Intelligent approval decision method and system for government affairs process
By quantifying the credibility scores of labels in the history of government approvals, the problem of unrecognized differences in the credibility of historical approval conclusion labels was solved. A weighted loss function was constructed to optimize the prediction network, thereby improving the noise resistance and prediction accuracy of the government approval model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-06
- Publication Date
- 2026-07-24
AI Technical Summary
In the existing government approval system, the credibility differences of historical approval conclusion labels are not quantitatively identified, which leads to undue interference during model training and affects the accuracy of prediction conclusions.
By calculating the consistency score of the approval basis text in terms of timeliness, the consistency score among approvers, and the consensus score among employees, a label credibility score is generated and used as a loss weight in the training process to construct a weighted loss function to optimize the prediction network.
This improves the government approval model's ability to resist label noise interference, ensuring that the decision reference data is based on information selected for credibility in terms of both prediction probability and historical case support, thereby enhancing the reliability of prediction conclusions.
Smart Images
Figure CN122453355A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to an intelligent approval decision-making method and system for government processes. Background Technology
[0002] With the deepening of e-government construction, local government approval systems have accumulated a large amount of historical approval data. Researchers have begun to introduce machine learning methods into the government approval process, building automated prediction systems for application materials by modeling and training on historical approval data. Existing technologies typically use approval conclusion labels in historical approval records as supervisory signals. The application material text is encoded through natural language processing and input into a classification model. The model parameters are optimized using the standard binary cross-entropy loss function to predict the approval conclusion of newly submitted application materials. This technical approach has been initially applied in various government business scenarios such as business registration and construction permits.
[0003] However, existing technologies have a fundamental flaw: they default to treating all historical approval conclusion labels as equivalent truth monitoring signals, ignoring the inherent differences in the credibility of these historical labels. Approval conclusion labels in historical government approval data are actually affected by multiple confounding factors: First, the text supporting approval changes substantially over time, significantly reducing the reference value of outdated historical approval conclusion labels for current model training; second, the same approver may issue contradictory approval conclusions for semantically similar application materials at different times, reflecting the internal instability of individual approver judgment standards; third, different approvers show significant differences in the pass rates for similar business transactions, reflecting inconsistencies in approval standards across employees. These three types of confounding factors are not identified or quantified in existing technologies, causing the model to assign the same gradient weights to low-credibility and high-credibility historical labels during training. Low-credibility historical labels have an impact on the direction of model parameter updates that is disproportionate to their actual reference value, ultimately leading to the systematic contamination of the trained model's predictions by historical approval behavior biases. Summary of the Invention
[0004] This application provides an intelligent approval decision-making method and system for government processes, which solves the technical problems in the prior art where the credibility differences of historical approval conclusion labels cannot be quantitatively identified, resulting in undue interference of low-credibility historical labels on model training, and the technical problem that the reference value of historical cases in the inference stage cannot be reflected in the output of decision reference data according to credibility weight.
[0005] Firstly, this application provides an intelligent approval decision-making method for government processes, the intelligent approval decision-making method for government processes including: Step S1: Extract each historical approval record from the government approval history database. Each historical approval record includes the application material text, approval conclusion label, approver number, approval basis text and business type number, to obtain a set of historical approval records. Step S2: For each historical approval record in the set of historical approval records, calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. Step S3: Substitute the label credibility score as the loss weight into the weighted loss function, use the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to train the prediction network to obtain the approval prediction network. Step S4: Input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability. Retrieve several historical approval records with the highest semantic similarity from the historical approval record set. Weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate. Output the prediction probability and the weighted historical pass rate together as decision reference data.
[0006] Secondly, this application provides an intelligent approval decision-making system oriented towards government processes, the intelligent approval decision-making system oriented towards government processes includes: The extraction module is used to extract historical approval records from the government approval history database. Each historical approval record contains the application material text, approval conclusion label, approver number, approval basis text and business type number, resulting in a set of historical approval records. The weighting module is used to calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number for each historical approval record in the set of historical approval records. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. The input module is used to substitute the label credibility score as the loss weight into the weighted loss function, and to train the prediction network with the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to obtain the approval prediction network. The retrieval module is used to input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability, retrieve the historical approval records with the highest semantic similarity from the historical approval record set, weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate, and output the prediction probability and the weighted historical pass rate together as decision reference data.
[0007] Thirdly, an intelligent approval decision-making device for government processes is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the intelligent approval decision-making device for government processes to execute the aforementioned intelligent approval decision-making method for government processes.
[0008] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned intelligent approval decision-making method for government processes.
[0009] The technical solution provided in this application retains three types of metadata fields—approver number, approval basis text, and business type number—simultaneously during the construction phase of the historical approval record set. This provides a complete data foundation for the subsequent calculation of tag credibility scores. This design ensures that historical approval conclusion tags are no longer isolated binary signals, but rather form a traceable data association with the identity of the approval entity that generated the tag, the content of the basis text, and the business category. Based on this, the timeliness consistency score quantifies the reliability of historical tags from the perspective of semantic offset of the approval basis text; the internal consistency score of approvers quantifies the strength of individual subjective bias from the perspective of the stability of a single approver's judgment on similar materials; and the cross-employee consensus score quantifies the impact of group standard divergence from the perspective of the dispersion of the pass rate of different approvers under the same business. These three dimensions establish corresponding quantitative channels for three types of independent interference sources that objectively exist in the government approval scenario. After being weighted and summed to synthesize the tag credibility score for each record, the approval conclusion tag of each record in the historical approval record set obtains a credibility measure that matches its actual reference value, fundamentally breaking the implicit assumption of treating all historical tags equally in existing technologies.
[0010] The training mechanism, which substitutes label credibility scores as loss weights into the weighted loss function, ensures that the parameter update direction of the approval prediction network is dominated by high-credibility historical approval records. The gradient contribution of low-credibility historical approval records is actively suppressed to a level commensurate with their actual reference value. This algorithmic feature is particularly crucial in the specific application domain of government approvals. Compared to general text classification scenarios, the source of label noise in historical government approval data has clear business semantic interpretability. It is this interpretability that allows the cosine similarity of text semantic vectors, the judgment dispersion in the approver number dimension, and the pass rate standard deviation in the business type number dimension to serve as effective credibility metrics in the loss function. The design constructs data rather than relying on domain-independent statistical assumptions. In the inference stage, the predicted probability and weighted historical pass rate are used together as decision reference data output. This design further extends the distinguishing ability of the label credibility score to the historical case retrieval stage, making the contribution of high credibility cases in semantically similar historical approval records to the weighted historical pass rate proportional to their credibility scores. The statistical weight of low credibility cases is compressed accordingly. The final output decision reference data is based on information that has been filtered for credibility in both the predicted probability and historical case support dimensions. This gives the application the ability to resist label noise interference that existing technologies do not have in the application scenario of government approval, which is highly sensitive to the quality of historical data. Attached Figure Description
[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a schematic diagram of one embodiment of the intelligent approval decision-making method for government affairs processes in this application. Figure 2 This is a schematic diagram illustrating the cumulative distribution of tag credibility scores under different business type numbers in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the joint distribution of the predicted probability and weighted historical pass rate of the application materials to be approved in this embodiment of the application. Detailed Implementation
[0013] This application provides an intelligent approval decision-making method and system for government processes. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.
[0014] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent approval decision-making method for government processes in this application includes: Step S1: Extract each historical approval record from the government approval history database. Each historical approval record includes the application material text, approval conclusion label, approver number, approval basis text and business type number, to obtain a set of historical approval records. Specifically, each historical approval record in the historical approval record set contains five fields: the application material text, which is a textual description of the original application submitted by the applicant; the approval conclusion label, which takes a value of 0 or 1, where 0 indicates that the approval conclusion of the historical approval record is rejection and 1 indicates that the approval conclusion is approval; the approver number, which is a unique identification code of the staff member who performed the approval; the approval basis text, which is the text content of the valid basis registered in the government affairs system at the time the historical approval record occurred; and the business type number, which is a unique index code of the government affairs business category to which the application belongs, used for subsequent grouping and filtering of historical approval records by business category. After word segmentation and stop word filtering of the application material text, it is input into a pre-trained language model for encoding. The resulting semantic vector of the application material text has a dimension of 768, which is consistent with the hidden layer output dimension of the pre-trained language model. The approval basis text is encoded in the same way, and the resulting semantic vector of the approval basis text also has a dimension of 768.
[0015] Step S2: For each historical approval record in the historical approval record set, calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. Specifically, the label credibility score is the core data object of this invention. Its physical meaning is a quantitative value representing the credibility of the approval conclusion labels of each historical approval record when used as training and supervision signals. The value ranges from 0 to 1; a higher score indicates less interference with the approval conclusion labels of that historical approval record and greater reliability as training signals. The timeliness consistency score is obtained by calculating the cosine similarity between the semantic vectors of historical approval basis text and the semantic vectors of currently valid approval basis text. A higher cosine similarity indicates that the content of the historical approval basis and the currently valid basis is closer, and the corresponding historical label's timeliness credibility is higher. The internal consistency score of approvers is obtained by statistically analyzing the proportion of different approval conclusions made by the same approver for semantically similar application materials in the same type of business and subtracting this proportion from 1. A lower proportion indicates that the approver's judgment is more stable, and the corresponding label credibility is higher. The cross-employee consensus score is obtained by calculating the standard deviation of the pass rate of each approver in the same type of business and subtracting this standard deviation from 1. A smaller standard deviation indicates that the standards of each approver for the same type of business are more consistent, and the corresponding label credibility is higher. In the weighted sum of the three scores, the preset weight coefficient for the timeliness consistency score is set to 0.5, the preset weight coefficient for the internal consistency score of approvers is set to 0.3, and the preset weight coefficient for the cross-employee consensus score is set to 0.2. The sum of the three weight coefficients is 1. The reason why the timeliness consistency score has the highest weight is that the change of the content of the approval basis in the government approval scenario has the most fundamental impact on the credibility of historical tags. The greater the deviation between the historical basis and the current valid basis, the lower the reference value of the corresponding historical tag for the current training.
[0016] Step S3: Substitute the label credibility score as the loss weight into the weighted loss function, use the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to train the prediction network, and obtain the approval prediction network. Specifically, the weighted loss function is a weighted binary cross-entropy loss function with label confidence scores as weights. Its calculation method is as follows: First, calculate the standard binary cross-entropy loss value for the predicted probability of the approval conclusion and the approval conclusion label for each historical approval record. Then, multiply the binary cross-entropy loss value of each historical approval record by the corresponding label confidence score. Finally, average the weighted loss values of all historical approval records to obtain the final weighted loss value. Historical approval records with label confidence scores close to 1 contribute to the gradient with almost complete weight, while historical approval records with label confidence scores close to 0 contribute more to the gradient. New contributions are actively suppressed to near zero, so that the parameter update direction of the approval prediction network is dominated by high-confidence historical approval records, while low-confidence historical approval records are still retained in the training set to maintain coverage of rare business scenarios but do not have a substantial impact on the parameter convergence direction. The prediction network adopts a two-layer fully connected structure. The first layer has an input dimension of 768, an output dimension of 256, and an activation function of ReLU. The second layer has an input dimension of 256, an output dimension of 1, and an activation function of Sigmoid. The training iteration stops when the weighted loss value no longer decreases for 5 consecutive rounds, thus obtaining the approval prediction network.
[0017] Step S4: Input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability. Retrieve the historical approval records with the highest semantic similarity from the historical approval record set. Weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate. Output the prediction probability and the weighted historical pass rate together as decision reference data.
[0018] Specifically, the text of the application materials to be approved undergoes the same word segmentation, stop word filtering, and pre-trained language model encoding processing as in step S1 to obtain a 768-dimensional semantic vector of the application materials to be approved. This vector is input into the approval prediction network and, after forward propagation, outputs a prediction probability ranging from 0 to 1. When retrieving the historical approval records with the highest semantic similarity from the historical approval record set, the cosine similarity of the semantic vector of the application materials to be approved with the semantic vector of the application materials of each historical approval record in the historical approval record set is calculated one by one. After sorting the records in descending order of cosine similarity, the top 10 records are selected as a subset of semantically similar historical approval records. The reason for setting the number of records to 10 is that this number covers sufficient historical reference cases while avoiding... A balance is achieved by introducing low-relevance historical records; the weighted historical pass rate is calculated by summing the product of the label credibility score of each historical approval record in the semantically similar historical approval record subset and the approval conclusion label, and then dividing by the sum of the label credibility scores of each historical approval record. The value ranges from 0 to 1. This value reflects the weighted pass rate of the historical approval case that is semantically closest to the application materials to be approved after excluding the interference of low-credibility labels; the decision reference data consists of two values: the predicted probability and the weighted historical pass rate. The former is the fitting output of the approval prediction network for the current application materials, and the latter is the statistical support of high-credibility historical cases for similar applications. Together, they provide objective data basis for the subsequent approval process.
[0019] In one specific embodiment, step S1 includes: The original application materials, original approval basis, approval conclusion label, approver number, and business type number of each historical approval record are extracted from the government approval history database to obtain the original data of each historical approval record. The original application material texts of each historical approval record are segmented and filtered for stop words. The filtered texts are then input into a pre-trained language model for encoding to obtain the semantic vectors of the application material texts of each historical approval record. The original approval basis text of each historical approval record is input into a pre-trained language model for encoding processing to obtain the semantic vector of the approval basis text of each historical approval record; By integrating the semantic vectors of the application materials, the semantic vectors of the approval basis, the labels of the approval conclusions, the approver numbers, and the business type numbers of each historical approval record, a set of historical approval records is obtained.
[0020] Specifically, the original application material text is the complete text content of the application submitted by the applicant to the government approval system, and the original approval basis text is the complete text content of the valid basis text registered in the government system at the time the historical approval record occurred. Both are stored in the government approval history database in the form of strings. When performing word segmentation on the original application material text, a dictionary-based forward maximum matching word segmentation algorithm is used to segment the original application material text into word sequences. Then, based on a preset stop word list, stop words in the word sequences are filtered to remove auxiliary words, modal words, and high-frequency meaningless words that do not contribute substantially to semantic features. The resulting effective word sequences are then input into a pre-trained language model. The pre-trained language model uses a BERT model fine-tuned on government affairs texts. The hidden layer output corresponding to the [CLS] marker is taken as the semantic representation of the text, with an output dimension of 768, resulting in a 768-dimensional semantic vector of the application material text. The same pre-trained language model is used to perform encoding processing on the original approval basis text. Similarly, the 768-dimensional hidden layer output corresponding to the [CLS] marker is taken to obtain a 768-dimensional semantic vector of the approval basis text. The original approval basis text does not need to undergo word segmentation and stop word filtering because the approval basis text is a highly structured and standardized text. Preserving the complete text input encoding helps to capture its complete semantic features.
[0021] The 768-dimensional semantic vectors of application materials, the 768-dimensional semantic vectors of approval basis, the approval conclusion labels (with values of 0 or 1), the string-type approver number, and the integer-type business type number of each historical approval record are integrated at the field level on a record-by-record basis. Each historical approval record is integrated to form a structured data object containing the above five types of fields. All structured data objects of historical approval records are organized in the order of record index to obtain a set of historical approval records. The dimensions of the semantic vectors of application materials and the semantic vectors of approval basis in the historical approval record set are unified to 768 dimensions. This dimension setting is consistent with the hidden layer output dimension of the pre-trained language model, and no additional dimension alignment processing is required. The approval conclusion label, approver number, and business type number are stored with the original field values.
[0022] In one specific embodiment, step S2, calculating the consistency score of the approval basis text based on the timeliness of approval, includes: Extract the text of the current valid approval basis from the current valid approval basis database, input the text of the current valid approval basis into the pre-trained language model for encoding processing, and obtain the semantic vector of the text of the current valid approval basis. Input the semantic vector of the approval basis text of each historical approval record and the semantic vector of the current valid approval basis text into the cosine similarity calculation unit. Perform a dot product operation on the semantic vector of the approval basis text of each historical approval record and the semantic vector of the current valid approval basis text, and divide by the product of their magnitudes to obtain the cosine similarity value between the semantic vector of the approval basis text of each historical approval record and the semantic vector of the current valid approval basis text. The cosine similarity value of each historical approval record is directly used as the basis timeliness consistency score for each historical approval record, thus obtaining the basis timeliness consistency score for each historical approval record.
[0023] Specifically, the currently valid approval basis database stores the complete content of the approval basis text registered and effective in the government approval system at the current moment. This database is synchronized in real time with the updates of government approval basis texts, ensuring that the extracted currently valid approval basis text always reflects the latest valid basis content. The currently valid approval basis text is input into the same government domain fine-tuned BERT model as in step S1 for encoding processing. The 768-dimensional hidden layer output corresponding to the [CLS] label is taken to obtain a 768-dimensional semantic vector of the currently valid approval basis text. The semantic vector of the currently valid approval basis text only needs to be calculated once in the same batch of processing. Together with the 768-dimensional approval basis text semantic vectors of each historical approval record in the historical approval record set, it constitutes the two inputs for cosine similarity calculation. Both have a dimension of 768, and the consistency of dimensions ensures the effectiveness of cosine similarity calculation.
[0024] The cosine similarity calculation unit performs a dot product operation on the semantic vector of the approval basis text of each historical approval record and the semantic vector of the current valid approval basis text. The dot product result is divided by the product of the magnitudes of the two vectors to obtain a cosine similarity value ranging from -1 to 1. Since the semantic vector of the approval basis text is obtained by the hidden layer representation of the [CLS] label output of the BERT model through ReLU activation, all its dimensions are non-negative. Therefore, the actual range of the calculated cosine similarity value is narrowed to 0 to 1. The technical reason for directly using the cosine similarity value of each historical approval record as the basis timeliness consistency score is that: the higher the cosine similarity value, the closer the historical basis text on which the historical approval record is based is to the current valid basis text in the semantic space, the smaller the difference between the two contents, the less the historical approval conclusion label is affected by the change of the basis text, and the higher the timeliness reliability as a training supervision signal. Therefore, the cosine similarity value and the basis timeliness consistency score are directly corresponding in physical meaning, without the need for additional mapping transformation. The range of the basis timeliness consistency score is consistent with the cosine similarity value, both being 0 to 1.
[0025] In one specific embodiment, step S2, calculating the internal consistency score of approvers with the same approver number and the same business type number, includes: From the historical approval record set, filter by approver number and business type number to obtain the subset of historical approval records of the same employee and business corresponding to each historical approval record; For each historical approval record in the subset of historical approval records for the same employee and the same business, calculate the cosine similarity of the semantic vectors of the application material texts pairwise. Select the application material texts with a cosine similarity exceeding the preset similarity threshold as semantically similar application material text pairs to obtain the set of semantically similar application material text pairs corresponding to each historical approval record. The number of text pairs with different approval conclusion labels in two historical approval records in the set of semantically similar application material text pairs is counted. The number of text pairs with different values is divided by the total number of text pairs in the set of semantically similar application material text pairs to obtain the inconsistency ratio of approval conclusion labels for each historical approval record. Subtract the proportion of inconsistent approval conclusion labels from 1 to obtain the internal consistency score of the approver for each historical approval record.
[0026] Specifically, the subset of historical approval records for the same employee and the same business refers to the subset of all historical approval records selected from the historical approval record set that have the same approver number as the current historical approval record and the same business type number as the current historical approval record. This selection operation limits both the approver number and the business type number to eliminate the interference of individual differences between different approvers and standard differences between different business types on the consistency calculation, so that the consistency score within the approver only reflects the stability of the same approver's judgment standard in the same type of business. For each historical approval record in the subset of historical approval records for the same employee and the same business, pairwise cosine similarity calculation is performed on the semantic vectors of the application materials. The preset similarity threshold is set to 0.85. This value is set on the basis that application material text pairs with a cosine similarity higher than 0.85 are highly similar in semantic space and can be identified as similar materials describing the same type of application matters. Text pairs with a cosine similarity lower than this threshold do not meet the prerequisite for comparing the consistency of approval conclusions due to large semantic differences. Application material text semantic vector pairs with a cosine similarity greater than 0.85 are selected as semantically similar application material text pairs. All semantically similar application material text pairs constitute a set of semantically similar application material text pairs.
[0027] The criterion for determining different values for approval conclusion labels is that in a semantically similar application material text pair, one historical approval record has an approval conclusion label value of 0 and the other has a value of 1. The number of text pairs that meet this condition is divided by the total number of text pairs in the semantically similar application material text pair set to obtain the approval conclusion label inconsistency ratio. This ratio reflects the frequency with which the same approver makes opposite approval conclusions for semantically highly similar application materials. The higher the ratio, the lower the stability of the approver's judgment. Subtracting the approval conclusion label inconsistency ratio from 1 yields the approver's internal consistency score, which ranges from 0 to 1. The higher the score, the more consistent the approver's approval conclusions are for similar application materials in the same business, the more stable the judgment criteria are, and the higher the credibility of the corresponding historical approval conclusion labels as training supervision signals. When the semantically similar application material text pair set is empty, that is, there are no text pairs with a cosine similarity exceeding 0.85 in the subset of historical approval records of the same employee and the same business, the approver's internal consistency score is set to 1 by default, indicating that no inconsistent judgment behavior can be detected under the current available data conditions, and no penalty is imposed on the label credibility of the historical approval record.
[0028] In one specific embodiment, step S2 calculates the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number, and obtains the label credibility score corresponding to each historical approval record by weighted summation of the three scores, including: Filter the historical approval records by business type number to obtain the subset of historical approval records for the same business; group the subset of historical approval records for the same business by approver number, and divide the number of records with the approval conclusion label value of "pass" in each group by the total number of records in each group to obtain the approver pass rate of each group; Calculate the mean of the approval rate for each group of approvers, square the difference between the approval rate of each group and the mean, take the mean and then take the square root to obtain the standard deviation of the approval rate for each historical approval record; subtract the standard deviation of the approval rate from 1 to obtain the cross-employee consensus score for each historical approval record. The consistency scores of the basis timeliness of each historical approval record, the consistency scores of the approvers within each historical approval record, and the consensus scores of the employees across each historical approval record are multiplied by their respective preset weight coefficients and then summed to obtain the tag credibility score for each historical approval record.
[0029] Specifically, the subset of historical approval records for the same business refers to the subset of all historical approval records whose business type number is the same as that of the current historical approval record, selected from the historical approval record set. This selection operation is limited to the business type number dimension, aiming to horizontally compare the differences in approval standards among different approvers within the same business scope. After grouping the subset of historical approval records for the same business by approver number, the approval rate of each group is calculated by dividing the number of records in the group with an approval conclusion label value of 1 by the total number of records in the group, with a value ranging from 0 to 1. After obtaining the arithmetic mean of the approval rates of each group, the difference between the approval rate of each group and the mean is squared one by one, and the square root is taken after calculating the arithmetic mean of all squared values to obtain the standard deviation of the approval rate. This calculation method is the same as the formula for calculating the population standard deviation in statistics. Since the approval rates of each group... The value ranges from 0 to 1. The absolute value of the difference between the approval rate of each group of approvers and the mean does not exceed 1. Therefore, the standard deviation of the approval rate ranges from 0 to 0.5. The cross-employee consensus score obtained by subtracting the standard deviation of the approval rate from 1 ranges from 0.5 to 1. The smaller the standard deviation, the closer the approval rate of each approver for the same type of business is, the more unified the cross-employee approval standards are, and the higher the corresponding cross-employee consensus score is. The historical approval conclusion label is less affected by the differences in individual standards of approvers and has a higher degree of credibility.
[0030] Of the three preset weighting coefficients, the preset weighting coefficient corresponding to the timeliness consistency score is set to 0.5, the preset weighting coefficient corresponding to the internal consistency score of approvers is set to 0.3, and the preset weighting coefficient corresponding to the cross-employee consensus score is set to 0.2. The sum of the three preset weighting coefficients is 1, which satisfies the normalization constraint of weighted summation. The reason for setting the preset weighting coefficient of the timeliness consistency score to be the highest is that the substantial changes in the approval basis text in the government approval scenario have the most fundamental impact on the credibility of historical approval conclusion labels. The greater the semantic deviation between the historical basis text and the current valid basis text, the lower the reference value of the corresponding historical label for the current prediction network training. The three factors are multiplied by 0.5 for the timeliness consistency score of each historical approval record, 0.3 for the internal consistency score of approvers, and 0.2 for the cross-employee consensus score, and then summed to obtain the label credibility score corresponding to each historical approval record. The value ranges from 0 to 1. The higher the score, the less the approval conclusion label of the historical approval record is affected by the combined impact of the three types of interference: timeliness mismatch of the basis text, unstable approval judgment, and cross-employee standard divergence.
[0031] Figure 2 This is a schematic diagram showing the cumulative distribution of tag credibility scores under different business type numbers in the embodiments of this application. Figure 2The figure illustrates the cumulative probability distribution curves of label credibility scores calculated using the three-dimensional weighted summation method described in step S2 for historical approval records with three different business type numbers in a specific embodiment of this application. The curve for business type A is generally biased to the left, indicating that the label credibility scores of historical approval records under this business type are generally low. The curve for business type C is generally biased to the right, indicating that the label credibility scores of historical approval records under this business type are generally high. The vertical dashed line in the figure is the threshold reference line at a label credibility score of 0.5. As can be seen from the figure, there are significant differences in the distribution of label credibility scores under different business types, verifying the ability of the three-dimensional weighted summation method to distinguish the differences in credibility of historical approval records under different business types.
[0032] In one specific embodiment, step S3 involves substituting the label credibility score as a loss weight into the weighted loss function, training the prediction network with the semantic vector of the application materials from each historical approval record as input and the approval conclusion label as a supervision signal, to obtain the approval prediction network, which includes: The semantic vectors of the application materials of each historical approval record are input into a fully connected layer for linear transformation, and then processed by an activation function to obtain the hidden layer feature vectors of each historical approval record. The hidden feature vectors of each historical approval record are input to the output layer and processed by linear transformation and Sigmoid function to obtain the predicted probability of the approval conclusion for each historical approval record. The weighted loss value is obtained by multiplying the binary cross-entropy loss value between the predicted probability of the approval conclusion and the approval conclusion label of each historical approval record by the corresponding label confidence score and taking the average value. The parameters of each layer of the prediction network are updated by gradient backpropagation based on the weighted loss value. The iteration continues until the weighted loss value converges, thus obtaining the approval prediction network.
[0033] Specifically, the prediction network adopts a two-layer fully connected structure. The first fully connected layer has an input dimension of 768, consistent with the dimension of the semantic vector of the application material text, and an output dimension of 256. The ReLU activation function is applied to the linear transformation result to truncate negative values to 0 and retain positive values, resulting in a 256-dimensional hidden layer feature vector. The second layer is the output layer with an input dimension of 256 and an output dimension of 1. The Sigmoid function is applied to the linear transformation result to map the output value to the interval between 0 and 1, obtaining the predicted probability of the approval conclusion for each historical approval record. This probability value reflects the prediction network's tendency to predict the approval of the current application material text. The hidden layer feature vector dimension is set to 256 because this dimension effectively reduces the number of parameters while retaining the main information of the semantic features of the application material text, avoiding overfitting under the condition of limited historical data of government approvals.
[0034] The binary cross-entropy loss value for each historical approval record is calculated as follows: The logarithm of the predicted probability of the approval conclusion is multiplied by the approval conclusion label, and the sum of this multiplication and the result of subtracting the logarithm of the predicted probability of the approval conclusion from 1, multiplied by 1, and then subtracting the approval conclusion label, is taken as the negative value. The weighted loss value is calculated by multiplying the binary cross-entropy loss value of each historical approval record by its corresponding label confidence score, and then taking the arithmetic mean of these values across all historical approval records. Historical approval records with lower label confidence scores contribute less to the weighted loss value, resulting in weaker gradient backpropagation signals and less impact on the update of parameters in each layer of the prediction network. Gradient backpropagation uses the Adam optimizer, with an initial learning rate set to... This value ensures a moderate parameter update step size and avoids loss value oscillations in government approval document classification tasks; the criterion for convergence of the weighted loss value is that the decrease in the weighted loss value is less than 1% for five consecutive iterations on the validation set. When this condition is met, the iteration stops, the parameters of each layer are fixed, and the approval prediction network is obtained.
[0035] In one specific embodiment, step S4 includes: The text of the application materials to be approved is segmented and filtered for stop words. The filtered text is then input into a pre-trained language model for encoding to obtain the semantic vector of the application materials to be approved. The semantic vector of the application materials to be approved is input into the approval prediction network for forward propagation to obtain the predicted probability corresponding to the application materials to be approved. The cosine similarity between the semantic vector of the application materials to be approved and the semantic vector of the application materials of each historical approval record in the historical approval record set is calculated one by one. After sorting them in descending order of cosine similarity, the first K records are truncated to obtain a subset of semantically similar historical approval records. The weighted historical pass rate is obtained by summing the product of the label credibility score of each historical approval record in the semantically similar historical approval record subset and the approval conclusion label of each historical approval record, and then dividing by the sum of the label credibility scores of each historical approval record in the semantically similar historical approval record subset. The predicted probability and the weighted historical pass rate are used together as decision reference data output.
[0036] Specifically, the text of the application materials to be approved is the complete text content of the new application materials currently submitted to the government approval system. After performing word segmentation and stop word filtering on it, the filtered text is input into the same government domain fine-tuned BERT model as in step S1 for encoding processing. The 768-dimensional hidden layer output corresponding to the [CLS] label is taken to obtain a 768-dimensional semantic vector of the application materials to be approved. The semantic vector of the application materials to be approved is then processed by the first fully connected layer of the approval prediction network through linear transformation and ReLU activation function to obtain a 256-dimensional intermediate vector. Then, it is processed by the output layer through linear transformation and Sigmoid function to obtain a prediction probability with a value range of 0 to 1. This prediction probability reflects the prediction tendency of the approval prediction network based on the parameters trained on high-reliability historical approval records for the current application materials to be approved.
[0037] The semantic vectors of the 768-dimensional application materials pending approval are compared one by one with the semantic vectors of the 768-dimensional application materials in each historical approval record in the historical approval record set. The calculation method is to divide the dot product of the two vectors by the product of their magnitudes. All cosine similarity values are sorted in descending order, and the top K corresponding historical approval records are selected. The value of K is set to 10. This value is chosen to strike a balance between covering sufficiently semantically similar reference cases and avoiding the introduction of records with low semantic relevance, resulting in a subset of semantically similar historical approval records. The weighted historical approval rate is calculated by comparing the historical approval records in the semantically similar historical approval record subset with the historical approval records from the subset. The product of the label credibility score of the batch record and the approval conclusion label of each historical approval record is summed one by one, and then divided by the sum of the label credibility scores of each historical approval record in the semantically similar historical approval record subset. The value ranges from 0 to 1. When the sum of the label credibility scores of each historical approval record is zero, the weighted historical pass rate is set to 0.5 by default to indicate that there is no effective historical reference information. The decision reference data is composed of two numerical fields: prediction probability and weighted historical pass rate. The former is the model inference output of the approval prediction network, and the latter is the statistical support value of semantically similar high-credibility historical approval cases. The two together serve as the data basis for subsequent approval processes.
[0038] Figure 3 This is a schematic diagram illustrating the joint distribution of the predicted probability and weighted historical pass rate of the application materials to be approved in this embodiment of the application. Figure 3The figure illustrates the joint distribution of the predicted probability obtained by inputting the semantic vector of the application material to be approved into the approval prediction network and the weighted historical pass rate calculated by retrieving semantically similar historical approval records from the historical approval record set in a specific embodiment of this application. The dashed line is the baseline when the predicted probability and the weighted historical pass rate are completely consistent. As can be seen from the figure, the sample points with a label confidence score of not less than 0.5 are more concentrated near the baseline, indicating that the weighted historical pass rate supported by high-confidence historical approval records has a higher consistency with the predicted probability of the approval prediction network.
[0039] The above describes the intelligent approval decision-making method for government affairs processes in the embodiments of this application. The following describes the intelligent approval decision-making system for government affairs processes in the embodiments of this application. One embodiment of the intelligent approval decision-making system for government affairs processes in the embodiments of this application includes: The extraction module is used to extract historical approval records from the government approval history database. Each historical approval record contains the application material text, approval conclusion label, approver number, approval basis text and business type number, resulting in a set of historical approval records. The weighting module is used to calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number for each historical approval record in the set of historical approval records. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. The input module is used to substitute the label credibility score as the loss weight into the weighted loss function, and to train the prediction network with the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to obtain the approval prediction network. The retrieval module is used to input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability, retrieve the historical approval records with the highest semantic similarity from the historical approval record set, weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate, and output the prediction probability and the weighted historical pass rate together as decision reference data.
[0040] This invention also provides an intelligent approval and decision-making device for government processes, which can be a server. This intelligent approval and decision-making device for government processes includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the intelligent approval and decision-making device for government processes includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the intelligent approval and decision-making device for government processes stores the corresponding data in this embodiment. The network interface of the intelligent approval and decision-making device for government processes is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.
[0041] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the intelligent approval decision-making method for government affairs processes.
[0042] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0043] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an intelligent approval and decision-making device (which may be a personal computer, server, or network device, etc.) oriented towards government processes to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intelligent approval decision-making method for government processes, characterized in that, The method includes: Step S1: Extract each historical approval record from the government approval history database. Each historical approval record includes the application material text, approval conclusion label, approver number, approval basis text and business type number, to obtain a set of historical approval records. Step S2: For each historical approval record in the set of historical approval records, calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. Step S3: Substitute the label credibility score as the loss weight into the weighted loss function, use the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to train the prediction network, and obtain the approval prediction network. Step S4: Input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability. Retrieve several historical approval records with the highest semantic similarity from the historical approval record set. Weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate. Output the prediction probability and the weighted historical pass rate together as decision reference data.
2. The intelligent approval decision-making method for government processes according to claim 1, characterized in that, Step S1 includes: The original application materials, original approval basis, approval conclusion label, approver number, and business type number of each historical approval record are extracted from the government approval history database to obtain the original data of each historical approval record. The original application material texts of each historical approval record are segmented and filtered for stop words. The filtered texts are then input into a pre-trained language model for encoding to obtain the semantic vectors of the application material texts of each historical approval record. The original approval basis text of each historical approval record is input into a pre-trained language model for encoding processing to obtain the semantic vector of the approval basis text of each historical approval record; The historical approval record set is obtained by integrating the text semantic vectors of the application materials, the text semantic vectors of the approval basis, the approval conclusion tags, the approver numbers, and the business type numbers of each historical approval record.
3. The intelligent approval and decision-making method for government processes according to claim 2, characterized in that, The calculation of the timeliness consistency score of the approval basis text in step S2 includes: Extract the text of the current valid approval basis from the current valid approval basis database, input the text of the current valid approval basis into the pre-trained language model for encoding processing, and obtain the semantic vector of the text of the current valid approval basis. Input the approval basis text semantic vector of each historical approval record and the current valid approval basis text semantic vector into the cosine similarity calculation unit respectively. Perform a dot product operation on the approval basis text semantic vector of each historical approval record and the current valid approval basis text semantic vector, and divide by the product of their magnitudes to obtain the cosine similarity value between the approval basis text semantic vector of each historical approval record and the current valid approval basis text semantic vector. The cosine similarity value of each historical approval record is directly used as the basis timeliness consistency score for each historical approval record, thus obtaining the basis timeliness consistency score for each historical approval record.
4. The intelligent approval decision-making method for government processes according to claim 3, characterized in that, Step S2, which calculates the internal consistency score for approvers with the same approver number and the same business type number, includes: From the set of historical approval records, filter by approver number and business type number to obtain a subset of historical approval records for the same employee and business corresponding to each historical approval record; For each historical approval record in the subset of historical approval records for the same employee and the same business, calculate the cosine similarity of the semantic vectors of the application material texts pairwise. Select the application material texts with a cosine similarity exceeding a preset similarity threshold as semantically similar application material text pairs to obtain a set of semantically similar application material text pairs corresponding to each historical approval record. The number of text pairs with different approval conclusion label values in the semantically similar application material text pair set is counted. The number of text pairs with different values is divided by the total number of text pairs in the semantically similar application material text pair set to obtain the inconsistency ratio of approval conclusion labels for each historical approval record. Subtracting the inconsistency rate of the approval conclusion label from 1 yields the internal consistency score of the approver for each historical approval record.
5. The intelligent approval and decision-making method for government processes according to claim 4, characterized in that, In step S2, the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number is calculated, and the label credibility score corresponding to each historical approval record is obtained by weighted summation of the three scores, including: Filter the historical approval record set by business type number to obtain a subset of historical approval records for the same business; group the subset of historical approval records for the same business by approver number, and divide the number of records with approval conclusion label value of "pass" in each group by the total number of records in each group to obtain the approver pass rate of each group; Calculate the mean of the approval rate for each group of approvers, square the difference between the approval rate of each group and the mean, take the mean and then take the square root to obtain the standard deviation of the approval rate for each historical approval record; subtract the standard deviation of the approval rate from 1 to obtain the cross-employee consensus score for each historical approval record. The consistency scores of the basis timeliness of each historical approval record, the consistency scores of the approvers within each historical approval record, and the consensus scores of the employees across each historical approval record are multiplied by their respective preset weight coefficients and then summed to obtain the tag credibility score for each historical approval record.
6. The intelligent approval decision-making method for government processes according to claim 5, characterized in that, In step S3, the label credibility score is used as a loss weight and substituted into the weighted loss function. The prediction network is trained using the semantic vector of the application materials text of each historical approval record as input and the approval conclusion label as a supervision signal, resulting in an approval prediction network, including: The semantic vectors of the application materials of each historical approval record are input into a fully connected layer for linear transformation, and then processed by an activation function to obtain the hidden layer feature vectors of each historical approval record. The hidden feature vectors of each historical approval record are input to the output layer and processed by linear transformation and Sigmoid function to obtain the predicted probability of the approval conclusion for each historical approval record. The weighted loss value is obtained by multiplying the binary cross-entropy loss value between the predicted probability of the approval conclusion of each historical approval record and the approval conclusion label of each historical approval record as the loss weight and the corresponding label confidence score by the corresponding label confidence score. The parameters of each layer of the prediction network are updated by gradient backpropagation based on the weighted loss value, and the iteration continues until the weighted loss value converges, thus obtaining the approval prediction network.
7. The intelligent approval decision-making method for government processes according to claim 1, characterized in that, Step S4 includes: The text of the application materials to be approved is segmented and filtered for stop words. The filtered text is then input into a pre-trained language model for encoding to obtain the semantic vector of the application materials to be approved. The semantic vector of the application material text to be approved is input into the approval prediction network for forward propagation processing to obtain the prediction probability corresponding to the application material text to be approved. The cosine similarity between the semantic vector of the application material to be approved and the semantic vector of the application material of each historical approval record in the historical approval record set is calculated one by one. After sorting them in descending order of cosine similarity, the first K records are truncated to obtain a subset of semantically similar historical approval records. The weighted historical pass rate is obtained by summing the product of the label credibility score of each historical approval record in the semantically similar historical approval record subset and the product of the approval conclusion label of each historical approval record, and then dividing by the sum of the label credibility scores of each historical approval record in the semantically similar historical approval record subset. The predicted probability and the weighted historical pass rate are used together as decision reference data output.
8. An intelligent approval and decision-making system for government processes, characterized in that: For implementing the intelligent approval decision-making method oriented towards government processes as described in any one of claims 1-7, the intelligent approval decision-making system oriented towards government processes includes: The extraction module is used to extract historical approval records from the government approval history database. Each historical approval record includes the application material text, approval conclusion label, approver number, approval basis text and business type number, resulting in a set of historical approval records. The weighting module is used to calculate the consistency score of the approval basis text, the consistency score of the approver within the same business type number under the same approver number, and the cross-employee consensus score corresponding to the standard deviation of the approval conclusion label pass rate across approver numbers under the same business type number for each historical approval record in the set of historical approval records. The three scores are weighted and summed to obtain the label credibility score corresponding to each historical approval record. The input module is used to substitute the label credibility score as the loss weight into the weighted loss function, and to train the prediction network with the semantic vector of the application material text of each historical approval record as input and the approval conclusion label as the supervision signal to obtain the approval prediction network. The retrieval module is used to input the semantic vector of the application materials to be approved into the approval prediction network to obtain the prediction probability, retrieve the historical approval records with the highest semantic similarity from the historical approval record set, weight the approval conclusion label with the label credibility score of each historical approval record to obtain the weighted historical pass rate, and output the prediction probability and the weighted historical pass rate together as decision reference data.
9. An intelligent approval and decision-making device for government processes, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the intelligent approval decision-making method for government processes as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the intelligent approval decision-making method for government processes as described in any one of claims 1 to 7.