A government affair service appeal intelligent processing system based on classification acceptance

Abstract

The application discloses a kind of based on classification acceptance's government affair service appeal intelligent processing system, specifically related to government affair service management field, for solving the problem that existing government affair service system only relies on appeal text to classify, it is difficult to reflect enterprise behavior difference and lead to classification deviation and unreasonable department assignment;By constructing enterprise fingerprint and silence tendency coefficient, the mapping relationship between silence behavior and appeal category distribution is established to generate category offset vector, and the classification score is corrected by gating in combination with abnormal semantic mark, to obtain the corrected classification probability distribution, and then according to the probability distribution, the appeal classification and department routing are completed, the joint use of enterprise behavior characteristics and semantic information is realized, the accuracy of appeal classification is improved, and the efficiency of government affair service processing is improved, and the cost of repeated circulation and collaborative processing between departments is effectively reduced, while the abnormal appeal recognition ability and decision stability in complex scenarios are improved.

Description

Technical Field

[0001] This invention relates to the field of government service management technology, and more specifically, to an intelligent processing system for government service requests based on classification and acceptance. Background Technology

[0002] In current government service platforms and enterprise service systems, enterprise requests are usually handled centrally through online channels, and automated classification models are used to identify request categories and assign them to departments in order to improve processing efficiency.

[0003] However, existing technologies generally rely solely on semantic analysis of the textual content of requests, neglecting the behavioral differences that enterprises develop over long-term operations and their impact on request types. This leads to discrepancies between classification results and actual processing needs. Specifically, different enterprises exhibit significant differences in interaction frequency within government systems. Some enterprises consistently lack proactive feedback, only submitting requests when problems accumulate to a certain extent or involve policy implementation or compliance supervision. This results in a noticeable shift in the distribution of their request categories compared to the overall enterprise group, which traditional text-based classification methods struggle to identify. Furthermore, when the expression of request texts deviates from historical corpus distributions, existing classification models often rely on existing semantic features for judgment, easily misclassifying abnormal or sudden issues into conventional categories, thus reducing the ability to identify abnormal events.

[0004] Furthermore, when there is uncertainty in the classification results or the probabilities of multiple categories are close, existing systems typically use a single maximum probability principle for departmental assignment. This lacks further verification of the reliability of the classification results and a decision-making mechanism based on distribution characteristics, easily leading to repeated order cancellations between departments or low collaboration efficiency. Therefore, how to incorporate corporate behavioral characteristics into the request classification process, characterize the correlation between silent behavior and the distribution of request categories, and effectively correct the classification results by incorporating semantic anomalies, thereby achieving more reasonable departmental routing, has become an urgent technical problem to be solved. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an intelligent processing system for government service requests based on classification to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A government service request intelligent processing system based on classification and acceptance includes: The enterprise coding module is used to construct enterprise attribute feature vectors based on enterprise industry classification and business scale, and to construct spatial codes based on the administrative division of the registered address. The enterprise attribute feature vectors and spatial codes are concatenated and then input into the autoencoder to reduce the dimensionality and obtain the enterprise fingerprint. The silence tendency assessment module is used to obtain the enterprise's historical request records, calculate the dispersion index of the time interval between adjacent requests and the time decay value of the most recent request, and generate a silence tendency coefficient by combining the dispersion index and the time decay value in a preset combination method. The preset combination method is obtained by fitting the trend of the frequency change of the enterprise's proactive reporting in the historical sample. The anomaly detection module is used to perform Chinese word segmentation on the text content of new enterprise requests and construct bag-of-words vectors. At the same time, it selects similar enterprises whose enterprise fingerprint similarity is greater than a preset threshold and extracts the bag-of-words vectors of the historical request texts of similar enterprises to form a comparison set. Calculate the cosine dissimilarity between the bag-of-words vector of the new claim and each bag-of-words vector in the comparison set, and select several samples with the smallest dissimilarity to form a semantic neighborhood. Calculate the dispersion of dissimilarity within the semantic neighborhood, and use the average vector of each bag-of-words vector in the semantic neighborhood as the neighborhood center. The cosine dissimilarity between the new appeal bag-of-words vector and the neighborhood center is calculated as the offset distance. When the offset distance is greater than the neighborhood dispersion, an abnormal semantic marker is generated. The offset mapping module is used to build a classifier. It takes the word segmentation result of the new appeal text as input and outputs the classification score of the preset appeal category. It selects historical samples, establishes a category offset mapping from the correspondence between the silence tendency coefficient and the appeal category distribution in the historical samples, and inputs the silence tendency coefficient of the current enterprise into the offset mapping to generate a category offset vector. The classification correction module is used to correct the classification probability distribution by superimposing the category offset vector and the classification score of the claim category with the abnormal semantic marker as the gating variable. The classification processing module is used to classify and process requests by department based on the corrected probability distribution, and output the optimal department to handle each request category.

[0007] As a further aspect of the present invention, the enterprise fingerprint obtained in the enterprise coding module specifically includes: All registered companies' industry classifications are converted into unique hot coding vectors, and their business scales are converted into level coding vectors according to registration type. The two are then concatenated to form a corporate attribute feature vector. At the same time, the administrative divisions of the registered addresses are converted into integer codes according to the province, city, and district levels, and then concatenated to form a spatial code. The enterprise attribute feature vector is concatenated with the spatial code and then input into the autoencoder. The number of hidden layer nodes in the autoencoder is less than the number of input layer nodes. The compressed vector output from the hidden layer is used as the enterprise fingerprint.

[0008] As a further aspect of the present invention, the silent tendency assessment module specifically includes generating a silent tendency coefficient by: Extract the timestamp of each request from the company's historical request records, sort them in ascending order by time, calculate the difference sequence of adjacent timestamps, and calculate the standard deviation of the sequence as an indicator of dispersion. The time interval is the difference between the current timestamp and the timestamp of the most recent request. This time interval is input into an exponential decay function and scaled by a preset decay coefficient. The function output value is used as the time decay value. We collected historical time-series samples of requests from companies in the same industry, using the rate of change in the frequency of proactive reporting by companies as the dependent variable and the dispersion index and time decay value as the two independent variables. We then used the least squares method to fit the combination coefficients of the two independent variables. The product of the dispersion index and the first combination coefficient, and the product of the time decay value and the second combination coefficient are added together, and the sum is used as the silencing tendency coefficient.

[0009] As a further aspect of the present invention, the offset mapping module specifically includes the following for constructing the classifier: Collect the word segmentation results of historical appeal texts and their corresponding appeal categories. For each text, count the frequency of each word to construct a bag-of-words vector. Use the bag-of-words vector as input and the category code as output to train a classifier based on a multinomial logistic regression model. The classifier assigns a set of weight coefficients to each category. The word segmentation results of the new appeal text are used to construct a bag-of-words vector in the same way, and the inner product of the weight coefficients of each category is calculated as the classification score of each appeal category.

[0010] As a further aspect of the present invention, the generation of the category offset vector in the offset mapping module specifically includes: By comparing the similarity of enterprise fingerprints, we select historical samples of similar enterprises that report new demands, extract the silence tendency coefficient and corresponding demand category for each demand from the historical samples, and count the frequency of each category in all samples as the baseline distribution. The range of the silent tendency coefficient is divided into multiple continuous sub-intervals. The frequency of occurrence of each claim category is counted in each sub-interval to obtain the conditional distribution of that sub-interval. The category-wise difference between the conditional distribution of each sub-interval and the baseline distribution is calculated to form the category offset vector of that sub-interval. The output is the category offset vector corresponding to the sub-interval of the current enterprise's silence tendency coefficient.

[0011] As a further aspect of the present invention, the classification correction module specifically includes generating the corrected classification probability distribution by: When the new claim does not have any abnormal semantic markers, the classification scores are adjusted by performing element-wise addition between the category offset vector and the classification scores of each category output by the classifier. When the new claim has abnormal semantic markers, the classification scores of each category remain unchanged, and the adjusted classification scores are normalized to obtain the corrected classification probability distribution.

[0012] As a further aspect of the present invention, the optimal receiving department for outputting the claim category in the classification processing module specifically includes: Based on historical work order flow records, a correspondence between request categories and responsible departments is established. At the same time, the corrected classification probability distribution is extracted, and the responsible department corresponding to the request category with the highest probability in the classification probability distribution is output as the optimal receiving department. If the deviation between the probability values ​​of multiple request categories is less than the preset distinction threshold, the corresponding multiple responsible departments are output as collaborative receiving departments.

[0013] The technical effects and advantages of the intelligent processing system for government service requests based on classification acceptance of the present invention are as follows: This invention structurally improves traditional text-based appeal classification mechanisms by introducing enterprise behavioral characteristics, effectively enhancing the consistency between classification results and actual business semantics. First, by modeling historical enterprise interactions, a silence tendency coefficient is constructed, and a mapping relationship between silence behavior and appeal category distribution is established. This ensures that classification results no longer rely solely on textual representations but reflect the true appeal structure of enterprises under different interaction modes, thereby reducing classification bias caused by ambiguous or incomplete semantic expressions. Second, by identifying anomalous semantics and using them as gating variables in the classification correction process, the impact of behavioral bias on classification results is suppressed when appeal semantics deviate from historical distributions. This improves the accuracy of identifying sudden or unconventional appeals and prevents anomalous issues from being misclassified as regular categories.

[0014] Furthermore, by making departmental routing decisions based on the corrected classification probability distribution, the system can select a single or collaborative approach to service delivery based on the concentration or dispersion of classification results. This reduces repeated order cancellations between departments, improving the efficiency of government service processing and the rationality of resource allocation. Overall, this invention represents a shift from single semantic classification to a joint decision-making model of "behavior-driven + semantic verification," demonstrating significant practical application value. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the structure of an intelligent processing system for government service requests based on classification acceptance, according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1 Figure 1 The present invention provides an intelligent processing system for government service requests based on classification and acceptance, comprising: The enterprise coding module is used to construct enterprise attribute feature vectors based on enterprise industry classification and business scale, and to construct spatial codes based on the administrative division of the registered address. The enterprise attribute feature vectors and spatial codes are concatenated and then input into the autoencoder to reduce the dimensionality and obtain the enterprise fingerprint. The silence tendency assessment module is used to obtain the enterprise's historical request records, calculate the dispersion index of the time interval between adjacent requests and the time decay value of the most recent request, and generate a silence tendency coefficient by combining the dispersion index and the time decay value in a preset combination method. The preset combination method is obtained by fitting the trend of the frequency change of the enterprise's proactive reporting in the historical sample. The anomaly detection module is used to perform Chinese word segmentation on the text content of new enterprise requests and construct bag-of-words vectors. At the same time, it selects similar enterprises whose enterprise fingerprint similarity is greater than a preset threshold and extracts the bag-of-words vectors of the historical request texts of similar enterprises to form a comparison set. Calculate the cosine dissimilarity between the bag-of-words vector of the new claim and each bag-of-words vector in the comparison set, and select several samples with the smallest dissimilarity to form a semantic neighborhood. Calculate the dispersion of dissimilarity within the semantic neighborhood, and use the average vector of each bag-of-words vector in the semantic neighborhood as the neighborhood center. The cosine dissimilarity between the new appeal bag-of-words vector and the neighborhood center is calculated as the offset distance. When the offset distance is greater than the neighborhood dispersion, an abnormal semantic marker is generated. The offset mapping module is used to build a classifier. It takes the word segmentation result of the new appeal text as input and outputs the classification score of the preset appeal category. It selects historical samples, establishes a category offset mapping from the correspondence between the silence tendency coefficient and the appeal category distribution in the historical samples, and inputs the silence tendency coefficient of the current enterprise into the offset mapping to generate a category offset vector. The classification correction module is used to correct the classification probability distribution by superimposing the category offset vector and the classification score of the claim category with the abnormal semantic marker as the gating variable. The classification processing module is used to classify and process requests by department based on the corrected probability distribution, and output the optimal department to handle each request category.

[0018] The enterprise fingerprint is obtained from the enterprise coding module.

[0019] Structured coding is performed on basic enterprise information. The industry classification field registered in the enterprise database is selected and mapped uniformly according to the national standard industry directory. For example, manufacturing, wholesale and retail, and information services each correspond to a unique category number. Based on this, a one-hot coding method is used for conversion. That is, a fixed-length positional coding vector is assigned to each category, and a coding space of length 32 is selected. For each industry category, only the corresponding position is set to 1, and the rest are set to 0, thus forming a discrete expression of industry classification. For enterprise operating scale, it is classified into levels according to the scale type in the registration information, and this level value is mapped to a continuous numerical code to reflect scale differences. The one-hot coding vector of industry classification and the operating scale level coding vector are concatenated in a fixed order to form the enterprise attribute feature vector. During the spatial information construction process, the administrative division information in the enterprise's registered address is read, and the provincial, municipal, and district-level administrative units are mapped to integer codes. For example, by pre-establishing an administrative division coding table, each administrative unit is mapped to a unique integer number, and concatenated in the order of "province-municipal-district" to form a spatial coding vector. In practical processing, to avoid coding conflicts, a segmented coding method is adopted. For example, the provincial code occupies 4 digits, the municipal code occupies 4 digits, and the district code occupies 4 digits. Fixed-length concatenation ensures that information at different levels can be distinguished. Finally, the enterprise attribute feature vector and the spatial coding vector are concatenated according to the beginning and end connection method to obtain an input feature vector of uniform length.

[0020] After constructing the input features, the concatenated enterprise attributes and spatial encoding vectors are input into the autoencoder model for dimensionality reduction. The autoencoder adopts a symmetrical structure, including an input layer, an encoding layer, and a decoding layer. The number of nodes in the input layer is consistent with the length of the concatenated vector, set to 48 dimensions. The encoding layer is set as a two-stage compression structure, with 24 nodes in the first hidden layer and 12 nodes in the second hidden layer, achieving feature extraction through layer-by-layer compression. The decoding layer structure is symmetrically set with the encoding layer and is used to reconstruct the original input. The model training adopts a reconstruction error minimization approach, with training samples being a historical enterprise feature dataset. During training, the difference between the input vector and the decoded output vector is used as the optimization objective, with a fixed learning rate for iterative updates. The number of training epochs is set to 50 to ensure the model converges to a stable state. After training, only the encoding layer portion is retained for feature extraction, i.e., the output of the second hidden layer is used as the compressed representation. For each enterprise, its input vector is fed into the trained encoding structure, and the output of this layer is directly extracted as the enterprise fingerprint. This fingerprint vector has a significantly lower dimensionality than the original input while retaining comprehensive feature information of industry attributes and spatial distribution. In practical applications, all enterprises use the same autoencoder model for processing to ensure that the fingerprints of different enterprises have a unified expression space, thereby enabling subsequent similarity calculation and grouping screening.

[0021] The silence tendency assessment module generates a silence tendency coefficient.

[0022] The system extracts the occurrence time of each request from the company's historical request records and converts it to a standardized time format, such as a timestamp in milliseconds or seconds, to avoid errors caused by differences in time formats across systems. All historical request timestamps from the same company are arranged in ascending order of chronological order to ensure the time series accurately reflects the rhythm of the company's requests. Based on this, the intervals between adjacent timestamps are calculated one by one to obtain a continuous time interval sequence. To avoid extreme values ​​interfering with the overall distribution, obviously abnormal time intervals are processed before calculation; for example, if an interval exceeds 365 days, it is truncated to 365 days to ensure sequence stability. Then, the dispersion of this time interval sequence is calculated. Specifically, the fluctuation of all time intervals is statistically analyzed to reflect the regularity of the company's request behavior. Standard deviation is used as a metric; a larger standard deviation indicates more drastic fluctuations in the request intervals, indicating unstable company behavior, while a smaller standard deviation indicates a more uniform periodicity. Simultaneously, the current system time is read and the difference between it and the company's most recent request time is calculated to obtain the current time interval, which directly reflects the length of time since the company's most recent request. Based on this, the time interval is input into an exponential decay function for conversion. The exponential decay function selects a decay form with the natural constant as the base and introduces a decay coefficient to control the decay rate. The decay coefficient is set to 0.05 by default as an example value. When the time interval is short, the function output value is close to 1. As the time interval gradually increases, the output value gradually decreases and approaches 0, thus forming a quantitative description of "recent activity". Through the above processing, a dispersion index reflecting historical behavior fluctuations and a time decay value reflecting the current activity state are obtained respectively. Together, they constitute the basic description of the enterprise's demand behavior characteristics.

[0023] After obtaining the dispersion index and time decay value, a calculation model for the silence tendency coefficient is established using industry benchmark samples. In practice, based on the target company's industry, multiple companies within the same industry are selected as benchmark samples. Historical complaint records of these companies within the same time span are selected and statistically analyzed according to a fixed time window. In this embodiment, a 30-day period is used as a statistical cycle. The number of complaints within each cycle is counted, and the difference in the number of complaints between two adjacent cycles is calculated. This difference is used as the change in reporting frequency. This change reflects the trend of the company's reporting behavior; a negative difference indicates a decrease in the number of reports, while a positive difference indicates an increase in the number of reports, thus characterizing the behavioral change of "tending towards silence" or "tending towards activity" in a specific numerical form. Based on this, the dispersion index and time decay value corresponding to each company are used as input factors, and a correlation analysis is performed with their corresponding change in reporting frequency. The least squares method is used for fitting calculation to obtain the combination coefficient corresponding to the two input factors. In this embodiment, the combination coefficient corresponding to the dispersion index is set to 0.6, and the combination coefficient corresponding to the time decay value is set to 0.4 as example values, indicating that the impact of historical rhythm fluctuations on silent behavior is slightly higher than the current activity level. After the model is determined, the dispersion index of the target enterprise is multiplied by its combination coefficient, and the time decay value is multiplied by its combination coefficient. The two results are then summed to obtain the final silence tendency coefficient. The larger the coefficient value, the more likely the enterprise is to be in a silent state, that is, more likely to report problems in a concentrated manner after they accumulate, thus providing stable and realistic behavioral feature inputs for subsequent category offset modeling.

[0024] In the anomaly identification module, anomaly semantic markers are generated.

[0025] The newly submitted appeal texts from enterprises are standardized. First, irrelevant symbols, repeated spaces, and common stop words are uniformly removed. Then, the text is segmented into Chinese words, dividing continuous sentences into semantically independent sets of terms. Word segmentation is performed using a combination of a fixed dictionary and statistical rules to ensure that industry terms and government keywords are fully recognized; for example, compound words such as "administrative license" and "environmental protection acceptance" are not incorrectly segmented. After segmentation, the frequency of each term in the text is counted, and a fixed-length bag-of-words vector is constructed according to a unified vocabulary. The vocabulary consists of a set of historical high-frequency words, with a predetermined number of high-frequency words selected as the vocabulary range to ensure coverage of the main semantic expressions. Low-frequency words outside the vocabulary are uniformly mapped to the "unknown word" position to avoid dimensionality expansion. After constructing the bag-of-words vector for the new appeal, the enterprise fingerprint database is read, and the similarity between the current enterprise fingerprint and the fingerprints of other enterprises in the database is calculated. Similarity is measured using the vector angle; in this embodiment, a similarity threshold of 0.75 is set, and higher values ​​are obtained. Enterprises with fingerprint similarity no lower than this threshold are selected as a set of similar enterprises. For the historical appeal texts of each company in this set, bag-of-words vectors are constructed in the same way as the current appeals, and these are then aggregated to form a comparison set. In practice, to avoid statistical instability due to insufficient sample size, it is stipulated that the number of similar companies should not be less than 20. If the initial screening results are insufficient, the similarity threshold is appropriately relaxed to 0.7 as a supplement, thereby ensuring that the subsequent semantic neighborhood construction has sufficient sample support.

[0026] After obtaining the new appeal bag-of-words vectors and the comparison set, the semantic difference between the new appeal vector and each historical appeal vector in the set is calculated one by one. The difference is measured using cosine dissimilarity, which reflects semantic differences by comparing the directional differences between two vectors; a larger value indicates greater semantic dissimilarity. All calculation results are sorted, and the set of samples with the smallest dissimilarity is selected as the semantic neighborhood. This set is fixed at 30 samples to ensure that the neighborhood is representative while avoiding the introduction of too much noise. Subsequently, the dissimilarity distribution within this semantic neighborhood is statistically analyzed. Specifically, the overall fluctuation of the differences between samples in the neighborhood is calculated, and this is used as the neighborhood dispersion to characterize the stability of the semantic region. A smaller dispersion value indicates that the neighborhood is semantically concentrated, while a larger value indicates that the semantics are dispersed. Based on this, the average of all bag-of-words vectors in the semantic neighborhood is calculated to obtain a central vector, which represents the typical semantic expression of the neighborhood. Then, the cosine dissimilarity between the new appeal vector and this central vector is calculated, and this value is used as the offset distance. When the offset distance is greater than the neighborhood dispersion, it indicates that the current request has deviated from the normal fluctuation range of the semantic neighborhood and is an abnormal expression. In this case, the abnormal semantic flag is set to 1. When the offset distance does not exceed the neighborhood dispersion, it indicates that the request is still within the normal semantic range. In this case, the abnormal semantic flag is set to 0.

[0027] In the offset mapping module, a classifier is constructed.

[0028] A historical sample set for classification training is constructed. Historical request records that have been processed in the government service system are selected, and the text content of each request and its corresponding final confirmed category are extracted. The category is derived from the processing results of the actual receiving department, such as policy consultation, safety supervision, and complaints / reports, ensuring the authenticity and reliability of the tag source. All historical request texts undergo unified Chinese word segmentation processing, with segmentation rules consistent with those in the aforementioned anomaly identification module, ensuring consistency in term expression across different processing stages. Based on the segmentation results, the frequency of each term in the text is counted, and a bag-of-words vector is constructed based on a unified vocabulary. The vocabulary consists of high-frequency terms from the historical samples. In this embodiment, terms with the highest frequency ranking within a predetermined range are selected as the vocabulary range. Terms not included in the vocabulary are uniformly mapped to the "low-frequency word" position, thus ensuring stable vector dimensions and avoiding excessive sparsity. For each historical request, its category is converted into a category code form, and a unique identifier is set for each category, with a value of 1 for the corresponding category position and 0 for the remaining positions. The bag-of-words vector is used as input data, and the category code is used as the output label to construct training sample pairs. During the model training phase, multinomial logistic regression was selected as the classification model structure. This model consists of an input layer and an output layer. The number of nodes in the input layer is consistent with the dimension of the bag-of-words vector, and the number of nodes in the output layer is consistent with the total number of categories. Each category corresponds to a set of weight coefficients to measure the contribution of each term to that category. During training, batch sample input is used for parameter updates. In this embodiment, the number of samples per batch is set to 128, and the number of training rounds is set to 60. Through iterative optimization round by round, the model output gradually approximates the true category distribution. During training, the input data is uniformly normalized, and the term frequency is scaled proportionally to the text length to avoid the bias effect of long texts on model training. After model training is completed, the weight coefficients and bias terms corresponding to each category are saved, forming a fixed parameter set that can be used for subsequent classification calculations.

[0029] After the classifier training is complete, the same text processing workflow as the training phase is applied to newly received enterprise requests. First, the new request text is segmented into words, and a corresponding bag-of-words vector is generated based on the constructed unified vocabulary, ensuring consistency between the new input data and the training data in the feature space. When generating the bag-of-words vector, each word in the text is counted, and its corresponding value is filled in according to its position in the vocabulary. Unregistered words are also assigned to low-frequency word positions to ensure complete matching of vector dimensions. After the bag-of-words vector is constructed, it is input into the trained classifier, and a classification score is calculated for each category. Specifically, the bag-of-words vector is multiplied term by term by the weight coefficient of each category and summed to obtain the original score value for that category. This score value reflects the degree of matching of the current request text within that category. During the calculation process, the same calculation method is used for all category scores to ensure comparability between different categories.

[0030] In the offset mapping module, a category offset vector is generated.

[0031] For each company corresponding to a new request, its fingerprint vector is first retrieved from the company fingerprint database. Then, its similarity is calculated against the fingerprints of all historical companies in the database. A uniform vector angle measurement method is used for similarity calculation to ensure comparability across different dimensions. To ensure stable and representative screening results, a fingerprint similarity screening threshold of 0.75 is set as an example value. Companies with similarity scores above this threshold are selected as the candidate set. When the number of companies meeting the criteria is insufficient, they are sorted from highest to lowest similarity, and companies with high similarity scores are added until the sample size reaches a preset lower limit, set at 30, to ensure sufficient sample support for subsequent statistics. For the selected set of similar companies, their historical request records are extracted, and the silence tendency coefficient and the final confirmed request category associated with each request are retrieved. The silence tendency coefficient is calculated uniformly using the aforementioned method, and the request category uses the actual processing results as the label source to avoid human labeling bias. After summarizing all extracted historical samples, the frequency of each category in all samples is counted, and the frequency of each category is divided by the total number of samples to obtain the frequency of each category. This set of frequencies constitutes the baseline distribution. During the statistical process, to avoid instability in subsequent calculations due to zero sample counts for a certain category, a very small frequency value is uniformly assigned to categories that do not appear. In this embodiment, it is set to 0.001 as an example value to ensure that all categories participate in subsequent difference calculations.

[0032] After obtaining the baseline distribution, the range of the silence tendency coefficient is divided into intervals to establish a correspondence between silence behavior and category distribution. Specifically, the minimum and maximum values ​​of the silence tendency coefficient in all historical samples are first counted, and intervals are divided based on this range. In this embodiment, an equal-frequency partitioning method is used, where samples are sorted by silence tendency coefficient from smallest to largest and divided into a set number of continuous sub-intervals, ensuring that the number of samples contained in each sub-interval is roughly the same, thus avoiding statistical bias caused by uneven data distribution. For each sub-interval, a set of samples falling within that interval is extracted, the frequency of each claim category in that set is counted, and its frequency of occurrence within that sub-interval is calculated to obtain the corresponding conditional distribution. During the calculation process, categories that do not appear are assigned minimum frequency values ​​to ensure distribution integrity. Subsequently, the conditional distribution of this sub-interval is compared category by category with the aforementioned baseline distribution, and the difference between the two is calculated for each category. The difference results reflect the degree of deviation of each category from the overall distribution under this silence interval, and all category differences constitute the category offset vector of this sub-interval. After constructing the offset vectors for all sub-intervals, for the current enterprise, the category offset vector of the corresponding sub-interval is directly selected as the output based on the position of its silence tendency coefficient within that interval. At the interval boundaries, a left-closed, right-open rule is used for assignment determination. For example, when the silence tendency coefficient equals the upper boundary of the interval, it is assigned to the next sub-interval, thus ensuring the uniqueness of the classification.

[0033] The classification correction module generates a corrected classification probability distribution.

[0034] The classification scores for each category output by the classifier are read. These scores are unnormalized results calculated based on bag-of-words features and classifier weights, and their values ​​reflect the degree of matching of the current appeal in each category. Simultaneously, the category offset vector and anomalous semantic markers generated in the preceding steps are read. The category offset vector and classification scores maintain the same dimension, with each dimension corresponding to the same category. The values ​​of the anomalous semantic markers are determined. In this embodiment, the anomalous semantic marker is defined as a binary variable: a value of 0 indicates that the current appeal semantics are within the historical semantic neighborhood, and a value of 1 indicates that the current appeal semantics deviate from the neighborhood center and are judged as anomalous. If the anomalous semantic marker is 0, a correction process is performed on the classification score. Specifically, the category offset vector and the classification score are added item by item at the same position, so that the original score of each category, while maintaining the original semantic matching, is superimposed with behavioral offset information, thereby adjusting the classification result. Before performing the addition operation, the category offset vector is constrained. In this embodiment, the offset value is limited to the interval [-1, 1] to avoid extreme offset values ​​from excessively affecting the classification result. This range is determined by statistically analyzing the offset distribution in historical samples. After the addition operation is completed, the adjusted classification score sequence is obtained. If the abnormal semantic flag is 1, no offset stacking processing is performed on the classification scores, and the classifier output result is directly retained. This ensures that the classification decision in the case of abnormal semantics relies entirely on the semantic information of the text itself, avoiding the introduction of offsets based on regular corpus statistics that may mislead abnormal samples. The abnormal semantic flag is only used as a control condition for whether to perform offset stacking and does not participate in the specific numerical calculation.

[0035] After adjusting or maintaining the classification scores, the resulting classification score sequence is normalized to convert it into probability distributions for each category. In practice, a uniform soft-maximum function is used to transform all category scores. First, an exponential transformation is performed on each category score. Then, the exponential values ​​of all categories are summed, and this sum is used to normalize the exponential value of each category, yielding the probability values ​​for each category. To ensure numerical stability, the classification score sequence is shifted before the exponential calculation, i.e., all scores are subtracted from the maximum value in the current sequence, thus avoiding overflow during exponential calculation. After normalization, the sum of the probabilities of each category in the resulting classification probability distribution is 1, and the probability value of each category is between 0 and 1, which can be directly used for subsequent decision-making. In actual implementation, to avoid instability caused by excessive concentration of a single category probability in extreme cases, a minimum probability lower bound is set for the probability distribution. In this embodiment, the minimum probability value is set to 0.01 as an example value. When the probability of a certain category falls below this value, it is adjusted to this lower bound, and the remaining probabilities are redistributed proportionally to the other categories to ensure that all categories have basic expressive capabilities. The entire normalization process applies uniformly to all category scores, regardless of whether they have undergone offset correction.

[0036] In the classification processing module, the optimal department to handle the request category is output.

[0037] The historical work order flow records are structured and organized to extract the final processing result and corresponding department information for each request. The request text classification results are then statistically correlated with the actual receiving departments. To ensure the accuracy of the statistical results, only stable work orders that have completed closed-loop processing and have not been returned or multiple times are selected as the basic sample. In this embodiment, the proportion of work orders processed by a single department is required to be no less than 70% of the total sample size to avoid interference from complex flow paths on the mapping relationship. For each preset request category, the number of times it has been handled by each department in the historical sample is counted, and the department with the highest number of handling is determined as the primary responsible department for that category. For cases where the number of handlings is close, such as when the difference in the number of handlings by the top two departments in a category is less than 10% of the total sample size, both departments are recorded as candidate responsible departments for that category to ensure the stability and coverage of the mapping relationship. In the actual construction process, a corresponding responsible department mapping table is established for each category, containing at least one primary responsible department and a list of optional candidate responsible departments. To avoid mapping distortion due to insufficient samples, for categories with fewer than 20 historical samples, mapping relationships are not established directly. Instead, they are assigned to the mapping results of similar categories according to business rules, thus ensuring that all mapping relationships are derived from sufficient sample statistics.

[0038] After obtaining the corrected classification probability distribution, the distribution is analyzed to determine the optimal receiving department. Specifically, the probability values ​​for all categories are first read from the distribution and sorted according to probability. The category with the highest probability is identified as the dominant category. In this embodiment, the primary responsible department corresponding to this dominant category in the established mapping table is directly selected as the priority receiving object. Simultaneously, the differences between the top-ranking categories are calculated, taking the difference between the highest and second-highest probabilities. Further, adjacent differences are calculated sequentially for subsequent categories. Based on this, a discrimination threshold is introduced to judge the difference. In this embodiment, the threshold is set to 0.1 as an example value. When the difference between the most probable category and the second most probable category is greater than the threshold, the classification result is considered to have obvious discrimination, and only the main responsible department is output as the optimal undertaking department. When the difference between the most probable category and the second most probable category does not exceed the threshold, the classification result is considered to have uncertainty. At this time, the difference between multiple categories is further judged, and all categories that meet the condition of "difference from the most probable category does not exceed the threshold" are selected together. The responsible departments corresponding to these categories are found in the mapping table, and these departments are uniformly output as collaborative undertaking departments. In the case of multiple department outputs, the department list is deduplicated to ensure that the same department appears only once. At the same time, the departments are arranged in order according to the probability of the category to reflect the priority processing order. The entire process makes decisions based on the structural characteristics of the classification probability distribution, without relying on additional rule settings, so that the department routing results can be dynamically adjusted with the classification uncertainty, thereby ensuring a stable and reasonable undertaking arrangement even in complex demand scenarios.

[0039] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0040] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0041] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0042] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0043] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0044] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0045] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion 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 a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. 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.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0047] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent processing system for government service requests based on categorized acceptance, characterized in that, include: The enterprise coding module is used to construct enterprise attribute feature vectors based on enterprise industry classification and business scale, and to construct spatial codes based on the administrative division of the registered address. The enterprise attribute feature vectors and spatial codes are concatenated and then input into the autoencoder to reduce the dimensionality and obtain the enterprise fingerprint. The silence tendency assessment module is used to obtain the enterprise's historical request records, calculate the dispersion index of the time interval between adjacent requests and the time decay value of the most recent request, and generate a silence tendency coefficient by combining the dispersion index and the time decay value in a preset combination method. The preset combination method is obtained by fitting the trend of the frequency change of the enterprise's proactive reporting in the historical sample. The anomaly detection module is used to perform Chinese word segmentation on the text content of new enterprise requests and construct bag-of-words vectors. At the same time, it selects similar enterprises whose enterprise fingerprint similarity is greater than a preset threshold and extracts the bag-of-words vectors of the historical request texts of similar enterprises to form a comparison set. Calculate the cosine dissimilarity between the bag-of-words vector of the new claim and each bag-of-words vector in the comparison set, and select several samples with the smallest dissimilarity to form a semantic neighborhood. Calculate the dispersion of dissimilarity within the semantic neighborhood, and use the average vector of each bag-of-words vector in the semantic neighborhood as the neighborhood center. The cosine dissimilarity between the new appeal bag-of-words vector and the neighborhood center is calculated as the offset distance. When the offset distance is greater than the neighborhood dispersion, an abnormal semantic marker is generated. The offset mapping module is used to build a classifier, which takes the word segmentation results of the new appeal text as input and outputs the classification score of the preset appeal category; Select historical samples, establish a category offset mapping based on the correspondence between the silence tendency coefficient and the distribution of demand categories in the historical samples, and input the silence tendency coefficient of the current enterprise into the offset mapping to generate a category offset vector; The classification correction module is used to correct the classification probability distribution by superimposing the category offset vector and the classification score of the claim category with the abnormal semantic marker as the gating variable. The classification processing module is used to classify and process requests by department based on the corrected probability distribution, and output the optimal department to handle each request category.

2. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, The enterprise coding module specifically obtains the enterprise fingerprint, including: All registered companies' industry classifications are converted into unique hot coding vectors, and their business scales are converted into level coding vectors according to registration type. The two are then concatenated to form a corporate attribute feature vector. At the same time, the administrative divisions of the registered addresses are converted into integer codes according to the three levels of province, city, and district, and then concatenated to form a spatial code. The enterprise attribute feature vector is concatenated with the spatial code and then input into the autoencoder. The number of hidden layer nodes in the autoencoder is less than the number of input layer nodes. The compressed vector output from the hidden layer is used as the enterprise fingerprint.

3. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, The silent tendency assessment module specifically includes the following steps to generate the silent tendency coefficient: Extract the timestamp of each request from the company's historical request records, sort them in ascending order by time, calculate the difference sequence of adjacent timestamps, and calculate the standard deviation of the sequence as an indicator of dispersion. The time interval is the difference between the current timestamp and the timestamp of the most recent request. This time interval is input into an exponential decay function and scaled by a preset decay coefficient. The function output value is used as the time decay value. We collected historical time-series samples of requests from companies in the same industry, using the rate of change in the frequency of proactive reporting by companies as the dependent variable and the dispersion index and time decay value as the two independent variables. We then used the least squares method to fit the combination coefficients of the two independent variables. The product of the dispersion index and the first combination coefficient, and the product of the time decay value and the second combination coefficient are added together, and the sum is used as the silencing tendency coefficient.

4. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, The offset mapping module specifically includes the following for constructing the classifier: Collect the word segmentation results of historical appeal texts and their corresponding appeal categories. For each text, count the frequency of each word to construct a bag-of-words vector. Use the bag-of-words vector as input and the category code as output to train a classifier based on a multinomial logistic regression model. The classifier assigns a set of weight coefficients to each category. The word segmentation results of the new appeal text are used to construct a bag-of-words vector in the same way, and the inner product of the weight coefficients of each category is calculated as the classification score of each appeal category.

5. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, In the offset mapping module, generating the category offset vector specifically includes: By comparing the similarity of enterprise fingerprints, we select historical samples of similar enterprises that report new demands, extract the silence tendency coefficient and corresponding demand category for each demand from the historical samples, and count the frequency of each category in all samples as the baseline distribution. The range of the silent tendency coefficient is divided into multiple continuous sub-intervals. The frequency of occurrence of each claim category is counted in each sub-interval to obtain the conditional distribution of that sub-interval. The category-wise difference between the conditional distribution of each sub-interval and the baseline distribution is calculated to form the category offset vector of that sub-interval. The output is the category offset vector corresponding to the sub-interval of the current enterprise's silence tendency coefficient.

6. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, The classification correction module specifically includes generating the corrected classification probability distribution, which includes: When the new claim does not have any abnormal semantic markers, the classification scores are adjusted by performing element-wise addition between the category offset vector and the classification scores of each category output by the classifier. When the new claim has abnormal semantic markers, the classification scores of each category remain unchanged, and the adjusted classification scores are normalized to obtain the corrected classification probability distribution.

7. The intelligent processing system for government service requests based on classification and acceptance as described in claim 1, characterized in that, In the classification processing module, the optimal receiving department for each type of request specifically includes: Based on historical work order flow records, a correspondence between request categories and responsible departments is established. At the same time, the corrected classification probability distribution is extracted, and the responsible department corresponding to the request category with the highest probability in the classification probability distribution is output as the optimal receiving department. If the deviation between the probability values ​​of multiple request categories is less than the preset distinction threshold, the corresponding multiple responsible departments are output as collaborative receiving departments.

Images