching to prevent ambiguity. Operating Instructions 0.5 Between precise operation and extended reference Notice 0.3 With its strong timeliness and interconnectedness, it should focus more on the user's current intent. Historical records 0.3 Primarily used for background queries, with high value in related extensions.

[0136] Meaning: Base_W reflects the system's initial level of trust in the "objective record association".

[0137] Sensitivity coefficient η; Value and basis: η = 0.2. This value ensures that the variation range of the fine-tuning term η*(R-H_U) is within ±0.2. The purpose is to introduce a moderate situational adaptability while respecting the preset weight of the file category (Base_W). When (R-H_U) is a large positive number, it indicates that the correlation of this calculation is very high and the user intent is concentrated. In this case, W increases and the decision depends more on R. Conversely, W decreases and the decision pays more attention to the dispersed user intent H_U.

[0138] The strategy decision threshold λ_1 is set to 0.7, and λ_2 is set to 0.4.

[0139] Furthermore, the service content generated by the control execution module includes:

[0140] When the precise push strategy is triggered, the service content is a recommendation card containing the target file D_x identifier and the comprehensive relevance R(D_a, D_x) value;

[0141] When the association expansion strategy is triggered, the service content is a visualization graph consisting of the K most frequently accessed file nodes and their association relationships, which are constructed with files as nodes and the comprehensive association degree R between files as edge weights in a relation graph. Starting from the focus file D_a, a few random walks are performed.

[0142] When the risk warning strategy is triggered, the service content is a warning report that clearly points out that there is a validity conflict between files D_a and D_x or that file D_x has expired;

[0143] The risk warning strategy is triggered when any of the following conditions are met: the timeliness coordination factor T(τ_a, τ_x) = -1, or the comprehensive correlation degree R(D_a, D_x) is lower than the preset risk threshold R_alert; wherein the risk threshold R_alert is 0.3.

[0144] When the system triggers the association expansion strategy, the specific method for generating the visualized knowledge graph includes the following steps:

[0145] Step 1: Construct a graph model of the relationship between archives.

[0146] The system abstracts each file in the knowledge base into a node. For any two files D_i and D_j, if their comprehensive correlation R(D_i, D_j) is greater than a set minimum correlation threshold (such as 0.05), an edge is established between their corresponding nodes, and the calculated comprehensive correlation R value is assigned to the edge as a weight. In this way, a weighted undirected graph (or directed graph) model with files as nodes and comprehensive correlation R as edge weight is formed.

[0147] Step 2: Execute a weighted path exploration starting from the focus file.

[0148] Based on the relationship graph constructed in the previous step, the system initiates an exploration process starting from D_a. This process simulates a traversal path, and its core rule is: in each exploration step, the probability of moving from the current file node to the next adjacent node is proportional to the weight of the edge connecting the two nodes (i.e., the comprehensive correlation degree R). This means that the higher the correlation degree, the greater the probability of the edge being selected as the exploration path. This exploration process will perform a large number of such steps, thereby forming multiple random walk paths in the graph that start from D_a and are guided by the correlation degree weight.

[0149] Step 3: Aggregate and filter key nodes based on the exploration results.

[0150] The system counts the total frequency of each archive node being accessed during the above path exploration process. The access frequency reflects the connectivity strength and hub status of the node with the focus archive D_a among many possible paths. Then, the system excludes the focus archive D_a itself and sorts them from high to low according to the access frequency, selecting the top 10 archive nodes as the set of key nodes that are most worth expanding and associating with the focus archive.

[0151] Step 4: Generate and output a visual map.

[0152] The system extracts the K key nodes, the focus file D_a node, and all edges with high weight between these nodes to form a concise and information-rich subgraph. The data of this subgraph is passed to the visualization engine and presented to the user in a graphical way. The size of the nodes in the graph can represent their access frequency, and the thickness of the edges can represent the strength of the correlation. This allows the user to intuitively understand the deep network of connections between files and realize the expansion of knowledge.

[0153] An artificial intelligence-based record management method includes:

[0154] S1: Extract the quantitative feature parameters of the archives, including the issuing agency level value L, the timeliness status coefficient τ, the subject feature vector F, and the archive category identifier C;

[0155] S2: When a user interacts with the focused file D_a, the comprehensive correlation degree R(D_a, D_x) between the focused file D_a and the target file D_x is calculated according to the formula R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c * J(C_a, C_x), where R(D_a, D_x) ∈ [0, 1];

[0156] S3: Calculate the user intent entropy value H_U based on the user's operation sequence within the current session window, where H_U ∈ [0, 1];

[0157] S4: Based on the category identifier C_a of the focus file D_a, query the preset category-basic weight mapping table to obtain its basic weight value Base_W(C_a), where Base_W(C_a) ∈ (0, 1);

[0158] S5: Calculate the dynamic decision weight W=max(0, min(1, Base_W(C_a)+η* (R(D_a, D_x) - H_U) ) ), where η is the preset sensitivity coefficient, and 0<η≤0.5;

[0159] S6: Substitute W into the decision function Decision = σ( W * R(D_a, D_x) - (1 - W) * H_U ), and trigger precise push, association expansion or risk warning strategies based on the comparison results of the output value with the thresholds λ_1 and λ_2, where 1 > λ_1 > λ_2 > 0;

[0160] S7: Generate and output the corresponding service content based on the triggered strategy.

[0161] Furthermore, in step S3, the user's current session window is determined by an adaptive time mechanism: starting from the current operation time and going back, if the time interval between two consecutive file operation behaviors exceeds a preset silence threshold T_idle, then the session window is interrupted with the silence threshold as the boundary; otherwise, the window continues to accumulate; the value range of the silence threshold T_idle is 5 to 30 minutes.

[0162] Furthermore, in step S7, the generated risk warning report classifies and labels the types of validity conflicts: when the validity status coefficient τ_x of the target file D_x is -1, it is labeled as "referring to an expired file" risk; when the validity status coefficient τ_a of the focus file D_a is 1 and the validity status coefficient τ_x of the target file D_x is -1, and L_a < L_x, it is labeled as "lower-level file referencing an expired upper-level file" high risk.

[0163] In practical use, based on existing technology, the archival feature module is used to process a single archival document and output a set of quantified archival feature parameters.

[0164] The associated scheduling module is used to perform the following operations when a user operates on a focused file D_a:

[0165] (a) Based on the archive feature parameters, calculate the comprehensive correlation degree R(D_a, D_x) between the focus archive D_a and any target archive D_x in the knowledge base, where R(D_a, D_x)∈[0, 1];

[0166] (b) Calculate the user intent entropy value H_U, which represents the degree of user intent focus or dispersion, based on the user's sequence of file operation behaviors in the current session window, where H_U ∈ [0, 1].

[0167] (c) A strategy decision function is used to map R(D_a, D_x), H_U, and the file category identifier C_a of the focus file D_a together into a strategy selection signal; the strategy decision function is defined as:

[0168] Decision =σ( W * R(D_a, D_x)-(1- W) * H_U);

[0169] Where σ is the Sigmoid activation function, and W is the dynamic decision weight, which is calculated as follows:

[0170] W = max( 0, min( 1, Base_W(C_a)+η*(R(D_a, D_x) - H_U)));

[0171] Where Base_W(C_a) is the basic weight value corresponding to the file category C_a obtained by querying a preset category-basic weight mapping table, and Base_W(C_a) ∈ (0, 1); η is a preset sensitivity coefficient, and 0<η≤0.5; the max and min functions are used to constrain the calculation results within the interval [0, 1];

[0172] Based on the output of the Decision function, one of the following three proactive service strategies will be dynamically triggered:

[0173] When Decision > λ_1, the precise push strategy is triggered;

[0174] When λ_2 < Decision ≤ λ_1, the association expansion strategy is triggered;

[0175] When Decision ≤ λ_2, the risk warning strategy is triggered;

[0176] Wherein, λ_1 and λ_2 are preset policy decision thresholds, and 1 > λ_1 > λ_2 > 0;

[0177] The control execution module is used to generate and output the service content corresponding to the specific policy that is triggered.

[0178] The above description is a further detailed explanation of the present invention in conjunction with specific embodiments, and it should not be considered that the specific implementation of the present invention is limited to this. For those skilled in the art to which the present invention pertains and related fields, any extensions, operation methods, and data substitutions made based on the technical solution concept of the present invention should fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based archive management system, characterized in that, include: The document feature module is used to process a single document and output a set of quantified document feature parameters. The associated scheduling module is used to perform the following operations when a user operates on a focused file D_a: (a) Based on the archive feature parameters, calculate the comprehensive correlation degree R(D_a, D_x) between the focus archive D_a and any target archive D_x in the knowledge base, where R(D_a, D_x)∈[0, 1]; (b) Calculate the user intent entropy value H_U, which represents the degree of user intent focus or dispersion, based on the user's sequence of file operation behaviors in the current session window, where H_U ∈ [0, 1]. (c) A strategy decision function is used to map R(D_a, D_x), H_U, and the file category identifier C_a of the focus file D_a together into a strategy selection signal; the strategy decision function is defined as: Decision =σ( W * R(D_a, D_x)-(1- W) * H_U); Where σ is the Sigmoid activation function, and W is the dynamic decision weight, which is calculated as follows: W = max( 0, min( 1, Base_W(C_a)+η*(R(D_a, D_x) - H_U))); Where Base_W(C_a) is the basic weight value corresponding to the file category C_a obtained by querying a preset category-basic weight mapping table, and Base_W(C_a) ∈ (0, 1); η is a preset sensitivity coefficient, and 0 <η≤0.5; the max and min functions are used to constrain the calculation results within the interval [0, 1]; Based on the output of the Decision function, one of the following three proactive service strategies will be dynamically triggered: When Decision > λ_1, the precise push strategy is triggered; When λ_2 < Decision ≤ λ_1, the association expansion strategy is triggered; When Decision ≤ λ_2, the risk warning strategy is triggered; Wherein, λ_1 and λ_2 are preset policy decision thresholds, and 1 > λ_1 > λ_2 > 0; The control execution module is used to generate and output the service content corresponding to the specific policy that is triggered.

2. The system according to claim 1, characterized in that, The quantitative archival feature parameters output by the archival feature module include: a) Issuing agency level value L: obtained by querying a pre-stored organizational tree structure table, where L is the depth of the issuing agency node in the tree structure, the depth of the root node is 1, and the depth of its child nodes increases layer by layer. b) Timeliness Status Coefficient τ: Based on the file's effective timestamp t_s, expiration timestamp t_e, current system timestamp t_now, and file status flag, it is assigned a value according to the following rules: If the file is marked as "revoked", then τ = -1; if the file is marked as "valid indefinitely" and not marked as "revoked", then τ = 1; if t_now < t_s, then τ = 0; if t_s ≤ t_now ≤ t_e, then τ = 1; if t_now > t_e, then τ = -1. c) Thematic feature vector F: It is generated by using a predefined vocabulary containing N core terms in the field of archives; segmenting the full text of the archives into words, counting the word frequency of each term and calculating its TF-IDF value to form an N-dimensional vector; and performing L2 norm normalization on the vector to obtain the thematic feature vector F. d) Archive Category Identifier C: Obtained by classifying the archive title and abstract using a pre-trained text classification model. The categories include: "Regulations and Policies", "Operational Guidelines", "Historical Records", "Contracts and Agreements", and "Notices and Announcements".

3. The system according to claim 2, characterized in that, The formula for calculating the comprehensive correlation degree R(D_a, D_x) is as follows: R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c *J(C_a, C_x); in: F_a and F_x are the topic feature vectors of the focus file D_a and the target file D_x, respectively; S(F_a, F_x) is the cosine similarity between vectors F_a and F_x; P(L_a, L_x, τ_a, τ_x) is the effectiveness correlation function; C_a and C_x are the file category identifiers for the focus file D_a and the target file D_x, respectively; J(C_a, C_x) represents the category association degree. When C_a = C_x, J = 1.0; when C_a and C_x belong to a preset association category group, J = 0.5; otherwise, J = 0. ω_s, ω_p, and ω_c are preset weight coefficients, and satisfy ω_s + ω_p + ω_c = 1.

4. The system according to claim 3, characterized in that, The formula for calculating the effectiveness correlation function P(L_a, L_x,τ_a,τ_x) is as follows: P(L_a, L_x,τ_a,τ_x) = [A(L_a, L_x)+T(τ_a, τ_x)] / 2; in: A(L_a, L_x) is a hierarchical association factor, which is obtained by querying a pre-defined authority relationship mapping table based on the organizational tree structure. This table defines the degree of influence of the validity of documents issued between different levels of institutions. T(τ_a, τ_x) is a time-dependent synergy factor, and its value takes the following rules: when τ_a = 1 and τ_x = 1, T = 1; when τ_a = 1 and τ_x = 0, T = δ, where δ is a preset constant between 0 and 1; in other cases, T = -1.

5. The system according to claim 1, characterized in that, The user intent entropy value H_U is calculated as follows: Let the set of files operated by the user within the current session window be {D_1, D_2, ..., D_M}, and the corresponding set of topic feature vectors be {F_1, F_2, ..., F_M}, then... H_U = [ 2 / (M*(M-1)) ] * Σ_{i=1}^{M-1} Σ_{j=i+1}^{M} (1 - S(F_i, F_j)) Where S(F_i, F_j) is the cosine similarity between vectors F_i and F_j.

6. The system according to claim 1, characterized in that, The preset model parameters are configured as follows: the category-base weight mapping table is as follows: for the "laws and regulations" or "contracts and agreements" category, Base_W is 0.7; for the "operation guide" category, Base_W is 0.5; for the "historical records" or "notices and announcements" category, Base_W is 0.3; the sensitivity coefficient η is 0.

2. The strategy decision threshold λ_1 is set to 0.7, and λ_2 is set to 0.

4.

7. The system according to claim 1, characterized in that, The service content generated by the control execution module includes: When the precise push strategy is triggered, the service content is a recommendation card containing the target file D_x identifier and the comprehensive relevance R(D_a, D_x) value; When the association expansion strategy is triggered, the service content is a visualization graph consisting of the K most frequently accessed file nodes and their association relationships, which are constructed with files as nodes and the comprehensive association degree R between files as edge weights in a relation graph. Starting from the focus file D_a, a few random walks are performed. When the risk warning strategy is triggered, the service content is a warning report that clearly points out that there is a validity conflict between files D_a and D_x or that file D_x has expired; The risk warning strategy is triggered when any of the following conditions are met: the timeliness coordination factor T(τ_a, τ_x) = -1, or the comprehensive correlation degree R(D_a, D_x) is lower than the preset risk threshold R_alert; wherein the risk threshold R_alert is 0.

3.

8. An artificial intelligence-based record management method, applied to the system described in any one of claims 1-7, characterized in that, include: S1: Extract the quantitative feature parameters of the archives, including the issuing agency level value L, the timeliness status coefficient τ, the subject feature vector F, and the archive category identifier C; S2: When a user interacts with the focused file D_a, the comprehensive correlation degree R(D_a, D_x) between the focused file D_a and the target file D_x is calculated according to the formula R(D_a, D_x) = ω_s * S(F_a, F_x) + ω_p * P(L_a, L_x, τ_a, τ_x) + ω_c * J(C_a, C_x), where R(D_a, D_x) ∈ [0, 1]; S3: Calculate the user intent entropy value H_U based on the user's operation sequence within the current session window, where H_U ∈ [0,1]; S4: Based on the category identifier C_a of the focus file D_a, query the preset category-basic weight mapping table to obtain its basic weight value Base_W(C_a), where Base_W(C_a) ∈ (0, 1); S5: Calculate the dynamic decision weight W = max(0, min(1, Base_W(C_a)+η* (R(D_a, D_x) - H_U)) ), where η is the preset sensitivity coefficient, and 0 <η≤ 0.5; S6: Substitute W into the decision function Decision = σ( W * R(D_a, D_x) - (1 - W) * H_U ), and trigger precise push, association expansion or risk warning strategies based on the comparison results of the output value with the thresholds λ_1 and λ_2, where 1 > λ_1 > λ_2 > 0; S7: Generate and output the corresponding service content based on the triggered strategy.

9. The method according to claim 8, characterized in that, In step S3, the user's current session window is determined by an adaptive time mechanism: starting from the current operation time and going back, if the time interval between two consecutive file operation behaviors exceeds a preset silence threshold T_idle, the session window is interrupted with the silence threshold as the boundary; otherwise, the window continues to accumulate. The value range of the silence threshold T_idle is 5 to 30 minutes.

10. The method according to claim 8, characterized in that, In step S7, the generated risk warning report classifies and labels the types of validity conflicts: when the validity status coefficient τ_x of the target file D_x is -1, it is labeled as "referring to an expired file" risk; when the validity status coefficient τ_a of the focus file D_a is 1 and the validity status coefficient τ_x of the target file D_x is -1, and L_a < L_x, it is labeled as "lower-level file referencing an expired upper-level file" high risk.

Images