Digital publishing process intelligent management system based on knowledge graph

Abstract

The invention discloses a digital publishing process intelligent management system based on a knowledge graph, and the system comprises a standardized data generation module which is used for obtaining digital publishing business data; the semantic and entity construction module is used for extracting publishing related semantic features to form a publishing knowledge graph; the semantic weight and feature generation module is used for generating a semantic weight matrix and a feature vector set; the causal structure learning module is used for determining a causal directed acyclic graph structure through an improved NOTEARS algorithm model under the action of semantic constraint, time dependence constraint and a causal attention self-regulation mechanism; the key task identification module is used for generating a task dependence path; the priority calculation module is used for calculating task execution priorities and generating a publishing process scheduling scheme; and the flow structure output module is used for outputting the publishing flow structure and the publishing flow scheduling scheme. According to the invention, structured presentation and intelligent optimization of the publishing process are realized.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph and business process optimization technology, and in particular to an intelligent management system for digital publishing processes based on knowledge graphs. Background Technology

[0002] With the continuous expansion of digital publishing operations and the accelerated iteration of content production processes, publishing institutions generate a large amount of complex and diverse data in manuscript processing, task allocation, personnel collaboration, and process scheduling. This data covers various task logs related to manuscript editing, proofreading, typesetting, and production, as well as collaboration records involving multiple roles. Existing publishing process management typically relies on static process templates or experience-driven manual scheduling methods, controlling the task sequence through simple rule engines or fixed flowcharts. However, when faced with complex process scenarios involving multiple manuscripts and multiple roles processing in parallel, problems often arise such as insufficient scheduling accuracy, inaccurate identification of process node dependencies, and accumulated execution delays.

[0003] Existing research attempts to introduce data-driven methods to analyze the publishing process, including inferring dependencies between tasks based on statistical correlation and classifying or clustering manuscript, task, and personnel information using shallow semantic models. However, publishing data commonly suffers from inconsistent semantic expressions, inconsistent task log granularity, and missing or non-standardized time records, making it difficult for traditional statistical methods to accurately identify the true dependencies between tasks. Furthermore, existing technologies often ignore the semantic connections between entities, treating different roles, tasks, or manuscripts as independent records, failing to effectively integrate semantic information to infer the process structure. Due to the lack of a linkage mechanism between semantic and temporal features, semantically similar nodes are frequently incorrectly associated, or nodes with temporal constraints are incorrectly modeled, leading to discrepancies between the learned process structure and the actual publishing process.

[0004] In the field of causal structure learning, existing technologies largely rely on traditional algorithms to learn the causal structure of processes from data. However, these algorithms mostly focus only on the statistical relationships between numerical features and lack the ability to handle multidimensional constraints such as text semantics and task time dependencies in publishing scenarios. They cannot solve the structural inference problem of complex semantic relationships and strong temporal dependencies in the publishing process. In addition, existing causal structure learning methods generally lack adaptive adjustment mechanisms for the edge weight update process. When the input data contains semantic noise or temporal shifts, causal structure learning is prone to getting trapped in local optima or even generating incorrect circular structures, affecting the reliability of the final causal directed acyclic graph.

[0005] In terms of process optimization and scheduling, existing technologies typically employ rule-based priority calculation or key task identification methods based on simple graph traversal. These methods fail to comprehensively optimize task execution order by incorporating multi-dimensional information such as task frequency, node continuity, and execution duration. Especially in complex publishing processes with numerous path branches and significant differences in task roles, existing scheduling methods struggle to balance process dependency constraints and resource allocation constraints. This results in a lack of overall optimization in the final scheduling scheme, failing to meet the business requirements of process visualization, optimizability, and searchability in digital publishing scenarios.

[0006] Therefore, how to provide an intelligent management system for digital publishing processes based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent management system for digital publishing processes based on knowledge graphs. This invention constructs a publishing knowledge graph, generates a semantic weight matrix, and introduces an improved NOTEARS causal structure learning method to achieve joint modeling of semantic, temporal, and causal relationships, significantly improving the accuracy of task dependency identification and key node location. Through priority calculation and scheduling generation mechanisms, a searchable optimized publishing process structure and scheduling scheme are formed, solving the problems of unclear process dependencies and inefficient scheduling in existing processes.

[0008] According to an embodiment of the present invention, a knowledge graph-based intelligent management system for digital publishing processes includes:

[0009] The standardized data generation module is used to acquire digital publishing business data and perform cleaning, formatting and standardization processes to generate standardized datasets.

[0010] The semantic and entity construction module is used to extract publishing-related semantic features from the standardized dataset, establish entity sets, determine the semantic relationships between entities, and form a publishing knowledge graph.

[0011] The semantic weight and feature generation module is used to generate a semantic weight matrix based on the published knowledge graph and to encode a feature vector set from the standardized dataset;

[0012] The causal structure learning module is used to input the semantic weight matrix and feature vector set into the improved NOTEARS algorithm model, and determine the causal directed acyclic graph structure under the action of semantic constraints, time dependency constraints and causal attention self-regulation mechanism.

[0013] The critical task identification module is used to generate task dependency paths based on the causal directed acyclic graph structure, calculate the frequency and continuity of task nodes, and determine critical task nodes and their relationships.

[0014] The priority calculation module is used to calculate the task execution priority based on the task dependency path and the task duration in the standardized dataset, and generate a publication process scheduling scheme.

[0015] The process structure output module is used to generate process structure records based on the publishing process scheduling scheme, and output the publishing process structure and the publishing process scheduling scheme.

[0016] Optionally, modules can be integrated using the following methods:

[0017] S1. Acquire digital publishing business data and perform cleaning, formatting and standardization processing to generate a standardized dataset;

[0018] S2. Extract publishing-related semantic features from the standardized dataset, establish an entity set, determine the semantic relationships between entities, and construct a publishing knowledge graph containing entity nodes and semantic relationship edges;

[0019] S3. Based on the published knowledge graph, calculate the semantic similarity between nodes to generate a semantic weight matrix, and perform feature encoding on the standardized dataset to generate a feature vector set;

[0020] S4. Input the feature vector set and the semantic weight matrix into the improved NOTEARS algorithm model. The improved NOTEARS algorithm performs causal structure learning under semantic constraints and time-dependent constraints, and sets a causal attention self-regulation mechanism to generate an adaptively optimized causal directed acyclic graph structure.

[0021] S5. Based on the causal directed acyclic graph structure, identify the dependency paths of task nodes in the publishing process, and determine the key task nodes and their mutual influence relationships.

[0022] S6. Calculate the task execution priority based on the dependency path and generate an optimized execution sequence for the publishing process;

[0023] S7. Output a searchable publishing workflow structure and scheduling scheme.

[0024] Optionally, S1 specifically includes:

[0025] S11. Collect manuscript data, task log data and personnel information data related to digital publishing. The manuscript data includes manuscript number, manuscript title, manuscript type and version information. The task log data includes task number, task type, task status, task start time and task end time. The personnel information data includes personnel number, personnel role type and department.

[0026] S12. Perform data association based on the manuscript number, task number, and personnel number to generate an initial associated data set with the manuscript number as the primary key;

[0027] S13. Perform unified encoding processing on the character fields of the initial associated data set, perform unified time format processing on the time fields, and record the field values ​​that cannot be determined according to preset rules.

[0028] S14. According to the field mapping rules, unify the data fields from different sources into consistent field names and field order to form a structured record set;

[0029] S15. Perform range scaling processing on the numerical fields in the structured record set, and transform the numerical fields proportionally according to a preset value range;

[0030] S16. Write the structured records, after character processing, time processing, field mapping, and range scaling, into the target data storage structure to form a standardized dataset.

[0031] Optionally, S2 specifically includes:

[0032] S21. Extract semantic fields related to publishing business from the standardized dataset. The semantic fields include manuscript title field, manuscript type field, task type field, task status field, and personnel role field.

[0033] S22. Perform word segmentation, part-of-speech tagging, and stop word cleanup on the semantic field to generate an initial lexical set for semantic representation;

[0034] S23. Construct a semantic representation vector based on the initial word set. The semantic representation vector is generated using a preset word vector model, and each word corresponds to a unique vector representation.

[0035] S24. Identify the semantic features of the manuscript, task node, and personnel node based on the semantic representation vector, generate manuscript entities, task entities, and personnel entities, and record them in the entity set;

[0036] S25. Compare the semantic representation vectors of each entity, and determine the semantic association between entities based on semantic similarity. The semantic association includes the association between the manuscript and the task, the association between the task and the personnel, and the order relationship between the tasks.

[0037] S26. Construct a publishing knowledge graph based on the entity set and semantic relationships, wherein the publishing knowledge graph includes a set of entity nodes and a set of semantic relationship edges.

[0038] Optionally, S3 specifically includes:

[0039] S31. Analyze the semantic relationship edges between entity nodes based on the published knowledge graph, and extract the semantic representation vectors of the entity nodes that constitute the relationship edges;

[0040] S32. Perform similarity calculation on the semantic representation vectors of any two entity nodes, and determine the semantic closeness between the nodes based on the similarity magnitude;

[0041] S33. According to the preset similarity grading rules, the semantic proximity is converted into a semantic association strength value, and a semantic strength label is assigned to the corresponding entity node pair;

[0042] S34. Arrange the semantic strength tags of all entity node pairs in order of node number to form a semantic weight matrix;

[0043] S35. Extract the value data corresponding to the manuscript field, task field and personnel field from the standardized dataset to form a field set;

[0044] S36. Perform field type determination on the field set, and convert character fields into character encoding representation and numeric fields into range scaling representation according to the field type using preset encoding rules to generate a set of field-level encoded values;

[0045] S37. The set of field-level encoded values ​​is combined according to the field order to form a feature vector set.

[0046] Optionally, S4 specifically includes:

[0047] S41. Input the feature vector set and semantic weight matrix into the improved NOTEARS algorithm model, and initialize the causal edge weight matrix. The rows and columns of the causal edge weight matrix correspond to the task nodes in the publishing process, respectively.

[0048] S42. Assign initial semantic weight coefficients to each candidate edge in the causal edge weight matrix based on the semantic weight matrix, and store the semantic weight coefficients in association with the initial edge weights of the corresponding candidate edges.

[0049] S43. Calculate the task time interval based on the task start time and task end time recorded in the standardized dataset, and generate a task time influence matrix according to the order of tasks in the publication process. The elements of the task time influence matrix represent the time influence intensity between task node pairs.

[0050] S44. Based on the values ​​of the semantic weight coefficients and the corresponding elements in the task time influence matrix, calculate the causal attention weights for each candidate edge. The causal attention weights are obtained by weighted summation and normalization of the semantic weight coefficients and the time influence intensity, and are recorded in the causal attention weight set.

[0051] S45. In each iteration update process, a basic edge weight update amount is generated based on the current value of the causal edge weight matrix, and the basic edge weight update amount is multiplied by the corresponding causal attention weight to obtain the actual edge weight update amount of each candidate edge. The edge weights in the causal edge weight matrix are adjusted by increasing or decreasing the actual edge weight update amount.

[0052] S46. After the edge weight adjustment is completed, the edge weights in the causal edge weight matrix that do not meet the semantic weight threshold condition are reduced, the edge weights that do not meet the task time sequence constraints are cleared, and the edge with the smallest edge weight in the set of edges forming a loop is deleted to obtain a causal edge weight matrix without loops.

[0053] S47. When the change in the causal edge weight matrix of two consecutive iterations is lower than the preset convergence threshold or the number of iterations reaches the preset upper limit, the iteration process ends, and the edges in the causal edge weight matrix with edge weights higher than the preset edge weight threshold are output as valid causal edges, forming an adaptively optimized causal directed acyclic graph structure.

[0054] Optionally, S5 specifically includes:

[0055] S51. Read the set of causal edges between task nodes from the adaptively optimized causal directed acyclic graph structure, and establish a mapping relationship between the preceding task node and the subsequent task node according to the direction of the causal edges.

[0056] S52. Based on the mapping relationship, take each task node without a preceding relationship as the starting node, generate multiple candidate dependency paths to each target task node by sequential traversal, and record the task nodes contained in the candidate dependency paths into the path sequence set in the order of appearance.

[0057] S53. Perform node sequence analysis on the path sequence set, count the frequency of each task node in different paths, and calculate the number of times adjacent task nodes appear consecutively in the same path to obtain the node continuity index.

[0058] S54. Generate the influence value and continuity value of the task node according to the occurrence frequency and node continuity index respectively, and sort the task nodes according to the combination result of the influence value and continuity value, and mark the task nodes in the top of the sorting results as key task nodes.

[0059] S55. Compare the occurrence of the key task nodes in the path sequence set, identify the upstream or downstream nodes shared between the key task nodes, and record the pairwise relationship between the shared nodes as key node association units.

[0060] S56. Construct a key task association structure based on the key node association unit. The key task association structure includes key task nodes and corresponding association edges.

[0061] Optionally, S6 specifically includes:

[0062] S61. Read all paths where each task node is located from the dependency path set, count the position number of each task node in its path, and record the position number in the position number list according to the path number order.

[0063] S62. Add the position numbers in the position number list one by one to obtain the cumulative position value of each task node, and combine the cumulative position value with the frequency of occurrence of the corresponding task node in the dependency path set according to a preset ratio to form the execution sorting base value of the corresponding task node.

[0064] S63. Read the task start time and task end time of each task node from the standardized dataset, calculate the task duration, and perform weighted processing on the task duration and the execution sorting base value according to the preset weight to generate the execution priority value of the task node.

[0065] S64. Sort the execution priority values ​​of all task nodes according to their numerical values ​​to generate a priority sorting sequence of task nodes, and determine the optimized execution sequence of the publishing process based on the priority sorting sequence.

[0066] S65. Record the sequential relationship between task nodes according to the optimized execution sequence, and read the corresponding task personnel information from the standardized dataset. Combine the task node sequence with the personnel information to form a scheduling node set.

[0067] S66. Generate a publishing process scheduling scheme based on the task node order and personnel information recorded in the scheduling node set.

[0068] Optionally, S7 specifically includes:

[0069] S71. Read the sorting results of task nodes and corresponding personnel information from the publishing process scheduling scheme, combine the task node order, personnel information and task duration according to the fixed field order to form a process structure record, and write the process structure record into the process structure record set.

[0070] S72. Traverse the process structure record set according to the task node order, set a record number for each process structure record, and store the record number in association with the task node number.

[0071] S73. The process structure records in the process structure record set are classified and split into a structured data set containing a task node sequence field, a personnel information field, and a task duration field according to the fields.

[0072] S74. Perform formatting processing on the structured data set to convert the character content, numerical content and time content of each field into a field format consistent with the target data storage structure;

[0073] S75. Write the formatted structured data set into the target data storage structure, and output the publication process structure and scheduling scheme.

[0074] The beneficial effects of this invention are:

[0075] This invention addresses the challenges of inconsistent semantic representations of multi-source heterogeneous data, non-standardized task time recording, and difficulty in accurately identifying task dependencies in digital publishing by deeply integrating knowledge graph construction, semantic weight modeling, and improved NOTEARS causal structure learning. It proposes a semantic entity construction method based on standardized datasets and generates a computable semantic relationship structure at the graph level using semantic similarity analysis and entity association extraction. In the feature generation stage, the simultaneous construction of a semantic weight matrix and field-level feature vectors provides joint constraints on semantic and temporal dimensions for subsequent causal structure learning, ensuring a quantifiable representation of the strength of associations between task nodes. In the feature generation phase, this invention simultaneously constructs a semantic weight matrix and a set of field-level feature vectors, enabling causal structure learning to possess constraints in both semantic and temporal dimensions. In the causal inference stage, while maintaining the basic structure of the NOTEARS framework, this invention introduces semantic constraints, temporal dependency constraints, and a causal attention self-regulation mechanism. By jointly adjusting the semantic weight coefficients and temporal influence during iterative updates, the candidate edge weight updates can dynamically reflect the strength of semantic relationships and the rationality of temporal logic. After loop detection and edge weight reduction, a stable causal directed acyclic graph structure is output. In the task identification and process optimization stage, this invention uses multi-path traversal of dependent paths and analysis of node occurrence frequency and continuity to identify key task nodes. By combining the cumulative value of path positions, occurrence frequency, and task duration, task execution priorities are generated, forming an optimized execution sequence that conforms to the publishing business logic. Finally, by normalizing and formatting the fields of the process structure record, a searchable output of the publishing process structure and scheduling scheme is achieved. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1This is a schematic diagram of the structure of an intelligent management system for digital publishing processes based on knowledge graphs proposed in this invention;

[0078] Figure 2 This is a flowchart of a knowledge graph-based intelligent management system for digital publishing processes proposed in this invention.

[0079] Figure 3 This is a schematic diagram of the causal structure learning process based on the improved NOTEARS algorithm model in this invention. Detailed Implementation

[0080] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0081] refer to Figure 1-3 A knowledge graph-based intelligent management system for digital publishing processes includes:

[0082] The standardized data generation module is used to acquire digital publishing business data and perform cleaning, formatting and standardization processes to generate standardized datasets.

[0083] The semantic and entity construction module is used to extract publishing-related semantic features from the standardized dataset, establish entity sets, determine the semantic relationships between entities, and form a publishing knowledge graph.

[0084] The semantic weight and feature generation module is used to generate a semantic weight matrix based on the published knowledge graph and to encode a feature vector set from the standardized dataset;

[0085] The causal structure learning module is used to input the semantic weight matrix and feature vector set into the improved NOTEARS algorithm model, and determine the causal directed acyclic graph structure under the action of semantic constraints, time dependency constraints and causal attention self-regulation mechanism.

[0086] The critical task identification module is used to generate task dependency paths based on the causal directed acyclic graph structure, calculate the frequency and continuity of task nodes, and determine critical task nodes and their relationships.

[0087] The priority calculation module is used to calculate the task execution priority based on the task dependency path and the task duration in the standardized dataset, and generate a publication process scheduling scheme.

[0088] The process structure output module is used to generate process structure records based on the publishing process scheduling scheme, and output the publishing process structure and the publishing process scheduling scheme.

[0089] In this embodiment, the modules are interconnected using the following method:

[0090] S1. Acquire digital publishing business data and perform cleaning, formatting and standardization processing to generate a standardized dataset;

[0091] S2. Extract publishing-related semantic features from the standardized dataset, establish an entity set, determine the semantic relationships between entities, and construct a publishing knowledge graph containing entity nodes and semantic relationship edges;

[0092] S3. Based on the published knowledge graph, calculate the semantic similarity between nodes to generate a semantic weight matrix, and perform feature encoding on the standardized dataset to generate a feature vector set;

[0093] S4. Input the feature vector set and the semantic weight matrix into the improved NOTEARS algorithm model. The improved NOTEARS algorithm performs causal structure learning under semantic constraints and time-dependent constraints, and sets a causal attention self-regulation mechanism to generate an adaptively optimized causal directed acyclic graph structure.

[0094] S5. Based on the causal directed acyclic graph structure, identify the dependency paths of task nodes in the publishing process, and determine the key task nodes and their mutual influence relationships.

[0095] S6. Calculate the task execution priority based on the dependency path and generate an optimized execution sequence for the publishing process;

[0096] S7. Output a searchable publishing workflow structure and scheduling scheme.

[0097] In this embodiment, S1 specifically includes:

[0098] S11. Collect manuscript data, task log data and personnel information data related to digital publishing. The manuscript data includes manuscript number, manuscript title, manuscript type and version information. The task log data includes task number, task type, task status, task start time and task end time. The personnel information data includes personnel number, personnel role type and department.

[0099] S12. Perform data association based on the manuscript number, task number, and personnel number to generate an initial associated data set with the manuscript number as the primary key;

[0100] S13. Perform unified encoding processing on the character fields of the initial associated data set, perform unified time format processing on the time fields, and record the field values ​​that cannot be determined according to preset rules.

[0101] S14. According to the field mapping rules, unify the data fields from different sources into consistent field names and field order to form a structured record set;

[0102] S15. Perform range scaling processing on the numerical fields in the structured record set, and transform the numerical fields proportionally according to a preset value range;

[0103] S16. Write the structured records, after character processing, time processing, field mapping, and range scaling, into the target data storage structure to form a standardized dataset.

[0104] In this embodiment, S2 specifically includes:

[0105] S21. Extract semantic fields related to publishing business from the standardized dataset. The semantic fields include manuscript title field, manuscript type field, task type field, task status field, and personnel role field.

[0106] S22. Perform word segmentation, part-of-speech tagging, and stop word cleanup on the semantic field to generate an initial lexical set for semantic representation;

[0107] S23. Construct a semantic representation vector based on the initial word set. The semantic representation vector is generated using a preset word vector model, and each word corresponds to a unique vector representation.

[0108] S24. Identify the semantic features of the manuscript, task node, and personnel node based on the semantic representation vector, generate manuscript entities, task entities, and personnel entities, and record them in the entity set;

[0109] S25. Compare the semantic representation vectors of each entity, and determine the semantic association between entities based on semantic similarity. The semantic association includes the association between the manuscript and the task, the association between the task and the personnel, and the order relationship between the tasks.

[0110] S26. Construct a publishing knowledge graph based on the entity set and semantic relationships, wherein the publishing knowledge graph includes a set of entity nodes and a set of semantic relationship edges.

[0111] In this invention, semantic fields are segmented using a general Chinese word segmentation tool, and an initial set of lexical units is generated using a pre-defined part-of-speech tagging model and a stop word list. For each lexical unit, a semantic representation vector is generated using a publicly available word vector model, with the vector dimension set to 128. Entity generation is performed as follows: manuscript entities, task entities, and personnel entities are generated based on manuscript records, task records, and personnel records in a standardized dataset, and a unique identifier is assigned to each entity. The semantic similarity between any two entities is calculated using cosine similarity, and the semantic relationship is determined based on a set threshold; the task order relationship is established by combining the start and end times in the task records. Finally, a publishing knowledge graph containing entity nodes and semantic relationship edges is constructed.

[0112] In this embodiment, S3 specifically includes:

[0113] S31. Analyze the semantic relationship edges between entity nodes based on the published knowledge graph, and extract the semantic representation vectors of the entity nodes that constitute the relationship edges;

[0114] S32. Perform similarity calculation on the semantic representation vectors of any two entity nodes, and determine the semantic closeness between the nodes based on the similarity magnitude;

[0115] S33. According to the preset similarity grading rules, the semantic proximity is converted into a semantic association strength value, and a semantic strength label is assigned to the corresponding entity node pair;

[0116] S34. Arrange the semantic strength tags of all entity node pairs in order of node number to form a semantic weight matrix;

[0117] S35. Extract the value data corresponding to the manuscript field, task field and personnel field from the standardized dataset to form a field set;

[0118] S36. Perform field type determination on the field set, and convert character fields into character encoding representation and numeric fields into range scaling representation according to the field type using preset encoding rules to generate a set of field-level encoded values;

[0119] S37. The set of field-level encoded values ​​is combined according to the field order to form a feature vector set.

[0120] In this invention, the semantic similarity of entity nodes is calculated using cosine similarity based on semantic representation vectors, with similarity values ​​ranging from zero to one. According to a preset grading rule, the similarity is divided into five intervals: high, relatively high, medium, relatively low, and low, each corresponding to a different semantic association strength value, which is recorded as a continuous numerical value. The semantic weight matrix is ​​constructed according to the entity node numbering order and is in two-dimensional matrix form, with diagonal elements set to zero. Field type determination is based on the format characteristics of the field content; fields containing letters or characters are identified as character fields, and fields that can be directly parsed into numerical values ​​are identified as numerical fields. Character fields generate encoded values ​​using dictionary encoding, while numerical fields are linearly scaled according to the minimum and maximum value ranges, and the encoded values ​​of each field are combined in a fixed order to generate a feature vector.

[0121] In this embodiment, S4 specifically includes:

[0122] S41. Input the feature vector set and semantic weight matrix into the improved NOTEARS algorithm model, and initialize the causal edge weight matrix. The rows and columns of the causal edge weight matrix correspond to the task nodes in the publishing process, respectively.

[0123] S42. Assign initial semantic weight coefficients to each candidate edge in the causal edge weight matrix based on the semantic weight matrix, and store the semantic weight coefficients in association with the initial edge weights of the corresponding candidate edges.

[0124] S43. Calculate the task time interval based on the task start time and task end time recorded in the standardized dataset, and generate a task time influence matrix according to the order of tasks in the publication process. The elements of the task time influence matrix represent the time influence intensity between task node pairs.

[0125] S44. Based on the values ​​of the semantic weight coefficients and the corresponding elements in the task time influence matrix, calculate the causal attention weights for each candidate edge. The causal attention weights are obtained by weighted summation and normalization of the semantic weight coefficients and the time influence intensity, and are recorded in the causal attention weight set.

[0126] S45. In each iteration update process, a basic edge weight update amount is generated based on the current value of the causal edge weight matrix, and the basic edge weight update amount is multiplied by the corresponding causal attention weight to obtain the actual edge weight update amount of each candidate edge. The edge weights in the causal edge weight matrix are adjusted by increasing or decreasing the actual edge weight update amount.

[0127] S46. After the edge weight adjustment is completed, the edge weights in the causal edge weight matrix that do not meet the semantic weight threshold condition are reduced, the edge weights that do not meet the task time sequence constraints are cleared, and the edge with the smallest edge weight in the set of edges forming a loop is deleted to obtain a causal edge weight matrix without loops.

[0128] S47. When the change in the causal edge weight matrix of two consecutive iterations is lower than the preset convergence threshold or the number of iterations reaches the preset upper limit, the iteration process ends, and the edges in the causal edge weight matrix with edge weights higher than the preset edge weight threshold are output as valid causal edges, forming an adaptively optimized causal directed acyclic graph structure.

[0129] In this invention, the causal edge weight matrix is ​​initialized as an asymmetric matrix with all elements set to zero, and the diagonal elements are fixed at zero. Semantic weight coefficients are obtained from the weight values ​​of corresponding node pairs in the semantic weight matrix and mapped to candidate edges according to edge direction. The task time influence is calculated based on the difference between task start and end times, determining the time interval between tasks, and recording the influence of earlier tasks on later tasks as positive intensity values ​​according to their order of occurrence. The causal attention weight is obtained by weighting the semantic weight coefficients and the time influence intensity according to a fixed ratio and then linearly normalizing. The basic edge weight update is performed using an iterative incremental method, generated based on the current edge weight difference; convergence is determined when the difference between adjacent iterations of the edge weight matrix is ​​below a threshold.

[0130] In this embodiment, S5 specifically includes:

[0131] S51. Read the set of causal edges between task nodes from the adaptively optimized causal directed acyclic graph structure, and establish a mapping relationship between the preceding task node and the subsequent task node according to the direction of the causal edges.

[0132] S52. Based on the mapping relationship, take each task node without a preceding relationship as the starting node, generate multiple candidate dependency paths to each target task node by sequential traversal, and record the task nodes contained in the candidate dependency paths into the path sequence set in the order of appearance.

[0133] S53. Perform node sequence analysis on the path sequence set, count the frequency of each task node in different paths, and calculate the number of times adjacent task nodes appear consecutively in the same path to obtain the node continuity index.

[0134] S54. Generate the influence value and continuity value of the task node according to the occurrence frequency and node continuity index respectively, and sort the task nodes according to the combination result of the influence value and continuity value, and mark the task nodes in the top of the sorting results as key task nodes.

[0135] S55. Compare the occurrence of the key task nodes in the path sequence set, identify the upstream or downstream nodes shared between the key task nodes, and record the pairwise relationship between the shared nodes as key node association units.

[0136] S56. Construct a key task association structure based on the key node association unit. The key task association structure includes key task nodes and corresponding association edges.

[0137] In this invention, the mapping relationship between preceding and subsequent task nodes is recorded in the form of an adjacency list, where each task node corresponds to all its subsequent nodes. Candidate dependency paths are generated based on a depth-first traversal, recursively visiting subsequent nodes starting from the point where no preceding node exists, and recording the visit sequence as a path sequence. The node continuity index is obtained by counting the number of consecutive occurrences of adjacent nodes on the same path, and the final node continuity value is obtained by summing the consecutive occurrence counts across all paths. Shared upstream or downstream nodes of key task nodes are identified based on their common positional relationships in the path sequence, and key node association units are constructed in the form of node pairs.

[0138] In this embodiment, S6 specifically includes:

[0139] S61. Read all paths where each task node is located from the dependency path set, count the position number of each task node in its path, and record the position number in the position number list according to the path number order.

[0140] S62. Add the position numbers in the position number list one by one to obtain the cumulative position value of each task node, and combine the cumulative position value with the frequency of occurrence of the corresponding task node in the dependency path set according to a preset ratio to form the execution sorting base value of the corresponding task node.

[0141] S63. Read the task start time and task end time of each task node from the standardized dataset, calculate the task duration, and perform weighted processing on the task duration and the execution sorting base value according to the preset weight to generate the execution priority value of the task node.

[0142] S64. Sort the execution priority values ​​of all task nodes according to their numerical values ​​to generate a priority sorting sequence of task nodes, and determine the optimized execution sequence of the publishing process based on the priority sorting sequence.

[0143] S65. Record the sequential relationship between task nodes according to the optimized execution sequence, and read the corresponding task personnel information from the standardized dataset. Combine the task node sequence with the personnel information to form a scheduling node set.

[0144] S66. Generate a publishing process scheduling scheme based on the task node order and personnel information recorded in the scheduling node set.

[0145] In this invention, the position number of a task node in a dependent path is determined by the arrangement order of nodes in the path, and the numbers are sequentially increased starting from the starting node. The path numbers of each dependent path are recorded according to the generation order. The execution sorting base value is a linear weighted combination of the cumulative position value and occurrence frequency of the task node, with the preset ratio set according to the sensitivity parameters of the publishing process, and can be a fixed ratio. The task duration is obtained by subtracting the task start time from the task end time, and the preset weight used in the weighting process is set by the process time consumption characteristics. The scheduling scheme includes the task node order, task execution time period, and personnel information, constructed according to the priority sorting results.

[0146] In this embodiment, S7 specifically includes:

[0147] S71. Read the sorting results of task nodes and corresponding personnel information from the publishing process scheduling scheme, combine the task node order, personnel information and task duration according to the fixed field order to form a process structure record, and write the process structure record into the process structure record set.

[0148] S72. Traverse the process structure record set according to the task node order, set a record number for each process structure record, and store the record number in association with the task node number.

[0149] S73. The process structure records in the process structure record set are classified and split into a structured data set containing a task node sequence field, a personnel information field, and a task duration field according to the fields.

[0150] S74. Perform formatting processing on the structured data set to convert the character content, numerical content and time content of each field into a field format consistent with the target data storage structure;

[0151] S75. Write the formatted structured data set into the target data storage structure, and output the publication process structure and scheduling scheme.

[0152] In this invention, the fixed field order of the process structure record is: task node sequence field, personnel information field, and task duration field, arranged sequentially. Field classification and splitting are performed based on field names, extracting the three types of fields from the record to form a structured scheduling field set. The output publishing process structure is represented as a task node sequence list, and the output publishing process scheduling scheme is represented as a structured record set containing task node sequence, personnel information, and task duration. The output results can be retrieved by record number.

[0153] Example 1:

[0154] To verify the feasibility and technical effectiveness of this invention in a real-world business scenario, it was applied to the book publishing process management of a large digital publishing institution. This institution processes approximately 32,000 manuscripts annually, covering multiple task stages including editing, proofreading, typesetting, and online publication, involving over 450 personnel. Due to the diverse sources of manuscripts, inconsistent task recording standards, and differing understandings of task order among different departments, it has long faced problems such as unclear process dependencies, inaccurate identification of key tasks, and low scheduling efficiency. Real-world business audits revealed that annual task delays due to unreasonable process scheduling exceed 175,000 hours, with an average publication cycle of 56.3 days, indicating significant room for optimization.

[0155] In this experiment, the invention first collected historical data covering 18 categories of book-related business over the past three years, including manuscript data, task log data, and personnel information data. This data was then cleaned, encoded, had fields standardized, and numerical ranges scaled using a standardized data generation module to form a standardized dataset with a consistent structure. Subsequently, a semantic and entity construction module extracted quantifiable semantic features from semantic fields such as manuscript titles, task types, and personnel roles, and constructed a publishing knowledge graph. This visualized and structured the semantic relationships between manuscripts, tasks, and personnel, laying the foundation for subsequent causal modeling.

[0156] In the causal structure learning stage of this invention, the generated feature vector set and semantic weight matrix are input into the improved NOTEARS algorithm. In this algorithm, semantic constraints ensure that tasks with similar semantics are more likely to form causal paths; temporal dependency constraints ensure that the task order conforms to historical temporal logic; and the causal attention self-regulation mechanism further enhances the relevance of edge weight updates, enabling the semantic and temporal importance to be automatically adjusted in different iterations. Finally, a stable causal directed acyclic graph structure is obtained, and the key task identification module identifies key nodes, generating the 12 most influential key tasks in the publishing process and their dependencies.

[0157] To quantitatively verify the effectiveness of this invention, the optimized execution sequence generated by this invention is compared with the process scheduling mechanism of traditional algorithms. Typical task stages such as editing, initial proofreading, secondary proofreading, final review, typesetting, prepress processing, and online release are selected as evaluation objects. The comparison indicators include average task waiting time, accuracy of key node identification, overall process duration, and number of personnel scheduling conflicts. The comparison experiment is shown in Table 1.

[0158] Table 1 Comparison of Publication Process Optimization Effects between the Invention Method and Traditional Methods

[0159] Comparison indicators Method of the present invention Traditional methods Average task waiting time (h) 3.1 6.8 Total process time (days) 41.7 56.3 Key task identification accuracy (%) 93.4 72.5 Task dependency identification accuracy (%) 91.2 68.1 Task priority ranking and actual execution matching degree (%) 89.3 64.7 Process anomaly node identification rate (%) 86.7 58.9 Overall on-time delivery rate of the project (%) 88.5 71.3

[0160] As can be seen from the comparison results in Table 1 above, the present invention significantly outperforms traditional methods in multiple indicators. Regarding the average task waiting time, the present invention achieves 3.1 hours, a reduction of 3.7 hours compared to the traditional method, representing a 54.4% reduction. This demonstrates that the optimized execution sequence constructed through semantic constraints and causal structure learning can significantly reduce invalid waiting, making the process more compact and efficient. In terms of total process duration, the present invention shortens the overall cycle from 56.3 days to 41.7 days, a reduction of 14.6 days, or 25.9%, fully illustrating that critical path identification based on causal directed acyclic graphs can effectively compress the overall publication cycle. In terms of critical task identification accuracy, the present invention achieves 93.4%, an improvement of 20.9 percentage points compared to the traditional method, indicating that the semantic weight matrix and causal attention self-regulation mechanism can more accurately identify important nodes affecting the publication process. In terms of task dependency identification accuracy, the present invention achieves 91.2%, an improvement of 23.1 percentage points, demonstrating the high-fidelity modeling capability of the improved NOTEARS for real dependency structures under both semantic and temporal constraints. In terms of the matching degree between task priority ranking and actual execution, this invention achieves 89.3%, an improvement of 24.6 percentage points compared to traditional methods, proving that the priority sequence generated by causal inference is more consistent with real business logic. The identification rate of abnormal nodes in the process is improved to 86.7%, an increase of 27.8 percentage points, indicating that the causal structure is more sensitive to abnormal propagation paths and can detect process bottlenecks and abnormal nodes earlier. The overall on-time delivery rate of projects increases from 71.3% to 88.5%, an increase of 17.2 percentage points, indicating that the intelligent scheduling mechanism constructed by this invention can significantly improve the on-time completion capability of publishing projects. In summary, this invention significantly outperforms existing technologies in terms of process identification accuracy, scheduling intelligence level, and overall efficiency improvement. The performance improvement is comprehensive and stable, proving that this invention has higher practical value and technological advancement in the field of intelligent management of digital publishing processes.

[0161] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent management system for digital publishing processes, characterized in that, include: The standardized data generation module is used to acquire digital publishing business data and perform cleaning, formatting and standardization processes to generate standardized datasets. The semantic and entity construction module is used to extract publishing-related semantic features from the standardized dataset, establish entity sets, determine the semantic relationships between entities, and form a publishing knowledge graph. The semantic weight and feature generation module is used to generate a semantic weight matrix based on the published knowledge graph and to encode a feature vector set from the standardized dataset; The causal structure learning module is used to input the semantic weight matrix and feature vector set into the improved NOTEARS algorithm model, and determine the causal directed acyclic graph structure under the action of semantic constraints, time dependency constraints and causal attention self-regulation mechanism. The critical task identification module is used to generate task dependency paths based on the causal directed acyclic graph structure, calculate the frequency and continuity of task nodes, and determine critical task nodes and their corresponding relationships. The priority calculation module is used to calculate the task execution priority based on the task dependency path and the task duration in the standardized dataset, and generate a publication process scheduling scheme. The process structure output module is used to generate process structure records based on the publishing process scheduling scheme, and output the publishing process structure and the publishing process scheduling scheme.

2. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 1, characterized in that, The modules are connected in the following way: S1. Acquire digital publishing business data and perform cleaning, formatting and standardization processing to generate a standardized dataset; S2. Extract publishing-related semantic features from the standardized dataset, establish an entity set, determine the semantic relationships between entities, and construct a publishing knowledge graph containing entity nodes and semantic relationship edges; S3. Based on the published knowledge graph, calculate the semantic similarity between nodes to generate a semantic weight matrix, and perform feature encoding on the standardized dataset to generate a feature vector set; S4. Input the feature vector set and the semantic weight matrix into the improved NOTEARS algorithm model. The improved NOTEARS algorithm performs causal structure learning under semantic constraints and time-dependent constraints, and sets a causal attention self-regulation mechanism to generate an adaptively optimized causal directed acyclic graph structure. S5. Based on the causal directed acyclic graph structure, identify the dependency paths of task nodes in the publishing process, and determine the key task nodes and their mutual influence relationships. S6. Calculate the task execution priority based on the dependency path and generate an optimized execution sequence for the publishing process; S7. Output a searchable publishing workflow structure and scheduling scheme.

3. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S1 specifically includes: S11. Collect manuscript data, task log data and personnel information data related to digital publishing. The manuscript data includes manuscript number, manuscript title, manuscript type and version information. The task log data includes task number, task type, task status, task start time and task end time. The personnel information data includes personnel number, personnel role type and department. S12. Perform data association based on the manuscript number, task number, and personnel number to generate an initial associated data set with the manuscript number as the primary key; S13. Perform unified encoding processing on the character fields of the initial associated data set, perform unified time format processing on the time fields, and record the field values ​​that cannot be determined according to preset rules. S14. According to the field mapping rules, unify the data fields from different sources into consistent field names and field order to form a structured record set; S15. Perform range scaling processing on the numerical fields in the structured record set, and transform the numerical fields proportionally according to a preset value range; S16. Write the structured records, after character processing, time processing, field mapping, and range scaling, into the target data storage structure to form a standardized dataset.

4. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S2 specifically includes: S21. Extract semantic fields related to publishing business from the standardized dataset. The semantic fields include manuscript title field, manuscript type field, task type field, task status field, and personnel role field. S22. Perform word segmentation, part-of-speech tagging, and stop word cleanup on the semantic field to generate an initial lexical set for semantic representation; S23. Construct a semantic representation vector based on the initial word set. The semantic representation vector is generated using a preset word vector model, and each word corresponds to a unique vector representation. S24. Identify the semantic features of the manuscript, task node, and personnel node based on the semantic representation vector, generate manuscript entities, task entities, and personnel entities, and record them in the entity set; S25. Compare the semantic representation vectors of each entity, and determine the semantic association between entities based on semantic similarity. The semantic association includes the association between the manuscript and the task, the association between the task and the personnel, and the order relationship between the tasks. S26. Construct a publishing knowledge graph based on the entity set and semantic relationships, wherein the publishing knowledge graph includes a set of entity nodes and a set of semantic relationship edges.

5. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S3 specifically includes: S31. Analyze the semantic relationship edges between entity nodes based on the published knowledge graph, and extract the semantic representation vectors of the entity nodes that constitute the relationship edges; S32. Perform similarity calculation on the semantic representation vectors of any two entity nodes, and determine the semantic closeness between the nodes based on the similarity magnitude; S33. According to the preset similarity grading rules, the semantic proximity is converted into a semantic association strength value, and a semantic strength label is assigned to the corresponding entity node pair; S34. Arrange the semantic strength tags of all entity node pairs in order of node number to form a semantic weight matrix; S35. Extract the value data corresponding to the manuscript field, task field and personnel field from the standardized dataset to form a field set; S36. Perform field type determination on the field set, and convert character fields into character encoding representation and numeric fields into range scaling representation according to the field type using preset encoding rules to generate a set of field-level encoded values; S37. The set of field-level encoded values ​​is combined according to the field order to form a feature vector set.

6. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S4 specifically includes: S41. Input the feature vector set and semantic weight matrix into the improved NOTEARS algorithm model, and initialize the causal edge weight matrix. The rows and columns of the causal edge weight matrix correspond to the task nodes in the publishing process, respectively. S42. Assign initial semantic weight coefficients to each candidate edge in the causal edge weight matrix based on the semantic weight matrix, and store the semantic weight coefficients in association with the initial edge weights of the corresponding candidate edges. S43. Calculate the task time interval based on the task start time and task end time recorded in the standardized dataset, and generate a task time influence matrix according to the order of tasks in the publication process. The elements of the task time influence matrix represent the time influence intensity between task node pairs. S44. Based on the values ​​of the semantic weight coefficients and the corresponding elements in the task time influence matrix, calculate the causal attention weights for each candidate edge. The causal attention weights are obtained by weighted summation and normalization of the semantic weight coefficients and the time influence intensity, and are recorded in the causal attention weight set. S45. In each iteration update process, a basic edge weight update amount is generated based on the current value of the causal edge weight matrix, and the basic edge weight update amount is multiplied by the corresponding causal attention weight to obtain the actual edge weight update amount of each candidate edge. The edge weights in the causal edge weight matrix are adjusted by increasing or decreasing the actual edge weight update amount. S46. After the edge weight adjustment is completed, the edge weights in the causal edge weight matrix that do not meet the semantic weight threshold condition are reduced, the edge weights that do not meet the task time sequence constraints are cleared, and the edge with the smallest edge weight in the set of edges forming a loop is deleted to obtain a causal edge weight matrix without loops. S47. When the change in the causal edge weight matrix of two consecutive iterations is lower than the preset convergence threshold or the number of iterations reaches the preset upper limit, the iteration process ends, and the edges in the causal edge weight matrix with edge weights higher than the preset edge weight threshold are output as valid causal edges, forming an adaptively optimized causal directed acyclic graph structure.

7. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S5 specifically includes: S51. Read the set of causal edges between task nodes from the adaptively optimized causal directed acyclic graph structure, and establish a mapping relationship between the preceding task node and the subsequent task node according to the direction of the causal edges. S52. Based on the mapping relationship, take each task node without a preceding relationship as the starting node, generate multiple candidate dependency paths to each target task node by sequential traversal, and record the task nodes contained in the candidate dependency paths into the path sequence set in the order of appearance. S53. Perform node sequence analysis on the path sequence set, count the frequency of each task node in different paths, and calculate the number of times adjacent task nodes appear consecutively in the same path to obtain the node continuity index. S54. Generate the influence value and continuity value of the task node according to the occurrence frequency and node continuity index respectively, and sort the task nodes according to the combination result of the influence value and continuity value, and mark the task nodes in the top of the sorting results as key task nodes. S55. Compare the occurrence of the key task nodes in the path sequence set, identify the upstream or downstream nodes shared between the key task nodes, and record the pairwise relationship between the shared nodes as key node association units. S56. Construct a key task association structure based on the key node association unit. The key task association structure includes key task nodes and corresponding association edges.

8. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, S6 specifically includes: S61. Read all paths where each task node is located from the dependency path set, count the position number of each task node in the path, and record the position number in the position number list according to the path number order. S62. Add the position numbers in the position number list one by one to obtain the cumulative position value of each task node, and combine the cumulative position value with the frequency of occurrence of the corresponding task node in the dependency path set according to a preset ratio to form the execution sorting base value of the corresponding task node. S63. Read the task start time and task end time of each task node from the standardized dataset, calculate the task duration, and perform weighted processing on the task duration and the execution sorting base value according to the preset weight to generate the execution priority value of the corresponding task node. S64. Sort the execution priority values ​​of all task nodes according to their numerical values ​​to generate a priority sorting sequence of task nodes, and determine the optimized execution sequence of the publishing process based on the priority sorting sequence. S65. Record the sequential relationship between task nodes according to the optimized execution sequence, and read the corresponding task personnel information from the standardized dataset. Combine the task node sequence with the personnel information to form a scheduling node set. S66. Generate a publishing process scheduling scheme based on the task node order and personnel information recorded in the scheduling node set.

9. The intelligent management system for digital publishing processes based on knowledge graphs according to claim 2, characterized in that, Specifically, S7 includes: S71. Read the sorting results of task nodes and corresponding personnel information from the publishing process scheduling scheme, combine the task node order, personnel information and task duration according to the fixed field order to form a process structure record, and write the process structure record into the process structure record set. S72. Traverse the process structure record set according to the task node order, set a record number for each process structure record, and store the record number in association with the task node number. S73. The process structure records in the process structure record set are classified and split into a structured data set containing a task node sequence field, a personnel information field, and a task duration field according to the fields. S74. Perform formatting processing on the structured data set to convert the character content, numerical content and time content of each field into a field format consistent with the target data storage structure; S75. Write the formatted structured data set into the target data storage structure, and output the publication process structure and scheduling scheme.

Images