File approval method based on NLP

Abstract

The invention relates to the technical field of administrative management, in particular to an NLP-based file approval method, which comprises the following steps: inputting an approval file, carrying out word segmentation processing on the content of the approval file, carrying out similarity calculation on a word segmentation result and a preset service type dictionary, and generating a file service classification result. According to the method, the file service classification result is formed through word segmentation processing of the approval file content, so that the approval process can be automatically classified based on text features, and dependence on fixed rules is reduced. In combination with semantic vector conversion and keyword semantic association calculation, the semantic understanding ability of approval contents is enhanced, accurate matching of approval files and organization structure responsibilities is ensured, and approval errors caused by insufficient keyword matching are avoided. By calculating the matching value, association is established between the approval file and the department responsibilities and between the approval file and the post responsibilities, so that allocation of the approval tasks is more targeted, and task execution difficulty caused by mismatching of the responsibilities of approvers is avoided.

Description

Technical Field

[0001] The present invention relates to the field of administrative management technology, and in particular to a document approval method based on NLP. Background Art

[0002] Existing technologies rely on fixed rules to classify approval documents, which is difficult to adapt to text changes in different business scenarios. The classification method of approval documents lacks dynamic adjustment capabilities, which affects the adaptability of the approval process. The allocation of approval tasks mainly relies on manual experience or preset rules, and lacks a data-driven dynamic optimization mechanism, making it difficult to ensure the fairness of task allocation. Some approval personnel are in a state of task overload or under-task for a long time, affecting the overall approval efficiency. Therefore, improvements are needed. Summary of the invention

[0003] The purpose of the present invention is to solve the shortcomings of the prior art and propose a document approval method based on NLP.

[0004] In order to achieve the above object, the present invention adopts the following technical solution, a document approval method based on NLP, comprising the following steps: Input the approval document, perform word segmentation on the content of the approval document, calculate the similarity between the word segmentation result and the preset business type dictionary, and generate the file business classification result; based on the file business classification result, perform semantic vector conversion on the keywords in the approval document content, calculate the semantic association strength between the keywords, and generate the file semantic feature vector; Based on the document semantic feature vector, the responsibilities of each department and the job responsibilities of each person in the organizational structure are called, the matching value between the vector content and the responsibilities of the department and the person is calculated, and a responsibility matching result is generated; based on the responsibility matching result, the current number of tasks and task processing time of each approval person are counted, and the matching index of the approval person is calculated and generated; Based on the approval personnel matching index, a task allocation threshold is set, candidate approvers are screened according to the task allocation threshold, and a task processing capability index is generated; based on the task processing capability index and combined with the task saturation of the candidate approvers, an approval task allocation sequence is generated; Based on the approval task allocation sequence, the information of the first approver in the sequence is extracted, the approval document is allocated to the first approver, the allocation time and the document identifier are recorded, and an approval task allocation record is generated.

[0005] Preferably, the steps of obtaining the file service classification result are: After the approval document is input, it is segmented through natural language processing, the text in the document is semantically analyzed, and the sentences are separated into independent vocabulary units to form the segmentation results; According to the word segmentation result, the similarity between each word segmentation element and an entry in a preset business type dictionary is calculated to obtain a similarity result; Based on the similarity result, the similarity scores of each business type are compared, and the one with the highest similarity score is selected as the type identifier of the file to generate a file business classification result.

[0006] Preferably, the steps of obtaining the file semantic feature vector are: Based on the file business classification result, the approval file content is parsed, keywords in the approval file content are extracted, keywords are screened, redundant keywords, synonyms or interference items are removed, and keywords are normalized to obtain a keyword list; According to the keyword list, the semantic association strength between the keywords is calculated, and the calculation formula is: ; in, Represents the strength of semantic association between keywords. Representative keywords The frequency of occurrence in the text, and Respectively represent and The semantic vector value of keywords, and Respectively represent and The relative position deviation of keywords in the context window, and Represents the semantic dependency distance of keywords in the text, and Represents the dependency weight of keywords in different syntactic structures, and Represents the number of times a keyword appears in different text fragments; Based on the semantic association strength, a file semantic feature vector is constructed.

[0007] Preferably, the steps of obtaining the responsibility matching result are: Based on the document semantic feature vector, call the responsibilities of each department and the responsibilities of each person in the organizational structure, parse the responsibilities description text, extract the responsibilities keywords, construct the responsibilities keyword vector, and obtain the responsibilities keyword vector matrix; According to the responsibility keyword vector matrix, the matching value between the document semantic feature vector and the responsibility keyword vector is calculated, and the expression is: ; in, Represents the matching value, is the first part of the semantic feature vector of the file Dimensional components, is the first Dimensional components, and Represent the relative distances between document keywords and responsibility keywords in the semantic space, For the The frequency value of the matching words, and Represents the position index of the document keywords and responsibility keywords in the syntactic structure, , and Represents the matching similarity coefficient between each job description text, Represents the number of dimensions of the document semantic feature vector, Represents the number of matching words in the responsibility keyword vector, The total number of job description texts; Based on the matching value, a matching evaluation is performed on the job descriptions of each department and the job descriptions of personnel positions, and the responsibility item with the highest matching value is selected as the belonging department and responsible person to obtain a responsibility matching result.

[0008] Preferably, the steps for obtaining the approval personnel matching index are: Based on the responsibility matching results, the current number of tasks of each approver is counted, and the cumulative task processing time of each approver's processed tasks is calculated. At the same time, the list of pending tasks of the approver is extracted to obtain the task statistics information of the approver; According to the task statistics of the reviewers, the task processing mode of each reviewer is analyzed, the average number of tasks completed per unit time is calculated, and the current task saturation degree of each reviewer is counted to obtain the reviewer task load index; Based on the task load index of the approval personnel, the matching index of the approval personnel is calculated. The calculation formula is: ; in, Represents the approval personnel matching index, The task processing time of the approved personnel who have completed the task. The current number of tasks for the approver. Represents the number of tasks to be processed by the approver. Represents the average completion time of historical tasks of the approver. Represents the task processing efficiency of the approver on different task types, Represents the standard deviation of task processing time of approvers in different task categories. and Represents the stability coefficient of task completion of the approver in different time periods, Represents the number of task records of the approver. Represents the number of task categories for approvers.

[0009] Preferably, the steps for obtaining the task processing capability indicator are: Based on the matching index of the approver, a task allocation threshold is set, the matching indexes of all approvers are screened, and the approvers whose matching indexes are higher than the task allocation threshold are extracted to obtain a list of candidate approvers; According to the list of candidate reviewers, the task completion time of each candidate reviewer is calculated, and the calculation formula is: ; in, Represents the task completion time of the candidate approver, is the total time that candidate approvers have completed tasks in the past. Represents the estimated duration of the current pending task. and Indicates the task switching time of candidate approvers. The standard processing time of the candidate approver's historical tasks. and A tuning parameter representing the complexity of the task, , and Represents the task processing efficiency of candidate approvers in different time periods; Based on the task completion time, the task processing capability of the candidate approver is evaluated, and the task processing capability index is obtained by combining the task category, processing timeliness and execution record.

[0010] Preferably, the steps for obtaining the approval task allocation sequence are: Based on the task processing capability indicator, the task processing capabilities of all candidate approvers are extracted, the task saturation of the candidate approvers is called, and the current number of tasks and task load of each candidate approver are analyzed to obtain the task status information of the candidate approver; According to the task status information of the candidate approvers, matching analysis is performed on the task processing capability index and task saturation of the candidate approvers, the acceptable number of tasks for each candidate approver is analyzed, and the candidate approvers are prioritized based on the acceptable number of tasks to generate a task allocation priority sequence for the candidate approvers; Based on the candidate approver task allocation priority sequence, candidate approvers are selected in turn according to the execution requirements of the approval tasks, the approval tasks are allocated in order of priority, and the task allocation order is recorded to generate an approval task allocation sequence.

[0011] Preferably, the steps for obtaining the approval task allocation record are: Based on the approval task allocation sequence, extract the information of the first approver in the approval task allocation sequence, call the identity information and position information of the approver, and verify the approval task allocation sequence to generate the first allocation information of the approval task; According to the approval task first allocation information, the file identifier of the to-be-approved file is matched, the approval file content and allocation time data are called, and the approval file is bound to the first approver, the allocation of the approval file is recorded, and the approval task allocation binding information is generated; Based on the approval task allocation binding information, an approval task allocation record is established to store the approval document allocation time, document identification and first approver information, and the approval task management database is updated synchronously.

[0012] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention forms a file business classification result through word segmentation processing of the content of the approval document, so that the approval process can be automatically classified based on text features, reducing the dependence on fixed rules. Combined with semantic vector conversion and keyword semantic association calculation, the semantic understanding ability of the approval content is enhanced, ensuring the accurate matching of the approval document with the organizational structure responsibilities, and avoiding approval errors caused by insufficient keyword matching. By calculating the matching value, the approval document is associated with the departmental responsibilities and job responsibilities, so that the allocation of approval tasks is more targeted, and the approval personnel are prevented from having difficulties in task execution due to mismatched responsibilities. Combined with the number of tasks and task processing time of the approval personnel, the matching index of the approval personnel is comprehensively calculated, so that the selection of the approval personnel is not only based on the adaptation of responsibilities, but also the current task load can be considered, reducing the approval delay caused by the imbalance of workload. Set the task allocation threshold, screen candidate approvers, and calculate the task processing capacity index, so that the task allocation has more dynamic adjustment capabilities. Combined with the task saturation of the approval personnel, the approval task allocation sequence is constructed to ensure the fairness and stability of task allocation, and avoid long-term task backlog or insufficient tasks for some personnel. By extracting the information of the first approver in the approval task allocation sequence, the allocation and recording of approval tasks are completed, forming a complete approval task tracking system, making the circulation of approval documents more transparent, and effectively optimizing the scheduling and execution efficiency of approval tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0015] See also Figure 1 The present invention provides a technical solution, a document approval method based on NLP, comprising the following steps: Input the approval document, perform word segmentation on the content of the approval document, calculate the similarity between the word segmentation result and the preset business type dictionary, and generate the file business classification result; based on the file business classification result, perform semantic vector conversion on the keywords in the approval document content, calculate the semantic association strength between the keywords, and generate the file semantic feature vector; Based on the document semantic feature vector, call the responsibilities of each department and the job responsibilities of each person in the organizational structure, calculate the matching value between the vector content and the responsibilities of the department and the person, and generate the responsibility matching result; based on the responsibility matching result, count the current number of tasks and task processing time of each approval personnel, and calculate and generate the approval personnel matching index; Based on the matching index of the approver, set the task allocation threshold, screen the candidate approvers according to the task allocation threshold, and generate the task processing capability index; based on the task processing capability index and the task saturation of the candidate approvers, generate the approval task allocation sequence; Based on the approval task allocation sequence, extract the information of the first approver in the sequence, assign the approval document to the first approver, record the allocation time and document ID, and generate the approval task allocation record.

[0016] The steps to obtain the file business classification results are as follows: After the approval document is input, it is segmented through natural language processing, the text in the document is semantically analyzed, and the sentences are separated into independent vocabulary units to form the segmentation results; According to the word segmentation results, the similarity between each word segmentation element and the entries in the preset business type dictionary is calculated to obtain a similarity result; Based on the similarity results, the similarity scores of each business type are compared, and the one with the highest similarity score is selected as the type identifier of the file to generate the file business classification result.

[0017] Specifically, after obtaining the approval document, its text content is first read, and the character sequence is retrieved and parsed line by line for the complete text, and each fragment is segmented. The word form and word frequency in the text are compared according to the pre-aggregated industry general vocabulary and common stop word list. The stop word list here is manually screened from similar approval documents accumulated over the years, and is formed by collecting words that appear frequently but have no actual meaning. Words that appear once are also included in the statistical scope. The minimum frequency threshold is specified as 1 and is inferred based on historical usage and a certain error tolerance. If a large number of invalid segmented words are found in the subsequent process, the value will be adjusted as appropriate, and then the processed segmented results will be combined with the character position. The information such as location, part-of-speech tagging and frequency tagging are sorted and separated by commas. When encountering words with unclear meanings or homophones, more appropriate word segmentation boundaries and parts of speech are determined according to their correlation in the corpus. In addition, additional comparisons and matches are also performed for common abbreviations such as abbreviations or industry-specific symbols. Words that may overlap multiple times in different contexts are checked one by one and the form that best fits the context is retained. If it is necessary to distinguish the scope of professional terms in specific scenarios, a list containing field keywords is set in advance. The list is formulated and periodically updated by several professionals in combination with approval specification documents and technical materials. Finally, all the annotated words are summarized and recorded in a unified manner to form the final word segmentation result.

[0018] After obtaining the word segmentation results, each word segmentation element is first encoded, and the word segmentation elements are converted into word vectors or calculated by string similarity, and then compared with the prepared business type dictionary one by one. The business type dictionary can be uniformly formatted first, and the key entries in the dictionary are arranged in alphabetical or pinyin order, and a category label is assigned to each entry. Then cosine similarity or character edit distance is selected as the main comparison indicator. At this time, a similarity threshold can be set to distinguish results with lower matching degrees. The value of the threshold can be estimated by the median matching accuracy of the industry text collected previously. For example, in recent sample statistics, it is found that when the similarity is lower than 0, .6 will lead to a higher error rate, so 0.6 can be used as the initial setting. If the error is still found to be high in subsequent verification, the threshold can be adjusted again. When comparing each word segmentation element with the dictionary entry, additional weight corrections can be made according to the word length and the difference between simplified and traditional Chinese characters. For example, professional terms with more characters can be regarded as higher priority, and the stop word list generated in the early stage can be combined to eliminate the frequently appearing words that do not correspond to specific business meanings. After the similarity calculation is completed, the matching degree of each word segmentation element to the dictionary entry is recorded, and a set of corresponding similarity scores are generated for all word segmentation elements. Finally, these score sets are integrated into the similarity result.

[0019] After obtaining the similarity results, the similarity scores of each business type contained therein will be summarized and compared. First, the similarity values ​​corresponding to all word segmentation elements will be weighted and merged, and the cumulative similarity of each business type in the overall text will be calculated. A difference threshold can be added to judge the degree of significant difference between the scores. The difference threshold can be determined by statistically analyzing the distribution of actual judgment results in historical approval documents. For example, in previous data, if the difference between the highest score and the second highest score is less than 0.05, there is often confusion in judgment. In this case, 0.05 can be selected as the initial reference value. If it is found in subsequent tests that the juxtaposition situation is still frequent, the threshold can be continuously adjusted. After confirming the ranking of the similarity scores, the business type ranked first is taken as the type identification of the file. If the scores of multiple business types are simultaneously higher than a set value and meet the minimum difference range limit, multiple labels can also be marked in the record. Finally, the obtained file type information is organized into file business classification results.

[0020] The steps to obtain the file semantic feature vector are: Based on the results of the document business classification, the approval document content is parsed, keywords in the approval document content are extracted, keywords are screened, redundant keywords, synonyms or interference items are removed, and keywords are normalized to obtain a keyword list; According to the keyword list, the semantic association strength between keywords is calculated using the following formula: ; in, Represents the strength of semantic association between keywords. Representative keywords The frequency of occurrence in the text, and Respectively represent and The semantic vector value of keywords, and Respectively represent and The relative position deviation of keywords in the context window, and Represents the semantic dependency distance of keywords in the text, and Represents the dependency weight of keywords in different syntactic structures, and Represents the number of times a keyword appears in different text fragments; Based on the semantic association strength, a file semantic feature vector is constructed.

[0021] Specifically, based on the document business classification results obtained previously, the text of the current approval document content is read line by line and scanned in character sequence. Then, all the words that appear are retrieved and compared with the existing synonym comparison table and basic terminology dictionary. The words that are marked as synonymous or only appear in specific fields are merged. If the cumulative number of appearances of some short words in the entire text is less than 3 and their meanings do not have conceptual reference, they can be identified as interference items and removed to reduce the impact of irrelevant content. During this period, the word form changes containing common roots will be checked and unified. For example, plural forms or mixed case English words will be processed, and redundancy will be controlled through character segmentation and word form restoration. In order to accurately distinguish For words with multiple meanings, the judgment will be made with reference to the position range of the word in the context and the high-frequency collocation information that has been collected. If the word is frequently associated with specific field terms or numerical indicators in the adjacent text range, it will be listed as a reserved category and recorded with its core meaning label. In order to define "low-frequency words" or "rare words", a corresponding frequency threshold will be set. The initial value of the threshold can be 2 and fine-tuned in combination with the text length distribution of historical approval documents. If it is found that there are still many residual low-frequency interference items in subsequent applications, it may be necessary to further lower or increase the threshold. Finally, all processed valid words are normalized to maintain their text consistency in different scenarios to obtain a keyword list.

[0022] The benefit of the formula is that it incorporates multiple factors such as keyword occurrence frequency, context position deviation, semantic dependency distance, and dependency weight in syntactic structure into the calculation, so that the association strength depends not only on the number of occurrences of the keyword itself, but also on its relative distribution and structural relevance in the text.

[0023] The steps to obtain the parameters are: First, count the number of times each keyword appears in the approval text, compare the number of occurrences with the total number of words in the current text, and then calculate the keyword ratio and form a frequency value. The frequency value at this time can be written as , QRR is the number of times the keyword appears, AL is the total number of words in the text. If a sample with a total number of words of about 2000 is extracted from the registered approval text, and the number of times this keyword appears is 50, then the corresponding That is In large-scale texts, the statistical results of multiple fragments can be integrated to form weighted frequencies and then normalized.

[0024] and The steps to obtain the parameters are: Each keyword needs to be vectorized, and the corresponding vector value can be obtained through the previously constructed word vector training. The specific method is to collect several words with known semantic labels from the same type of approval texts in previous years and perform vector training. The training model can use a fixed window scanning method to evenly extract several words within the context of each word to construct training samples, and then use maximum likelihood estimation to continuously update the word vector parameters. Assume that when the model finally converges, the vector value corresponding to the keyword "supply chain" is (0.12, 0.35, -0.09), and the vector value corresponding to the keyword "logistics" is (0.10, 0.30, -0.05). At this time It can be (0.12, 0.35, -0.09), The value can be (0.10, 0.30, -0.05), and the same is true for other keywords.

[0025] and The steps to obtain the parameters are: Record the context position of each keyword when it appears in the text, and set a context window. For example, define the window span as 5 words, and count the average position coordinates of the keywords in each place within the range of 5 words on both sides. Then calculate the difference between this average position coordinate and the first position of the sentence where the keyword is located, and then map the result to the interval of -1 to 1. If the average coordinate of the context window of the keyword "project" at its appearance is 12, and the beginning of the sentence is 0, then we can get , ZC is the total number of words in the sentence. If the total number of words in the sentence is 30, then .

[0026] and The steps to obtain the parameters are: We need to first record the distance difference between each keyword in sentences or paragraphs. The distance can be defined as the difference in the number of words or symbols between keywords, or it can be obtained by calculating the number of jumps from node m to node n in the dependency tree. If we observe that the keywords "purchase" and "production" are separated by 3 edges in the dependency tree, we can or Assign a value of 3. If the spacing changes to 2 or 4 in other text paragraphs, record it accordingly.

[0027] and The steps to obtain the parameters are: Determine the position of keywords in the grammatical structure, such as subject, predicate, object, etc. Keywords in the subject position can be given a higher dependency weight, while those in the modifier position in the clause can be regarded as lower weights. When quantifying these weight values, they can be selected in the range of 0.5 to 2.0 and calibrated in combination with statistical experience. If it is observed in a large number of monitored approval texts that keywords appearing in the subject position are more likely to be related to global semantics, the weight of the subject position can be set to 1.8, and the weight of the modifier position can be set to 0.9. Finally, the dependency weight ratio of the two keywords is calculated. When the actual measured values ​​are input, the comparison can be completed.

[0028] and The steps to obtain the parameters are: It is necessary to identify the number of times a keyword appears in different text segments. The approval document is divided into several segments through the previous paragraph segmentation, such as processing each 500 words or each natural paragraph separately, and then summarize the frequency of each keyword appearing in different segments. If a keyword appears in all 10 segments, it will be recorded as 10 occurrences. If the number of occurrences in 3 of the segments is obviously higher, you can also add differentiation counting logic to further mark the density of these keywords in local paragraphs. For example, if the keyword "budget" appears in 8 paragraphs and the keyword "contract" appears in 12 paragraphs in the specific example, then the corresponding =8, =12.

[0029] Calculation process: In an approval text scenario, a pair of keywords A and B has the following parameters: , , , , ; Denominator ; Substituting these values ​​into the numerator: Numerator = ; Do the multiplication first: ; Then add up: ; Denominator = 20; therefore: ; The results show that when the keyword The final correlation strength is about 0.1103 after combining the semantic vector difference with the context position deviation, dependency weight ratio and other factors. A value greater than 0.60 indicates that their correlation is more obvious, while the currently calculated value of 0.1103 indicates that the correlation between keyword A and keyword B is weak.

[0030] In the previous section, we calculated the semantic association strength between keywords and obtained After the numerical sequence of all keyword combinations, the association strength data corresponding to the keyword combinations are collected, and the distribution statistics of the association strength values ​​in the range of 0 to 1 are performed. If the association strength of some keyword pairs exceeds 0.70, they can be classified as a high association group. If it is between 0.30 and 0.70, it is marked as a moderate association range. If it is lower than 0.30, it is considered to be weakly associated. In order to refine its applicability in subsequent text descriptions, it is necessary to aggregate the keywords belonging to the high association group and record them in the same keyword cluster. Here, the keywords can be compared again within the high association group by comparing them pair by pair. If the association values ​​are both greater than 0.60, they are classified into the same cluster. Otherwise, if they are greater than 0.30 and less than 0.60, they are temporarily classified as adjacent clusters. This hierarchical grouping method can be used to map the themes of large-scale approval texts at the keyword level, and then compare the main type labels in the business classification results one by one to see if there is a keyword cluster corresponding to a specific business category. If it is determined that there is a high degree of overlap, the keyword cluster can be bound to the category for subsequent tracing and reference. Finally, all keywords and their cluster affiliations are summarized, and the document semantic feature vector is constructed with the support of the association strength results.

[0031] The steps to obtain the responsibility matching results are as follows: Based on the document semantic feature vector, call the responsibilities of each department and the responsibilities of each person in the organizational structure, parse the responsibilities description text, extract the responsibilities keywords, construct the responsibilities keyword vector, and obtain the responsibilities keyword vector matrix; According to the responsibility keyword vector matrix, the matching value between the document semantic feature vector and the responsibility keyword vector is calculated. The expression is: ; in, Represents the matching value, is the first part of the semantic feature vector of the file Dimensional components, is the first Dimensional components, and Represent the relative distances between document keywords and responsibility keywords in the semantic space, For the The frequency value of the matching words, and Represents the position index of the document keywords and responsibility keywords in the syntactic structure, , and Represents the matching similarity coefficient between each job description text, Represents the number of dimensions of the document semantic feature vector, Represents the number of matching words in the responsibility keyword vector, The total number of job description texts; Based on the matching value, the job descriptions of each department and the job descriptions of personnel are matched and evaluated, and the responsibility items with the highest matching values ​​are selected as the affiliated departments and responsible persons to obtain the responsibility matching results.

[0032] Specifically, based on the document semantic feature vector, the text content of the responsibilities of each department and the job responsibilities of personnel in the organizational structure is first read, and the proper nouns, job titles and functional description items in the text are determined by continuous sentence-by-sentence retrieval. Then, representative words in these contents are extracted one by one and merged. Synonymous expressions or responsibilities with a large degree of overlap will be merged. Here, the responsibilities content text will be divided into several small sections, each covering a range of approximately 200 to 300 words, and it will be checked whether the key verbs and nouns appearing therein are included in the previously constructed job responsibilities vocabulary. When it is found that some words and common responsibilities or authority description words in the job responsibilities vocabulary appear more than 3 times, they are identified as possible responsibilities keywords, and the number of the paragraph where the corresponding word is located and the text distance between it and other possible responsibilities keywords are recorded. The text distance can be accumulated and counted by the number of characters or words. Every time a verb with job-specific meaning such as "review" and "approval" is encountered, the distance between it and the current keyword is checked. When the distance reaches the range of 50 to 80 words, a distant distribution is recorded, and within 50 When the interval is within the word, a closer distribution is recorded. The closeness between the keyword and the job description is judged by comparing multiple distribution results. When the text distance value distribution tends to be concentrated in a smaller range, it indicates a higher correlation. At the same time, semantic recognition is performed on keywords that appear more than twice. When referring to the organizational structure information, it can be found whether the keyword corresponds to the department's responsibility field. If some keywords have a high correlation with multiple departments, they will be further subdivided and compared with other description details in the job description text, such as whether it contains a specific scope of authority or a specific work scenario. Finally, a group of responsibility keywords that have been removed from the redundancy and dominate the job description text are obtained, and these responsibility keywords are numbered in the order of text records. When converting them one by one into vector form based on the numbering order, the previously completed vector training method will be referred to, and the corresponding position component or authority component in the department information will be superimposed to form multiple dimensions of the responsibility keyword vector. The vectors corresponding to all keywords are grouped according to the position or department category, and stored in a unified data structure according to the collected text distribution. After merging and sorting, the responsibility keyword vector matrix is ​​obtained.

[0033] The benefit of the formula is that it incorporates the multi-dimensional differences between the file semantic feature vector and the responsibility keyword vector, word frequency, position index, and the similarity coefficient of each responsibility description text into the comprehensive calculation process, and combines information such as vector dimension difference, word frequency ratio, syntactic position difference, and cross-text similarity to obtain more targeted matching values.

[0034] The steps to obtain the parameters are: This parameter is the semantic feature vector of the file in The dimensional components are obtained through the text semantic vector training process mentioned above. During vector training, frequently appearing keywords and phrases are extracted from a large number of approval documents, their contextual dependence and contextual distribution are learned, and each keyword is given a vector representation of several dimensions. Then, the keyword vectors contained in the entire approval document are weighted or aggregated in other forms to generate the final document semantic feature vector. That is in The specific value of the dimension is determined according to the density of the keyword in the text, the scope of occurrence, and the relevance to the context. For example, 500 approved documents are collected to construct a semantic space with a word vector dimension of 100. The keyword set obtained above is selected and aggregated one by one. In the aggregation process, keywords with a frequency of more than 3 times and a relevance to the text topic of more than 60% are given a higher weight. Finally, the semantic feature vector of the file is synthesized. Assuming the value of the 10th dimension is After statistics and normalization, it is approximately 0.27.

[0035] The steps to obtain the parameters are: This parameter indicates the responsibility keyword vector in the The values ​​on the dimensions are obtained based on the same vector training principle, but the selected corpus is all from the text of departmental responsibilities and job responsibilities. For the core keywords in each responsibility text, their occurrence frequency is recorded first, and then the same dimension aggregation idea is used to generate a unique responsibility keyword vector. It is this vector that In order to make the document semantic feature vector and the responsibility keyword vector comparable, the training schemes of the two need to be consistent, so that the semantic components can correspond to each other on the same dimension. For example, when parsing the responsibilities of the human resources department, statistics are collected for keywords such as "recruitment", "attendance", and "performance", and vectors are generated by matching them with the previously set training model. A 100-dimensional vector representation may be obtained, in which the 10th dimension is 0.20. =0.20.

[0036] and The steps to obtain the parameters are: These two parameters represent the relative distance between the file keywords and the responsibility keywords in the semantic space. It is necessary to extract the vector coordinates corresponding to the keywords, calculate the Euclidean distance or cosine angle between them in the vector space, and then perform appropriate normalization to map the result to a range of 0 to 1. The larger the value, the farther the distance between the two in the semantic space, and the smaller the value, the closer they are. When collecting a batch of common department responsibility keywords, it may be found that the distance between "audit" and "approval" in the vector space is small, so The distance between keywords related to “market” and “market” may be close to 0.10, while the distance between keywords related to “market” and “market” may be 0.35. After measuring the distance between each keyword pair, the corresponding and For difference calculation in the formula, for example, select keyword pairs from 200 job description texts and corresponding approval documents, and calculate the distance one by one. It is found that the distance is 0.25. After normalization, the final value is 0.25. , and the other distance is 0.30, which can be taken as .

[0037] The steps to obtain the parameters are: This parameter indicates the The frequency value of a matching word in the responsibility text will be counted based on the total number of words in the departmental responsibility description and the position responsibility description and the number of times the matching word appears. For example, if the key responsibility word "finance" appears 150 times in a 30,000-word responsibility text, its frequency value can be recorded as , and then subdivided according to the corresponding job level or sentence weight coefficient. In order to obtain a more practical reference value, this result can be compared with other key responsibility words and normalized. If the final calculation finds that the frequency of "finance" in the entire text ranks in the top 10%, it can be Assign a weight value of 0.90. If it ranks in the bottom 10%, you can consider assigning a weight value of 0.30. Finally, bring these word frequency values ​​into the formula for product operation. The acquisition example given here is: when the matching word "compliant" appears 120 times and the length of the text in which it is located is about 24,000 words, its initial word frequency is 0.005. If the subsequent comparison shows that its frequency of occurrence ranks 30% among similar keywords, it will eventually be recorded as =0.70.

[0038] and The steps to obtain the parameters are: These two parameters are used to indicate the position index of the file keyword and the responsibility keyword in the syntactic structure. You need to split the corresponding sentence according to the word order and set a fixed sequence position for each word. Then, record the index numbers of the file keyword and the responsibility keyword respectively. When the syntactic structure is relatively complex, you can first build a dependency tree or sentence analysis tree, traverse the nodes in turn to obtain the keyword position, and then map this position value to a reasonable interval. For example, in a sentence with a length of 40 words, if the fifth word is the file keyword, you can =5, the 10th word is the responsibility keyword, then =10, then do a range normalization from 0 to 1, for example, map 5 to , 10 is mapped to Etc., and finally stored in subsequent calculations. If in a longer description, the index number can be distributed in a larger range accordingly. Take a practical example: a sentence has 50 words, the file keyword appears in the 8th word, and the responsibility keyword appears in the 20th word. After normalization , .

[0039] , , The steps to obtain the parameters are: These three parameters represent the matching similarity coefficients between each responsibility description text and the key content of the file, which are used to measure the similarity extracted from multiple departments or multiple job descriptions. It is necessary to compare the key paragraphs of different responsibility texts with the keywords in the file. You can choose to use the word vector model built in the previous corpus training to measure the cosine similarity between the sentence or paragraph vector in each responsibility text and the file vector, and then divide the results into segments and finally map them to the range of 0 to 1. For example, when analyzing the responsibility text of Department A, a total average similarity of 0.73 is obtained, which can be Assign a value of 0.73, and the average similarity of the responsibility text of department B is 0.45, then Take 0.45, and get 0.67 for the text of department C. =0.67. These values ​​form a list in the multi-department and multi-text scenario, which is used for the cumulative operation of the denominator of the formula to distinguish the overall matching degree between different responsibility texts and the current file. For example, when a certain approval document is compared with the responsibility texts of 5 different departments in sequence, the similarity values ​​are 0.70, 0.63, 0.54, 0.77, and 0.58 respectively. =0.70, =0.63, =0.54…etc. and use it in calculations.

[0040] The steps to obtain the parameters are: This parameter indicates the number of dimensions of the document semantic feature vector, which is usually fixed during the corpus training phase. For example, if 100 or 150 dimensions are selected when constructing word vectors, then The corresponding value should be determined. When using a 100-dimensional word vector, the entire document semantic feature vector is the projection on this 100-dimensional space, so =100. In order to obtain a specific feasible dimension, it is necessary to conduct multiple experiments on the previously collected industry texts. The dimensions can be compared from 50 to 200 to see the matching accuracy and computing consumption under each dimension. Then, a number of dimensions is finally selected on the premise of balancing accuracy and efficiency. For example, after monitoring 120 simulation tests, the word vector with 100 dimensions is the best, so Substitute =100 into the relevant calculations.

[0041] The steps to obtain the parameters are: This parameter is the number of matching words in the responsibility keyword vector. It is necessary to first retrieve a series of keywords that play a major role in the responsibility text and count them when constructing the responsibility keyword vector matrix. For example, if 20 high-frequency or high-weight keywords are retrieved from the responsibility description of a department, only 10 of them may be matched after intersection comparison with the document keyword list. =10, when expanding the search scope or merging multiple job responsibilities, The keywords will also change accordingly. For example, if the keywords of the three job texts are combined into 50 keywords, and 25 keywords are selected to match the files, then =25, and finally these values ​​are substituted into the corresponding formula for subsequent multiplication operations.

[0042] The steps to obtain the parameters are: This parameter represents the total number of job description texts. When parsing an organizational structure, it is necessary to take into account all the job descriptions of each department, and then count each text separately. In some scenarios, there are 10 departments, and each of the HR, Finance, Logistics, Sales and other departments has one or more job description documents. After merging all of them, there may be a total of 15 to 20 texts. This is the total number. For example, if the organizational structure check determines that there are 18 job description documents, then =18. In the statistical stage, these texts will be analyzed for keywords and constructed as vectors in turn, and then participate in the combined calculation of the denominator of this formula.

[0043] Calculation process: When known , , , and after obtaining the relevant parameters in the previous step, bring them into the formula, then: ; ; ; Then the numerator part = ; Denominator = ; so: ; Since the normalization interval mentioned above is from 0 to 1, when the calculated value exceeds 1.0, it can be marked as 1.0 in the matching evaluation to reflect a higher degree of responsibility matching.

[0044] The result shows that the matching value obtained in this example is greater than 1.0, which means that the document and the selected job description text are relatively close in terms of dimension component difference, keyword position index, and cross-text similarity, which means that the two have a high degree of overlap in job description content. If the actual normalization result is =1.0 indicates the highest matching degree, and this responsibility can be considered as a priority item for matching evaluation in subsequent steps.

[0045] Based on the matching value, the corresponding departmental responsibilities description is first retrieved according to the acquired file semantic feature vector information, and the matching value is divided into different intervals and recorded one by one. When the matching value is greater than 0.75, it is regarded as a high matching degree, which can be marked and summarized into a candidate set. When the matching value is between 0.40 and 0.75, it is regarded as a moderate matching degree, which is also marked and stored uniformly. If the matching value is lower than 0.40, it is recorded as a relatively weak match and directly attached to the corresponding department or position name in numerical form. Then, the items with higher matching values ​​are concentrated for secondary screening. Here, the distribution of matching values ​​can be compared with the focus of departmental responsibilities description. For example, when it is found that most of the keywords in the same position responsibility text If the degree of consistency with the key themes of the document is high, this text will be regarded as a more suitable option. Then several job descriptions with a matching degree of 0.80 or above will be compared one by one to see whether they involve specific management categories or professional categories in the document theme. If the number of fully matched categories reaches 3 or more, the job responsibilities will be put into the final selection list. During this period, the hierarchical relationship in the organizational structure will be referred to for layer-by-layer comparison. All job texts will be sorted and marked with their matching value scores, listed in order from highest to lowest. The first job with the highest matching value will be extracted from the list and associated with its department in order to determine the department and the person in charge. Finally, the results will be merged into a centralized record to obtain the responsibility matching results.

[0046] The steps to obtain the approval personnel matching index are as follows: Based on the responsibility matching results, the current number of tasks of each approver is counted, and the cumulative task processing time of each approver's processed tasks is calculated. At the same time, the list of pending tasks of the approver is extracted to obtain the task statistics of the approver; According to the task statistics of the reviewers, the task processing mode of each reviewer is analyzed, the average number of tasks completed per unit time is calculated, and the current task saturation degree of each reviewer is counted to obtain the reviewer task load index; Based on the task load index of the approval personnel, the matching index of the approval personnel is calculated. The calculation formula is: ; in, Represents the approval personnel matching index, The task processing time of the approved personnel who have completed the task. The current number of tasks for the approver. Represents the number of tasks to be processed by the approver. Represents the average completion time of historical tasks of the approver. Represents the task processing efficiency of the approver on different task types, Represents the standard deviation of task processing time of approvers in different task categories. and Represents the stability coefficient of task completion of the approver in different time periods, Represents the number of task records of the approver. Represents the number of task categories for approvers.

[0047] Specifically, based on the responsibility matching results, the completed task records are first retrieved from the work information database of each approver, and the current number of tasks of each approver is counted by comparing the identification number and completion time of the task one by one, and then the actual execution time of each task is accumulated and summarized. The duration of each completed task will be extracted here, and its start time and end time will be recorded. The processing time of a single task is obtained by subtraction, and then the duration of all completed tasks of the same approver is added to get the cumulative value. In order to ensure the accuracy of the duration data, it is necessary to compare such records with the punch-in or operation traces in the log system to confirm whether the corresponding task is associated with the approver. If duplicate registrations or task numbers transferred across departments are found in the retrieval, it is necessary to determine the start and end points of these records according to the situation, and calculate the duration in a unified format and then accumulate it. During this period, the current The list of unfinished tasks is included in the statistics. The pending tasks of each approver will be marked as "unfinished" or "in process" in the system, and the estimated processing time limit for the task will be recorded. If the processing time limit exceeds the benchmark value set in advance by the department, for example, there is still no result after 48 hours, a reminder mechanism may be triggered. The reason for choosing 48 hours as the benchmark value is that in nearly 100 approval cases, the average time required for most approval processes is about 36 hours. An additional 12 hours of margin is given to prevent delays caused by multi-department collaboration factors. While recording these pending tasks, it is also necessary to check their task category and priority. If the priority mark is higher, it may be listed and sorted separately later. Finally, the cumulative time of all completed tasks, the current unfinished list of pending tasks, and the number of tasks are merged to form centralized data to obtain the task statistics of the approver.

[0048] According to the task statistics of the reviewers, it is necessary to analyze the past processing cycle and task structure of each reviewer. By calculating the number of tasks completed by each reviewer in a fixed period of time, the average number of tasks completed per unit time can be obtained. Then, this average value can be compared with the current task list to observe the correspondence between the actual total number of tasks being processed and the average processing rate per person. Here, a standard working week can be defined as 40 hours, and the overtime frequency can be determined according to the department's habits or the rules stipulated by the company. For example, some departments do not process on weekends, while other departments have a 24-hour duty system. Therefore, when calculating the processing efficiency per unit time, the available time periods should be distinguished according to the department's work and rest conditions, and then the processing efficiency should be compared with the current number of tasks. If the current number of tasks for an reviewer is 8, combined with the average processing rate of the person in the past two weeks of 0.2 per hour, that is, about 8 tasks can be completed within a week, then it can be determined that this reviewer If the task saturation of an approval personnel approaches the critical level, adding new approval tasks at this time may cause queues or delays. If under the same conditions, someone has only two tasks in backlog or a higher processing rate per unit time, it means that his current work saturation is relatively lower. In addition to the speed index, the differences in task types should also be checked. If an approval personnel currently has all high-priority tasks involving cross-departmental signatures, an additional mark can be given during statistics to avoid simply equating it with the number of low-priority tasks with loose time limits. If it is found in the search that someone's work saturation exceeds the threshold set by the department, for example, more than 80% of the working time is occupied by high-priority tasks, it is recorded as a high saturation state. The specific value of the threshold can be derived from the load statistics of nearly 100 approval processes. By observing the actual operation efficiency of the approval in segments, the most suitable 80% interval is obtained through continuous testing between 70% and 90%. Finally, the saturation calculation results of all approval personnel are integrated into the task load index of the approval personnel.

[0049] The benefit of the formula is that when evaluating the matching degree of approval personnel, it not only integrates the processing time of completed tasks and the current number of tasks, but also introduces multiple factors such as the number of tasks to be processed, the historical average duration, the efficiency differences of different task types, and the stability of task completion. Through the quantitative processing of multiple parameters and the coordinated matching of the numerator and denominator, it provides more detailed data support for the subsequent automatic allocation of approval tasks.

[0050] The steps to obtain the parameters are: This parameter indicates the task processing time of the approved personnel's completed tasks. When counting each completed task, it is necessary to read the start time and end time from the system log, and accurately obtain the task processing time by the difference between the two. Then, the duration of all completed tasks of the same approved personnel is recorded in the order of task numbers. These records form a duration array, and each element corresponds to a task duration. When the task is completed, the elements in the array will be directly added. In order to make the values ​​comparable, invalid or interrupted tasks need to be excluded after the preliminary statistics are completed. For example, in a few large-scale collaborative tasks that span months, the actual time invested by the approvers may be segmented, and multiple durations need to be added and then merged into the final duration of the task. Finally, a set of accurate completed task durations is obtained. For example, by monitoring 5 closed tasks in a month, it is recorded that they consumed 2.5 hours, 3.0 hours, 1.8 hours, 2.0 hours, and 4.5 hours respectively, then the corresponding The acquisition example here is as described above. Anyone can check the historical task list of the approver and accumulate the start and end time differences to obtain The specific value of each item.

[0051] The steps to obtain the parameters are: This parameter is a detailed list of the current number of tasks for the approver. First, check the list of all unfinished tasks for the person in the current state, and then record the number of tasks of different types or priorities separately. For example, there are 7 ordinary approval tasks, 2 expedited approval tasks, and 1 cross-departmental collaborative task. These 10 tasks can be counted as a unified current task number list. The task category number can also be marked internally, such as 0 for ordinary, 1 for expedited, 2 for cross-departmental, etc. When doing product operations, such as 1 to 3 correspond to three categories respectively, and the specific values ​​are read from the list. For example, when an approval staff counts, he finds 7 ordinary tasks, 2 expedited tasks, and 1 cross-departmental collaboration task, then , , , if different categories are quantified or decentralized, other parameters can also be used in subsequent formulas to balance the expression. For example: retrieve a person's daily task list from the backend database and compare the status labels one by one to confirm that there are 3 categories and 10 tasks in total, so they are assigned one by one. etc. location.

[0052] The steps to obtain the parameters are: This parameter represents the number of pending tasks for the approver. It needs to be distinguished from the difference between the "current tasks" in the previous step. "Pending" focuses more on those tasks that have not yet officially entered the processing stage but have been assigned to the approver by the system. If the system finds that there are still several tasks that have not been opened or confirmed in the to-do list of a certain person, the number of these tasks should be accumulated to form , and check it with the assignment information in the database. If it is found that the task has been shared and assigned by multiple people during multi-department collaboration, it can be counted into the personal to-do quantity only after the approver confirms the order, so as to avoid repeated measurement. In actual statistics, it can be seen that some approvers' to-do lists may contain 2 to 5 to-do tasks. If there are more than 5, short-term congestion often occurs. Example: The system queries someone's to-do list and finds that 4 of the records have the status of "not clicked to confirm" and another 2 are in the status of "need to be countersigned" and have not officially started. The person has been assigned by the system, so the total .

[0053] The steps to obtain the parameters are: This parameter indicates the average completion time of the historical tasks of the approver. It is necessary to extract the processing time of several completed tasks from the historical data and then perform the average calculation. In order to make the result representative, the historical samples of the past three or six months can be selected, and obviously abnormal extremely short or extremely long tasks can be eliminated. Then, all the processing times included in the statistics are added up and divided by the total number of tasks to get the average completion time. If 20 completed tasks are collected in the past three months, and their processing time is 60 hours in total, then Example: Randomly select task records of multiple reviewers in a department and calculate the average value for each person. If someone finally finds that the average completion time is 3.2 hours, then 3.2 is used as the Write the subsequent formula.

[0054] The steps to obtain the parameters are: This parameter is used to reflect the task processing efficiency of the reviewers on different task types. The completion time of different types of tasks, such as ordinary tasks, urgent tasks, and complex collaborative tasks, can be counted separately, and then the output of these types can be compared with the average level of tasks of the same type to form a rate or efficiency score. For example, for ordinary tasks, an average of 0.4 items are completed per hour, and for urgent tasks, an average of 0.2 items are completed per hour. These sub-items can be combined and weighted or compared with the benchmarks of the same position in the industry to obtain the final In order to make To be used directly by the formula, it needs to be mapped from 0 to 1. For example, if after comprehensive evaluation, it is found that someone's efficiency is in the top 20% of all personnel, then it can be assigned a value of 0.80, and if it is in the bottom 20%, then it can be assigned a value of 0.20. Example: Statistics on the completion rate of various types of tasks of a certain approval personnel in the past month, after comparing with the department average, it is judged that its efficiency coefficient is 0.75, that is, .

[0055] The steps to obtain the parameters are: This parameter represents the standard deviation of task processing time of the approver in different task categories. It is necessary to list the completion time records of different types of tasks one by one, and then calculate the standard deviation corresponding to each type of task, and finally merge them into a unified value. Here, the root mean square or weighted mean method can be used to combine the standard deviations of different categories. If the standard deviation of ordinary approval records is 1.2 hours and the standard deviation of expedited records is 1.8 hours, then the corresponding number of tasks can be used as the weight for aggregation. If the proportion of ordinary tasks is as high as 70% and the proportion of expedited tasks is 30%, then It may be between 1.3 and 1.5 hours. It can also be mapped to the range of 0 to 1 through normalization, such as 1.3 hours corresponds to 0.65, 1.5 hours corresponds to 0.75, etc. For example, the processing time of a person's past 10 ordinary tasks fluctuated between 3 and 5 hours, and the calculated standard deviation was 0.7 hours. There were 5 urgent tasks with a larger fluctuation, with a standard deviation of 1.5 hours. After summarizing the two according to the task volume weight, it may be 1.1 hours, which is then mapped to 0.55, that is, .

[0056] and The steps to obtain the parameters are: These two parameters represent the stability coefficient of task completion of the reviewer in different time periods. It is necessary to quantify the task completion performance of the reviewer in different time periods such as working days and non-working days, or during the day and at night. It is possible to count the total number of tasks, completion efficiency, delay, etc. in the time period, and then convert the result into a stability index between 0 and 1. If a stable processing rate can be maintained during most of the working day, then will be higher, for example, 0.80 to 0.90, if the processing efficiency fluctuates significantly at night. The stability coefficients are lower, such as 0.60 to 0.70. These values ​​can be combined into two independent stability coefficient records. For example, if we observe the tasks handled by a certain approval staff in the department from 8:00 to 18:00 in the daytime for the past two months and find that his on-time completion rate is 85%, then Set to 0.85, and the on-time completion rate after 18:00 at night is only 60%, then =0.60, which is finally used for reference in subsequent formulas.

[0057] The steps to obtain are: This parameter represents the number of task records of the approver, which is obtained by counting the completed tasks mentioned above. It is necessary to ensure that all records are approval tasks completed independently or in a leading manner by the person in order to be included. If the person has completed 15 approval tasks with complete duration records in the past period of time, then If some of the tasks are in the co-organized or participated status but without the leading authority, they are usually not included in the main completion list of the person to avoid duplication or deviation. For example: query the completed task flow of the approver in the background, remove 3 co-organized status records and find that there are 12 tasks left, so .

[0058] The steps to obtain are: This parameter represents the number of task categories of the approver. It is necessary to classify the types of tasks that the approver has handled currently or in the past. For example, the approval process can be divided into multiple types such as contract approval, expense reimbursement, and purchase application. If the approver is involved in 4 different types of approval tasks, then , which may also occur for those who only handle a single type of approval at a time For example, the system pre-sets four types of approval business, namely expense reimbursement, purchase application, legal review, and asset management. If an approval officer has a handling record for all of them, the number of his task categories is 4.

[0059] Calculation process: In a specific scenario, the historical and current task data of a certain approver has been obtained. The following are example values: , the duration of the five completed tasks of the person is counted and obtained Hour, Hour, Hour, Hour, Hour.

[0060] The current tasks of this person are divided into two types: urgent and ordinary, corresponding to and .

[0061] (2 urgent items), (3 pieces for general category), then .

[0062] , 4 unopened to-do tasks were found in the person’s to-do list.

[0063] hours, and 3 hours is obtained by calculating the average completion time after selecting several completed tasks in the past three months.

[0064] After comparing with the average efficiency of the department and normalizing it, the efficiency of this person is 0.80.

[0065] , the standard deviation of the processing time on different task types is counted and summarized as 0.60.

[0066] , , determine its stability coefficient during weekdays and nighttime.

[0067] Calculate the numerator first: ; ; So the numerator = = ; Then calculate the denominator: ; ; ; Add these values ​​together: ; Then add 1: ; So the denominator = ; final: ; when When it exceeds 2.00, it is marked as a high load matching degree, indicating that the person is in a sensitive stage of work allocation based on the comprehensive factors such as the time of completed tasks, the number of current tasks, the number of tasks to be processed, and historical efficiency. A value below 1.00 indicates that the person has a better balance between task load and efficiency.

[0068] The results show that this person is currently overloaded in multiple dimensions. If approval tasks are to be assigned to this person, it may cause a certain degree of task accumulation. When the value becomes larger and larger, it indicates that there is a more obvious contradiction between the current load and the historical efficiency, which requires careful consideration in the subsequent task allocation stage.

[0069] The steps to obtain the task processing capability index are as follows: Based on the matching index of the approvers, the task allocation threshold is set, the matching index of all approvers is screened, and the approvers whose matching index is higher than the task allocation threshold are extracted to obtain a list of candidate approvers; According to the list of candidate reviewers, calculate the task completion time of each candidate reviewer. The calculation formula is: ; in, Represents the task completion time of the candidate approver, is the total time that candidate approvers have completed tasks in the past. Represents the estimated duration of the current pending task. and Indicates the task switching time of candidate approvers. The standard processing time of the candidate approver's historical tasks. and A tuning parameter representing the complexity of the task, , and Represents the task processing efficiency of candidate approvers in different time periods; Based on the task completion time, the task processing ability of candidate approvers is evaluated, and the task processing ability index is obtained by combining the task category, processing timeliness and execution record.

[0070] Specifically, based on the matching index of the reviewers, first specify an applicable task allocation threshold and load the matching index of all reviewers. After reading the data of each matching index one by one, refer to the previously collected approval process execution samples to determine a demarcation standard that can distinguish between high-load and medium- and low-load personnel. If it is necessary to form a specific threshold in the early collection, it can be processed in segments in combination with statistical methods. For example, when the past 50 approval process records are selected and the final matching index range of the corresponding reviewers is between 1.50 and 3.00, there is often a serious task congestion. If the matching index is below 1.00, the general load is relatively light, so 1.50 can be used as the empirical threshold. The value is used to screen out personnel with higher matching indicators, and then the matching indicators of all approval personnel are compared with the threshold, those with lower values ​​are screened out and the remaining records are retained. For those who have special work patterns in different time periods, their average task processing efficiency and overtime hours and other information are also checked. If they meet the specific requirements of the department and their matching degree is higher than the set standard, they will be listed as candidates. For those who meet the standards, their job qualifications or approval authority scope must be further checked to confirm that they have the ability to handle the corresponding task types. If their job responsibilities are incompatible, they will be excluded even if the matching indicators meet the standards. After the screening is completed, the approval personnel who pass the matching threshold are summarized to generate a list of candidate approval personnel.

[0071] The benefit of the formula lies in that it comprehensively considers multiple factors such as the total duration of tasks completed in the past, the estimated duration of the current tasks to be processed, the task switching time, task complexity and processing efficiency. By taking the form of absolute difference, square root operation and the superposition of multiple efficiencies on the denominator, the impact of different time periods and different task characteristics are included in a unified measurement, providing more detailed data support for the subsequent evaluation of the overall task completion time of candidate approvers.

[0072] Steps to obtain the P parameter: This parameter indicates the total duration of tasks that the candidate approver has completed in the past. It is necessary to retrieve all closed or archived tasks of the person in the system, and then extract the actual execution duration for each task record. These duration values ​​are added together to get the preliminary value of P. To avoid interference, abnormal records need to be excluded, such as time periods caused by long interruptions in the log or tasks with missing records. Multiple cross-checks are performed to ensure that only effective processing times are included. For example, a candidate approver completed a total of 8 tasks during the statistical period, which took 1.2 hours, 3.0 hours, 2.5 hours, 2.0 hours, 4.0 hours, 3.5 hours, 2.2 hours and 2.6 hours respectively. These numbers can be added together to get P=1.2+3.0+2.5+2.0+4.0+3.5+2.2+2.6=21.0 hours.

[0073] Steps to obtain Q parameters: This parameter represents the estimated duration of the current pending task. It needs to be estimated before the task is assigned based on the task's own operating procedures, workload, existing historical processing data and other information. For example, in the approval system, a typical processing time is set for a process. The median value of similar tasks can be used to represent the average time required for the task from receipt to completion. Then, it can be appropriately increased or decreased based on the priority of the task or cross-departmental needs. For example, a cross-departmental financial approval generally takes an average of 2.5 hours, but the current task involves contract confirmation, which requires additional query of supplier qualifications, which may add 1.0 hour to the basic time, so Q can be 3.5 hours.

[0074] Steps to obtain R parameters and S parameters: These two parameters represent the task switching time of candidate approvers. Usually, it is necessary to monitor the interval between the completion of the previous task and the formal processing of the next task between different tasks. It can be calculated based on the timestamp recorded in the system operation log. Once the previous task enters the archive state and the time difference between the "start processing" mark of the next task is a switching duration. R and S can be reflected in the switching average values ​​of different types of tasks respectively. For example, R refers to the average switching time from ordinary approval to expedited approval, and S refers to the average switching time from expedited approval to cross-departmental approval. These average values ​​can be obtained by summarizing multiple switching processes during the statistical period and dividing them by the number of switches. If the candidate approver has experience in multiple types of approvals and their common switching times are between 15 minutes and 30 minutes, these times can be converted into hours and then recorded as R and S respectively. For example, the calculated value is R=0.3 hours and S=0.4 hours.

[0075] Steps to obtain the W parameter: This parameter is the standard processing time of the candidate approver's historical tasks. It is necessary to count the actual processing time of the candidate approver for several tasks of the same or similar type in the past, extract these processing times and make an arithmetic average or median. Determine the treatment method for outliers based on statistical principles. If it is found that some tasks are affected by external factors and the processing time is abnormally long or short, they need to be eliminated or corrected by mean regression. After selecting the qualified task records, add their processing time and divide it by the total number of records to get a basic average time. Then, fine-tune the value based on the timeliness assessment given by the department or enterprise. For example, a large number of statistics show that the median processing time of this person for ordinary approval tasks is about 2.5 hours, then W can be set to 2.5, or it can be corrected through more samples.

[0076] Steps to obtain X and Y parameters: These two parameters represent the adjustment parameters of task complexity. The current pending tasks need to be disassembled, and factors such as the number of main sub-links, document review requirements, or the number of cross-departmental collaborations are listed. Numerical values ​​are assigned to these factors in a quantitative manner and accumulated or normalized to obtain X or Y. For example, X is specifically for the number of sub-links for document verification, and Y is for the number of cross-departmental or external collaboration links. The larger the sum of the two, the higher the overall difficulty of the task. For comparison, refer to the 50 similar tasks in the sample. If it is found that most tasks require checking 2 to 3 documents and departmental cooperation 1 to 2 times, it can be regarded as a normal range of difficulty. If this task requires checking 5 documents and cross-departmental communication 4 times, the sum of X plus Y will increase significantly. When the statistics are 2 or 3, it can be understood as medium complexity, and 5 or 6 are higher complexity.

[0077] Steps to obtain Z parameters, A parameters and B parameters: These three parameters represent the task processing efficiency of candidate approval personnel in different time periods and are used to add up the denominators. It is necessary to distinguish and count the completion efficiency of personnel in the morning shift, ordinary working hours, and evening or overtime hours. The number of completed tasks in each time period is divided by the total effective working time of the time period to obtain a time period efficiency value, which is finally mapped to the same interval and recorded in Z, A, and B respectively. For example, Z represents the efficiency from 8:00 to 12:00, A represents the efficiency from 13:00 to 18:00, and B represents the efficiency after 19:00 or on weekends. If the person completes tasks quickly from 8:00 to 12:00, and can complete 0.4 items per hour within 2 weeks, it can be mapped to 0.75 after comparing the average efficiency of the department. If 0.3 items can be completed per hour from 13:00 to 18:00, it can be assigned a value of 0.60. If 0.2 items can be completed per hour in the evening, it can be assigned a value of 0.45.

[0078] Calculation process: P = 21.0 hours (the total time spent by the candidate reviewer on completing tasks during the statistical period); Q = 3.5 hours (the estimated time required for the current pending tasks); R=0.3 hours, S=0.4 hours (average duration of switching between different types of tasks); W = 2.5 hours (standard processing time calculated from historical records); X=2, Y=1 (quantitative parameters obtained after decomposing the task complexity); Z=0.75, A=0.60, B=0.45 (efficiency mapping values ​​for each period); First calculate the numerator: ; ; ; Add these three terms together: ; Then calculate the denominator: ; so: ; The result shows that the candidate approver estimates the overall task completion time to be about 10.38 hours after taking into account the total time of previously completed tasks, the time spent on the current pending tasks, the task switching time, and the processing efficiency in different time periods. If another candidate approver is found in the subsequent comparison process A lower value indicates that the person may be able to complete the task faster and may be given priority when it is assigned.

[0079] Based on the task completion time, first check the history records of each candidate approver and compare them with the previously calculated Numerical value, all candidate reviewers are arranged in order from low to high. At this time, it is necessary to conduct additional review on the stability of the same person in different types of tasks. For example, if someone is very familiar with contract approval work, then when handling the same category It may drop significantly if another person is good at financial reimbursement process and faces financial related approvals. The data is too low, so when collecting statistics, we need to pull out their professional proficiency descriptions of the corresponding task categories, and check the completion time period data of the corresponding task categories in their historical approvals. If someone has a long processing time or cross-departmental communication that exceeds the predetermined standard for many consecutive times, it is necessary to track the time-consuming frequency of their communication. If the average communication delay of a person is within 0.5 hours each time, it is considered relatively stable, but if it exceeds 1.0 hours, it is marked as communication delay behavior. After confirming this information, combined with the processing timeliness requirements, if some approvals have urgent needs within 48 hours, we will focus on those For candidates whose task time is less than 10 hours, we will prioritize their overall performance on such tasks in their past execution records. Finally, we will integrate various analyses together, compare the candidate approvers’ task category adaptability, past records and corresponding time requirements, and thus obtain task processing capability indicators.

[0080] The steps to obtain the approval task allocation sequence are: Based on the task processing capability index, the task processing capabilities of all candidate approvers are extracted, the task saturation of the candidate approvers is called, and the current number of tasks and task load of each candidate approver are analyzed to obtain the task status information of the candidate approver; According to the task status information of candidate approvers, match and analyze the task processing capability index and task saturation of candidate approvers, analyze the acceptable number of tasks for each candidate approver, and prioritize the candidate approvers based on the acceptable number of tasks to generate a priority sequence for task allocation for candidate approvers; Based on the priority sequence of candidate approver task assignment, candidate approvers are selected in turn according to the execution requirements of the approval tasks, the approval tasks are assigned in order of priority, and the task assignment sequence is recorded to generate the approval task assignment sequence.

[0081] Specifically, based on the task processing capability index, first centrally read the historical and current work records of all candidate approvers, and check their average processing time for each task type and the recent processing speed based on existing statistical data. Then, compare the task saturation of each candidate approver. The saturation generally needs to be summarized from aspects such as the number of ongoing tasks, the number of entrusted tasks, and the proportion of expedited tasks in the recent period of time. To this end, you can first query each person's current to-do list in the database and mark the tasks that have been assigned but not yet started, and then summarize the task completion status in each time period based on the data from the past few weeks or months. If the average number of ongoing tasks for the approver in the selected reference period is around 5 and at least 2 of them are in high priority or cross-departmental collaboration scenarios, then its saturation is defined in a higher range. If you need to specifically divide it into high, medium, and low saturation levels, you can refer to statistical methods. Perform quantile analysis on the data of all approvers. For example, take 1.2 times the median of the past processing records as the dividing line, and classify the approvers who exceed this value as high saturation, those who fall within a certain range below this value as medium saturation, and those who are significantly lower than this value as low saturation. If an approver remains in the high saturation segment for a long time, it means that the workload he is currently undertaking is relatively large. After the statistical saturation is completed, the existing processing capacity value of the corresponding approver and its saturation will be viewed side by side. If the capacity value of an approver is high but the saturation is close to the critical value, it may not be appropriate to arrange more tasks. If there are other approvers with medium capacity values ​​and low saturation, they may be more suitable to undertake some new tasks in the current period. After the data of all candidate approvers are aggregated by personnel number or name, they are brought into the software for comprehensive sorting of their current number of tasks, time distribution of completed tasks, and saturation indicators, and finally the task status information of the candidate approvers is obtained.

[0082] According to the task status information of the candidate approvers, the ability index value and saturation value of each candidate approver are first extracted and matched and analyzed. Specifically, the ability index of a single approver can be combined with its current saturation by comparing them one by one. If it is in the area of ​​high ability index and low saturation, it can be preliminarily determined that it has the potential to accept more tasks. If it is in the area of ​​medium ability index and high saturation, it indicates that the approver has a lot of work pressure in the near future and needs to be cautious in adding tasks. In order to clarify the number of tasks that each approver can accept, the processing rate of the tasks completed in the past three weeks can be referred to, divided into hours or days, and the task processing rate can be compared with the sum of the remaining workload of the current tasks. If the rate is significantly greater than the median value of the current remaining workload, it means that The number of acceptable tasks can be maintained at around 2 to 3, otherwise, only one more task may be accepted or no new tasks may be accepted. Such a judgment process can be cross-validated in combination with several sample intervals. For example, the processing rate curve of the approver can be observed in the last two weeks and the two weeks before that. If the difference between the two is less than 10%, the evaluation conclusion of the acceptable number of tasks is confirmed to be relatively stable. If the difference exceeds 20%, an additional correction should be set in the calculation. After completing this judgment, the acceptable number of tasks for all candidate approvers shall be recorded in sequence, and these approvers shall be prioritized from high to low according to the acceptable number of tasks. If multiple approvers are ranked in a tie, their processing speeds in key category tasks can be compared to finally generate a priority sequence for the task allocation of candidate approvers.

[0083] Based on the priority sequence of candidate approvers' tasks, first match the candidate approvers one by one according to the execution requirements of the approval tasks, retrieve the highest priority approver list from the first in the sequence and compare their available time and remaining energy for the same type of tasks in the current cycle. If the number of tasks they can accept is still available and consistent with the target task category, assign this task to the approver and register the assignment. Then check the next approver in the sequence, continue to match and assign the same or similar types of approval items. When encountering tasks that require cross-departmental co-location or involve special data review, the completion time records of the approver's previous handling of such tasks will be called for confirmation. If it is found that the approver has worked overtime frequently, If the completion time is too high or often higher than the benchmark time set by the department for the same type of tasks, it will be appropriately avoided or the priority will be lowered in the allocation. You can also pre-set a number of hours to mark this benchmark time limit. For example, in most ordinary approval processes, the benchmark time limit is 24 hours, but in cross-departmental processes, it can be appropriately increased to 36 hours to see whether the approval person's previous similar completion time has obviously deviated from this benchmark. If the deviation is more than 6 hours for multiple times, its sequence priority will be automatically lowered. According to this matching process, all approval tasks that have yet to be assigned will select the corresponding approvers in turn. After all tasks have successfully completed the allocation process, the corresponding assignment order of each task will be recorded, and finally the approval task allocation sequence will be generated.

[0084] The steps to obtain approval task assignment records are as follows: Based on the approval task allocation sequence, extract the information of the first approver in the approval task allocation sequence, call the identity information and position information of the approver, and verify the approval task allocation sequence to generate the first allocation information of the approval task; According to the first allocation information of the approval task, match the file identification of the file to be approved, call the approval file content and allocation time data, bind the approval file to the first approver, record the allocation of the approval file, and generate the approval task allocation binding information; Based on the approval task allocation binding information, an approval task allocation record is established to store the approval document allocation time, document identification and first approver information, and the approval task management database is updated synchronously.

[0085] Specifically, based on the approval task allocation sequence, first read the list of approved personnel in order and locate the approved personnel at the top of the sequence, then compare the approved personnel's identity number, position number, and related department code registered in the system, and retrieve their verified records in the identity database in character form. If the matching degree of the approved personnel's identity information is greater than the pre-established benchmark value, it is considered that the information is consistent. If the corresponding benchmark is not reached, it is marked as identity pending confirmation. The benchmark value is generally obtained by continuously observing the identity proofreading results of multiple approved personnel and calculating the similarity. For example, when comparing the identity tags of 50 approved personnel with the system records, it is found that more than 90% of the matches can be successfully identified. Therefore, 90% can be selected as this benchmark, and then continue to retrieve the corresponding position name and authority range of the approved personnel in the position description data set, and combine the authority mapping table collected in the early stage to determine whether they have the processing authority involved in this approval process. If the authority code is consistent with the process requirements, it will be recorded as an executable position. If If it is found that there are restrictive clauses in the job description, such as only being able to handle financial tasks while the current task is contract approval, it will be marked as inappropriate for the position and the assignment will be abandoned in subsequent execution. After the position confirmation is completed, the approval task allocation sequence needs to be further verified, and the number of the approver is distinguished from the remaining subsequent personnel in the sequence to prevent duplicate assignments. Then, the internal contact file is used to query the available time period in the current working day or week. If there is a serious conflict with the task requirements, it will be prompted in the analysis results. The conflict threshold can be obtained from the scheduling tolerance range agreed upon within the department. For example, some departments stipulate that any overlap with the duration of a high-priority project of 80% or more is considered a serious conflict. The statistical method can be combined with the task schedule of the approver. If it is found that the overlap reaches or exceeds 80%, it is listed as unavailable. If it is less than 80%, it can still be assigned. After all verifications are completed, the number, identity description and position information of the approver used in this assignment are sorted and recorded to obtain the first allocation information of the approval task.

[0086] According to the first allocation information of the approval task, first retrieve the file identification list in the current system that is in the pending approval status. Each file has a unique identification number and its basic attributes in the archive record, such as file type, information about the person to be in charge, etc. Then, use the file identification number as the primary key to query the corresponding approval file content in the database, and confirm whether it contains fields such as contract attachments or amount clauses. If it is detected that these fields are consistent with the corresponding relationship of job permissions, further procedures can be entered. If the job permissions do not support processing such fields, they should be marked as unbindable in subsequent records. After the matching is completed, the allocation time data of the current system is extracted. The allocation time can be combined with the allocation timestamp of the superior scheduling center to form an accurate task assignment. At that moment, if it is found that the stored timestamp deviates greatly from the regular time allocation interval (for example, more than 2 hours, while the acceptable interval set within the department is usually between 0 and 2 hours, and this value is obtained by taking the mean plus the standard deviation of the actual time consumed when dozens of tasks were issued in the past), then this record will be marked as an abnormal time point. If the difference between it and the acceptable interval specified by the department does not exceed 2 hours, it will be regarded as a normal allocation operation. Subsequently, the approval document will be associated one-to-one with the first allocation information of the current approval task, and its key information in the binding process will be summarized, including the approval personnel number, job title, file identification and specific allocation time. After the summary is completed, the corresponding database records will be updated to obtain the approval task allocation binding information.

[0087] Based on the binding information of the approval task allocation, we first collect multiple data items including the approval document allocation time, document identification and the number of the first approver, and put them together in an approval process tracking table for sequential arrangement. We also classify and mark these binding information according to different stages such as being processed and completed. If the file has entered a certain collaborative stage in the previous link, its temporary storage status will also be translated into "cross-departmental in progress". If it is currently in this state, it will not be updated repeatedly to prevent overwriting of existing process traces. Then, using the system serial number as the retrieval clue, we will associate all the newly created binding information with the identity ID of the approver, and write these merged data into the approval task management system. In the management database, each piece of data carries its effective time to indicate the precise recording point. The recording method can be to design an extended field in the database to store the time. If it is necessary to summarize all the allocation behaviors within a certain period of time during statistics, the corresponding list can be directly obtained through time query. Here, a comparison is made based on the previously stored allocation time and the current server time. If the time difference is between 0 and 24 hours, it is listed as a new allocation for the day. If it exceeds 24 hours, it is classified as a historical allocation. After grouping the allocation of each category, it can also be provided to management personnel for review. When the above grouping and sorting are completed, the database index list is updated and this round of operations is ended to obtain the final approval task allocation record.

[0088] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A document approval method based on NLP, characterized in that: The following steps are involved: Input the approval document, perform word segmentation on the content of the approval document, calculate the similarity between the word segmentation result and the preset business type dictionary, and generate the file business classification result; based on the file business classification result, perform semantic vector conversion on the keywords in the approval document content, calculate the semantic association strength between the keywords, and generate the file semantic feature vector; Based on the document semantic feature vector, the responsibilities of each department and the responsibilities of each person in the organizational structure are called, the matching values ​​of the vector content and the responsibilities of the department and the person are calculated, and the responsibility matching results are generated; Based on the responsibility matching results, the current number of tasks and task processing time of each approval personnel are counted, and the approval personnel matching index is calculated and generated; Based on the approval personnel matching index, a task allocation threshold is set, candidate approvers are screened according to the task allocation threshold, and a task processing capability index is generated; based on the task processing capability index and combined with the task saturation of the candidate approvers, an approval task allocation sequence is generated; Based on the approval task allocation sequence, the information of the first approver in the sequence is extracted, the approval document is allocated to the first approver, the allocation time and the document identifier are recorded, and an approval task allocation record is generated.

2. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the file service classification result are: After the approval document is input, it is segmented through natural language processing, the text in the document is semantically analyzed, and the sentences are separated into independent vocabulary units to form the segmentation results; According to the word segmentation result, the similarity between each word segmentation element and an entry in a preset business type dictionary is calculated to obtain a similarity result; Based on the similarity result, the similarity scores of each business type are compared, and the one with the highest similarity score is selected as the type identifier of the file to generate a file business classification result.

3. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the file semantic feature vector are as follows: Based on the file business classification result, the approval file content is parsed, keywords in the approval file content are extracted, keywords are screened, redundant keywords, synonyms or interference items are removed, and keywords are normalized to obtain a keyword list; According to the keyword list, the semantic association strength between the keywords is calculated, and the calculation formula is: ； in, Represents the strength of semantic association between keywords. Representative keywords The frequency of occurrence in the text, and Respectively represent and The semantic vector value of keywords, and Respectively represent and The relative position deviation of keywords in the context window, and Represents the semantic dependency distance of keywords in the text, and Represents the dependency weight of keywords in different syntactic structures, and Represents the number of times a keyword appears in different text fragments; Based on the semantic association strength, a file semantic feature vector is constructed.

4. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the responsibility matching result are as follows: Based on the document semantic feature vector, call the responsibilities of each department and the responsibilities of each person in the organizational structure, parse the responsibilities description text, extract the responsibilities keywords, construct the responsibilities keyword vector, and obtain the responsibilities keyword vector matrix; According to the responsibility keyword vector matrix, the matching value between the document semantic feature vector and the responsibility keyword vector is calculated, and the expression is: ； in, Represents the matching value, is the first part of the document semantic feature vector Dimensional components, is the first Dimensional components, and Represent the relative distances between document keywords and responsibility keywords in the semantic space, For the The frequency value of the matching words, and Represents the position index of the document keywords and responsibility keywords in the syntactic structure, , and Represents the matching similarity coefficient between each job description text, Represents the number of dimensions of the document semantic feature vector, Represents the number of matching words in the responsibility keyword vector, The total number of job description texts; Based on the matching value, a matching evaluation is performed on the job descriptions of each department and the job descriptions of personnel positions, and the responsibility item with the highest matching value is selected as the belonging department and responsible person to obtain a responsibility matching result.

5. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the approval personnel matching index are as follows: Based on the responsibility matching results, the current number of tasks of each approver is counted, and the cumulative task processing time of each approver's processed tasks is calculated. At the same time, the list of pending tasks of the approver is extracted to obtain the task statistics information of the approver; According to the task statistics of the reviewers, the task processing mode of each reviewer is analyzed, the average number of tasks completed per unit time is calculated, and the current task saturation degree of each reviewer is counted to obtain the reviewer task load index; Based on the task load index of the approval personnel, the matching index of the approval personnel is calculated. The calculation formula is: ； in, Represents the approval personnel matching index, The task processing time of the approved personnel who have completed the task. The current number of tasks for the approver. Represents the number of tasks to be processed by the approver. Represents the average completion time of historical tasks of the approver. Represents the task processing efficiency of the approver on different task types, Represents the standard deviation of task processing time of approvers in different task categories. and Represents the stability coefficient of task completion of the approver in different time periods, Represents the number of task records of the approver. Represents the number of task categories for approvers.

6. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the task processing capability indicator are as follows: Based on the matching index of the approver, a task allocation threshold is set, the matching indexes of all approvers are screened, and the approvers whose matching indexes are higher than the task allocation threshold are extracted to obtain a list of candidate approvers; According to the list of candidate reviewers, the task completion time of each candidate reviewer is calculated, and the calculation formula is: ； in, Represents the task completion time of the candidate approver, is the total time that candidate approvers have completed tasks in the past. Represents the estimated duration of the current pending task. and Indicates the task switching time of candidate approvers. The standard processing time of the candidate approver's historical tasks. and A tuning parameter representing the complexity of the task, , and Represents the task processing efficiency of candidate approvers in different time periods; Based on the task completion time, the task processing capability of the candidate approver is evaluated, and the task processing capability index is obtained by combining the task category, processing timeliness and execution record.

7. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the approval task allocation sequence are as follows: Based on the task processing capability indicator, the task processing capabilities of all candidate approvers are extracted, the task saturation of the candidate approvers is called, and the current number of tasks and task load of each candidate approver are analyzed to obtain the task status information of the candidate approver; According to the task status information of the candidate approvers, matching analysis is performed on the task processing capability index and task saturation of the candidate approvers, the acceptable number of tasks for each candidate approver is analyzed, and the candidate approvers are prioritized based on the acceptable number of tasks to generate a task allocation priority sequence for the candidate approvers; Based on the candidate approver task allocation priority sequence, candidate approvers are selected in turn according to the execution requirements of the approval tasks, the approval tasks are allocated in order of priority, and the task allocation order is recorded to generate an approval task allocation sequence.

8. The NLP-based document approval method according to claim 1, characterized in that: The steps for obtaining the approval task allocation record are as follows: Based on the approval task allocation sequence, extract the information of the first approver in the approval task allocation sequence, call the identity information and position information of the approver, and verify the approval task allocation sequence to generate the first allocation information of the approval task; According to the approval task first allocation information, the file identifier of the to-be-approved file is matched, the approval file content and allocation time data are called, and the approval file is bound to the first approver, the allocation of the approval file is recorded, and the approval task allocation binding information is generated; Based on the approval task allocation binding information, an approval task allocation record is established to store the approval document allocation time, document identification and first approver information, and the approval task management database is updated synchronously.

Images