A file flow management system optimization method

Abstract

The application relates to the technical field of intelligent data processing and office automation, and discloses a file process management system optimization method, which comprises the following steps: cleaning official documents, performing sliding window blocking, and generating a global semantic vector by naming entity density weighting; deciding a decoding path between differential extraction and full-amount generation modes based on a local semantic density index; generating an initial task by correcting an original logarithmic probability vector of a model by using a granularity mapping matrix; performing multi-dimensional redundancy checking, calculating an execution damping coefficient based on an official document reference map, and controlling a task state output final list; and calculating a granularity entropy change index according to an operation log to reversely update the granularity mapping matrix. The application utilizes a heterogeneous feedback closed loop to construct a dynamically updated granularity mapping matrix, realizes adaptive approximation of task generation granularity from extensive to accurate, and solves the problems that a model cannot continuously adapt to the evolution of specific department business habits and is difficult to solve model drift.

Description

Technical Field

[0001] This invention relates to the field of intelligent data processing and office automation technology, specifically to an optimization method for a document workflow management system. Background Technology

[0002] In the daily operations and management of large enterprises, document flow is the core link connecting the decision-making and execution levels. From safety hazard investigation notices to quarterly equipment maintenance instructions, a massive amount of unstructured documents carry the company's specific business actions. To improve execution efficiency, rapidly converting these documents written in natural language into traceable and quantifiable structured task data in a lean management system has become a key requirement for enterprise digital transformation. This not only reduces the tediousness of manual data entry but also ensures the timeliness and accuracy of instructions.

[0003] Current technologies primarily rely on large language models combined with optical character recognition (OCR) to process such documents. By pre-designing general prompt word templates or fine-tuning the model for specific domains, the system can automatically identify key entities in the document. This approach performs exceptionally well when processing entirely new document types or providing summary overviews, quickly locating responsible parties and deadlines from the entire text. It has initially achieved automated extraction from unstructured text to semi-structured data, reducing the reading and transcription burden on business personnel to some extent.

[0004] However, this generalized approach reveals significant logical flaws in deep business scenarios. The task granularity generated by the model is often rigid, failing to adapt to the vastly different operational habits of different departments. Frontline teams may need to break down long sentences into multiple actions for item-by-item check-ins, while functional departments tend to manage them in a consolidated manner. The system lacks the memory and adaptability to such frequent user corrections, resulting in high friction costs in human-computer interaction. For routine documents with fixed formats, full parsing not only wastes computing power, but also dilutes key change information with a large amount of repetitive background text, easily inducing the model to generate non-factual results. Furthermore, existing extraction technologies typically view each document in isolation, lacking awareness of the overall business process status. The system often blindly generates tasks without considering pre-existing dependencies, leading to logically unexecutable instructions or conflicts with existing processes. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an optimization method for a document workflow management system. This method solves the technical problems of existing technologies in the document task extraction process, such as rigid generation granularity that is difficult to match with users' differentiated work habits, full parsing of high-frequency routine documents leading to dilution of key incremental information by background noise, and deadlock in task flow logic due to lack of awareness of dependencies on preceding business processes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing a document workflow management system, comprising the following steps:

[0007] The system receives official document data and performs text cleaning. It performs sliding window block processing and weighted encoding on the plain text sequence and generates a global semantic vector of the official document by combining the named entity density weighted algorithm.

[0008] Based on the local semantic density index of the global semantic vector in the historical document vector database, the subsequent decoding path is determined between the differential extraction mode and the full generation mode.

[0009] The corresponding granularity mapping matrix is ​​loaded according to the determined decoding path. During the decoding process, the granularity mapping matrix is ​​used to linearly correct the original log probability vector of the model output layer to generate an initial task list.

[0010] Perform multi-dimensional redundancy verification on the initial task list based on joint determination of responsible subject, time window and semantic content, and calculate the execution damping coefficient based on the document citation map to control the task status, and output the final task list;

[0011] Collect heterogeneous operation logs for the final task list, calculate the granularity entropy change index, and use the granularity entropy change index to update the granularity mapping matrix in reverse.

[0012] Preferably, the step of receiving official document data and performing text cleaning, performing sliding window block processing and weighted encoding on the plain text sequence, and generating a global semantic vector of the official document by combining a named entity density weighted algorithm includes:

[0013] The system receives official document data streams or layout files using a pre-defined data interface protocol, performs layout cleaning operations to remove headers, footers, and format control characters, and extracts the plain text sequence.

[0014] The plain text sequence is sliced ​​using an overlapping sliding window mechanism to obtain an ordered set of text fragments;

[0015] Each segment in the text segment set is encoded using a pre-trained encoder model to identify the number of responsible entities, time nodes, and action instruction entities contained in each segment;

[0016] The weight of each segment is calculated based on the number of identified entities, and the feature vectors of each segment are weighted and synthesized based on the weights to obtain the global semantic vector.

[0017] Preferably, determining the subsequent decoding path between the differential extraction mode and the full generation mode based on the local semantic density index of the global semantic vector in the historical document vector database includes:

[0018] The global semantic vector is used as the query vector to retrieve a set of candidate vectors within a preset similarity range from the historical document vector database;

[0019] The number of neighbor samples that meet the lower similarity limit in the candidate vector set is counted, and the density of the distribution of the number of neighbor samples in the neighborhood of a unit hypersphere is calculated to obtain the local semantic density index.

[0020] The local semantic density index is compared with a preset density determination threshold;

[0021] If the local semantic density index is higher than the density determination threshold, then the differential extraction mode is selected;

[0022] If the local semantic density index is lower than or equal to the density determination threshold, then the full generation mode is selected.

[0023] Preferably, the step of loading the corresponding granularity mapping matrix according to the determined decoding path, and using the granularity mapping matrix to linearly correct the original log probability vector of the model output layer during the decoding process to generate the initial task list includes:

[0024] Identify the responsible department to which the current document belongs, and retrieve the granularity mapping matrix corresponding to the responsible department from the system configuration library. The granularity mapping matrix is ​​a sparse vector with the same dimension as the model vocabulary.

[0025] During the time step of the model autoregressive decoding, the original log probability vector generated by the model is obtained;

[0026] The original log probability vector is weighted and superimposed with the granularity mapping matrix to obtain the modified log probability vector.

[0027] Calculate the token generation probability distribution based on the modified logarithmic probability vector, and generate a text sequence.

[0028] The text sequence is parsed using a predefined quintuple extraction template to extract a standardized set of quintuples containing the responsible party, action instructions, operation object, deadline, and constraints as the initial task list.

[0029] Preferably, the step of performing a multi-dimensional redundancy check on the initial task list based on a joint determination of the responsible party, time window, and semantic content, and calculating an execution damping coefficient based on a document citation graph to control the task status, and outputting the final task list includes:

[0030] Calculate the responsible party, deadline, and action instruction semantic vector for each task in the initial task list;

[0031] The calculated responsible party, deadline, and action instruction semantic vector are compared with the data in the existing task library to intercept tasks with the same responsible party, deadline within a preset overlapping window, and action instruction semantic similarity exceeding the duplication judgment threshold.

[0032] Analyze the explicit reference relationships in the official document text and construct the document reference graph containing document nodes and directed reference edges;

[0033] Based on the document citation graph, query the set of predecessor nodes for the current task, and calculate the execution damping coefficient according to the execution status of the predecessor nodes in the set of predecessor nodes;

[0034] If the execution damping coefficient meets the suspension condition, the current task is marked as suspended; if the execution damping coefficient does not meet the suspension condition, the current task is marked as active.

[0035] Preferably, the step of collecting heterogeneous operation logs for the final task list, calculating the granularity entropy change index, and using the granularity entropy change index to update the granularity mapping matrix includes:

[0036] Capture the splitting, merging, rejection, and retention operations performed by the user on the final task list, and generate the structured heterogeneous operation log;

[0037] The frequency of various operations within a specific time window is statistically analyzed, and the granular entropy change index, which measures the uncertainty of user operation distribution, is calculated based on the principle of information entropy.

[0038] Determine the dominant operation type and calculate the confidence decay factor and direction adjustment coefficient in conjunction with the granularity entropy change index;

[0039] The weight parameters in the granularity mapping matrix are corrected using a momentum-based gradient descent algorithm, based on the confidence decay factor and the direction adjustment coefficient.

[0040] Preferably, if the local semantic density index is higher than the density determination threshold, then selecting the differential extraction mode includes:

[0041] The document corresponding to the historical vector with the highest similarity score is selected from the candidate vector set as the benchmark reference document;

[0042] The string difference algorithm is used to compare the current document with the benchmark reference document to extract the changed text set and the newly added text set;

[0043] A prompt word is constructed that includes the changed text set and the newly added text set. The instruction model extracts task elements only for the differences in content.

[0044] Preferably, the granularity mapping matrix is ​​constructed in the following ways:

[0045] Analyze the punctuation and conjunction usage habits in the historical task data of the responsible department;

[0046] If the responsible department tends to break down the data into finer-grained parts, then positive weight values ​​are assigned to the index positions corresponding to semicolons, newlines, and numeric list characters in the granularity mapping matrix.

[0047] If the responsible department tends to engage in coarse-grained merging, then a positive weight value is assigned to the index position corresponding to the conjunction in the granularity mapping matrix, or a negative weight value is assigned to the index position corresponding to the semicolon and newline character.

[0048] Preferably, calculating the execution damping coefficient based on the execution state of the predecessor nodes in the predecessor node set includes:

[0049] Identify the keyword references between the current task and the predecessor node;

[0050] If the cited context keyword contains a strongly dependent word, then the first-level weight coefficient is set; if the cited context keyword contains a weakly dependent word, then the second-level weight coefficient is set.

[0051] Obtain the current status of each of the preceding nodes. If the current status is completed or archived, it is mapped to a completed value; otherwise, it is mapped to an incomplete value.

[0052] The execution damping coefficient is obtained by weighting the state mapping values ​​of each predecessor node using the weighting coefficient.

[0053] Preferably, the step of using a momentum-based gradient descent algorithm to correct the weight parameters in the granularity mapping matrix according to the confidence decay factor and the direction adjustment coefficient includes:

[0054] Compare the difference between the frequency of split operations and the frequency of merge operations;

[0055] If the frequency of the splitting operation is higher than the frequency of the merging operation and the difference exceeds the significance threshold, then a positive direction adjustment coefficient is set to increase the weight of the position corresponding to the splitting granularity indicator vector.

[0056] If the frequency of the merging operation is higher than the frequency of the splitting operation and the difference exceeds the significance threshold, then a negative direction adjustment coefficient is set to reduce the weight of the position corresponding to the splitting granularity indicator vector;

[0057] The confidence decay factor is calculated using the granularity entropy change index. When the granularity entropy change index increases, the update magnitude of the granularity mapping matrix is ​​suppressed by the confidence decay factor.

[0058] This invention provides an optimization method for a document workflow management system. It has the following beneficial effects:

[0059] 1. This invention utilizes a heterogeneous feedback closed loop to construct a dynamically updated granular mapping matrix, directly applying bias intervention to the logarithmic probability of the model decoding layer. This allows the system to automatically fine-tune the generation weights of segmentation symbols and conjunctions based on changes in the information entropy of user operations, achieving an adaptive approximation of task generation granularity from coarse to precise. Compared to traditional solutions that rely solely on static prompts or fixed parameter fine-tuning, this solution addresses the problem of models failing to continuously adapt to the evolving business habits of specific departments and the difficulty in resolving model drift, ensuring that the generated results consistently align with frontline operational preferences.

[0060] 2. This invention designs a policy-adaptive routing mechanism based on local semantic density. For high-frequency routine documents, it automatically switches to differential extraction mode, inputting only text fragments that differ from the benchmark document into the model. This compresses the length of the input context and focuses on core change information. Compared with the existing technology that treats all types of documents equally for full parsing, this solution effectively avoids noise interference and non-factual generation risks introduced by long contexts, and significantly improves the computing power utilization efficiency and key information extraction accuracy in high repetition business scenarios.

[0061] 3. This invention introduces execution damping state control logic based on document citation graphs. By calculating the completion status of predecessor nodes, it quantifies the execution resistance of the current task, ensuring that the generated task instructions strictly meet the pre-dependency conditions in terms of business timing. Unlike conventional methods that only extract text semantics from a single document, this solution overcomes the model's lack of global state awareness of the business chain, eliminates logical deadlocks or unexecutable task issuance due to missing pre-conditions, and ensures the compliance and continuity of business flow. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the system architecture and process of the document workflow management system optimization method according to an embodiment of the present invention;

[0063] Figure 2 This is a logical flowchart of a file workflow management system optimization method according to an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of the execution flow of the multidimensional constraint verification module according to an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the execution flow of the heterogeneous feedback closed-loop module according to an embodiment of the present invention;

[0066] Figure 5 This is a comparison chart of the probability distribution of token prediction before and after granular intervention according to an embodiment of the present invention.

[0067] in:

[0068] 10. Data collection and vectorization module; 20. Policy adaptive routing module; 30. Task granularity adaptive generation module; 40. Multidimensional constraint verification module; 50. Heterogeneous feedback closed loop module. Detailed Implementation

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Reference Figure 1 The document workflow management system optimization method of this invention runs on a computing device, which establishes communication connections with the upstream office automation system and the downstream lean management system through data interfaces.

[0071] The system architecture includes a data collection and vectorization module 10, a policy adaptive routing module 20, a task granularity adaptive generation module 30, a multi-dimensional constraint verification module 40, and a heterogeneous feedback closed-loop module 50. The modules interact with each other via an internal bus or memory sharing mechanism, collaboratively completing the transformation from unstructured documents to structured tasks and the adaptive updating of system parameters.

[0072] The data collection and vectorization module 10 is used to establish a mapping relationship between unstructured official documents and a high-dimensional semantic space. The module periodically pulls document data streams from the upstream office automation system via a preset interface protocol, or receives document files in response to user upload commands. The module performs optical character recognition and layout cleaning operations on the acquired raw document files, removing headers, footers, and format control characters, and extracting the plain text sequence. The module then uses a pre-trained deep neural network model to encode the plain text sequence and, combined with a named entity density weighting algorithm, calculates and generates a global semantic vector for the document, while simultaneously extracting the issuing unit, document number, and publication time as metadata indexes.

[0073] The policy-adaptive routing module 20 is configured downstream of the data collection and vectorization module 10, and is used to dynamically decide the processing mode based on the vector space distribution characteristics. The policy-adaptive routing module 20 receives the global semantic vector output by the data collection and vectorization module 10, and retrieves neighbor vectors within a preset radius from the historical document vector database. The policy-adaptive routing module 20 calculates the local semantic density index of the current document based on the retrieval results. The policy-adaptive routing module 20 compares the calculated local semantic density index with a preset threshold. When the local semantic density index is higher than the preset threshold, the policy-adaptive routing module 20 selects the differential extraction mode; when the local semantic density index is lower than or equal to the preset threshold, the policy-adaptive routing module 20 selects the full generation mode.

[0074] The task granularity adaptive generation module 30 is connected to the policy adaptive routing module 20 and is used to convert unstructured text into structured task data that meets specific granularity requirements. The task granularity adaptive generation module 30 loads the prompt word template and corresponding decoding strategy determined by the policy adaptive routing module 20. The task granularity adaptive generation module 30 reads the granularity mapping matrix parameters of the target responsible department. During the large language model decoding process, the task granularity adaptive generation module 30 uses the granularity mapping matrix parameters to linearly correct the logarithmic probability of the model output layer, intervening in the character generation probability representing segmentation or merging semantics. The task granularity adaptive generation module 30 outputs a standardized task quintuple set containing the responsible subject, action, object, deadline, and constraints.

[0075] The multi-dimensional constraint verification module 40 is configured at the output of the task granularity adaptive generation module 30 and is used to verify the uniqueness and logical feasibility of task data. The multi-dimensional constraint verification module 40 receives a set of task quintuples generated by the task granularity adaptive generation module 30. The multi-dimensional constraint verification module 40 calculates the spatiotemporal feature fingerprint of the task quintuple set and searches it in the pending task database, intercepting duplicate tasks that meet the spatiotemporal collision conditions. The multi-dimensional constraint verification module 40 parses the explicit citation relationships between documents and constructs a citation graph. Based on the citation graph, the multi-dimensional constraint verification module 40 queries the execution status of preceding related tasks and calculates the execution damping coefficient of the current task. When the execution damping coefficient meets the suspension condition, the multi-dimensional constraint verification module 40 marks the current task as suspended.

[0076] The heterogeneous feedback closed-loop module 50 connects to both the downstream lean management system and the task granularity adaptive generation module 30, and is used to construct a parameter evolution closed loop. The heterogeneous feedback closed-loop module 50 collects user operation logs from the downstream lean management system, including splitting, merging, rejecting, and accepting operations. Based on the operation logs, the heterogeneous feedback closed-loop module 50 statistically analyzes the operation probability distribution within a specific time window and calculates the granularity entropy change index. According to the granularity entropy change index and the dominant operation type, the heterogeneous feedback closed-loop module 50 uses a gradient update algorithm to correct the granularity mapping matrix parameters in the task granularity adaptive generation module 30, completing the system iteration.

[0077] Reference Figure 2 This invention provides a method for optimizing a document workflow management system, comprising the following steps:

[0078] S100, the data collection and vectorization module 10 performs data access and representation; receives and cleans official document data, performs sliding block division and weighted encoding on the text sequence, generates multi-dimensional semantic vectors and constructs metadata index;

[0079] S200, the policy adaptive routing module 20 executes processing policy decisions; it detects the local semantic density of the current document vector and dynamically switches the subsequent decoding path between differential extraction mode and full generation mode accordingly;

[0080] S300, the task granularity adaptive generation module 30 performs structured task generation; loads the granularity mapping matrix of the target department, intervenes in the segmentation granularity by superimposing bias weights on the model output layer, and generates a standardized initial task list;

[0081] S400, the multi-dimensional constraint verification module 40 performs task compliance verification; it eliminates duplicate tasks based on spatiotemporal feature fingerprints, and calculates the execution damping coefficient based on the document citation map, suspending tasks whose preconditions have not been completed;

[0082] S500, the heterogeneous feedback closed-loop module 50 performs adaptive parameter updates; it collects downstream user operation logs and calculates granular entropy change indicators, and uses gradient algorithms to update the granular mapping matrix in reverse, thereby realizing the iterative evolution of the generation strategy.

[0083] In step S100, the data collection and vectorization module 10 performs the data access and multi-granularity vector representation process. In this embodiment, considering that general large language models or encoders usually have input sequence length limitations (e.g., 512 or 1024 tokens), directly truncating long official documents would lead to the loss of semantics in the latter half, while a simple average pooling algorithm would cause key business instruction information to be diluted by a large amount of background text. Therefore, this embodiment adopts a sliding window segmentation combined with entity density weighting strategy to highlight key business features while preserving local semantic coherence. The specific implementation process includes:

[0084] S101, the data collection and vectorization module 10 performs standardized cleaning of multi-source heterogeneous data. The system reads document data streams in real time from upstream office automation systems through a pre-built data interface protocol, or receives OFD and PDF format files in response to external input commands. For format files, the data collection and vectorization module 10 calls an optical character recognition engine to perform image-to-text conversion and obtain the original character sequences. During this process, the data collection and vectorization module 10 performs page layout analysis, identifying and removing headers, footers, page numbers, and semantically meaningless layout control characters based on page coordinate information. Specifically, the filtering logic involves obtaining the bounding box coordinates of the characters. Compare it with the overall page margin threshold. The comparison is performed; the threshold for the text area is set according to the national standard format of official documents (such as GB / T 9704), for example, the page margin threshold is set to 10% to 15% of the page width; when the character coordinates fall into the page margin area defined by the threshold, they are judged as non-text noise and are removed, thus obtaining a pure text sequence containing only the semantic content of the main text.

[0085] S102, the data collection and vectorization module 10 performs sliding window segmentation on the plain text sequence. To ensure that sentences spanning multiple segments are preserved intact and to avoid comprehension gaps caused by segmentation points located in the middle of semantic units, the data collection and vectorization module 10 uses an overlapping sliding window mechanism to slice the text. In this embodiment, a baseline window length is set. and sliding step size .in, The value range is set to 256 to 512 tokens to adapt to the input dimensions of mainstream encoders; The value range is set to to This overlapping region ensures the continuity of the context. After segmentation, the original text sequence is transformed into an ordered set of text fragments.

[0086] S103, the data collection and vectorization module 10 constructs a weighted global semantic vector based on entity density. The data collection and vectorization module 10 loads a pre-trained encoder model (such as BERT or RoBERTa) to independently encode each segment in the text fragment set, generating a corresponding segment embedding vector. To address the issue of insufficient weighting of key business information (such as specific rectification deadlines and responsible departments) in long texts, the data collection and vectorization module 10 introduces a named entity recognition algorithm to identify the number of key entities such as "responsible entity," "time node," and "action instruction" in each segment, and calculates the weight of each segment accordingly. The formula for calculating the global semantic vector of the official document is as follows:

[0087] ;

[0088] Normalization factor The calculation formula is as follows:

[0089] ;

[0090] in, This represents the synthesized global semantic vector of the official document; Indicates the total number of segments into which the document has been divided; For segment number; Indicates the first A text fragment; Indicates the encoder model for the th The high-dimensional feature vectors output by each segment; Indicates the first The density of key entities within each segment is calculated as follows: This is the ratio of the total number of named entities identified within the fragment to the total number of tokens in the fragment; This is a preset entity weighting coefficient, with a value range of [0.5, 2.0]. It can be increased when the business scenario has high sensitivity to details. value; This is a normalization factor used to prevent numerical drift in the synthesized vector. Using this formula, text fragments containing specific task instructions or deadlines will occupy a larger proportion of the global vector.

[0091] S104, the data collection and vectorization module 10 extracts metadata and constructs a hybrid index. The data collection and vectorization module 10 uses a combination of regular expression matching and layout structure analysis to extract the full name of the issuing unit, the unique font size of the document, and the publication timestamp from the first page of the document. The data collection and vectorization module 10 then combines the above structured metadata with the global semantic vector generated in step S103. The data is stored in a vector database. Metadata is used for hard rule filtering in subsequent stages, while global semantic vectors are used for similarity calculation and density detection of soft rules. Together, they constitute a complete data index of the document object.

[0092] In step S200, the policy adaptive routing module 20 executes a processing strategy decision based on local semantic density. In this embodiment, based on statistical analysis of a large amount of official document circulation data, it was found that official document business exhibits obvious uneven data distribution characteristics: that is, most routine documents (such as monthly security reports and quarterly reports) are highly similar in text structure and main content, with only key parameters differing; while a small number of sudden documents have entirely new semantic structures. If all documents are subjected to full-volume deep parsing indiscriminately, it will not only increase the computational resource overhead, but also easily introduce irrelevant noise due to long contexts. Therefore, this embodiment distinguishes between "routine" and "sudden" business by detecting the distribution density of the vector space, and then matches differentiated processing paths. The specific implementation process includes:

[0093] S201, the policy-adaptive routing module 20 performs spatial probing of the historical document vector database. The policy-adaptive routing module 20 maintains a historical document vector database. This library stores the global semantic vectors of all documents previously processed by the system. When the global semantic vector of the current document output in step S100 is received... At that time, the policy adaptive routing module 20... To query a vector, in the vector library Nearest neighbor retrieval is performed within this process. This process uses cosine similarity as the distance metric to calculate... With existing vectors in the library The similarity score is used to improve retrieval efficiency. This embodiment employs an approximate nearest neighbor search algorithm based on Hierarchical Navigation Small World (HNSW) or Inverted Index (IVF), and then filters out similarity scores higher than a preset similarity lower limit. The set of candidate vectors. Wherein, The value range is set to [0.85, 0.95]. This threshold is determined based on the cluster compactness of the vector space and is used to ensure that the selected neighbor samples belong to the same category in terms of business type.

[0094] S202, the policy-adaptive routing module 20 calculates the local semantic density index of the current document. To quantify the frequency of the current document in the business space, the policy-adaptive routing module 20 calculates the local semantic density based on the retrieval results of S201. This index characterizes the density of historical samples within a unit hypersphere neighborhood defined by the aforementioned similarity lower limit. The density calculation formula is as follows:

[0095] ;

[0096] in, This represents the local semantic density value, which ranges from [0,1]. This represents the total number of historical vector samples currently used as a reference benchmark, or the number of samples within the most recent sliding time window (such as the last 30 days). Using sliding window counting can avoid interference from outdated historical data. Represents the first in the history vector library One sample vector; The cosine similarity function between the current vector and historical vectors; The similarity threshold set for the aforementioned steps; This is an indicator function; it takes a value of 1 when the condition within the parentheses is true (i.e., the similarity is greater than the threshold), and a value of 0 otherwise. Through this formula, the system can convert the business familiarity with official documents into a calculable numerical indicator.

[0097] S203, the policy-adaptive routing module 20 executes the dual-mode routing switching logic. The policy-adaptive routing module 20 presets a density determination threshold. This threshold is determined based on the system's historical throughput and service repetition rate statistics, and typically ranges from [0.05, 0.15]. The policy adaptive routing module 20 will calculate the threshold... and A comparison is performed, and the routing function is used based on the comparison result. Determine the subsequent task generation strategy. The routing function expression is as follows:

[0098] ;

[0099] in, This indicates the differential extraction mode. When the density value is greater than a threshold, it indicates that the current document belongs to a high-frequency, routine, standardized document, and the system calls this mode. In differential extraction mode, the policy adaptive routing module 20 first selects the document corresponding to the historical vector with the highest similarity score from the candidate vector set filtered by S201 as the "baseline reference document". Subsequently, the policy adaptive routing module 20 uses a string difference algorithm (such as the Myers difference algorithm) to compare the current document with the baseline reference document and extract the "changed text set" and "new text set". The policy adaptive routing module 20 constructs lightweight prompt words that only contain the differing content. The instruction large language model only updates or extracts task elements for the information in the above-mentioned changed set, while ignoring the unchanged templated text. This mode can significantly reduce the number of input tokens and reduce the risk of non-factual generation caused by excessively long context in the model. This indicates a full-volume generation mode. When the density value is less than or equal to the threshold, it indicates that the current document belongs to a low-frequency, sudden, or entirely new business type, with low historical reference value. The system invokes this mode, and the policy adaptive routing module 20 retains the entire plain text sequence extracted in step S100, and inputs it into the large language model along with a complete task extraction prompt word template. The instruction model then performs deep semantic parsing and task element extraction on the entire text. Through this dual-modal switching, the system achieves adaptive optimization of computational cost and extraction accuracy when processing massive amounts of documents.

[0100] In step S300, the task granularity adaptive generation module 30 performs the structured task generation process. In this embodiment, considering that the text generation of a large language model is essentially a probability-based word-by-word prediction process, the model output layer generates a non-normalized numerical vector (Logits) covering all tokens in the vocabulary before Softmax normalization. The magnitude of this value directly determines the probability of the corresponding token being sampled. Addressing the issue of significant differences in task granularity among different business departments (for example, the engineering department tends to use semicolons to split compound instructions into independent tasks, while the administrative department tends to use conjunctions to retain them as a whole task), this embodiment introduces an external bias signal into the Logits layer to directly intervene in the model's prediction probability of "splitters" or "mergers," thereby forcing the generated results to conform to a preset granularity standard without changing the model weights. The specific implementation process includes:

[0101] S301, the task granularity adaptive generation module 30 constructs or loads the granularity mapping matrix. The task granularity adaptive generation module 30 first identifies the responsible department to which the currently processed document belongs (e.g., the metadata extracted via S104), and retrieves the corresponding granularity mapping matrix from the system configuration library. The data structure of this matrix is ​​designed as a sparse vector with the same dimension as the vocabulary of a large language model. Specifically, let the size of the model vocabulary be... ,but For length is A one-dimensional array. In this embodiment, the task granularity adaptive generation module 30 analyzes the punctuation usage habits of the department's historical task data. If the department tends to use fine-grained splitting, then the Token index position corresponding to the "segmentation semantics" (such as semicolon ";", newline character "") is used in the array. Positive weight values ​​are assigned to the numeric list symbol "1."; if the department tends to use coarse-grained merging, negative weight values ​​are assigned to the above positions, or positive weight values ​​are assigned to the token index positions of "merging semantics" (such as the conjunctions "and" and "as well as"). All other general vocabulary positions unrelated to granularity are set to zero. The specific weight values ​​are calculated by normalizing the frequency of occurrence of the corresponding symbols in the department's historical data, and the values ​​are usually in the range of [-1.0, 1.0].

[0102] S302, the task-granularity adaptive generation module 30 performs linear correction on the Logits layer. At each time step of the autoregressive decoding in the large language model, the original output layer of the model generates an unnormalized log probability vector (Logits), denoted as... At this point, the task granularity adaptive generation module 30 does not directly... Instead of feeding it into the Softmax layer, a granular mapping matrix is ​​introduced. A linear bias is applied to it. The correction formula is as follows:

[0103] ;

[0104] in, This represents the corrected log probability vector, used for subsequent sampling selection; This represents the original log probability vector predicted by the large language model based on the current context, with dimension . ; This is the granularity mapping matrix loaded in step S301, with the same dimension. ; This is the granularity intervention coefficient, used to control the intensity of the intervention, and its value range is set to [1.0, 5.0]. When When set to a lower value (e.g., 1.0), only the generation tendency is fine-tuned; when When set to a higher value (such as 5.0), it can be significantly amplified. The value of the specified symbol in the Softmax algorithm makes the probability distribution after the calculation exhibit polarization characteristics, thereby forcing the model to select or avoid specific splitting symbols and achieving hard constraints on the number and length of tasks.

[0105] S303, the task granularity adaptive generation module 30 generates a standard task set based on the redistribution probability. The task granularity adaptive generation module 30 will then modify the vector... The Softmax activation function is input to calculate the final token generation probability distribution. Since the probabilities of splitters or mergers have been adjusted in S302, the model tends to generate text sequences that conform to a preset granularity during decoding. The task granularity adaptive generation module 30 performs regular expression parsing on the generated text sequence, using a predefined quintuple extraction template to extract a standardized set of task quintuples. These quintuples specifically include: responsible party (the specific department or position performing the task), action instructions (specific verb descriptions, such as "inspect" or "write"), operation object (the entity affected by the action, such as "equipment" or "report"), deadline (a specific time point or period), and constraints (standards or criteria to be followed when performing the task). If a certain element is missing in the generated text, the task granularity adaptive generation module 30 marks it as "to be completed," ultimately outputting a task list that conforms to the structured definition.

[0106] Reference Figure 3 In step S400, the multi-dimensional constraint verification module 40 performs post-processing verification and status control of task generation. In this embodiment, considering that task instructions generated by large language models are often based on semantic inference from static text and lack awareness of the existing business system's operational status, they are prone to duplicate task assignments or generating unexecutable tasks without preconditions. To address these issues, this embodiment introduces a redundancy filtering mechanism based on spatiotemporal attributes and an execution damping mechanism based on causal logic to ensure that tasks sent to business terminals are physically unique and logically meet execution timing requirements. The specific implementation process includes:

[0107] S401, the multi-dimensional constraint verification module 40 performs multi-dimensional redundancy verification. The multi-dimensional constraint verification module 40 accesses the standard task quintuple set generated in step S300 and reads the existing task library in the system that is in a "pending" or "in progress" state. To accurately identify duplicate tasks, the multi-dimensional constraint verification module 40 does not rely solely on text similarity but employs a three-dimensional joint judgment logic that includes the responsible party, time window, and semantic content. The multi-dimensional constraint verification module 40 defines a multi-dimensional collision detection function. For newly generated tasks With existing tasks The comparison and calculation formula are as follows:

[0108] ;

[0109] in, The collision detection result is set to 1, which indicates that the task is a duplicate, and 0, which indicates that the task is a valid new task. and These represent the responsible parties (departments or positions) for new tasks and existing tasks, respectively. This is an indicator function that returns 1 if the condition is met, and 0 otherwise. and These represent the deadline timestamps for the two tasks, respectively. The preset time overlap window threshold is set based on the average business processing cycle (e.g., 72 hours). That is, if two tasks are within the same business cycle and other dimensions are consistent, they are considered duplicate instructions. and These represent the semantic vectors obtained by the encoder mapping of the "action instruction" text in the new task and the existing task, respectively. This represents the cosine similarity between the two vectors mentioned above. The threshold for determining repetition ranges from [0.90, 0.98]. A higher threshold indicates a lower tolerance for semantic differences. When the calculation result is 1, the system automatically discards it. Alternatively, it can be marked as "already exist" to avoid sending duplicate orders to users.

[0110] S402, the multi-dimensional constraint verification module 40 constructs a document citation graph. In document circulation, "explicit citation" relationships often exist (e.g., newly issued documents explicitly state "implementing the requirements of document [202X] XX"). This citation relationship implies pre-dependent logic between tasks. The multi-dimensional constraint verification module 40 is configured with specific regular expression extraction rules (e.g., matching standard document number formats like "

[2024] No. 15") to scan the entire document and extract all historical document identifiers cited. The multi-dimensional constraint verification module 40 constructs a directed citation graph using documents or their contained tasks as nodes and citation relationships as directed edges. .in, Represents a set of official documents or task nodes. This represents the set of directed edges pointing from the referenced document (parent node) to the current document (child node). Through this graph, the multi-dimensional constraint verification module 40 can trace all the preceding parent nodes of the current task and determine the topological position of the task in the business chain.

[0111] S403, the multi-dimensional constraint verification module 40 calculates the execution damping coefficient and controls the task state. To prevent subsequent tasks from starting before the preceding tasks are completed, the multi-dimensional constraint verification module 40 uses a referenced directed graph. Query current task The set of all direct predecessor nodes If set If the set is empty, the damping coefficient is directly determined to be 0, and the task state is unhindered; if the set is not empty, the multidimensional constraint verification module 40 obtains the current execution state of these predecessor nodes and calculates the execution damping coefficient. The calculation formula is as follows:

[0112] ;

[0113] in, This represents the execution damping coefficient, with a value range of [0,1]. The larger the value, the greater the execution resistance, meaning the more severe the lack of preconditions. This refers to the set of all preceding dependent tasks of the current task. Represents the first element in the set. One prerequisite task; For the first The weight coefficient of each prerequisite task is determined by the system based on the keywords in the reference context: if the reference context contains strong dependency words such as "based on" or "according to", it is set to 1.0; if it contains weak dependency words such as "refer to" or "reference", it is set to 0.5. For the first The status of each preceding task; For the state mapping function, when The value is 1 if the condition is "Completed" or "Archived", otherwise it is 0. After the calculation is completed, the multidimensional constraint verification module 40 will... With the preset damping threshold (In this embodiment, the value is set to 0.1) for comparison. If... This indicates that the prerequisites were not fully satisfied, and the multi-dimensional constraint verification module 40 will check the current task. The status is forcibly set to "suspended," and the reason for the suspension and a list of unfinished prerequisite tasks are displayed to the user; if If so, the task status will be set to "active", allowing it to be pushed to the user's workbench.

[0114] Reference Figure 4 In step S500, the heterogeneous feedback closed-loop module 50 performs iterative updates to the system parameters. In this embodiment, since the document processing standards in business scenarios are not static, the parameters of the pre-trained large language model or the initially set rule matrix often cannot continuously adapt to the gradually evolving business habits of a specific department (for example, a department gradually tends to more fine-grained task breakdown). To solve the model drift problem caused by static parameters, this embodiment constructs a reverse feedback channel from the business terminal to the generation backend. This channel calculates the information entropy of user interaction behavior, judges the clarity of user needs, and then fine-tunes the granular mapping matrix in a gradient descent manner to achieve dynamic approximation of business preferences by the system. The specific implementation process includes:

[0115] S501, the heterogeneous feedback closed-loop module 50 collects heterogeneous operation logs. The heterogeneous feedback closed-loop module 50 deploys a data collection interface at the front end of the downstream lean management system to capture user intervention operations on automatically generated tasks in real time. The heterogeneous feedback closed-loop module 50 focuses on collecting three types of heterogeneous operation events: splitting operations (cutting a long task into multiple sub-tasks), merging operations (combining multiple fragmented tasks into a single overall task), and rejection operations (directly deleting or rewriting the task). The heterogeneous feedback closed-loop module 50 encapsulates these operation events into structured log records. Each record includes: the original generated task text, the user-modified task text, the operation type identifier, and the corresponding timestamp. The system sets a statistical period (such as weekly or monthly), and at the end of the period, aggregates and generates an operation distribution dataset.

[0116] S502, Heterogeneous Feedback Closed-Loop Module 50 quantifies the granularity entropy change index. To assess the deviation between the current model's generated granularity and the user's actual expectations, the heterogeneous feedback closed-loop module 50 calculates the granularity entropy change index. This metric, drawing on the principle of information entropy, measures the uncertainty of user action distribution. Physically, when users consistently perform "split" operations on the vast majority of tasks, or consistently execute "split" operations, the entropy value is low, indicating clear user needs. Conversely, when user actions are randomly distributed between splitting, merging, and rejecting, the entropy value increases, indicating divergent or noisy user needs. The calculation formula is as follows:

[0117] ;

[0118] in, This represents the granularity entropy change index; the higher the value, the lower the consistency of user operations. This represents a set of operation types, specifically including {keep, split, merge, reject}. In this embodiment, the number of elements in the set is... ; For the set of Types of operations; This indicates that within the statistical period, the [number]th [year]... The frequency of a particular operation type is the ratio of the number of operations of that type to the total number of operations. This formula allows the system to quantify the dispersion of feedback from business departments to the current task generation strategy from a macroscopic perspective.

[0119] S503, the heterogeneous feedback closed-loop module 50 performs gradient-based matrix updates. To achieve system parameter iteration, the heterogeneous feedback closed-loop module 50 uses the calculated entropy change index and dominant operation type to update the granularity mapping matrix defined in step S301. Perform a reverse update. This embodiment uses a variant of the momentum-based gradient descent algorithm to correct the matrix weights, and the update formula is as follows:

[0120] ;

[0121] in, and These represent the granularity mapping matrices after the update and before the update, respectively; The learning rate is used to control the step size of parameter updates. In this embodiment, it is set to [0.01, 0.05] to avoid drastic parameter oscillations. The current entropy value calculated in step S502; The maximum possible entropy value, i.e., the entropy value when the probabilities of the four operations are equal, is calculated as follows: ;item As a confidence decay factor, its physical function is: when user operations are extremely chaotic ( near When this value approaches 0, the system automatically suppresses updates to avoid introducing noise; when the user's operational tendency is clear ( When the value is low, the system increases the update increment. The system compares the "split" frequency, which is the direction adjustment coefficient. With the frequency of "merging" The difference, if ( (Assuming a significance threshold of 0.1), then... ;like ,but ;otherwise ; This is a segmentation granularity indicator vector, which is only applied at token index positions corresponding to tokens with strong segmentation semantics (such as semicolons ";", periods ".", and newlines "). The value of "and numeric list symbol" is 1 in each position, and 0 in all other positions. The iterative update logic works as follows: when... When the value is positive, the weight of the splitter increases, and the task granularity generated by the model becomes finer; when... When the value is negative, the weight of the segmentation symbol decreases, and the model tends to generate long and complex sentences or use conjunctions to connect them, thereby achieving coarsening and merging of task granularity, so that the generation strategy can automatically adapt to the business preferences of the department.

[0122] To further clarify the collaborative working process of the technical solution of this invention, a specific work scenario example will be used below. Assume that the document workflow management system of this embodiment is applied to the administrative office environment of a large energy transportation company. The upstream office automation system issues an unstructured PDF document titled "Notice on Rectification of Safety Hazard Investigation of Traction Power Supply Equipment in Spring 2026".

[0123] The data collection and vectorization module 10 first receives the PDF file through an interface and extracts the text content using OCR technology. After cleaning the layout to remove the official document formatting and page numbers / footers, the data collection and vectorization module 10 inputs the plain text sequence into a sliding window algorithm. During this process, the data collection and vectorization module 10 identifies high-density entity information in the document, such as "power supply section maintenance workshop" (responsible entity), "substation main transformer" (operation object), and "before April 30th" (deadline). Based on this entity density, the data collection and vectorization module 10 uses a weighted algorithm to generate a global semantic vector for the document. This vector mathematically represents that the document belongs to the business semantic category of "equipment operation and maintenance."

[0124] Subsequently, the policy-adaptive routing module 20 receives the global semantic vector and projects it into the historical document vector space. The policy-adaptive routing module 20 calculates and finds that a large number of historically similar vectors exist around this vector, and the local semantic density index... The calculated result is 0.28, which is higher than the preset threshold of 0.15. This indicates that such seasonal inspection notices are high-frequency routine documents within the company. Therefore, the strategy adaptive routing module 20 automatically activates the "differential extraction mode," retrieves historical benchmark documents of the same type for comparison, and extracts only the differentiated segments related to "new infrared temperature measurement projects" as input to the backend, instead of inputting the entire lengthy document. This significantly reduces computational overhead and focuses on the key points of generation.

[0125] Next, the task granularity adaptive generation module 30 intervenes. This module identifies the responsible party for the current task as the "power supply section maintenance workshop." System configuration data shows that the workshop's frontline work teams are accustomed to executing short, checklist-style instructions, which are printed as work tickets for item-by-item checking, rather than reading long narrative texts. Therefore, the task granularity adaptive generation module 30 loads the corresponding granularity mapping matrix for this workshop. During text generation, although the large language model tends to generate complex long sentences like "check the main transformer oil level and verify the chromatographic analysis records" (containing the conjunction "and") based on contextual probabilities, the granularity mapping matrix assigns positive bias weights to semicolons ";" and newline characters at the Logits layer, while assigning a negative weight to the conjunction "and." Ultimately, the task granularity adaptive generation module 30 forces the model to output a fine-grained structured task list such as "1. Check the main transformer oil level; 2. Verify the chromatographic analysis records."

[0126] To visually demonstrate the impact of the Logits layer intervention on the generated results, such as Figure 5As shown, after the intervention of the granular mapping matrix, the generation probability of tokens representing segmentation semantics (such as semicolons and newlines) is significantly improved, and numerically significantly exceeds that of the previously higher probability conjunction tokens. This mathematically constrains the model to output discretized task instructions that conform to the operating habits of business personnel, and realizes physical-level control over task granularity.

[0127] After generating the initial task list, the multi-dimensional constraint verification module 40 immediately performs a compliance scan on the tasks. Assume that a task issued by the safety supervision department two days ago already exists in the system, with the responsible entity being the "power supply section maintenance workshop," the semantic similarity of the action instructions being as high as 0.98, and both targeting the same batch of equipment. The multi-dimensional constraint verification module 40 uses a multi-dimensional collision detection function... The newly generated task is determined to be a duplicate instruction and is automatically intercepted to avoid duplicate order assignment to the workshop. Simultaneously, the multi-dimensional constraint verification module 40 parses the document and finds that it references "

[2025] XX document". By constructing a reference graph, it discovers that the status of the preceding document is "completed," and calculates the execution damping coefficient. Since the value is 0, the multidimensional constraint verification module 40 sets the status of the remaining valid tasks to "active" and pushes it to the downstream lean management system.

[0128] Finally, when the workshop foreman received the task on the lean management system terminal, he found that one of the tasks, "Replacing Old Insulators," had an overly broad scope. In practice, it was necessary to distinguish between "rigid cross-span insulators" and "suspension insulators" and report them separately. Therefore, he manually performed a "split" operation on the system, breaking it down into two sub-tasks. The heterogeneous feedback closed-loop module 50 captured this "split" behavior log and calculated the operational entropy value of the department when handling such equipment tasks. The lower value indicates a very clear demand. Within the preset parameter update cycle, the heterogeneous feedback closed-loop module 50 automatically updates the granularity mapping matrix for "power supply section maintenance workshop" in the task granularity adaptive generation module 30 using the gradient descent algorithm, further increasing the weight of relevant delimiters. Therefore, when the system processes similar equipment documents for that workshop again, it will automatically generate finer-grained tasks, achieving adaptive evolution of the system.

[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing a document workflow management system, characterized in that, Includes the following steps: The system receives official document data and performs text cleaning. It performs sliding window block processing and weighted encoding on the plain text sequence and generates a global semantic vector of the official document by combining the named entity density weighted algorithm. Based on the local semantic density index of the global semantic vector in the historical document vector database, the subsequent decoding path is determined between the differential extraction mode and the full generation mode. The corresponding granularity mapping matrix is ​​loaded according to the determined decoding path. During the decoding process, the granularity mapping matrix is ​​used to linearly correct the original log probability vector of the model output layer to generate an initial task list. Perform multi-dimensional redundancy verification on the initial task list based on joint determination of responsible subject, time window and semantic content, and calculate the execution damping coefficient based on the document citation map to control the task status, and output the final task list; Collect heterogeneous operation logs for the final task list, calculate the granularity entropy change index, and use the granularity entropy change index to update the granularity mapping matrix in reverse.

2. The document workflow management system optimization method according to claim 1, characterized in that, The process of receiving official document data and performing text cleaning, including sliding window block processing and weighted encoding on the plain text sequence, and generating a global semantic vector for the official document using a named entity density weighted algorithm, includes: The system receives official document data streams or layout files using a pre-defined data interface protocol, performs layout cleaning operations to remove headers, footers, and format control characters, and extracts the plain text sequence. The plain text sequence is sliced ​​using an overlapping sliding window mechanism to obtain an ordered set of text fragments; Each segment in the text segment set is encoded using a pre-trained encoder model to identify the number of responsible entities, time nodes, and action instruction entities contained in each segment; The weight of each segment is calculated based on the number of identified entities, and the feature vectors of each segment are weighted and synthesized based on the weights to obtain the global semantic vector.

3. The document workflow management system optimization method according to claim 1, characterized in that, The determination of the subsequent decoding path between the differential extraction mode and the full generation mode based on the local semantic density index of the global semantic vector in the historical document vector database includes: The global semantic vector is used as the query vector to retrieve a set of candidate vectors within a preset similarity range from the historical document vector database; The number of neighbor samples that meet the lower similarity limit in the candidate vector set is counted, and the density of the distribution of the number of neighbor samples in the neighborhood of a unit hypersphere is calculated to obtain the local semantic density index. The local semantic density index is compared with a preset density determination threshold; If the local semantic density index is higher than the density determination threshold, then the differential extraction mode is selected; If the local semantic density index is lower than or equal to the density determination threshold, then the full generation mode is selected.

4. The document workflow management system optimization method according to claim 1, characterized in that, The step of loading the corresponding granularity mapping matrix according to the determined decoding path, and using the granularity mapping matrix to linearly correct the original log probability vector of the model output layer during the decoding process, and generating the initial task list includes: Identify the responsible department to which the current document belongs, and retrieve the granularity mapping matrix corresponding to the responsible department from the system configuration library. The granularity mapping matrix is ​​a sparse vector with the same dimension as the model vocabulary. During the time step of the model autoregressive decoding, the original log probability vector generated by the model is obtained; The original log probability vector is weighted and superimposed with the granularity mapping matrix to obtain the modified log probability vector. Calculate the token generation probability distribution based on the modified logarithmic probability vector, and generate a text sequence. The text sequence is parsed using a predefined quintuple extraction template to extract a standardized set of quintuples containing the responsible party, action instructions, operation object, deadline, and constraints as the initial task list.

5. The document workflow management system optimization method according to claim 1, characterized in that, The initial task list undergoes a multi-dimensional redundancy check based on a joint determination of the responsible party, time window, and semantic content. An execution damping coefficient is calculated based on the document citation graph to control the task status. The final task list output includes: Calculate the responsible party, deadline, and action instruction semantic vector for each task in the initial task list; The calculated responsible party, deadline, and action instruction semantic vector are compared with the data in the existing task library to intercept tasks with the same responsible party, deadline within a preset overlapping window, and action instruction semantic similarity exceeding the duplication judgment threshold. Analyze the explicit reference relationships in the official document text and construct the document reference graph containing document nodes and directed reference edges; Based on the document citation graph, query the set of predecessor nodes for the current task, and calculate the execution damping coefficient according to the execution status of the predecessor nodes in the set of predecessor nodes; If the execution damping coefficient meets the suspension condition, the current task is marked as suspended; if the execution damping coefficient does not meet the suspension condition, the current task is marked as active.

6. The document workflow management system optimization method according to claim 1, characterized in that, The step of collecting heterogeneous operation logs for the final task list, calculating granularity entropy change indicators, and using these indicators to inversely update the granularity mapping matrix includes: Capture the splitting, merging, rejection, and retention operations performed by the user on the final task list, and generate the structured heterogeneous operation log; The frequency of various operations occurring within a specific time window is statistically analyzed, and the granular entropy change index, which measures the uncertainty of user operation distribution, is calculated based on the principle of information entropy. Determine the dominant operation type and calculate the confidence decay factor and direction adjustment coefficient in conjunction with the granularity entropy change index; The weight parameters in the granularity mapping matrix are corrected using a momentum-based gradient descent algorithm, based on the confidence decay factor and the direction adjustment coefficient.

7. The document workflow management system optimization method according to claim 3, characterized in that, If the local semantic density index is higher than the density determination threshold, then selecting the differential extraction mode includes: The document corresponding to the historical vector with the highest similarity score is selected from the candidate vector set as the benchmark reference document; The string difference algorithm is used to compare the current document with the benchmark reference document to extract the changed text set and the newly added text set; A prompt word is constructed that includes the changed text set and the newly added text set. The instruction model extracts task elements only for the differences in content.

8. The document workflow management system optimization method according to claim 4, characterized in that, The granularity mapping matrix is ​​constructed in the following ways: Analyze the punctuation and conjunction usage habits in the historical task data of the responsible department; If the responsible department tends to break down the data into finer-grained parts, then positive weight values ​​are assigned to the index positions corresponding to semicolons, newlines, and numeric list characters in the granularity mapping matrix. If the responsible department tends to engage in coarse-grained merging, then a positive weight value is assigned to the index position corresponding to the conjunction in the granularity mapping matrix, or a negative weight value is assigned to the index position corresponding to the semicolon and newline character.

9. The document workflow management system optimization method according to claim 5, characterized in that, The calculation of the execution damping coefficient based on the execution status of the predecessor nodes in the predecessor node set includes: Identify the keyword references between the current task and the predecessor node; If the cited context keyword contains a strongly dependent word, then the first-level weight coefficient is set; if the cited context keyword contains a weakly dependent word, then the second-level weight coefficient is set. Obtain the current status of each of the preceding nodes. If the current status is completed or archived, it is mapped to a completed value; otherwise, it is mapped to an incomplete value. The execution damping coefficient is obtained by weighting the state mapping values ​​of each predecessor node using the weighting coefficient.

10. The document workflow management system optimization method according to claim 6, characterized in that, The step of using a momentum-based gradient descent algorithm to correct the weight parameters in the granularity mapping matrix according to the confidence decay factor and the direction adjustment coefficient includes: Compare the difference between the frequency of split operations and the frequency of merge operations; If the frequency of the splitting operation is higher than the frequency of the merging operation and the difference exceeds the significance threshold, then a positive direction adjustment coefficient is set to increase the weight of the position corresponding to the splitting granularity indicator vector. If the frequency of the merging operation is higher than the frequency of the splitting operation and the difference exceeds the significance threshold, then a negative direction adjustment coefficient is set to reduce the weight of the position corresponding to the splitting granularity indicator vector. The confidence decay factor is calculated using the granularity entropy change index. When the granularity entropy change index increases, the update magnitude of the granularity mapping matrix is ​​suppressed by the confidence decay factor.

Images