Intelligent auxiliary decision-making method and device based on multi-source data

Abstract

The embodiment of the application provides a kind of intelligent auxiliary decision-making method and device based on multi-source data, realizes the accurate identification of information by semantic analysis and feature extraction.Construct decision mechanism, combine flow model and scheduling model, establish reliable task management strategy.Introduce efficiency optimization, through completion degree evaluation and model retraining, ensure the continuous improvement of decision-making.This method effectively solves the shortcomings of traditional technology in information extraction, task management and efficiency evaluation, etc., and provides technical support for intelligent auxiliary decision-making.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to an intelligent auxiliary decision-making method and device based on multi-source data. Background Technology

[0002] Existing intelligent decision-making support methods have significant shortcomings. Traditional systems perform poorly in document processing and semantic analysis, failing to effectively extract accurate document information and thus impacting decision-making outcomes.

[0003] Furthermore, existing technologies suffer from bottlenecks in task allocation and workflow management. Most systems lack robust prioritization mechanisms and analysis strategies, resulting in suboptimal processing efficiency.

[0004] Existing systems have technical shortcomings in performance evaluation. They lack in-depth analysis of execution completion, making it difficult to achieve efficient model updates through feedback optimization, thus affecting decision-making accuracy. Solving these problems is crucial for improving intelligent decision-making capabilities. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides an intelligent auxiliary decision-making method and device based on multi-source data, which can effectively solve the shortcomings of traditional technologies in information extraction, task management and performance evaluation, and provide technical support for intelligent auxiliary decision-making.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides an intelligent auxiliary decision-making method based on multi-source data, including: The document management system is accessed to obtain a file data stream. The file data stream is processed by a file parser to generate a standard information table. Text segmentation is performed on the standard information table to obtain a word sequence set. Based on the word sequence set, topic recognition rules are constructed to generate a topic mapping table. The topic mapping table is processed by a semantic analysis module to obtain a semantic feature matrix. Keyword extraction is performed on the semantic feature matrix to obtain an entity label group. Based on the entity label group, a deep classifier is constructed to generate a type annotation set. The type annotation set is matched with a historical database to obtain a feature vector table. Based on the feature vector table, a flow model is constructed to generate a task allocation matrix. Priority calculation is performed on the task allocation matrix to obtain a processing order table. The processing order table is matched with the business rule base to generate a set of judgment factors. Scoring calculation is performed on the set of judgment factors to obtain a suggested solution value. Based on the suggested solution value, a scheduling model is constructed to generate a task decomposition table. The task decomposition table is compared with the progress data to obtain a group of prompt signals. Timeliness sorting is performed on the group of prompt signals to generate a push suggestion. The push notification is written into the office terminal to obtain an execution list. The completion rate of the execution list is evaluated to generate a feedback index table. Based on the feedback index table, an efficiency analyzer is constructed to obtain a training dataset. The training dataset is then used to update the flow model and scheduling model through model retraining.

[0007] Furthermore, it also includes: obtaining the original file set by accessing the document management system database based on the interface protocol; parsing the original file set according to document type to generate a structured data table; performing format unification processing on the structured data table to obtain a standardized text group; constructing text cleaning rules based on the standardized text group to generate a preprocessed text library; processing the preprocessed text library through a word segmentation model to generate an initial word sequence list; and performing stop word filtering on the initial word sequence list to obtain an effective word order set. The effective word order set is reorganized according to the semantic relationship of the text to generate a topic feature matrix. Topic clustering analysis is performed on the topic feature matrix to obtain a topic distribution map. A mapping rule base is constructed based on the topic distribution map to generate a category correspondence table. The category correspondence table is compared with the word frequency statistics results to generate a weight coefficient group. Normalization is applied to the weight coefficient group to obtain a topic mapping table.

[0008] Furthermore, it also includes: converting the topic mapping table into a semantic vector group based on the word vector model, performing similarity calculation on the semantic vector group to obtain an association strength matrix, constructing a semantic network based on the association strength matrix to generate a node connection graph, processing the node connection graph through path analysis to obtain a semantic feature matrix, performing feature selection on the semantic feature matrix to obtain a core word set, and constructing an entity extractor based on the core word set to generate an entity label group; The entity label group is processed by a deep neural network to generate a classification feature table. Label encoding is performed on the classification feature table to obtain a category mapping set. Matching rules are constructed based on the category mapping set to generate a query vector group. The query vector group is compared with sample records in the historical database to generate a similarity table. Threshold filtering is applied to the similarity table to obtain a feature vector table.

[0009] Furthermore, it also includes: splitting the feature vector table into feature subset groups according to a preset dimension, performing normalization processing on the feature subset groups to generate a standardized vector set, constructing a neural network classifier based on the standardized vector set to generate a category probability matrix, processing the category probability matrix through a decision tree model to obtain a flow path graph, performing path optimization on the flow path graph to obtain a task allocation matrix, and constructing a timeliness evaluator based on the task allocation matrix to generate a processing order table; The processing order table is processed by a rule matching engine to generate a rule adaptation table. Feasibility verification is performed on the rule adaptation table to obtain a set of constraints. A scoring model is constructed based on the set of constraints to generate a weight distribution table. The weight distribution table is weighted and calculated with the set of judgment factors to obtain a score sequence. Normalization is applied to the score sequence to generate a suggested solution value.

[0010] Furthermore, it also includes: converting the suggested solution values ​​according to task type to generate a solution element set; performing task decomposition on the solution element set to obtain a sub-task sequence; constructing a resource allocator based on the sub-task sequence to generate a scheduling rule table; processing the scheduling rule table through a constraint solver to obtain an execution plan group; applying feasibility verification to the execution plan group to generate a task decomposition table; and constructing a monitoring model based on the task decomposition table to obtain a state tracking matrix. The status tracking matrix and real-time progress data are compared to generate a deviation vector group. A threshold judgment is performed on the deviation vector group to obtain an alarm signal table. A priority evaluator is constructed based on the alarm signal table to generate an urgency matrix. The urgency matrix is ​​processed by sorting rules to obtain a push sequence. A template mapping is applied to the push sequence to generate a push proposal.

[0011] Furthermore, it also includes: parsing the push suggestion according to the terminal type to generate a push instruction set, performing a format conversion on the push instruction set to obtain a terminal command table, constructing a task dispatcher based on the terminal command table to generate a deployment parameter group, processing the deployment parameter group through an interface adapter to obtain an execution instruction stream, applying a status check to the execution instruction stream to generate an execution list, and constructing a monitoring model based on the execution list to obtain a task status table; The task status table is compared with the preset completion standard to generate an achievement degree matrix. Multi-dimensional analysis is performed on the achievement degree matrix to obtain an evaluation feature group. Based on the evaluation feature group, a performance calculator is constructed to generate an indicator score table. The indicator score table is normalized to obtain a feedback indicator table.

[0012] Furthermore, it also includes: decomposing the feedback indicator table according to the evaluation dimension to generate indicator feature groups, performing data cleaning on the indicator feature groups to obtain an effective sample set, constructing a feature extractor based on the effective sample set to generate a training feature table, processing the training feature table through label mapping to obtain training data pairs, applying data augmentation to the training data pairs to generate a training dataset, and constructing a validation segmenter based on the training dataset to obtain a training validation set. The training and validation sets are processed by a parameter optimizer to generate a model parameter set. Iterative training is performed on the model parameter set to obtain an updated weight table. A model updater is constructed based on the updated weight table to generate a parameter configuration table. The parameter configuration table is written into the flow model and the scheduling model to complete the model update.

[0013] Secondly, this application provides an intelligent auxiliary decision-making device based on multi-source data, comprising: The document processing module is used to access the document management system to obtain document data streams, process the document data streams through a document parser to generate a standard information table, perform text segmentation on the standard information table to obtain a word sequence set, construct topic recognition rules based on the word sequence set to generate a topic mapping table, process the topic mapping table through a semantic analysis module to obtain a semantic feature matrix, perform keyword extraction on the semantic feature matrix to obtain entity label groups, construct a deep classifier based on the entity label groups to generate a type annotation set, and match the type annotation set with a historical database to obtain a feature vector table. The suggestion generation module is used to construct a flow model based on the feature vector table to generate a task allocation matrix, perform priority calculation on the task allocation matrix to obtain a processing order table, match the processing order table with the business rule base to generate a set of judgment factors, perform scoring calculation on the set of judgment factors to obtain a suggested solution value, construct a scheduling model based on the suggested solution value to generate a task decomposition table, compare the task decomposition table with the progress data to obtain a group of prompt signals, and perform timeliness sorting on the group of prompt signals to generate a push suggestion. The auxiliary decision-making module is used to write the push suggestion into the office terminal to obtain an execution list, evaluate the completion rate of the execution list to generate a feedback index table, construct an efficiency analyzer based on the feedback index table to obtain a training dataset, and update the flow model and scheduling model through model retraining of the training dataset.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent auxiliary decision-making method based on multi-source data.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent auxiliary decision-making method based on multi-source data.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the intelligent auxiliary decision-making method based on multi-source data.

[0017] As can be seen from the above technical solution, this application provides an intelligent auxiliary decision-making method and device based on multi-source data. Through semantic analysis and feature extraction, it achieves accurate information identification. A decision-making mechanism is constructed, combining a flow model and a scheduling model to establish a reliable task management strategy. Performance optimization is introduced, ensuring continuous improvement of decisions through completion evaluation and model retraining. This method effectively solves the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, providing technical support for intelligent auxiliary decision-making. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the intelligent auxiliary decision-making method based on multi-source data in the embodiments of this application; Figure 2 This is a structural diagram of the intelligent auxiliary decision-making device based on multi-source data in the embodiments of this application; Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0022] In view of the problems existing in the prior art, this application provides an intelligent auxiliary decision-making method and device based on multi-source data. Through semantic analysis and feature extraction, it achieves accurate information identification. A decision-making mechanism is constructed, combining a flow model and a scheduling model to establish a reliable task management strategy. Performance optimization is introduced, ensuring continuous improvement of decisions through completion evaluation and model retraining. This method effectively solves the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, providing technical support for intelligent auxiliary decision-making.

[0023] To effectively address the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, and to provide technical support for intelligent decision-making assistance, this application provides an embodiment of an intelligent decision-making assistance method based on multi-source data. See [link to embodiment]. Figure 1 The intelligent auxiliary decision-making method based on multi-source data specifically includes the following: Step S101: Access the document management system to obtain the document data stream, process the document data stream through the document parser to generate a standard information table, perform text segmentation on the standard information table to obtain a word sequence set, construct topic recognition rules based on the word sequence set to generate a topic mapping table, process the topic mapping table through the semantic analysis module to obtain a semantic feature matrix, perform keyword extraction on the semantic feature matrix to obtain entity label groups, construct a deep classifier based on the entity label groups to generate a type annotation set, and match the type annotation set with the historical database to obtain a feature vector table; First, after connecting to the document management system, the file data stream is read. Source identifiers and timestamps are aligned according to the interface protocol, and the documents are split into three categories: text, attachments, and metadata. Then, the text and metadata are sent to the file parser, where character set unification, paragraph boundary recognition, and field mapping are performed. A standard information table is output, with fields including title, body text, originating unit, and processing time limit. This standard information table maintains a reference relationship with the original attachments under the same document key for subsequent semantic processing and backtracking.

[0024] Based on the standard information table, a word segmentation pipeline is constructed to perform joint processing of domain dictionary-priority matching and statistical segmentation on the text fragments, resulting in a word sequence set. To reduce noise, low-information-weight words are filtered out according to the stop word list and character category rules, while retaining the in-situ indexes of numerical and time expressions to facilitate subsequent topic identification and location backfilling. The word sequence set processed in this way is written into the same batch identifier as the originating unit, ensuring traceability during cross-document clustering.

[0025] The word sequence set is input into the topic recognition module. First, a word co-occurrence matrix is ​​constructed using a word co-occurrence window. Then, it is aggregated according to textual semantic relationships to form a topic candidate set, and a topic mapping table is generated based on this set. The mapping rule uses word cluster density and cross-document stability as dual constraints, merging high-frequency word clusters from the same source within similar time windows, while reducing the weight of low-frequency clusters with excessively large spans. The topic mapping table generates several topic tags for each document, retaining weights and covered sentence indexes to facilitate subsequent semantic feature alignment.

[0026] Based on the topic mapping table, the semantic analysis module fuses the topic tags and contextual syntactic dependencies of the same document, outputting a semantic feature matrix. This matrix uses sentence segments as rows and semantic slots as columns, with slots covering five categories: behavior, object, quota, time limit, and region. The weights of the source topics are also backfilled as side information. To control cross-slot interference, consistency checks are applied to mutually exclusive slots, eliminating conflicting tags appearing in the same position.

[0027] The semantic feature matrix was used for keyword extraction, employing a joint scoring system of location information, slot weights, and syntactic roles to filter entity label groups. These labels were categorized into five types: institutions, behavioral elements, quantity metrics, timeframes, and geographical units, and bound to character ranges from the original text fragments. During extraction, homonyms were disambiguated using local context windows, and the disambiguation results, along with confidence markers, were written back to the label entries for reference in subsequent classification stages.

[0028] Based on the entity label group, a deep classifier, denoted as the "Document Type Discrimination Network," is constructed. This network takes label sequences and their slot features as input and outputs a type label set covering three aspects: matter category, urgency, and processing path. During training, the consistency between historical processed samples and rule verification is used as a joint constraint. In the inference phase, only the label combination with the highest consistency is retained, and the side confidence value of each label is output as the retrieval weight for subsequent matching.

[0029] The type-label set is used for historical database matching. The matching process first maps the labels to the source unit and time range as query keys, then compares them with the structured index of historical samples, and generates a feature vector table by taking similar entries. To avoid a single signal dominating, the matching scores are integrated according to the label category, missing categories are filled with placeholders, and the source reference key of each vector is retained to ensure that subsequent processes can be verified.

[0030] In this process, an optional internal consistency measure can be jointly provided for type identification and historical similarity, serving as a threshold reference for subsequent stages. The scoring function is defined as follows: Z = θ1·U + θ2·V - θ3·W Wherein, Z is the consistency measure; U is the side confidence mean of the main labels in the type label set; V is the weighted average of similarity returned by historical matches; W is the conflict ratio of mutually exclusive slots in the semantic feature matrix; θ1, θ2, and θ3 are non-negative weights, satisfying that their sum is limited by the interval range and is not fixed as a constant. Z is read in the next step of constructing the flow model and used as a priori weight for path selection, but it does not participate in the final external display.

[0031] Finally, the feature vector table and the consistency metric are registered together on the data bus as input for the subsequent step S102. The feature vector table will then be used to calculate the category probability and flow path, and the consistency metric will be used to limit the search space of candidate paths, thereby reusing the semantic and historical information accumulated in this step during the path generation and task decomposition stages.

[0032] Step S102: Based on the feature vector table, construct a flow model to generate a task allocation matrix, perform priority calculation on the task allocation matrix to obtain a processing order table, match the processing order table with the business rule base to generate a set of judgment factors, perform scoring calculation on the set of judgment factors to obtain a suggested solution value, construct a scheduling model based on the suggested solution value to generate a task decomposition table, compare the task decomposition table with the progress data to obtain a group of prompt signals, and perform timeliness sorting on the group of prompt signals to generate a push suggestion. First, the feature vector table and consistency metric output from the previous steps are read, aggregated into samples of the same batch by document key, and missing slots are filled with placeholders. Based on this input, a "document circulation model" is constructed, using the source unit, matter category, urgency, and processing time limit as core features. The consistency metric is used to limit the candidate path set, and a task allocation matrix is ​​generated using layer-by-layer path expansion and conflict pruning. Rows in the matrix correspond to matter instances, columns correspond to undertaking nodes, and elements record the assigned weight and expected start time for subsequent priority calculation.

[0033] Based on the task allocation matrix, the load and available time periods of each node are read to generate a time-based resource view. The waiting time and cross-node migration cost for each task at the candidate nodes are calculated accordingly to obtain a processing order list. During sorting, the allocation weight and waiting time are weighted and combined, and a minimum passing threshold is set using a consistency metric. Paths with low consistency are downweighted to avoid occupying critical time periods when resources are scarce.

[0034] The processing sequence list is sent to the rule matching module, where each item is mapped against the business rule base to generate a set of evaluation factors. The mapping process aligns the item categories with statutory deadlines, and the lists of required and parallel steps, while also detecting cooperation requirements related to the source unit, forming structured factors including deadline urgency, step dependency, and cross-departmental collaboration strength. This factor set retains reference keys to the processing sequence list on each item, ensuring the scoring phase is traceable.

[0035] Based on the aforementioned set of evaluation factors, a "solution scoring model" is constructed. This model uses four quantitative items—deadline urgency, dependency, load status, and consistency measurement—to perform a composite calculation and output a suggested solution value. During scoring, higher weights are assigned to the initial stages of strong dependency chains, and buffers are reserved for paths involving high-intensity cross-departmental collaboration to avoid infeasible solutions during subsequent scheduling. The suggested solution value, along with the previously ranked results, serves as the scheduling entry point.

[0036] The suggested solution value is read by the "task scheduling model," combined with the node resource view and time slots, and expanded into a sequence of subtasks to generate a task decomposition table. The decomposition follows the pre- and post-constraints of non-parallelizable steps and the resource mutual exclusion constraints of parallelizable steps, providing start and completion time windows and recording the required positions, locations, and tools. For items with low consistency but high urgency, alternative nodes and backup time windows are registered as fallback paths.

[0037] Based on the task breakdown table, real-time progress data is integrated, aligned according to subtask keys, and the deviation between the planned and actual results is calculated to generate a set of alert signals. Deviation signals are categorized into three types: start delay, execution lag, and resource shortfall, and include window margin and impact range. To prevent false alarms triggered by momentary jitter, signal generation employs short-window smoothing for single-point anomalies and consistency checks with adjacent processes.

[0038] The notification signal group is read by the priority evaluator, which sorts it by timeliness based on window remaining capacity, task level, and dependency chain position, and outputs a push notification. The push content is summarized by task, including the task-handling nodes that need to be reminded, the expected subsequent steps, and suggested actions, and retains a reference to the task breakdown table to facilitate quick location of specific subtasks on the terminal side for processing.

[0039] First, the push notification generated in the previous steps is read, expanded by task key and subtask key, and mapped into a deployable instruction structure according to the terminal protocol. Specifically, the undertaking node, start time window, resource requirements, and handling actions are encoded into terminal command fields, while retaining the reference key with the task breakdown table for subsequent status feedback alignment. After format validation and field completion, the instruction stream is written to the office terminal to generate an execution list; the execution list is organized in chronological order, and the reporting frequency and anomaly reporting channel are recorded for each item.

[0040] Based on the execution list, the status feedback channel of the access terminal is used to aggregate start, progress, completion, and abnormal events according to the sub-task key, forming a process trajectory. To control noise, duplicate reports are merged, and the most recent valid status is retained using a timestamp priority strategy. At the same time, duplicate entries across terminals are deduplicated. The trajectory is aligned with the original planned window to calculate the actual occupied time and resource availability, providing input for subsequent completion evaluation.

[0041] The process trajectory is read by the "Completion Assessment Module," which calculates three metrics based on planned milestones: achievement rate, delay amount, and number of rollback triggers, and outputs a feedback indicator table. During the assessment, for items with alternative execution nodes, the combined duration of the original path and the alternative path is used as the verification standard, and delays transmitted across stages are attributed and decomposed to avoid duplicate counting. The feedback indicator table retains weight labels and data integrity labels for each item dimension to facilitate subsequent sample selection.

[0042] Based on the feedback metric table, an "efficiency analyzer" is constructed, merging metrics and contextual features into training samples. Specifically, the task category, urgency level, node load, and consistency metric are read, and input-label pairs are generated by mapping them to achievement rate and delay. To reduce bias, entries with insufficient data integrity are removed, oversampling is performed on sparse category entries for balancing, and the original reference keys are preserved to support error backtracking after training.

[0043] The training samples are divided into training and validation groups and input into the "flow model retraining pipeline". This pipeline takes item features and resource views as input and learns the mapping relationship between path selection and weight allocation. During training, a penalty term is set for latency, and the contribution of samples with high achievement rates is increased. The updated parameters are used to replace the path scoring module in the document flow model. The replacement is carried out under version control, and the old version is retained for rollback.

[0044] Based on the retraining, the "task scheduling model retraining pipeline" is invoked, using the subtask-level start window, resource allocation, and actual completion time as inputs, and the deviation as a label, to learn the impact of time window and resource allocation on execution results. After training, the time buffer and resource mutual exclusion threshold conditions in the scheduling rules are updated. For entries that have triggered rollback, the success rate of the backup path is used as a priori and written back into the candidate sequence scoring of subsequent decomposition stages.

[0045] To ensure consistency between parameter updates and the execution side, the outputs of the two retraining pipelines simultaneously generate a validation summary, including validation set performance, sample coverage, and satisfaction of key constraints. This summary is written to the configuration center as a checklist before the next round of push proposal generation. If the validation fails, the update is frozen, and only the feedback metric table is recorded for accumulation in the next cycle.

[0046] Finally, the updated flow model and scheduling model are directly invoked when the feature vector table is input for the next batch. The former generates a new task allocation matrix with the corrected path scores, and the latter generates a new task decomposition table with the updated buffer and mutual exclusion threshold conditions. Thus, push, distribution, evaluation, and retraining form a closed loop, and the feedback indicator table and training dataset are referenced cyclically in each cycle, ensuring that the push suggestions for subsequent batches continue to rely on real execution performance.

[0047] Step S103: Write the push suggestion into the office terminal to obtain an execution list, evaluate the completion rate of the execution list to generate a feedback index table, construct an efficiency analyzer based on the feedback index table to obtain a training dataset, and update the flow model and scheduling model by retraining the training dataset.

[0048] First, the push notification output from the previous steps is read, expanded into issueable commands by the task key and subtask key, and the fields are mapped according to the terminal communication protocol. Specifically, the undertaking node, start time window, location, and resource list are encoded into terminal commands, and the reference keys in the task breakdown table are filled back to ensure that subsequent status feedback can correspond one-to-one. A separate channel and frequency constraints are set for scenarios where abnormalities must be reported. After format verification, the commands are written to the office terminal, an execution list is generated, and the batch number and timestamp are recorded.

[0049] Based on the execution list, status events returned by the receiving terminal are aligned according to the subtask key, and four types of records—start, in progress, completed, and abnormal—are aggregated to form a process trajectory. For duplicate events, a time-priority strategy is used to retain the latest record, and cross-terminal reports are deduplicated. The process trajectory is compared with the original planned time window, the difference between the starting offset and the execution duration is calculated, and it is noted whether a backup path was triggered, serving as the raw metric input for evaluation.

[0050] The raw metrics are read by the "Completion Assessment Module," which calculates four types of indicators based on planned milestones: achievement rate, delay, number of rollback triggers, and resource gap ratio, and outputs a feedback indicator table. During the assessment, the longest path metric is used to measure delay for parallel processes, while a serial cumulative metric is used for non-parallel processes to avoid double counting. For subtasks completed by alternative nodes, the difference labels between the original target node and the alternative node are recorded to facilitate subsequent attribution.

[0051] Based on the feedback metric table, an "efficiency analyzer" is constructed, merging contextual features and evaluation results to generate training samples. Contextual features include event category, urgency level, node load, and prior consistency measure; evaluation results include these four categories of metrics and their window positions. Samples lacking key fields are removed, sparse category entries are resampled and balanced, and reference keys are retained in the samples to support error backtracking. The cleaned sample set is organized into training and validation datasets.

[0052] To quantify the contribution of samples to the training phase, this embodiment defines a weighted objective function to control the intensity of subsequent updates: Q = ρ1·A + ρ2·B - ρ3·C Wherein, Q represents the sample contribution; A is the intra-batch quantile mapping value of the achievement rate, used to reflect the completion status; B is the inverse normalized value of the latency, used to measure time performance; C is the joint penalty of the number of backoff triggers and the resource gap ratio; ρ1, ρ2, and ρ3 are non-negative weights, whose range is limited to a preset interval but not fixed. Q is read during the model retraining phase and used as sample-level weights, and is not publicly displayed.

[0053] The training dataset first enters the "flow model retraining pipeline," using the context of items and nodes as input and path selection and weight allocation as learning objectives. During training, Q is used as the sample weight to enhance the influence of high-quality samples, and penalty constraints are introduced on delay-related features to suppress high-risk paths. After training is complete, new path scoring parameters are generated and written into the document flow model using a versioning strategy, while retaining the old version for rollback.

[0054] Based on the retraining, the "task scheduling model retraining pipeline" is invoked, using the time window, resource allocation, and historical deviation of the subtask as input, and minimizing the deviation as the objective. After training, the time buffer and resource mutual exclusion threshold conditions are updated, and the success rate of alternative paths is registered as prior knowledge for candidate ranking in subsequent decomposition stages. The updated results are also versioned and stored in the database, and a validation summary is generated to record the validation set performance and sample coverage.

[0055] Finally, the updated flow model and scheduling model are directly invoked when the next batch of feature vector tables arrives. The former generates a task allocation matrix based on the new path scoring, while the latter generates a task decomposition table based on the updated threshold conditions; both work together to generate a new round of push suggestions. The feedback indicator table and training dataset continue to accumulate in subsequent cycles, providing a basis for continuous correction of the subsequent processing order and resource allocation.

[0056] As described above, the intelligent decision-making assistance method based on multi-source data provided in this application can achieve accurate information identification through semantic analysis and feature extraction. It constructs a decision-making mechanism, combining a flow model and a scheduling model to establish a reliable task management strategy. Performance optimization is introduced, ensuring continuous improvement of decisions through completion evaluation and model retraining. This method effectively addresses the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, providing technical support for intelligent decision-making assistance.

[0057] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S201: Obtain the original file set by accessing the document management system database based on the interface protocol, parse the original file set according to document type to generate a structured data table, perform format unification processing on the structured data table to obtain a standardized text group, construct text cleaning rules based on the standardized text group to generate a preprocessed text library, process the preprocessed text library through a word segmentation model to generate an initial word sequence list, and perform stop word filtering on the initial word sequence list to obtain an effective word order set; Step S202: Reorganize the effective word order set according to the semantic relationship of the text to generate a topic feature matrix, perform topic clustering analysis on the topic feature matrix to obtain a topic distribution map, construct a mapping rule base based on the topic distribution map to generate a category correspondence table, compare the category correspondence table with the word frequency statistics results to generate a weight coefficient group, and apply normalization calculation to the weight coefficient group to obtain a topic mapping table.

[0058] First, after connecting to the official document management system database, the original document set is obtained according to the existing interface protocol. The source system, collection time, and document unique key are registered by batch number. For each document, four types of fields are read: body text, title, issuing unit, and attachment list. The attachments are referenced by the type identifier in the document header to ensure that subsequent parsing can trace back to the original document. After the retrieval is completed, the parsing process is driven by a type mapping table. The types such as meeting minutes, requests, and notices are respectively entered into the corresponding parsing templates, and a structured data table is output, with fields including body text fragments, paragraph numbers, tag fields, and time fields.

[0059] Based on the structured data table, a format standardization process is performed, unifying line breaks, indentation, and numbering styles from different sources into a unified symbol system. Regularization is applied to dates, amounts, and organizational abbreviations to generate standardized text groups. For table-like fragments, row and column coordinates and cell merging information are retained, and readable text and position indices are written together to facilitate subsequent thematic analysis by extracting key table points beyond the sentence / segment level. A one-to-one mapping is established between the standardized text groups and the original attachment reference keys to ensure that cross-textual relationships are not lost.

[0060] The standardized text set was used to construct text cleaning rules. A rule set was generated based on common noise patterns, covering four cleaning paths: headers and footers, fixed signatures, scanned residual characters, and annotation numbers. During cleaning, a whitelist was created for numbers and chapter names with retention value to avoid accidental deletion. After cleaning, a preprocessed text library was generated, with each record containing a summary of the differences before and after cleaning, used for sampling and rollback. The preprocessed text library was sorted by document key and paragraph number to ensure semantic continuity during subsequent word segmentation.

[0061] Based on the preprocessed text library, a word segmentation model is loaded, and a joint strategy of domain dictionary-based priority matching and statistical segmentation is applied to each paragraph to generate an initial word sequence list. To maintain the traceability of time and quantity expressions, date, amount, and proportion segments are segmented as a whole and annotated with the original position index. Subsequently, stop word filtering is performed on the initial word sequence list to remove function words, format markers, and low-information short words, resulting in an effective word sequence set. The paragraph number and character range of each word are retained for subsequent topic reorganization and backfilling.

[0062] Based on the effective word order set, the text is reorganized according to semantic relationships to construct a topic feature matrix. Specifically, word co-occurrence pairs are extracted using a fixed window, and the word pair connection strength is established by combining dependency syntactic relations. Paragraphs are used as row indices, and topic candidates are used as column indices, filling the two dimensions of connection strength and position density. To reduce cross-paragraph diffusion, a decay factor is set for cross-window connections of non-contiguous paragraphs to retain the dominant influence of neighboring contexts, forming a topic feature matrix that can be used for clustering.

[0063] The topic feature matrix is ​​then used for cluster analysis, outputting a topic distribution map. During the analysis, sparse candidates are first filtered out using a density threshold, and then merged based on topic similarity to obtain several topic clusters for each document. Each topic cluster is labeled with core phrases and the range of paragraphs it covers. The topic distribution map retains the mapping relationship between document keys and topic cluster numbers, providing a basis for subsequent rule generation. Simultaneously, stability is calculated for similar topics across documents, serving as a confidence source for mapping rules.

[0064] Based on the aforementioned topic distribution map, a mapping rule base is constructed, generating a category correspondence table. The left side of each rule represents a combination of core phrases and contextual trigger words, while the right side shows category labels and applicable conditions. For ambiguous core phrases, the issuing unit and time window are added as supplementary criteria. The category correspondence table is compared with word frequency statistics to calculate the word frequency percentage of each category within the document and its consistency with the context. Weight coefficient groups are generated, and categories with low coverage are given minimum weight protection to prevent them from being overwhelmed by high-frequency words.

[0065] The weight coefficients are normalized to obtain a topic mapping table. Normalization prioritizes category coverage and contextual consistency within the document, with cross-document stability as an adjustment term, ensuring comparable weights for similar texts from different sources. The topic mapping table records category labels, core phrases, covered paragraphs, and weight values ​​at the entry level, and retains reference keys to the effective word order set and topic distribution map. This table is directly read by the semantic analysis module in subsequent steps to generate a semantic feature matrix and trigger entity extraction. It also serves as prior information input to the classifier for path selection and subsequent flow modeling.

[0066] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S301: The topic mapping table is transformed into a semantic vector group based on the word vector model. Similarity calculation is performed on the semantic vector group to obtain the association strength matrix. A semantic network is constructed based on the association strength matrix to generate a node connection graph. The node connection graph is processed through path analysis to obtain a semantic feature matrix. Feature selection is performed on the semantic feature matrix to obtain a core word set. An entity extractor is constructed based on the core word set to generate an entity label group. Step S302: The entity label group is processed by a deep neural network to generate a classification feature table. Label encoding is performed on the classification feature table to obtain a category mapping set. Matching rules are constructed based on the category mapping set to generate a query vector group. The query vector group is compared with sample records in the historical database to generate a similarity table. Threshold filtering is applied to the similarity table to obtain a feature vector table.

[0067] First, the topic mapping table obtained in the previous steps is read. The core phrases and covered paragraphs are aggregated by pressing the document key. The terms and their context windows are then input into the word vector model and converted into semantic vector groups. To maintain source consistency, the vectors of the same core word in different paragraphs are weighted by position weights, and the topic weights are then filled back into the side information field, forming batch data blocks that facilitate similarity calculation.

[0068] Based on the semantic vector set, pairwise similarity is calculated to obtain the association strength matrix. Similarity is primarily measured using cosine similarity, and a paragraph distance decay coefficient is introduced to reduce the impact of long-distance co-occurrence across paragraphs. Subsequently, weak edges are filtered out according to threshold conditions and sparsity rules, and the remaining strong associations are used to construct a semantic network, outputting a node connection graph. Nodes in the graph correspond to terms, and edge weights record the association strength and source topic number, facilitating subsequent path-level analysis.

[0069] The node connection graph is fed into the path analysis process, which prioritizes searching for high-weight continuous paths and statistically analyzes the segment range covered by each path to generate a semantic feature matrix. This matrix uses paths as rows and semantic slots as columns, with columns including five categories: behavior, object, quantity, time limit, and region. Within each cell, the path score and coverage density are recorded. To suppress synonymous and near-synonymous redundancy, paths with a large number of shared nodes are merged, retaining only the representative path with the highest score.

[0070] Based on the semantic feature matrix, feature selection is performed to obtain the core word set. The selection strategy integrates three factors: path score, slot distribution balance, and cross-document stability, eliminating pseudo-topic words formed solely by high-frequency word accumulation. A one-to-one mapping is established between the core word set and the original paragraph index, serving as candidate trigger words for the entity extraction stage, ensuring that subsequent extraction can trace back to specific text positions for disambiguation.

[0071] The core vocabulary set is used to construct the entity extractor. This extractor combines trigger words, neighboring part-of-speech sequences, and dependency arc types to output entity tag groups in a rule-based, statistically complete order. Tags are stratified by organization, behavioral element, quantity, timeframe, and geographical unit, and carry character ranges and confidence markers. For multi-word phrases, the extractor uses the shortest containing segment as the boundary and registers ambiguity candidates for reference in the classification stage.

[0072] Based on the entity label group, a deep neural network, denoted as the "Document Semantic Classification Network," is loaded to generate a classification feature table. The network input consists of a label sequence and its slot embeddings, and the output is an intermediate representation of the matter category, urgency level, and processing path. Subsequently, label encoding is performed on the classification feature table to obtain a category mapping set. The encoded content includes category labels, side confidence values, and optional path candidate location information to align with the historical sample structure.

[0073] The category mapping set is used to construct matching rules and generate query vector sets. The rules map category tags to the sending unit and time window as search keys, and convert side confidence values ​​into search weights. The query vector sets are compared line by line with the structured index of the historical database to generate a similarity table. In addition to category consistency, the similarity calculation also considers the relevance of the processing path and the closeness of the time limit, forming parallel records of individual scores and overall scores.

[0074] Based on the similarity table, a threshold filtering method is applied to obtain a feature vector table. During filtering, a minimum retention ratio is set for categories with low coverage to avoid sample bias; records with high confidence but path conflicts are weighted less than directly removed, and source citation keys and sub-scores are written into the vector. This feature vector table serves as direct input for subsequent flow modeling. Its row-level keys maintain a traceable relationship with the entity label groups and topic mapping table from the previous stage, supporting source tracing and verification for subsequent path selection and task allocation.

[0075] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S401: Split the feature vector table into feature subset groups according to a preset dimension, perform normalization processing on the feature subset groups to generate a standardized vector set, construct a neural network classifier based on the standardized vector set to generate a class probability matrix, process the class probability matrix through a decision tree model to obtain a flow path graph, perform path optimization on the flow path graph to obtain a task allocation matrix, and construct a timeliness evaluator based on the task allocation matrix to generate a processing order table. Step S402: The processing order table is processed by the rule matching engine to generate a rule adaptation table. Feasibility verification is performed on the rule adaptation table to obtain a set of constraints. A scoring model is constructed based on the set of constraints to generate a weight distribution table. The weight distribution table and the set of judgment factors are weighted to obtain a score sequence. Normalization is applied to the score sequence to generate a suggested solution value.

[0076] First, the feature vector table generated in the previous steps is read and split into three categories according to preset dimensions: source features, event semantic features, and deadline and resource features, forming feature subset groups. After splitting, normalization processing is performed on each subset: for quantitative fields, distributed stretching is applied and scaling parameters are recorded; for enumerated fields, one-hot encoding is performed and positional order is maintained, outputting a standardized vector set. To ensure consistency across batches, the scaling parameters of the current batch are compared with historical parameters for drift verification. If the offset exceeds the tolerance range, the parameters are re-estimated within the current batch and a version number is recorded for subsequent model reading.

[0077] Based on the standardized vector set, a "document item classification network" is constructed. The input is a concatenation of three sub-vectors, and the output is a category probability matrix. This matrix uses item instances as rows, predefined categories as columns, and probability values ​​as cells, with weights from the preceding consistency measure used as loss weighting factors to suppress the influence of noisy samples. After inference, only the sparse representations of the top few high-probability categories are retained to reduce the branch size of subsequent path search.

[0078] The category probability matrix is ​​fed into the "path generator." This component, based on a decision tree model, maps high-probability categories to combinations of undertaking nodes and processing steps, resulting in a flow path graph. Nodes in the graph represent processing steps, edges represent feasible preceding and following relationships, and the edges record the source category and estimated time consumption. To avoid loops and invalid branches, the path generation stage incorporates constraint pruning based on legally mandated prerequisite relationships, and simultaneously removes nodes that are temporarily unreachable based on their current load, ensuring the graph structure is executable.

[0079] Based on the aforementioned workflow path graph, path optimization is performed to generate a task allocation matrix. The optimization objective is to minimize waiting time and cross-departmental migration costs under given time constraints, and the output matrix employs a layer-by-layer expansion and parallel selection strategy. Rows in the matrix correspond to task instances, columns to candidate nodes, and cells record allocation weights, expected start windows, and backup candidate flags. In cases where multiple equally effective candidates exist, two parallel allocation paths (primary and secondary) are retained to support subsequent rollback.

[0080] The task allocation matrix is ​​read by the "timeliness evaluator," which, combined with the node load view and the statutory deadline, calculates the completion probability and overdue risk of each task, and outputs a processing order table. During sorting, the allocation weight and expected start window are read first, and tasks nearing their deadlines are promoted in the order; tasks with low source consistency are kept in a conservative order and marked as items requiring verification. The processing order table is organized by task key and retains references to the path graph and standardized vector set for easy backtracking later.

[0081] Based on the processing sequence table, the "rule matching engine" aligns each item with the business rule base to generate a rule adaptation table. The matching content includes a list of required materials, co-signature dependencies, and regional permission boundaries. For items with missing materials, they are marked as insufficient and supplementary suggestions are generated. Subsequently, a feasibility verification is performed on the rule adaptation table, integrating node availability time periods and resource mutual exclusion conditions to output a constraint set, clearly defining non-parallel steps and the earliest possible start time.

[0082] The constraint set is read by the "scheme scorer" and synthesized into a weight distribution table. During synthesis, time urgency, process dependency, and resource conflict risk are treated as three components, with historical node fulfillment stability introduced as an adjustment term. To facilitate subsequent calculations, the weight distribution table uses nodes as columns, items as rows, and cells as weight values. Entries with insufficient conditions are assigned a minimum retention weight and a completion flag to prevent them from being mistakenly entered into the execution queue before completion.

[0083] Based on the weight distribution table, the preceding "judgment factor set" is read, and the weights of each node for the same matter are weighted and calculated to obtain a score sequence. During the calculation, the initial link with strong dependency chain is weighted and amplified, the parallel link with high conflict risk is penalized, and the order of the processing order is converted into time-side weights and superimposed to output a score vector that takes into account both timeliness and resource feasibility.

[0084] The score sequence is normalized to obtain suggested solution values. Normalization is performed on a task-by-task basis, using distribution scale and stability as benchmarks, mapping scores to execution tendency values ​​in the range of zero to one, while preserving the relative differences between primary and backup nodes. The suggested solution values, along with window information from the processing sequence table, are passed to the scheduling stage as direct input for subtask deployment and resource allocation, and serve as the baseline for deviation determination when comparing with real-time progress.

[0085] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S501: Convert the suggested solution values ​​according to task type to generate a solution element set, perform task decomposition on the solution element set to obtain a sub-task sequence, build a resource allocator based on the sub-task sequence to generate a scheduling rule table, process the scheduling rule table through a constraint solver to obtain an execution plan group, apply feasibility verification to the execution plan group to generate a task decomposition table, and build a monitoring model based on the task decomposition table to obtain a state tracking matrix. Step S502: Calculate the difference between the status tracking matrix and the real-time progress data to generate a deviation vector group. Perform threshold judgment on the deviation vector group to obtain an alarm signal table. Construct a priority evaluator based on the alarm signal table to generate an urgency matrix. Process the urgency matrix through sorting rules to obtain a push sequence. Apply template mapping to the push sequence to generate a push proposal.

[0086] First, the suggested solution values ​​output from the previous steps are read and mapped to solution element sets according to task type. During mapping, the task category, processing path, primary and backup nodes, and time window are broken down into element items, and the sorting position and deadline boundaries are filled in according to the processing order table, forming a four-element structure of "type-node-window-preference". To maintain consistency, items with missing windows are placed using the historical average time period and marked as requiring verification for processing in subsequent decomposition stages.

[0087] Based on the set of scheme elements, task decomposition is performed to obtain a sequence of subtasks. The decomposition rules divide the original items according to the constraints and parallelism conditions of the processing path, expanding each node into a subtask and inheriting the corresponding window and execution tendency. For nodes with mandatory prerequisite materials, a "materials supplementation" subtask is generated in advance with an earliest start time set; for parallelizable steps, resource mutual exclusion markers are assigned to prevent simultaneous occupation of the unique resource. The subtask sequence retains the reference key to the original item for easy subsequent summarization.

[0088] The subtask sequence is read by the "resource allocator" and, combined with the node resource view and three capability pools (personnel, venue, and tools), a scheduling rule table is generated. The rule table provides the number of positions, minimum duration, and required equipment list for each subtask, and marks the set of alternative resources. When the target resource is temporarily unavailable, the rule table registers alternative resources and suboptimal time periods to reduce the probability of subsequent solutions being infeasible.

[0089] Based on the scheduling rule table, the constraint solver is invoked to form an execution plan group. The solver searches for scheduling solutions that satisfy the path order, using time windows, resource exclusivity, and cross-department migration costs as constraints. It assigns start and end times to each subtask and selects primary or backup resources. To handle conflict scenarios, the solver allows limited window shifting and records the shift magnitude as input for subsequent feasibility verification.

[0090] The execution plan group is sent to the feasibility verification process, where each plan is verified item by item in conjunction with the statutory deadline and node reachability, and a task breakdown table is output. During verification, plans with incomplete prerequisite materials are returned for reordering, and plans whose window shift exceeds the threshold due to resource occupation are marked as risk items. The task breakdown table is organized by subtasks, recording time windows, resource assignments, and risk markings, and aggregated back into an event-level index for easy tracking by link in the monitoring process.

[0091] Based on the task decomposition table, a monitoring model is constructed, generating a status tracking matrix. The monitoring model subscribes to terminal reports and business system progress, aligns them by subtask keys, and samples them at fixed intervals. The matrix columns include planned time, actual time, resource availability, anomaly tags, and rollback status. To suppress occasional noise, the monitoring model applies short-window smoothing to single-point anomalies and records confidence markers, retaining the original reports as a basis for traceability.

[0092] The state tracking matrix is ​​compared with the real-time progress data to generate a deviation vector set. The deviation calculation distinguishes between three types of indicators: start delay, execution lag, and resource gap, and propagates along the subtask chain to assess the impact on subsequent stages. For entries with rollbacks, the deviation vector simultaneously records the rollback trigger reason and switching time, used for subsequent priority evaluation.

[0093] Based on the aforementioned deviation vector group, a threshold judgment is performed, and an alarm signal table is output. The threshold is mainly determined by the deadline boundary and window margin. Stricter judgments are applied to high-level issues, and alarms are prioritized for stable deviations across two consecutive sampling periods to reduce false alarms caused by momentary jitter. The alarm signal includes a subtask key, deviation type, and suggested action placeholders, providing input for urgency assessment.

[0094] The alarm signal table is read by the "priority evaluator," which calculates an urgency matrix by combining the event level, link location, and window margin. The evaluator assigns higher weights to alarms located at critical starting points, increases the level of alarms nearing their deadlines, and appropriately reduces the weight of alarms with alternative resources, resulting in a sortable urgency matrix. The matrix is ​​organized with subtasks as rows and urgency as columns, retaining references to the task decomposition table.

[0095] Finally, the urgency matrix is ​​processed using sorting rules to obtain a push sequence, which is then mapped to a push suggestion. The push sequence is arranged from highest to lowest urgency. Multiple alarms within the same issue are merged into a single handling suggestion, clearly defining the responsible node, the time window to be adjusted, and alternative resource solutions. The push suggestion includes original deviations and risk markers in each item, facilitating rapid location and execution by the terminal, while also providing a traceable reference key for subsequent execution checklist writing and completion assessment.

[0096] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S601: Parse the push suggestion according to the terminal type to generate a push instruction set, perform format conversion on the push instruction set to obtain a terminal command table, build a task dispatcher based on the terminal command table to generate a deployment parameter group, process the deployment parameter group through an interface adapter to obtain an execution instruction stream, apply status verification to the execution instruction stream to generate an execution list, and build a monitoring model based on the execution list to obtain a task status table; Step S602: Compare the task status table with the preset completion standard to generate an achievement degree matrix. Perform multi-dimensional analysis on the achievement degree matrix to obtain an evaluation feature group. Based on the evaluation feature group, construct a performance calculator to generate an indicator score table. Normalize the indicator score table to obtain a feedback indicator table.

[0097] First, the push notification output from the previous steps is read and parsed to generate a push instruction set according to the terminal type. During parsing, the undertaking node, subtask key, time window, and action are broken down into basic fields, and the display style and interactive items are selected according to the terminal capability table. For example, mobile terminals only retain necessary fields and add a quick feedback button, while desktop terminals retain the complete resource alternative and a remarks area. After parsing, the fields of different terminals are aligned to ensure that the identification of the same subtask is consistent across multiple terminals.

[0098] Based on the aforementioned push instruction set, a format conversion is performed to obtain a terminal command table. The format conversion includes field renaming, time format standardization, and resource list structuring to ensure that the instructions can be directly parsed by the programs on each terminal side. Subsequently, a task dispatcher is constructed based on the terminal command table, aggregating instructions from the same undertaking node and assigning batch numbers and receipt time limits to generate a deployment parameter group. This parameter group records the sending channel, retry policy, and receipt verification code, providing a basis for subsequent status verification.

[0099] The deployment parameter set is processed by the interface adapter and converted into an execution command stream. The adapter encapsulates the message body according to the communication protocols of different terminals and writes a batch number and checksum in the header to ensure the integrity of cross-system transmission. To prevent duplicate delivery, the adapter implements mutual exclusion control for concurrent transmission of the same subtask and registers the transmission timestamp. After transmission, the execution command stream enters the status verification stage, generating an initial status flag through acknowledgment comparison and timeout detection.

[0100] Based on the execution instruction flow, an execution list is generated using application status checks. The checks include three aspects: receipt verification, terminal online status, and message delivery verification. Qualified instructions are registered as "issued," while failed instructions have their failure reasons recorded and are added to the retry queue. The execution list is organized by subtasks, recording the responsible node, status flag, and latest receipt time, and maintains consistency with the task decomposition table to facilitate button alignment during monitoring.

[0101] Based on the execution list, a monitoring model is constructed to generate a task status table. The monitoring model subscribes to process events transmitted back from the terminals, merges start, in progress, completed, and abnormal records at fixed intervals, and deduplicates cross-terminal reports. The task status table maintains the latest status, status duration, and abnormal label for each subtask, and also records the mapping with batch numbers to support sample selection and traceability in subsequent evaluation phases.

[0102] Based on the task status table, it is compared with preset completion standards to generate an achievement matrix. The comparison criteria are based on the subtask plan window and the statutory deadline. For parallel processes, the maximum completion time is used as the link criterion, and for serial processes, the cumulative completion time is used as the criterion. The matrix calculates the achievement rate, delay, and number of backoff triggers, and marks the data integrity, removing undeterminable entries due to missing reports.

[0103] The achievement matrix undergoes multi-dimensional analysis, outputting an evaluation feature set. The analysis dimensions include three categories: task type, implementation node, and time window. It also identifies high-frequency anomaly tags and their distribution within the process chain. For samples with long-tail categories, aggregation is performed by document source and time segmentation to mitigate bias caused by sample sparsity. The evaluation feature set retains item-level statistics on achievement rate and delay amount, serving as direct input for subsequent score calculations.

[0104] Based on the aforementioned evaluation feature set, a performance calculator is constructed to generate an indicator score table. The calculator assigns a positive weight to the achievement rate, penalizes delays and rollback triggers, reduces the weight of items with insufficient data integrity, and uses the historical stability of the undertaking node as an adjustment factor to form a comprehensive score. To maintain comparability between different batches, scores for the same dimension are standardized within the batch before being merged to avoid extreme samples affecting the overall scale.

[0105] Finally, the index score table is normalized to obtain the feedback index table. During normalization, a zero-to-one mapping is performed for both the item dimension and the node dimension, and key sub-item scores and anomaly labels are retained as data entry points for subsequent performance analysis and retraining. The reference keys of the feedback index table and the task status table are consistent, and it is directly read into the performance analyzer in the next loop to support model updates and strategy adjustments.

[0106] In one embodiment of the intelligent auxiliary decision-making method based on multi-source data in this application, it may further include the following: Step S701: Decompose the feedback indicator table according to the evaluation dimension to generate indicator feature groups, perform data cleaning on the indicator feature groups to obtain an effective sample set, construct a feature extractor based on the effective sample set to generate a training feature table, process the training feature table through label mapping to obtain training data pairs, apply data augmentation to the training data pairs to generate a training dataset, and construct a validation segmenter based on the training dataset to obtain a training validation set. Step S702: The training and validation set is processed by the parameter optimizer to generate a model parameter set. Iterative training is performed on the model parameter set to obtain an updated weight table. A model updater is constructed based on the updated weight table to generate a parameter configuration table. The parameter configuration table is written into the flow model and the scheduling model to complete the model update.

[0107] First, the feedback indicator table generated in the previous steps is read and decomposed into three categories according to the evaluation dimensions: item dimension, node dimension, and time window dimension, forming indicator feature groups. During decomposition, achievement rate, delay amount, number of rollbacks, and data integrity are used as basic metrics, and the reference relationships with item keys, subtask keys, and batch numbers are preserved to ensure that subsequent sample construction is traceable. To avoid cross-dimensional confusion, dimension labels are explicitly registered in each decomposition record, ensuring that the same entry belongs to only one main dimension.

[0108] Based on the aforementioned indicator feature group, data cleaning is performed to obtain a valid sample set. The cleaning process includes three steps: missing field removal, outlier truncation, and duplicate record removal. Entries with uncertain definitions are excluded based on data integrity markers. Fields showing significant cross-batch drift are recorded with drift summaries for use as a reference for subsequent feature robustness screening. After cleaning, groups are established according to event keys and time windows to ensure that validation segmentation can be performed chronologically during the training phase.

[0109] The valid sample set is read by the "feature extractor" to generate a training feature table. The extractor encodes the item category, processing path ranking, node historical stability, window margin, and urgency level as numerical features, and uses delay, rollback count, and resource gap ratio as context for target-related features. To suppress multicollinearity, highly correlated inputs are compressed in the main direction, and the version number of the compression mapping is retained in the feature table for easy restoration during deployment.

[0110] Based on the training feature table, label mapping is performed to obtain training data pairs. The mapping strategy constructs two labels in parallel: the comprehensive score and the overdue status. The former serves as the regression target, and the latter as the classification target, to support subsequent multi-task training. To prevent label noise amplification, samples with low data integrity are assigned low-weight labels, while the original sub-scores are retained for the loss function to weight each sub-score.

[0111] The training data is fed into the data augmentation process to generate a training dataset. Augmentation methods include two types: time window perturbation and mild noise injection. The former fine-tunes the planning window while keeping the labels unchanged to improve robustness to schedule shifts, while the latter applies small random perturbations to non-critical fields to improve generalization ability. The augmented data is sliced ​​according to event key and time order, and slice numbers are labeled to provide time stability constraints for validating the segmenter.

[0112] Based on the training dataset, a validation segmenter is constructed to obtain training and validation sets. The segmenter uses time slice crossover as a principle to ensure that the validation set comes from time slices later than the training set, avoiding information leakage; at the same time, long-tail categories are extracted proportionally to maintain category coverage. The training and validation sets maintain reference keys at the sample level to ensure that the original feedback index table can be located for error analysis during the optimization phase.

[0113] The training and validation sets are read by the parameter optimizer to generate model parameter sets. The optimizer outputs initial parameters for the two types of models: path scoring parameters for the document circulation model and time buffers and mutual exclusion thresholds for the task scheduling model. During the optimization phase, the training dataset is read, and iterative training is performed using multi-task loss, with regression and classification objectives participating with different weights. A penalty term is introduced to address the error caused by window shifting, and an adaptive learning rate is applied to high-weight samples to shorten the convergence path.

[0114] After iteration, an updated weight table is obtained. This weight table is subdivided by model and module, including the weight set of the path scoring submodule and the buffer coefficients in the scheduling rules. To ensure deployment security, the updated weight table includes a validation set performance summary and constraint satisfaction flags. If any critical constraint fails, only the observation mode configuration is output, without triggering an overwrite.

[0115] Based on the updated weight table, a model updater is constructed, generating a parameter configuration table. The configuration table records the target model, parameter namespace, version number, and activation conditions, and provides rollback anchor points. The updater is injected in batches according to a canary strategy. First, the last two batches of input are reproduced in a shadow environment to verify whether the output task allocation matrix and task decomposition table meet the existing boundary conditions. Only after successful verification is the updater switched to the primary version.

[0116] Finally, the parameter configuration table is written into the document circulation model and task scheduling model to complete the model update. The updated document circulation model generates a new task allocation matrix when a new batch of feature vector tables enters, and the updated task scheduling model outputs a new task decomposition table accordingly. The training and validation sets continue to expand in the next cycle, and the new feedback index table will flow back to the decomposition and cleaning stages. The closed loop remains within the same reference key system, supporting continuous iteration and steady-state operation.

[0117] To effectively address the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, and to provide technical support for intelligent decision-making assistance, this application provides an embodiment of an intelligent decision-making assistance device based on multi-source data for implementing all or part of the aforementioned intelligent decision-making assistance method based on multi-source data. See [link to embodiment]. Figure 2 The intelligent auxiliary decision-making device based on multi-source data specifically includes the following components: The document processing module 10 is used to access the document management system to obtain document data streams, process the document data streams through a document parser to generate a standard information table, perform text segmentation on the standard information table to obtain a word sequence set, construct topic recognition rules based on the word sequence set to generate a topic mapping table, process the topic mapping table through a semantic analysis module to obtain a semantic feature matrix, perform keyword extraction on the semantic feature matrix to obtain entity label groups, construct a deep classifier based on the entity label groups to generate a type annotation set, and match the type annotation set with a historical database to obtain a feature vector table. The suggestion generation module 20 is used to construct a flow model based on the feature vector table to generate a task allocation matrix, perform priority calculation on the task allocation matrix to obtain a processing order table, match the processing order table with the business rule base to generate a set of judgment factors, perform scoring calculation on the set of judgment factors to obtain a suggested solution value, construct a scheduling model based on the suggested solution value to generate a task decomposition table, compare the task decomposition table with the progress data to obtain a group of prompt signals, and perform timeliness sorting on the group of prompt signals to generate a push suggestion. The auxiliary decision-making module 30 is used to write the push suggestion into the office terminal to obtain an execution list, evaluate the completion rate of the execution list to generate a feedback index table, construct an efficiency analyzer based on the feedback index table to obtain a training dataset, and update the flow model and scheduling model through model retraining of the training dataset.

[0118] As described above, the intelligent auxiliary decision-making device based on multi-source data provided in this application can achieve accurate information identification through semantic analysis and feature extraction. It constructs a decision-making mechanism, combining a flow model and a scheduling model to establish a reliable task management strategy. Performance optimization is introduced, ensuring continuous improvement of decisions through completion evaluation and model retraining. This method effectively addresses the shortcomings of traditional technologies in information extraction, task management, and performance evaluation, providing technical support for intelligent auxiliary decision-making.

[0119] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the intelligent auxiliary decision-making method based on multi-source data.

[0120] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent auxiliary decision-making method based on multi-source data.

[0121] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described intelligent auxiliary decision-making method based on multi-source data.

[0122] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent auxiliary decision-making method based on multi-source data, characterized in that, The method includes: The document management system is accessed to obtain a file data stream. The file data stream is processed by a file parser to generate a standard information table. Text segmentation is performed on the standard information table to obtain a word sequence set. Based on the word sequence set, topic recognition rules are constructed to generate a topic mapping table. The topic mapping table is processed by a semantic analysis module to obtain a semantic feature matrix. Keyword extraction is performed on the semantic feature matrix to obtain an entity label group. Based on the entity label group, a deep classifier is constructed to generate a type annotation set. The type annotation set is matched with a historical database to obtain a feature vector table. Based on the feature vector table, a flow model is constructed to generate a task allocation matrix. Priority calculation is performed on the task allocation matrix to obtain a processing order table. The processing order table is matched with the business rule base to generate a set of judgment factors. Scoring calculation is performed on the set of judgment factors to obtain a suggested solution value. Based on the suggested solution value, a scheduling model is constructed to generate a task decomposition table. The task decomposition table is compared with the progress data to obtain a group of prompt signals. Timeliness sorting is performed on the group of prompt signals to generate a push suggestion. The push notification is written into the office terminal to obtain an execution list. The completion rate of the execution list is evaluated to generate a feedback index table. Based on the feedback index table, an efficiency analyzer is constructed to obtain a training dataset. The training dataset is then used to update the flow model and scheduling model through model retraining.

2. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The document management system is accessed to obtain a file data stream. This file data stream is processed by a file parser to generate a standard information table. Text segmentation is performed on the standard information table to obtain a word sequence set. Based on the word sequence set, topic recognition rules are constructed to generate a topic mapping table, including: The original file set is obtained by accessing the document management system database based on the interface protocol. The original file set is parsed according to document type to generate a structured data table. The structured data table is subjected to format unification processing to obtain a standardized text group. Based on the standardized text group, text cleaning rules are constructed to generate a preprocessed text library. The preprocessed text library is processed by a word segmentation model to generate an initial word sequence list. Stop word filtering is performed on the initial word sequence list to obtain an effective word sequence set. The effective word order set is reorganized according to the semantic relationship of the text to generate a topic feature matrix. Topic clustering analysis is performed on the topic feature matrix to obtain a topic distribution map. A mapping rule base is constructed based on the topic distribution map to generate a category correspondence table. The category correspondence table is compared with the word frequency statistics results to generate a weight coefficient group. Normalization is applied to the weight coefficient group to obtain a topic mapping table.

3. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The process involves processing the topic mapping table through a semantic analysis module to obtain a semantic feature matrix, extracting keywords from the semantic feature matrix to obtain entity label groups, constructing a deep classifier based on the entity label groups to generate a type annotation set, and matching the type annotation set with a historical database to obtain a feature vector table, including: The topic mapping table is transformed into a semantic vector group based on the word vector model. Similarity calculation is performed on the semantic vector group to obtain the association strength matrix. A semantic network is constructed based on the association strength matrix to generate a node connection graph. The node connection graph is processed through path analysis to obtain a semantic feature matrix. Feature selection is performed on the semantic feature matrix to obtain a core word set. An entity extractor is constructed based on the core word set to generate an entity label group. The entity label group is processed by a deep neural network to generate a classification feature table. Label encoding is performed on the classification feature table to obtain a category mapping set. Matching rules are constructed based on the category mapping set to generate a query vector group. The query vector group is compared with sample records in the historical database to generate a similarity table. Threshold filtering is applied to the similarity table to obtain a feature vector table.

4. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The process of constructing a flow model based on the feature vector table to generate a task allocation matrix, performing priority calculation on the task allocation matrix to obtain a processing order table, matching the processing order table with the business rule base to generate a set of evaluation factors, and performing scoring calculation on the set of evaluation factors to obtain a suggested solution value includes: The feature vector table is split into feature subsets according to a preset dimension. Normalization is performed on the feature subsets to generate a standardized vector set. A neural network classifier is built based on the standardized vector set to generate a class probability matrix. The class probability matrix is ​​processed through a decision tree model to obtain a flow path graph. Path optimization is performed on the flow path graph to obtain a task allocation matrix. A timeliness evaluator is built based on the task allocation matrix to generate a processing order table. The processing order table is processed by a rule matching engine to generate a rule adaptation table. Feasibility verification is performed on the rule adaptation table to obtain a set of constraints. A scoring model is constructed based on the set of constraints to generate a weight distribution table. The weight distribution table is weighted and calculated with the set of judgment factors to obtain a score sequence. Normalization is applied to the score sequence to generate a suggested solution value.

5. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The process of constructing a scheduling model based on the suggested scheme values ​​to generate a task decomposition table, comparing the task decomposition table with progress data to obtain a group of prompt signals, and performing timeliness sorting on the prompt signal group to generate a push recommendation includes: The suggested solution values ​​are converted according to task type to generate a solution element set. The solution element set is decomposed into a sub-task sequence. A resource allocator is built based on the sub-task sequence to generate a scheduling rule table. The scheduling rule table is processed by a constraint solver to obtain an execution plan group. Feasibility verification is applied to the execution plan group to generate a task decomposition table. A monitoring model is built based on the task decomposition table to obtain a state tracking matrix. The status tracking matrix and real-time progress data are compared to generate a deviation vector group. A threshold judgment is performed on the deviation vector group to obtain an alarm signal table. A priority evaluator is constructed based on the alarm signal table to generate an urgency matrix. The urgency matrix is ​​processed by sorting rules to obtain a push sequence. A template mapping is applied to the push sequence to generate a push proposal.

6. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The step of writing the push suggestion to the office terminal to obtain an execution list, and evaluating the completion rate of the execution list to generate a feedback indicator table includes: The push suggestion is parsed according to the terminal type to generate a push instruction set. The push instruction set is then converted to obtain a terminal command table. A task dispatcher is built based on the terminal command table to generate a deployment parameter group. The deployment parameter group is processed through an interface adapter to obtain an execution instruction stream. A status check is applied to the execution instruction stream to generate an execution list. A monitoring model is built based on the execution list to obtain a task status table. The task status table is compared with the preset completion standard to generate an achievement degree matrix. Multi-dimensional analysis is performed on the achievement degree matrix to obtain an evaluation feature group. Based on the evaluation feature group, a performance calculator is constructed to generate an indicator score table. The indicator score table is normalized to obtain a feedback indicator table.

7. The intelligent auxiliary decision-making method based on multi-source data according to claim 1, characterized in that, The process of constructing a performance analyzer based on the feedback index table to obtain a training dataset, and updating the flow model and scheduling model through model retraining using the training dataset, includes: The feedback indicator table is decomposed into indicator feature groups according to the evaluation dimensions. Data cleaning is performed on the indicator feature groups to obtain an effective sample set. A feature extractor is built based on the effective sample set to generate a training feature table. The training feature table is processed through label mapping to obtain training data pairs. Data augmentation is applied to the training data pairs to generate a training dataset. A validation segmenter is built based on the training dataset to obtain a training validation set. The training and validation sets are processed by a parameter optimizer to generate a model parameter set. Iterative training is performed on the model parameter set to obtain an updated weight table. A model updater is constructed based on the updated weight table to generate a parameter configuration table. The parameter configuration table is written into the flow model and the scheduling model to complete the model update.

8. An intelligent auxiliary decision-making device based on multi-source data, characterized in that, The device includes: The document processing module is used to access the document management system to obtain document data streams, process the document data streams through a document parser to generate a standard information table, perform text segmentation on the standard information table to obtain a word sequence set, construct topic recognition rules based on the word sequence set to generate a topic mapping table, process the topic mapping table through a semantic analysis module to obtain a semantic feature matrix, perform keyword extraction on the semantic feature matrix to obtain entity label groups, construct a deep classifier based on the entity label groups to generate a type annotation set, and match the type annotation set with a historical database to obtain a feature vector table. The suggestion generation module is used to construct a flow model based on the feature vector table to generate a task allocation matrix, perform priority calculation on the task allocation matrix to obtain a processing order table, match the processing order table with the business rule base to generate a set of judgment factors, perform scoring calculation on the set of judgment factors to obtain a suggested solution value, construct a scheduling model based on the suggested solution value to generate a task decomposition table, compare the task decomposition table with the progress data to obtain a group of prompt signals, and perform timeliness sorting on the group of prompt signals to generate a push suggestion. The auxiliary decision-making module is used to write the push suggestion into the office terminal to obtain an execution list, evaluate the completion rate of the execution list to generate a feedback index table, construct an efficiency analyzer based on the feedback index table to obtain a training dataset, and update the flow model and scheduling model through model retraining of the training dataset.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent auxiliary decision-making method based on multi-source data as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent auxiliary decision-making method based on multi-source data as described in any one of claims 1 to 7.

Images