Intelligent text analysis method and system for government affair service system

Abstract

The invention provides an intelligent text analysis method and system for a government affair service system, and belongs to the technical field of data processing, and the method specifically comprises the steps: carrying out the determination of a portrait updating method of a user according to a file push user under a user tag corresponding to an interest portrait of the user, the method comprises the following steps of: updating interested portraits of users on the basis of a portrait updating method, determining updating delay types of different users on the basis of the portrait updating method, determining exception analysis and identification schemes of a model under different user tags by utilizing the updating delay types, and identifying the interested portraits of the users by utilizing the exception analysis and identification schemes. According to the method and the device, the identification deviation condition of the natural language processing model under different user tags is determined, and the user for real-time updating processing of the interest portrait is determined according to the identification deviation condition of the natural language processing model under different user tags, so that the exposure and the matching property of file pushing are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to an intelligent text analysis method and system for a government affair service system. BACKGROUND

[0002] With the deepening of the construction of the digital government, a large amount of text data is generated by government departments at all levels, including policy documents, notices, announcements, service guides, work summaries, etc. At present, the distribution of government texts mainly relies on hierarchical forwarding, portal website publishing or unified group sending, which has significant pain points: 1) Information overload and low efficiency: government personnel need to manually filter the content related to their own responsibilities from a large amount of information, which is inefficient and easy to miss key information; 2) Insufficient push precision: important policies are difficult to accurately reach all relevant executive personnel, supervisors or consulting service personnel; 3) Lack of personalization: the existing system lacks deep understanding of user roles, responsibilities and interests, and cannot achieve intelligent push of "thousand faces".

[0003] Therefore, an intelligent text analysis method and system for a government affair service system are urgently needed. SUMMARY

[0004] To achieve the object of the application, the application adopts the following technical solutions: Specifically, the application provides an intelligent text analysis method for a government affair service system, which specifically comprises: S1 determining the recommendation data under different user tags based on the text recommendation data of the government affair service system, determining the update demand type of the user's interest portrait according to the recommendation data under different user tags, and determining the portrait update method of the user according to the file push user under the user tag corresponding to the user's interest portrait when the update demand type does not belong to the target demand type; S2 updating the user's interest portrait based on the portrait update method, determining the update delay type of different users based on the portrait update method, and determining the abnormal analysis identification scheme of different user tags by using the update delay type determination model; S3 determining the identification deviation of the natural language processing model under different user tags by using the abnormal analysis identification scheme, and determining the user for real-time update processing of the interest portrait according to the identification deviation of the natural language processing model under different user tags.

[0005] The application has the following beneficial effects: According to the recommendation data under different user tags, the update requirement type of the user's interest portrait is determined, so as to realize the number of file recommendations under different user tags, determine the exposure rate of the file, and update the user's interest portrait according to the difference of the exposure rate, so as to improve the update timeliness of the interest portrait in the case of poor exposure rate, and then improve the exposure rate and contact surface of the file.

[0006] According to the recognition deviation of the natural language processing model under different user tags, the user for real-time updating of the interest portrait is determined, so as to realize the update delay user, and in the case that the user tag with a large label error probability exists in the user tag with multiple natural language processing models, the update timeliness of the interest portrait of the update delay user is adjusted in time, so as to improve the matching degree of the recommended file under the user tag and the real user tag of the user, and on the basis of identifying the abnormal user tag based on the click data, the influence of the user tag on the click data due to the non-timely update of the user tag can be effectively excluded, and the reliability of the identification processing of the user tag is further improved.

[0007] Further, the recommendation data under the user tag includes the number of recommended texts under the user tag.

[0008] Further, the method for determining the update requirement type of the user's interest portrait is: determining the recommended text under the user tag based on the recommendation data under different user tags; determining the number of recommended users of different recommended texts based on the user tag of the user and the recommended text under the user tag; determining the update requirement type of the user's interest portrait based on the number of recommended users of different recommended texts.

[0009] Further, the method for determining the user's portrait update method is: determining the number of file pushing users under the user tag corresponding to the user's interest portrait based on the file pushing user under the user tag corresponding to the user's interest portrait; determining the effective user tag in the user tag of the user based on the number of file pushing users; determining the portrait update method of the user based on the effective user tag in the user tag of the user and the constituting data of the effective user tag of all users.

[0010] Further, the effective user tag is a user tag with a number of file pushing users greater than a preset recommended user number threshold.

[0011] Further, the method for determining the user of the real-time updating process of the interest portrait is: With the identification bias of the natural language model under different user tags, the number of identification bias of the recommended text under different user tags is determined, and based on the proportion of the number of identification bias of the recommended text under the user tags in the number of recommended text under the user tags, the identification bias factor of the user tags is determined; Based on the identification bias factor of the user tags, the identification bias risk label in the user tags is determined; With the identification bias risk label data, and combined with the user data of the identification bias risk label, it is determined whether the user is the user of the real-time updating process of the interest portrait.

[0012] In a second aspect, the present application provides a computer system, comprising a memory and a processor connected in communication, and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the computer program to perform the above-mentioned intelligent text analysis method for government service system.

[0013] Other features and advantages will be set forth in the following description, and the objectives and other advantages of the present application will be achieved and obtained in the structures specifically pointed out in the description and the accompanying drawings.

[0014] In order to make the above-mentioned objectives, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and other features and advantages of the present application will become more apparent from the detailed description of example embodiments thereof with reference to the attached drawings, in which: Figure 1 is a flowchart of an intelligent text analysis method for a government service system; Figure 2 is a flowchart of a method for determining the updating demand type of the user's interest portrait; Figure 3 is a flowchart of a method for determining the portrait updating method; Figure 4 is a flowchart of a method for determining the abnormal analysis identification scheme under the user tags. DETAILED DESCRIPTION

[0016] In order for those skilled in the art to better understand the technical solutions in the specification, the technical solutions in the specification will be clearly and completely described in the specification below in combination with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all. Based on the embodiments of the specification, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the specification.

[0017] Embodiment 1 As Figure 1 shown, the present application provides an intelligent text analysis method for a government service system, which specifically comprises: S1 determines the recommendation data under different user tags based on the text recommendation data of the government service system, determines the update demand type of the user's interest portrait according to the recommendation data under different user tags, and determines the user's portrait update method according to the file push user under the user tag corresponding to the user's interest portrait when the update demand type does not belong to the target demand type; The core decision target of the method is to intelligently judge whether the knowledge portrait of the relevant staff group needs to start high-priority real-time update by analyzing the distribution breadth and coverage balance of the internal files of the government system under different staff tags. The core logic is a multi-level progressive decision-making process: first, identify whether there are widely distributed files, second, evaluate the universality of such files, then analyze the concentration of these files, and finally determine whether there is a "distribution solidification" phenomenon according to the concentration, so as to decide whether to update the knowledge portrait in real time to optimize the file distribution strategy and improve the coverage and fairness of the government information transmission, and ensure that internal policy files can be evenly and effectively reached all relevant personnel.

[0018] Further, the recommendation data under the user tag includes the number of recommended texts under the user tag.

[0019] Specifically, as Figure 2 shown, the method for determining the update demand type of the user's interest portrait is: S11 determines the recommended text under the user tag with the recommendation data under different user tags; The keywords of this step are "staff label" and "distributed files". "Staff label" is a classification label identified by the system based on the job responsibilities, management authority, business field, historical processing file type, and interest profile of internal staff in the government system, such as "financial approval officer", "project construction administrator", "policy and regulation researcher", etc. "Distributed files" refers to the set of government files, policy notifications, work guidelines, etc. that the system actively distributes to all internal staff with the label within the statistical period, regardless of whether the staff actually consults them. This step is set up in this way because in the government system, ensuring that relevant personnel receive files related to their duties in a timely manner is a basic requirement, and distribution is a direct reflection of the system's fulfillment of this duty. The significance lies in aggregating distribution behavior data by label dimension, allowing for structured analysis of the execution of the system's file distribution strategy in different duty groups, focusing on evaluating the coverage completeness of the "file delivery" key administrative link, providing accurate "system side" distribution data basis for subsequent analysis, and ensuring objective evaluation of the government information transmission mechanism.

[0020] Example: Staff Zhang, the director, is labeled as the "safety production supervision" label. The system extracts 50 relevant files (including safety inspection notices, accident reports, regulation updates, etc.) that have been distributed to all "safety production supervision" label staff in the past week, which are the "distributed files" determined in this step.

[0021] S12 determines the number of recommended users for different recommended texts according to the user label of the user and the recommended text under the user label; The keyword of this step is "number of distributed personnel". It refers to the number of internal personnel with the staff label to which each distributed file determined in step S11 is distributed by the system. This is a measure of the "distribution coverage" of a single file, focusing on the range of potential recipients to which the system delivers the file, rather than the actual consultation situation. The reason for setting up this step is that in the transmission of government files, "delivery" is the premise of "awareness" and "execution". The number of distributed personnel directly reflects the effort and initial range of the system's attempt to reach the relevant staff. The significance lies in quantifying the system's distribution actions into analyzable data, allowing us to evaluate from the source whether the file has obtained the necessary delivery coverage and the allocation of the system's distribution resources on different files, which is the key basis for diagnosing whether there are "coverage blind spots" or "insufficient distribution of key files" in the file transmission mechanism.

[0022] Example: For the above 50 safety production files, the system counts the number of distributed personnel for each file. It is found that an emergency notice on safety production during flood season is distributed to 190 of the 200 workers under the label, while a technical rule on a certain special inspection is only distributed to 15 people.

[0023] S13 determines the update requirement type of the interest portrait of the user based on the number of recommended users of different recommended texts.

[0024] It can be understood that, based on the number of recommended users of different recommended texts, the update requirement type of the interest portrait of the user is determined, specifically including: S131 determines whether there is a recommended text with a number of recommended users greater than a preset number of recommended user thresholds based on the number of recommended users of the recommended text, and if so, proceeds to the next step, and if not, determines that the update requirement type of the interest portrait of the user is a target requirement type; This step is a general decision step, and its specific logic is implemented through the following sub-steps (S131-S134), which aims to finely determine whether the knowledge portrait update of the target requirement type (i.e. real-time update) is needed through a series of threshold judgments from the distribution breadth and receiving distribution dimensions, to solve the potential problem of uneven distribution of files.

[0025] S131: Determine whether there is a widely distributed file: This step introduces the keywords "preset number of distributed personnel threshold" and "target requirement type". The "preset number of distributed personnel threshold" is a threshold value set based on the proportion of the total number of workers under the label, which is used to judge whether the file has obtained sufficient basic distribution (for example, 60% of the total number of personnel under the label). The "target requirement type" specifically refers to the processing mode that needs to trigger real-time and high-frequency knowledge portrait update. The reason for this judgment is that if the number of distributed personnel of all distributed files under a certain label is lower than this threshold, it means that the system's file distribution strategy for this group is seriously flawed, and a large number of files that should be known have not reached enough workers, and the knowledge portrait of relevant personnel may not be updated with necessary policy information, resulting in work disconnection. Its significance lies in serving as an early warning mechanism to immediately identify the serious problem of "systematic distribution deficiency", directly triggering intensive real-time update and distribution strategy adjustment, to ensure full coverage of basic government information in a more proactive manner, and prevent information from affecting administrative execution due to lack of information.

[0026] Example: Continuing with the previous example, there are 200 workers under the "safety production supervision" label, and the preset threshold is 120 people (60%). It is found that among the 50 files, there are several files with more than 190 distributed personnel, so it does not meet the condition of "all below threshold", and the process proceeds to S132.

[0027] S132 determines the effective recommendation text as the recommendation text whose recommended user quantity is greater than a preset recommended user quantity threshold value, judges whether the quantity proportion of the effective recommendation text in the recommendation text is greater than a preset quantity proportion threshold value, if yes, determines that the update demand type of the interest portrait of the user does not belong to the target demand type, and if not, enters the next step; Judge the proportion of effective distribution files: The keywords of this step are "effective distribution files" and "preset quantity proportion threshold value". "Effective distribution files" are the files screened out in S131, whose distribution personnel number exceeds the preset threshold value. "Preset quantity proportion threshold value" is the proportion threshold value of the number of effective files to the total number of distribution files (such as 60%). The reason for setting this step is: if the proportion of widely covered "effective distribution files" is too high, it means that the system will concentrate most of the distribution resources on a small number of "universal" or "higher emergency" files, which may reflect the nature of the file or a conservative distribution strategy, which may lead to insufficient types of received files for the staff, and some special or deep files may not be covered. The significance lies in evaluating the concentration degree of the system distribution behavior. If the proportion is too high, it means that the current distribution mode is stable but may not meet the diversified professional information needs, and further analysis is needed; if the proportion is moderate or low, it means that most files are not fully distributed, and there may be a risk of insufficient distribution.

[0028] Example: In the above example, assume that there are 8 effective distribution files whose distribution personnel number exceeds 120, accounting for 16% (8 / 50), which is lower than the 30% proportion threshold value, so the condition is not established, and the process enters S133.

[0029] S133 determines the quantity proportion of the effective recommendation text in all effective recommendation texts in different users according to the recommended user data of the effective recommendation text, judges whether there is a user whose quantity proportion is greater than a preset proportion threshold value, if yes, enters the next step, and if not, determines that the update demand type of the interest portrait of the user does not belong to the target demand type; The keywords of this step are "personnel receiving concentration proportion" and "preset concentration threshold". The "personnel receiving concentration proportion" here specifically refers to calculating the number of times a single staff member is distributed with effective distribution files (i.e. widely covered files) by the system, accounting for the proportion of the total distribution times of all effective distribution files. The "preset concentration threshold" is a critical value (such as 3%) to determine whether a single staff member is over-concentrated in distributing widely covered files. The reason for setting this step is that even if there are not many widely covered files, if the system repeatedly and concentratedly distributes these files to a small number of the same staff members, it indicates that the distribution logic of the system may have personal or post-based bias, creating excessive "high-priority file" exposure for a small number of people, while other staff members with the same label are relatively lacking. The significance lies in revealing the imbalance of distribution resource allocation from the individual level of staff, identifying whether there is a "distribution privileged post", which is an important micro perspective for evaluating the fairness of information transmission within the government system.

[0030] Example: Suppose only the aforementioned 8 out of 50 are effective distribution files, and statistics show that regional regulatory director Wang has been distributed 4 of the 8 files. Then the concentration proportion of Wang in this effective distribution is 4 / 8 times ≈ 50%, which exceeds the assumed threshold of 40%. Therefore, there is an over-concentrated personnel in distribution, and the process goes to S134.

[0031] S134 takes the user whose number proportion is greater than the preset proportion threshold as the overlapping user, and determines the update demand type of the interest portrait of the user according to the proportion of the overlapping user in the recommended users of the effective recommendation text.

[0032] It should be noted that when the proportion of the overlapping user in the recommended users of the effective recommendation text is greater than the preset proportion threshold, the exposure rate of the text is poor, and the update demand type of the interest portrait of the user is determined as the target demand type.

[0033] Final decision according to the proportion of overlapping personnel: The keywords of this step are "coincidence personnel" and "preset proportion threshold". "Coincidence personnel" refers to the staff identified in S133 whose proportion of personnel receiving concentration exceeds the preset concentration threshold. The "preset proportion threshold" is the proportion threshold of these coincidence personnel in all staff who have received the effective distribution file (e.g. 5%). The reason for this step is ultimately set: even if there are a few "high-frequency receiving personnel", if the proportion of such personnel in the entire audience is very low, it may belong to special duties; on the contrary, if the proportion is too high, it means that the distribution audience of the widely-covered file has formed a highly coincident and solidified group, and the distribution diversity of the system has failed at the receiving level, leading to the rigidification of the overall file transmission mode, and the similarity of the file combination received by most staff with the same tag is too high, which may miss professional files that are more suitable for their specific duties. The significance lies in making the final administrative decision: when the proportion of coincidence personnel is too high, it indicates that the file distribution under the tag is not healthy, and there is "systematic distribution solidification", which must be updated through real-time updating of the knowledge portrait of relevant staff, optimizing the tag system or distribution rules, and breaking this solidified distribution mode to improve the relevance and comprehensiveness of file distribution among different staff.

[0034] Example: Continue the previous example, the total number of staff who have received effective files is 180, among which there are 12 coincidence personnel like Director Wang, accounting for 6.67% of the entire audience. If the preset proportion threshold is 5%, then 6.67% > 5%. At this time, it is determined that the file distribution mode has a solidification risk, and the knowledge portrait update requirement type of this staff group is the target requirement type, which needs to start real-time updating to optimize the distribution strategy.

[0035] It should be noted that when the update requirement type of the user's interest portrait is the target requirement type, the real-time updating method is used to update the interest portrait of all users, thereby improving the exposure of the text.

[0036] It should be noted that the file pushing user under the user tag is the user who exists in the user tag.

[0037] The core decision goal of the method is to intelligently decide whether to adopt real-time update or delayed update of the user's portrait maintenance strategy by analyzing which of the user's interest tags are "effective" (i.e., have a sufficient audience size) and combining the universality of the user's effective tags among all users. The core logic is: first, filter out "effective user tags" with a large-scale audience base, then conduct double evaluation from the individual level (number of effective tags) and the group level (tag composition universality), and finally make a final decision through a comprehensive factor. This logic aims to balance update efficiency and resource consumption - users with niche interests or special tag combinations are given priority for real-time update to quickly explore optimization; users with mainstream interests and stable tags are given delayed update to save system resources.

[0038] Specifically, as shown in Figure 3 The determined method of the user portrait update method is: S21 determines the number of file push users under the user tag corresponding to the user's interest portrait. Key word explanation: "user tag" refers to a work attribute identifier constructed based on the legal responsibility description of government personnel and their historical actual browsing and processing file records and other dynamic behavior data, such as "high-frequency processing of cross-department coordination files for policy researchers" and "frequent review of environmental regulations for frontline law enforcement personnel". Unlike traditional fixed tags, user tags automatically adjust with changes in personnel behavior. "File push personnel" refers to the set of all staff currently marked with the user tag in the government system. "Number of file push personnel" refers to the total number of staff currently covered by each user tag.

[0039] This step is the basis for evaluating the "actual influence" or "real universality" of the user tag. The number of personnel covered by a user tag reflects not only the legal basis of the responsibility combination or work mode, but also the extent to which it is actually adopted in the current organization's actual operation. It is a key indicator for distinguishing between "core work modes" and "marginal or emerging work modes". The significance of this step is to combine dynamic and personalized user portraits with group analysis, providing an objective basis for determining whether a tag is "effective" based on real behavior data. This ensures that the analysis respects the stability of the organizational structure while capturing the dynamic evolution of work practices.

[0040] Example: Zhang, the head of the department, has the following user tags: "responsible for infrastructure project approval and frequently consults green building standards" (covers 180 people), "participates in financial budget supervision and frequently reviews performance reports" (covers 120 people), "concerned about coastal zone protection and browses related scientific research literature" (covers 25 people), and "occasionally handles historical building renovation consultations" (covers 8 people). The system separately obtains the number of file push personnel for each of these four user tags.

[0041] S22 determines the valid user tags among the user tags of the user according to the number of file push users. In the above step, it is determined whether the user has valid user tags. If yes, proceed to the next step. If no, determine that the user's portrait updating method is a real-time updating method.

[0042] Key explanations: "Valid user tags" are defined as user tags with a number of file push personnel greater than a preset push personnel quantity threshold. This threshold is set based on statistical analysis value, aiming to filter out those with sufficient behavior sample size and reliable pattern analysis. "Real-time updating method" refers to high-frequency, near-real-time incremental updating of staff knowledge portraits, enabling user tags to quickly respond to changes in user behavior.

[0043] The concept of "valid user tags" is set to focus on those work patterns that have formed a certain scale in practice. For user tags with very few personnel coverage, their exposure is small. The significance of this step is to perform key diversion: if all user tags of a staff member are invalid (i.e., their work patterns are very unique or emerging in the system), the exposure of user tags related to them is small, therefore the updating efficiency needs to be improved.

[0044] Example: Suppose the system sets the preset push personnel quantity threshold to 50. Comparing Zhang's user tags: "responsible for infrastructure project approval and frequently consults green building standards" (180 people > 50 people) and "participates in financial budget supervision and frequently reviews performance reports" (120 people > 50 people) are valid user tags; "concerned about coastal zone protection and browses related scientific research literature" (25 people < 50 people) and "occasionally handles historical building renovation consultations" (8 people < 50 people) are invalid user tags. Since Zhang has valid user tags, the process proceeds to S23.

[0045] S23 determines the user's portrait updating method based on the valid user tags among the user tags of the user and the constituent data of all users' valid user tags.

[0046] It should be noted that the valid user tags are user tags with a number of file push users greater than a preset recommended user quantity threshold.

[0047] It can be understood that the portrait updating method of the user is determined based on the valid user tags in the user tags of the user and the constituent data of the valid user tags of all users, and specifically includes: S231 determines whether the number of valid user tags of the user is greater than a preset useful tag number threshold, if yes, determines that the portrait updating method of the user is a delayed updating method, if not, proceeds to the next step; determining whether the number of valid user tags of the staff is greater than a preset valid tag number threshold, the keyword explanation: "preset valid tag number threshold" is a critical value for measuring the breadth of the mainstream work mode of the staff, indicating whether there are enough user tags for exposure, for example, set to 1 or 2.

[0048] This step evaluates from the individual level of the staff. The staff with a large number of valid user tags means that the file exposure under the user tags is large. The significance of setting this step is to identify those "standardized workers" with large exposure of user tags, which itself has important support for the exposure of files. Adopting a delayed updating method (such as daily or weekly batch updating) for them is sufficient to meet their knowledge synchronization needs, while saving a lot of resources for real-time calculation and pushing. This reflects the resource optimization configuration for stable work mode.

[0049] Example: Zhang, the chief, has 2 valid user tags. Assuming that the preset valid tag number threshold is 1, then 2>1, the condition is established. According to the rules, it can be directly determined as delayed updating. But to show the complete process, assume that the threshold is set to 3, then 2<3, the condition is not established, and the process enters S232.

[0050] S232 determines whether the valid user tags of the user are greater than a preset constituent ratio threshold in the constituent ratio of the valid user tags of all users, if yes, determines that the portrait updating method of the user is a delayed updating method, if not, proceeds to the next step; Keyword explanation: "constituent ratio" specifically refers to the proportion of the number of personnel covered by a certain valid user tag of the staff to the total number of personnel covered by all valid user tags of all staff. "Preset constituent ratio threshold" is to judge the degree of contribution to high exposure.

[0051] This step selects from the group level. If the valid user tags of a staff cover a very high proportion of personnel, at this time the influence of the staff on the improvement of the exposure of the file is larger, so the updating timeliness can be reduced.

[0052] Example: Calculate the proportion of Zhang's effective tags. Assume that the number of all valid user tags in the whole system is 8. "Then the tag proportion is 0.25, which is less than 0.3, the condition is not established, and the process enters S233.

[0053] S233 determines the update demand factor of the user based on the number of valid user tags of the user and the number of user tags, and determines the portrait update method of the user based on the update demand factor.

[0054] It should be noted that the update demand factor is related to the number of valid user tags of the user and the number of user tags, and the more the number of valid user tags of the user and the number of user tags, the smaller the update demand factor.

[0055] The update demand factor " is a quantitative index for comprehensively evaluating the urgency of dynamic portrait update of the staff. The design principle is: the more the number of valid user tags of the staff and the total number of user tags (valid + invalid), the smaller the update demand factor. The logic is that both more valid tags and more total tags have a significant effect on the improvement of exposure, so the update demand factor is smaller.

[0056] The first two steps (S231, S232) are quick decisions from "whether there are enough mainstream modes" and "whether it is the most common mode". S233 introduces a more smooth and comprehensive index to handle more extensive intermediate cases. The core of the design of the update demand factor is that the urgency of portrait update is inversely proportional to the "determinacy" and "diversity" of the working mode it reflects. The significance of this step is to achieve fine-grained resource grading, which can scientifically allocate different strategies from millisecond-level real-time update to day-level delayed update according to the complexity and maturity of the staff's behavior mode, maximize the overall operation efficiency of the government affairs system on the premise of ensuring the timely synchronization of key personnel knowledge.

[0057] Example: Assume that the update demand factor U = K / (α * N_valid + β * N_total), where N_valid is the number of valid user tags (2), N_total is the total number of user tags (4), α and β are weight coefficients, and K is a constant. Calculate U as a numerical value. Assume that the preset demand factor threshold is T. If U > T, the update demand is high, and the real-time update method is adopted; if U ≤ T, the delayed update method is adopted. Assume that U > T is calculated here, and the final portrait update method for Zhang is the real-time update method.

[0058] Specifically, when the update demand factor is greater than a preset demand factor threshold, it is determined that the user's portrait update method is a real-time update method, and otherwise, it is determined that the user's portrait update method is a delayed update method.

[0059] S2 updates the interest portrait of the user based on the portrait update method, determines different user update delay types based on the portrait update method, and determines an abnormality analysis identification scheme of the user under different user labels by using the update delay types; Further, the method for determining the update delay type of the user is: When the user's portrait update method is a real-time update method, it is determined that the user's update delay type is a real-time update user. When the user's portrait update method is a delayed update method, it is determined that the user's update delay type is an update delay user.

[0060] The core decision target of the method is to intelligently select the most matched abnormality analysis identification scheme of the label by comprehensively analyzing the portrait update delay of the user matched by the user label and combining the logic of the previous embodiments such as user label construction and portrait update demand determination. The core logic is a multi-level decision-making process closely connected. First, determine the "label matching user" based on the dynamic label constructed by the user portrait, and then combine Figure 3 The embodiment determines the update delay type to identify "update delay user", and finally selects from two abnormality analysis schemes according to system global risk, label heat, delay pollution degree and other indicators through four-step conditional judgment. The logic aims to establish a complete management chain from label construction to update decision to effect monitoring, ensuring the coordinated operation and resource optimization allocation of the recommendation system at each link.

[0061] Specifically, as Figure 4 shown, the method for determining the abnormality analysis identification scheme under the user label is: S31 takes the user in the interest portrait who has the user label as a label matching user; "User label" refers to a dynamic label constructed based on the user portrait, such as "personnel handling high-frequency new energy subsidy approval" (constructed based on responsibility and browsing files). "Interest portrait" is a dynamic portrait constructed by the system based on user behavior data, which contains a real-time adjusted label set. "Label matching user" refers to all user sets in the current system whose interest portrait is marked with the specific dynamic label.

[0062] Why this setup is used and its significance: This step is a direct application of the logic used in the previous embodiments to build tags based on user responsibilities and browsing behavior. By transforming abstract user profile tags into specific user groups, it provides clear target objects for subsequent anomaly analysis. Its significance lies in ensuring that the objects of anomaly analysis monitoring are real, dynamic user groups within the business scenario, rather than static, pre-defined categories. This allows monitoring to align with actual business changes, reflecting the dynamism and accuracy of user profile construction. Simultaneously, this step also completes the transition from "personalized tag definition" to "group effect evaluation."

[0063] Example: Based on the implementation of the government system, after the user tag "frontline law enforcement personnel who frequently consult the latest environmental regulations" is constructed, the system determines that there are currently 800 staff members who match this tag, and these 800 staff members become the "tag matching users" of this tag.

[0064] S32 determines the users with update delays based on the different update delay types of each user; "Update Delay Type" refers to... Figure 3 The user profile update method is used to classify users into different types based on the implementation examples. When the user profile update method is a real-time update method, the update delay type is real-time update user; when the user profile update method is a delayed update method, the update delay type is delayed update user. "Delayed update user" specifically refers to those who update through... Figure 3 In the implementation example, the decision process determines that the user adopts the delayed update method.

[0065] This step enables a seamless transition between previous and subsequent implementations, directly transforming the results of profile update decisions into risk inputs for anomaly analysis and monitoring. Figure 3 The example, by analyzing metrics such as the number and proportion of valid user tags, has scientifically determined which users need real-time updates and which can have delayed updates. This decision is continued here, treating "users with delayed updates" as potential sources of recommendation risk because their profiles may be somewhat outdated. The significance lies in constructing a complete governance chain: first deciding "how to update user profiles," then deciding "how to monitor recommendation effectiveness" based on the update status, forming a closed-loop decision-making process. This avoids redundant analysis and improves the overall efficiency of the system.

[0066] Example: In Figure 3 In this embodiment, User A is determined to require a real-time update method due to the small number of valid tags and their unique composition. User B is determined to require a delayed update method due to the large number and stability of valid tags. In this step, User B is categorized as an "update-delayed user" and will be considered for risk assessment in subsequent anomaly analysis.

[0067] S33 determines the abnormal analysis identification scheme under the user label according to the updated delay user data and the label matching user of the user label.

[0068] It can be understood that, according to the updated delay user data and the label matching user of the user label, determining the abnormal analysis identification scheme under the user label specifically includes: S331 determines the proportion of the number of updated delay users in all users based on the updated delay user data, and takes it as the delay user proportion, judges whether the delay user proportion is greater than the preset delay user proportion threshold, if yes, the number of updated delay users is relatively large at this time, in order to improve the accuracy of recommendation processing, the abnormal analysis identification scheme under all user labels is determined as the preset analysis scheme, if not, the next step is entered; This step is a general decision step, and its specific logic is realized through the following sub-steps (S331-S334), which aims to select from two abnormal analysis schemes according to the data provided by the previous steps. The preset analysis scheme adopts a lower trigger threshold (the second preset browsing user number proportion threshold), and the second preset analysis scheme adopts a higher trigger threshold (the preset browsing user number proportion threshold).

[0069] Judge whether the global delay user proportion is too high, this step evaluates the overall health of the system. High delay user proportion means that a large number of user portraits are updated with lag, which directly affects the reliability of the entire recommendation system. In this case, a more sensitive preset analysis scheme (low threshold) needs to be used for comprehensive monitoring. This reflects the linkage mechanism of abnormal analysis and system operation state (update delay), which automatically improves the monitoring intensity when the system has a global problem.

[0070] Example: suppose the system determines according to the data provided by the previous steps that there are 450 updated delay users (15%) in the total of 3000 users. Figure 3 If the preset delay user proportion threshold is 10%, then 15% > 10%. At this time, it is determined that the system has a serious delay risk, and the abnormal analysis scheme under all user labels is determined as the preset analysis scheme (adopting a lower trigger threshold), entering the comprehensive high-sensitivity monitoring state.

[0071] S332 judges whether the label matching user of the user label is less than the preset matching user number threshold, if yes, the number of label matching users of the user label is relatively small at this time, and therefore the abnormal analysis identification scheme under the user label is determined as the preset analysis scheme, if not, the next step is entered; Judge whether the number of label matching users is too small, this step combines the user label construction logic to identify long-tail or emerging business scenarios. If the number of matching users of a certain label is very small (such as Figure 1The small business label in the embodiment) indicates that the business field has few participants or is just emerging, the data is sparse, and the model stability is poor. It is necessary to use a preset analysis scheme (low threshold) to strengthen monitoring, which is consistent with the principle of paying more attention to small businesses in the previous embodiment.

[0072] Example: In the government system, the label "responsible for data cross-border security review" may have only 50 matching users (far less than the preset matching user number threshold of 200). In this case, regardless of other conditions, the abnormal analysis scheme under this label is directly determined as the preset analysis scheme, ensuring strict monitoring of this emerging, important but small business field.

[0073] S333 determines whether there is an update delay user in the label matching user of the user label. If yes, go to the next step, if no, the update timeliness of the portrait of the label matching user of the user label is higher at this time, so the abnormal analysis under the user label can realize accurate evaluation of the recommendation reliability of the model, and the abnormal analysis identification scheme under the user label is determined as the preset analysis scheme; Determine whether there is an update delay user in the label matching user. This step checks whether the specific business label is "polluted" by portrait delay. If all matching users of a certain label are real-time update users, it means that the portraits of relevant personnel in this business field are kept fresh, and the recommendation environment is good. The preset analysis scheme can be directly used for regular monitoring. This reflects the idea of precise policy: only the user group with real risks is subject to corresponding monitoring intensity.

[0074] Example: For the label "daily document processing personnel", there may be 5000 matching users and all of them are real-time update users. At this time, it is directly determined to use the preset analysis scheme, but note that the "preset analysis scheme" here is real-time analysis, but since the user portraits are fresh, the conditions for triggering analysis may be relatively loose (low threshold), which is more of a security rather than a correction monitoring.

[0075] S334 determines whether the proportion of update delay users in the label matching user of the user label is greater than the preset delay user proportion threshold. If yes, the abnormal analysis identification scheme under the user label is determined as the second preset analysis scheme, if no, the abnormal analysis identification scheme under the user label is determined as the preset analysis scheme; This step is the key to fine management. When the proportion of intra-label delay users exceeds the threshold, it means that the portrait of part of the key personnel in this business field is seriously lagging behind, and the reliability of the abnormal analysis processing based on browsing data is not high at this time, so the second preset analysis scheme is adopted. On the contrary, if the proportion is controllable, the preset analysis scheme can be used, and in-depth analysis is only performed when the recommendation effect is obviously poor (very low browsing rate). This is in line with the logic of allocating resources according to risk levels in the previous embodiment.

[0076] Example: For the label "emergency management personnel", 1000 users are matched, of which 150 users are update delay users (accounting for 15%). If the preset delay user proportion threshold is 10%, then 15% > 10%, and it is determined to adopt the second preset analysis scheme (high threshold).

[0077] It should be noted that the preset analysis scheme is to analyze the recommended text under the user label when the ratio of the number of browsing users of the recommended text to the number of recommended users is less than the second preset browsing user number proportion threshold (for example, less than 10%), to determine whether the model has analysis abnormalities, that is, whether the recommended text is related to the user label.

[0078] In addition, it can be understood that the second preset analysis scheme is to determine whether analysis is needed according to the number of browsing users of the recommended text under the user label. Specifically, when the ratio of the number of browsing users of the recommended text under the user label to the number of recommended users is less than the preset browsing user number proportion threshold (for example, less than 20%), the recommended text is analyzed to determine whether the model has analysis abnormalities, that is, whether the recommended text is related to the user label.

[0079] It should be noted that the second preset browsing user number proportion threshold is less than the preset browsing user number proportion threshold.

[0080] S3 determines the identification deviation of the natural language processing model under different user labels using the abnormal analysis identification scheme, and determines the user for real-time update processing of the interest portrait according to the identification deviation of the natural language processing model under different user labels.

[0081] The core decision goal of the method is to identify high-risk business labels and corresponding staff by analyzing the recognition bias of natural language models under different business labels in the government system, so as to accurately determine which staff portraits need real-time updating. The core logic is a three-layer decision-making process based on risk transmission: first, quantify the model recognition bias degree of each business label; second, filter out "recognition bias risk labels" that exceed the risk threshold; and finally, based on these risk labels and the personnel involved, determine the scope of staff that need real-time updating through multi-condition judgment. This logic aims to establish a linkage mechanism between model performance monitoring and portrait updating decision-making, ensuring that when a file recommendation model in a specific business domain has problems, the portrait updating strategy for related personnel can be adjusted in a timely manner to improve the efficiency of user portrait updating and the recognition reliability of recognition bias labels.

[0082] Specifically, the method for determining the user of the real-time updating of the interest portrait is: S41, determine the number of recognition biases of recommended texts under different user labels based on the recognition bias of natural language models under different user labels, and determine the recognition bias factor of the user label based on the proportion of the number of recognition biases of recommended texts under the user label in the number of recommended texts under the user label. "Recognition bias of natural language model" refers to the error or bias of a natural language processing model used to understand file content and match business labels in a government system when processing files in a specific business domain. For example, the model fails to correctly identify that a file about "old community renovation" should belong to the "urban construction" label, and incorrectly classifies it into other labels. "Number of recognition biases of recommended texts" refers to the number of files that are incorrectly identified or recommended by the model under a certain business label within a certain statistical period. "Proportion" refers to the proportion of the number of recognition biases in the total number of recommended texts under the label. "Recognition bias factor" is a quantitative indicator that measures the severity of model recognition bias under the business label, usually calculated based on the proportion.

[0083] This step is the basis for risk quantification. In the government system, the accuracy of natural language models directly affects the accurate push of policy files to relevant staff. By monitoring and quantitatively evaluating the performance of models under each business label, we can identify where the model has significant problems in which business domains. The purpose of calculating the "recognition bias factor" is to convert the qualitative "model may have bias" into a quantitative, comparable risk value, providing an objective, data-driven decision basis for subsequent risk label screening, ensuring that resources are prioritized for the most problematic business domains.

[0084] Example: In the government system, for the "flood control and drought resistance" business tag, the system has recommended 100 related files in the past week. Through manual sampling or system verification, it is found that 8 of the files have a recommendation deviation (such as the flood control plan being incorrectly recommended to non-related personnel). The number of identified deviations is 8, and the proportion is 8% (8 / 100). Assuming that the identified deviation factor directly adopts this proportion, the identified deviation factor of the "flood control and drought resistance" tag is 0.08.

[0085] S42 determines the identified deviation risk tag in the user tag based on the identified deviation factor of the user tag; The "preset deviation factor threshold" is a critical value set by the system to determine whether the model deviation of a business tag has reached a level that requires special attention. The "identified deviation risk tag" refers to those business tags whose identified deviation factor is greater than the preset deviation factor threshold.

[0086] This step is a key link in risk focusing. Not all tags with some deviation need immediate action, only those with deviation beyond the acceptable range are considered high risk. The purpose of setting the "preset deviation factor threshold" and screening the "identified deviation risk tag" is to achieve risk classification management and avoid overreaction to minor deviations. It helps the system quickly locate a small number of high-risk areas that most urgently need intervention from among numerous business tags, enabling subsequent portrait update decisions to be targeted and resources to be concentrated on solving the most critical problems.

[0087] Example: The system's preset deviation factor threshold is 0.05. According to the calculation in step S41, the identified deviation factor of the "flood control and drought resistance" tag is 0.08 > 0.05, so it is determined as an identified deviation risk tag. The identified deviation factor of the "daily document circulation" tag is 0.02 < 0.05, so it is not a risk tag.

[0088] S43 determines whether the user is a user of real-time update processing of the interest portrait based on the identified deviation risk tag data and in combination with the user data of the identified deviation risk tag.

[0089] This step is the final decision step, and its specific logic is implemented through the following sub-steps (S431-S433), which aims to determine which staff need to be included in the scope of real-time update processing based on factors such as the size of the risk tag and the proportion of staff involved. Special note: The "user" focused on in this step refers specifically to the original update delay users (determined by the system to use the delay update method according to Figure 3 / 4 embodiment), because the real-time update users are already in real-time update.

[0090] Further, the identification bias risk label is a label of a user whose identification bias factor is greater than a preset bias factor threshold.

[0091] Further, the user is the original update delay user.

[0092] Specifically, if the number of identification bias risk labels is greater than a preset risk label number threshold, it is determined that all users belong to the real-time update processing user of the interest portrait.

[0093] Further, in S431, if the number of identification bias risk labels is not greater than the preset risk label number threshold, the number of users with identification bias risk labels in all update delay users is determined based on the user data with identification bias risk labels, and if the proportion of the users with identification bias risk labels in all update delay users is less than a preset bias risk user proportion, it is determined that all users belong to the real-time update processing user of the interest portrait, and if not, the next step is entered. The "preset risk label number threshold" is a critical value for judging whether there are too many high-risk business fields in the system. The "user with identification bias risk label" refers to a staff member whose business portrait contains at least one identification bias risk label. The "preset bias risk user proportion" refers to the proportion threshold of these high-risk personnel in all update delay personnel.

[0094] This step evaluates the overall risk breadth faced by the system. If the number of risk labels is too large, it means that model bias problems exist widely and are not a local problem, and need to be addressed systematically - all staff members are included in real-time updating. If the number of risk labels is not large, but the proportion of high-risk personnel involved is low, it means that the risk is concentrated in a small range, and the same "all real-time updating" conservative strategy can be adopted to quickly contain the risk spread. Both of these cases adopt the most thorough intervention method, embodying the management principle of "taking decisive measures when risks spread".

[0095] Example: Suppose the system analyzes 100 business labels and identifies 5 risk labels (such as "flood prevention and drought resistance", "data security", etc.). The preset risk label number threshold is 3, and since 5 > 3, the condition is met. At this time, it is determined that all staff members belong to the real-time update processing user of the interest portrait. This is case one: too many risk labels.

[0096] Another case: if the number of risk labels is 2 (≤3), it is found that there are only 10 update delay users with these risk labels, and the total number of update delay users in the entire system is 100, accounting for 10%. If the preset bias risk user proportion is 5%, then 10% is not less than 5%, and the condition is not met.

[0097] S432 judges whether the number of the identification bias risk labels of the user is greater than a risk label number threshold value, if yes, determines that the user belongs to the user of the real-time updating processing of the interest portrait, if not, enters the next step; The "risk label number threshold value" (note that this is different from the "preset risk label number threshold value" of S431) is a critical value for judging the personal risk of a single staff member, which refers to the number threshold of the identification bias risk labels contained in the business portrait of the staff member.

[0098] This step evaluates from the individual level of the staff member. If the business responsibilities of a staff member involve multiple high-risk business fields (i.e., multiple risk labels are contained in his / her portrait), it means that the work content is complex and in multiple unreliable fields, and the accuracy of the individual portrait is crucial for correct document reception. Therefore, even if the global risk is controllable, these "high-risk positions" personnel need to be singled out for real-time updating. This reflects the fine management and key protection of key positions.

[0099] Example: (branch of "no" of S431) The system has 5 risk labels, and the high-risk personnel involved account for 8% of the update delay personnel (> 5%). Now evaluate the staff member Li. Li's business portrait contains 2 risk labels: "flood control and drought resistance" and "data security". Assuming that the risk label number threshold for individuals is 1, then 2 > 1, the condition is met. Therefore, it is determined that Li belongs to the real-time updating processing user of the interest portrait.

[0100] S433 if the number of real-time updating processing users among the users with the identification bias risk labels is not greater than a preset real-time updating user number threshold value, all the users with the identification bias risk labels are treated as real-time updating processing users of the interest portrait.

[0101] The "preset real-time updating user number threshold value" is a proportional or absolute number threshold set for the group of "users with identification bias risk labels", which is used to judge whether the high-risk group has been adequately covered by real-time updating.

[0102] This step is a bottom-up and safeguard mechanism. After S432 screening, some high-risk personnel may still not be included in real-time updating (because their personal risk label number does not meet the standard). But if the proportion or number of personnel in real-time updating among these high-risk personnel is too low, it means that the high-risk group as a whole lacks timely protection. In order to avoid the risk of spreading within the group, all remaining high-risk personnel are included in real-time updating at this time. The significance lies in ensuring the overall protection of high-risk business fields and preventing group risk from getting out of control due to individual screening omissions.

[0103] Example: (branch of S432 "No") Staff Wang has a risk label, but does not exceed the personal threshold (assuming the threshold is 2). Check all 50 staffs with the risk label corresponding to Wang, and find that only 5 of them are currently real-time update users. If the preset threshold of the number of real-time update users is 10 (or 20%), then 5 < 10, the condition is not established. At this time, all 50 staffs with the risk label (including Wang) are treated as real-time update processing users of the interest portrait.

[0104] Embodiment 2 In a second aspect, the present application provides a computer system, comprising a memory and a processor connected in communication, and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the computer program to perform the above-mentioned intelligent text analysis method for a government service system.

[0105] Each of the embodiments in the specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0106] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different than the order in the embodiments and still achieve the desired result. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous or possible.

[0107] The above only describes one or more embodiments of the specification and does not limit the specification. One or more embodiments of the specification can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of one or more embodiments of the specification shall be included in the scope of the claims of the specification.

Claims

1. An intelligent text analysis method for government service systems, characterized in that, Specifically, it includes: Based on the text recommendation data of the government service system, recommendation data under different user tags is determined. Based on the recommendation data under different user tags, the update requirement type of the user's interest profile is determined. When the update requirement type does not belong to the target requirement type, the method for updating the user's profile is determined based on the user tag corresponding to the user's interest profile and the file push user. The user's interest profile is updated based on the aforementioned profile update method. Based on the profile update method, different update delay types for different users are determined. The update delay types are then used to determine anomaly parsing and identification schemes for different user tags. Using the aforementioned anomaly analysis and identification scheme, the identification deviation of the natural language processing model under different user tags is determined. Based on the identification deviation of the natural language processing model under different user tags, users whose interest profiles are updated in real time are identified.

2. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, The method for determining the update request type of the user's interest profile is as follows: Based on recommendation data under different user tags, determine the recommended text under the user tags; Based on the user's user tags and the recommended text under the user tags, determine the number of users recommended for different recommended texts; Based on the number of users recommending different recommended texts, the update requirement type of the user's interest profile is determined.

3. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, When the user's interest profile update request type is the target request type, the interest profile of all users will be updated in real time to increase the exposure of the text.

4. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, The file push users under the user tags are those whose user tags exist.

5. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, The method for determining the user profile update method is as follows: The number of users who push files to a user under the user's user tag corresponding to the user's interest profile is determined. Based on the number of users to whom the file was pushed, determine the valid user tags among the user tags of the users; Based on the valid user tags in the user's user tags and the composition data of the valid user tags of all users, the user's profile update method is determined.

6. The intelligent text analysis method for government service systems as described in claim 5, characterized in that, If it is determined that the user does not have a valid user tag, then the user's profile update method is determined to be a real-time update method.

7. The intelligent text analysis method for government service systems as described in claim 5, characterized in that, The effective user tags are those of users whose number of file push users exceeds a preset threshold for the number of recommended users.

8. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, The method for determining the user's update delay type is as follows: When the user's profile update method is a real-time update method, the user's update delay type is determined to be a real-time update user; When the user's profile update method is a delayed update method, the user's update delay type is determined to be a delayed update user.

9. The intelligent text analysis method for government service systems as described in claim 1, characterized in that, The method for determining users in the real-time update process of the interest profile is as follows: Based on the recognition deviation of the natural language model under different user tags, the number of recognition deviations of the recommended text under different user tags is determined. Based on the proportion of the number of recognition deviations of the recommended text under the user tag in the total number of recommended texts under the user tag, the recognition deviation factor of the user tag is determined. Based on the identification bias factor of the user tag, identify the identification bias risk tag in the user tag; Using the identification bias risk label data and combined with user data containing the identification bias risk label, it is determined whether the user is a user subject to real-time update processing of the interest profile.

10. A computer system, comprising: A memory and processor connected by communication, and a computer program stored in the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes an intelligent text analysis method for a government service system as described in any one of claims 1-9.

Images