Community robot conversation routing and abnormal user identification method

Abstract

This invention relates to a method for routing dialogues with social bots and identifying abnormal users. The method involves receiving a target user's message in a social conversation; routing the message to a first social bot that outputs a clarification prompt and obtains a first response; if the first response does not provide verifiable information that meets preset verification criteria, subsequent messages are routed to a second social bot that outputs a specification prompt and obtains a second response; if the second response does not restate the request within the discussion boundaries as specified in the specification prompt, subsequent messages are routed to a third social bot that outputs a governance consequence prompt and obtains a third response; if the third response still does not meet the verification criteria or still does not restate the request within the discussion boundaries, the target user is determined to be an abnormal user; and a routing constraint strategy bound to the determination is applied to subsequent messages. This achieves coordinated linkage between social bot dialogue routing and abnormal user identification, improving system processing efficiency and governance stability while ensuring social order.

Description

Technical Field

[0001] This invention relates to the field of computer network technology, specifically to a method for routing dialogues in social robots and identifying abnormal users. Background Technology

[0002] Based on the technical solutions disclosed in patent document CN116010551A, which describes a method and apparatus for detecting chat text, existing technologies primarily revolve around identifying abnormal chat identities and detecting sensitive content. Their approach uses identity screening as a prerequisite. In cases of abnormal user identities, they then perform sensitive content matching and detection on historical chat texts, pending chat texts, and global historical chat texts, determining corresponding violation management measures based on the detected sensitivity level. The core of this solution remains based on sensitive word database matching, rule detection, and neural network model recognition. Although this solution enhances its ability to identify fragmented sensitive information through text combination and first character concatenation, its overall processing logic still follows a linear pattern of text content detection—risk level classification—violation handling. This pattern has several significant shortcomings. First, its anomaly judgment focuses on sensitive identification at the content level, lacking dynamic analysis of the user's behavioral evolution process and failing to continuously observe the user's ability to correct behavior after being reminded or restricted. Therefore, it cannot distinguish between malicious, continuous violations and occasional risky behaviors caused by inappropriate expression.

[0003] Secondly, this scheme uses identity anomalies as a trigger condition, which may fail to detect potential risks in a timely manner when the user's identity is normal but their behavior is abnormal, thus creating a blind spot in the identification process. Thirdly, while its violation management methods differentiate between low, medium, and high risk levels, the measures are mostly outcome-based actions such as alerts, muting, or silent bans, lacking a phased guidance mechanism before formal judgment, which can easily cause significant interference to legitimate users in cases of misjudgment. Furthermore, this technology still relies on sensitive text libraries and model recognition as its core. When facing complex dialogue scenarios such as semantic avoidance, contextual escaping, and role-playing expressions, the detection effectiveness still depends on the accuracy of the algorithm, lacking a logical verification mechanism based on the interaction process.

[0004] Furthermore, while this patent adds detection of text sequence combinations and first character combinations to prevent users from circumventing detection by splitting text, it still belongs to the extended detection of the text content itself, rather than the structured verification of user behavior patterns. Its technical logic remains at the level of judging text results, without constructing a closed-loop process for dialogue guidance and behavior confirmation. When the system detects highly sensitive content, although a manual review mechanism is introduced, this review is a post-event verification and does not set up a reversible correction path during the dialogue. Its silent banning mechanism, while covert, is a one-way restriction measure, without providing clear recovery conditions and process control. Therefore, in complex community environments, this solution is prone to problems such as over-reliance on content detection algorithms, rigid handling methods, lack of gradual guidance, and insufficient governance transparency. Overall, existing technologies focus more on text detection itself, with insufficient design of procedural processes for dialogue routing control, behavior verifiability, and anomaly judgment, making it difficult to ensure community order while simultaneously considering the interpretability of user expression correction and system governance. Summary of the Invention

[0005] The purpose of this invention is to provide a method for routing dialogues in social robots and identifying abnormal users, thereby addressing some of the drawbacks and shortcomings pointed out in the background art.

[0006] The present invention addresses the aforementioned technical problems by employing the following technical solution: a method for routing dialogues between social robots and identifying abnormal users, comprising: receiving a target user's speech in a social conversation; routing the speech to a first social robot to output a clarification prompt and obtain a first response; if the first response does not provide verifiable information that meets preset verification elements, then routing subsequent speech to a second social robot to output a specification prompt and obtain a second response;

[0007] If the second response does not restate the request within the discussion boundaries as prompted by the specifications, the subsequent speech will be routed to the third community robot to output the governance consequences and obtain the third response;

[0008] If the third response still fails to meet the verification elements or fails to restate the request within the discussion boundaries, the target user is determined to be an abnormal user; and a routing constraint policy bound to the determination is applied to subsequent statements.

[0009] Furthermore, the first community robot generates a list of missing elements based on the preset verification elements and outputs a field template for each missing element. The field template includes a field identifier and a value position. When the first response is missing any field, the value position is empty, or the value does not pass the preset verification rule corresponding to the field, it is determined that no verifiable information that meets the preset verification elements has been provided, and routing to the second community robot is triggered.

[0010] Furthermore, the second community robot issues a boundary receipt identifier bound to the discussion boundary in the specification prompt, and requires the target user to include the identifier and satisfy the discussion boundary when restating the request; when the second response does not include the identifier or includes it but does not restate it according to the specification, it is determined that the request has not been restated within the discussion boundary and routing to the third community robot is triggered.

[0011] Furthermore, the routing constraint strategy includes a reversible constraint mechanism: routing subsequent messages from the target user to the isolation channel and outputting a verification recovery prompt, requiring the user to complete the verification elements and restate the request within the discussion boundary within a preset time limit; if the requirement is met, the isolation is lifted and the normal routing is restored; if the requirement is not met, the isolation constraint is maintained or upgraded.

[0012] Furthermore, the preset verification rules include cross-field consistency verification: when the missing element list contains both a time field and an object identifier field, the time field value and the object identifier field value in the first response are required to be true simultaneously under the preset coexistence condition; when either field passes the verification individually but the two do not meet the coexistence condition, it is determined that the value has not passed the preset verification rules, and routing to the second community robot is triggered.

[0013] Furthermore, the field template includes a clarification count identifier corresponding to each field. When the first community robot fails the preset verification rule in the first response, it only outputs the field template again for the fields that fail the verification and updates the clarification count identifier. When the number of clarifications for the same field reaches the preset upper limit and still fails the preset verification rule, it is directly determined that no verifiable information that meets the preset verification elements has been provided, and routing to the second community robot is triggered.

[0014] Furthermore, the verification recovery prompt includes a recovery receipt identifier bound to the isolation channel, requiring the target user to submit restatement content carrying the identifier within a preset time limit; when the restatement content completes the verification elements and restates the request within the discussion boundary, the isolation is lifted and the normal route is restored; if the identifier is not carried, it is determined that the verification recovery prompt has not been met and the isolation constraints are maintained or upgraded.

[0015] Furthermore, the reversible constraint mechanism sets recovery bifurcation conditions: the verification recovery prompt provides at least two preset recovery paths, corresponding to the priority of verification elements and the priority of discussion boundary restatement, respectively; if the target user selects a path but fails to complete the corresponding priority condition within the preset time limit, the isolation constraint is upgraded; if the corresponding priority condition is completed and another condition is fulfilled, the isolation is lifted and the normal route is restored.

[0016] Furthermore, the verification recovery prompt configures a path receipt identifier for each preset recovery path; when the target user selects a path, the corresponding receipt identifier is returned, and subsequent submissions carrying this receipt identifier are included in the priority conditions of the selected path to complete the determination; submissions without a receipt identifier do not trigger the conditions for lifting isolation.

[0017] Furthermore, the upgraded isolation constraint includes a one-way convergence constraint bound to the selected recovery path: when selecting the verification element priority path, the isolation channel only allows the submission of content corresponding to the verification element field until the priority condition is met; when selecting the discussion boundary restatement priority path, only the submission of restatement content that meets the discussion boundary is allowed until the priority condition is met; if content from non-selected paths is blocked before completion, and the preset time limit has expired without completion, the isolation constraint is upgraded.

[0018] The beneficial effects of this invention are as follows: By setting up a phased dialogue routing mechanism with clarification prompts, standardization prompts, and governance consequence prompts, this invention establishes abnormal user identification on the basis of a progressively guided and clearly informed process, avoiding the risk of false positives from triggering abnormal judgments with a single statement. Through preset verification elements, field templates, and cross-field consistency verification mechanisms, structured verification of the content of statements is achieved, providing objective standards for judging information integrity and logical consistency. The introduction of boundary confirmation markers provides verifiable evidence for target users' confirmation of discussion boundaries, thereby improving the accuracy and traceability of abnormal judgments. The overall process is driven by routing behavior, embedding abnormal identification into the dialogue interaction process, improving the transparency and procedural legitimacy of community governance.

[0019] Furthermore, this invention, by setting reversible constraint mechanisms and recovery bifurcation conditions, preserves the recovery channel while performing isolation processing, making anomaly handling correctable and controllable. Through recovery receipt identifiers, path receipt identifiers, and a one-way convergence constraint mechanism, it ensures that the recovery behavior corresponds one-to-one with the selected path, avoiding interference from invalid or chaotic submissions to the system's judgment. By using clarification count identifiers and preset time limit controls, clear termination conditions and escalation rules are established, forming a closed-loop management structure for the isolation and recovery processes. This achieves coordinated linkage between community robot dialogue routing and abnormal user identification, improving system processing efficiency and governance stability while ensuring community order. Attached Figure Description

[0020] Figure 1 This is a logic diagram for the phased dialogue verification and anomaly detection of the social robot in this invention.

[0021] Figure 2 This is a schematic diagram of the multi-layer routing structure of the e-commerce after-sales complaint community robot in Embodiment 1 of the present invention.

[0022] Figure 3 This is a routing statistics chart of complaint messages for each day of a week in the three-stage community robot in Embodiment 1 of the present invention.

[0023] Figure 4 This is a comparison chart of the completeness scores of verification elements after penalties for multiple users in different rounds in Embodiment 1 of the present invention.

[0024] Figure 5 This is a two-dimensional distribution diagram of the verification score and cross-field conflict for anomaly determination in Embodiment 1 of the present invention.

[0025] Figure 6 This is a waterfall diagram illustrating the process of suspicious users entering isolation and dual-path recovery within one week in Embodiment 2 of the present invention.

[0026] Figure 7 This is a schematic diagram illustrating the evolution of the recovery ratio of abnormally isolated users over time under the state transition model in Embodiment 2 of the present invention.

[0027] Figure 8 This is a schematic diagram of the optimal strategy frontier of the cost function J in Embodiment 2 of the present invention. Detailed Implementation

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Combined with appendix Figure 1 This invention relates to a method for dialogue routing and abnormal user identification in social chatbots, applicable to online social conversation environments with multiple members. The system first receives the target user's speech in real time within the social conversation and parses the speech to extract its descriptive elements and intended meaning. The method does not directly make an abnormal judgment on the target user, but rather verifies and guides the target user's speech through a phased dialogue routing mechanism.

[0030] Upon receiving a message from the target user, the system routes the message to the first community bot. The first community bot generates a corresponding clarification prompt based on preset verification elements. These preset verification elements define the basic information necessary for a message to form a valid discussion within the community, including the subject of the event, time information, factual evidence, or verifiable sources. The first community bot then prompts the target user to supplement any missing or incomplete elements in a structured manner and obtains the target user's first response to the clarification prompt.

[0031] The system then verifies the first response. If the first response does not contain all the preset verification elements, or although it contains relevant elements but cannot form verifiable information content, it is determined that the first response does not provide verifiable information that meets the preset verification elements. In this case, the system does not immediately implement restrictive measures, but instead routes the target user's subsequent statements to the second community bot. The second community bot outputs a guidance prompt, which is generated based on the current community rules and discussion scope. This prompt clarifies the discussion boundaries and requires the target user to restate their demands within those boundaries. Through this phased routing mechanism, the target user gradually completes information completion and expression convergence during the process of structured clarification and guidance.

[0032] After completing the second community bot's guidelines and obtaining the second response, the system performs a boundary compliance check on the second response. The discussion boundaries are determined based on community rules, the current topic scope, and announcements already published by the administrator, and are used to limit the target user's topic range and expression. The system checks each item in the second response to determine whether it states the request within the defined topic, whether it removes irrelevant or extended accusations, and whether it restates the original request in a structured manner according to the guidelines.

[0033] When a second response fails to restate the request within the discussion boundaries, the system determines that the target user has not completed the normative correction. Failure to complete the normative correction includes deviating from the current discussion topic, further expanding the scope of accusations, repeating unverified information, or refusing to clearly express the request according to the guidelines. In this case, the system automatically routes the target user's subsequent comments to a third-party community bot.

[0034] The third-party community bot is used to output governance consequence prompts. These prompts are generated based on community governance rules and clearly inform the target user of potential governance measures if they continue to fail to express their demands within the discussion boundaries. These measures may include limiting speaking frequency, transferring to an isolation channel, or suspending public discussion privileges. After outputting the governance consequence prompts, the third-party community bot obtains the target user's third response and records this response as a basis for subsequent abnormal user identification. By routing the dialogue step-by-step to the third-party community bot, the system completes a clear notification and confirmation process before formally implementing governance measures.

[0035] After the third community robot outputs the governance consequences prompt and obtains a third response, the system performs a comprehensive verification and judgment on the third response. This comprehensive verification and judgment includes both a judgment on the completeness of verification elements and a judgment on the compliance with discussion boundaries. The system first checks whether the third response has supplemented the previously missing preset verification elements. Verification elements are used to ensure that the content has a basic factual basis and verifiable information. If the third response still does not contain all the preset verification elements, or the supplemented content cannot form verifiable information, it is deemed that the verification element requirements are not met. The system also reviews the third response for compliance with discussion boundaries, judging whether it restates the demands within the established discussion scope, whether it eliminates irrelevant extended content, and whether it converges its expression according to the guidelines. If the third response still does not restate the demands within the discussion boundaries, it is deemed that it has not completed the normative correction.

[0036] When a third response fails to meet verification criteria or fails to restate the request within the discussion boundaries, the system determines the target user as an abnormal user. Abnormal user determination is based on persistent non-compliant behavior following phased dialogue guidance, and is process-oriented and traceable. After determining an abnormal user, the system applies routing constraint policies linked to the determination to the target user's subsequent statements. These policies include adjusting the routing priority of the target user's statements, prioritizing routing subsequent statements to governance-related community bots or isolation channels, and limiting the frequency or scope of their statements. The routing constraint policies can be set to different strength levels based on the degree of abnormality, and the system continuously monitors the target user's behavior in subsequent dialogues. When the target user's subsequent behavior meets the verification criteria and expresses the request within the discussion boundaries, the system can remove or reduce the routing constraints.

[0037] After receiving a message from a target user, the first community robot does not simply output generalized clarification statements. Instead, it performs structured analysis of the message content based on preset verification elements. These preset verification elements define the basic information elements that a message should possess to constitute a valid discussion in a community environment, including the main subject of the event, the time of occurrence, the parties involved, and the factual basis or source of information. Based on the analysis results, the first community robot identifies missing or incomplete elements and generates a corresponding list of missing elements.

[0038] For each missing element in the list, the first community robot generates a field template. The field template includes a field identifier and a value field. The field identifier clarifies the category of information to be supplemented, while the value field carries the specific content to be supplemented by the target user. The field template is presented in a structured format, allowing the target user to fill in or supplement specific elements, thus avoiding general statements.

[0039] After receiving the first response, the system performs field-level checks. These checks include whether the information corresponding to the field identifier is returned completely, whether the value is empty, and whether the value conforms to the preset validation rules for that field. The preset validation rules are used to determine the basic validity of the supplementary content, such as whether it contains a clear time expression, an identifiable object name, or meets predetermined information format requirements. If the first response has any missing field, an empty value, or fails the corresponding preset validation rules, the system determines that no verifiable information meeting the preset validation criteria has been provided. After making this determination, the system triggers subsequent dialogue routing, transferring the target user's subsequent statements to a second community bot for processing.

[0040] The preset verification rules further include a cross-field consistency verification mechanism. After generating the missing element list, the first community robot performs independent verification on each field and simultaneously checks the correlation between fields. When the missing element list contains both a time field and an object identifier field, the system requires that the corresponding time field value and object identifier field value in the first response meet preset coexistence conditions. The coexistence conditions are based on the system's built-in event logic constraints and are used to determine whether an object has the rationality or possibility of engaging in relevant behavior at a specific time point. During the verification process, the system first checks whether the time field and object identifier field meet their respective preset verification rules, and then performs a consistency comparison between the two. If any field passes the verification individually, but the time field and object identifier field do not meet the coexistence conditions, the combined information is deemed to have a logical conflict, the value is determined to have failed the preset verification rules, and routing to the second community robot is triggered.

[0041] The field template also includes a clarification count identifier for each field. This identifier records the number of times a field has been requested for supplementation or correction during the clarification phase. When the first community robot fails the preset verification rules in its first response, it only re-outputs the field template for the failed field and simultaneously updates the clarification count identifier for that field. Each time the field template is re-output, the system ensures that other fields that have passed verification are not repeatedly prompted to avoid irrelevant interference. When the number of clarifications for the same field reaches the preset limit and still fails the preset verification rules, the system directly determines that the target user has not provided verifiable information that meets the preset verification elements, discontinues the clarification process, and triggers routing to the second community robot.

[0042] When the second community bot outputs the guidelines prompt, it simultaneously generates a boundary receipt identifier corresponding to the current discussion boundary. The discussion boundary is determined based on the community rule text, the current conversation topic, and the management policy, and is used to limit the scope and structure of the content expressed by the target user. The boundary receipt identifier is a unique identifier generated by the system, corresponding one-to-one with the guidelines prompt, and is used to confirm that the target user is aware of and accepts the discussion boundary.

[0043] The second community bot explicitly requires target users to include a boundary confirmation flag when restating their requests and to express them within the discussion boundaries. When a target user submits a second response, the system first checks whether it includes the boundary confirmation flag. If the second response does not include the flag, it is considered as a failure to effectively acknowledge the guidelines, and the system directly determines that the request was not restated within the discussion boundaries.

[0044] When the second response carries a boundary confirmation flag, the system further verifies the boundary compliance of the restated content. Verification includes whether it is confined to the current topic, whether irrelevant extended accusations have been removed, and whether the request is expressed in a structured manner according to the guidelines. If the restated content carries a boundary confirmation flag but is not expressed according to the guidelines, the system determines that the target user has not restated their request within the discussion boundaries. In any of the above situations, the system triggers subsequent dialogue routing, transferring the target user's subsequent statements to a third-party community bot for processing.

[0045] Once the system identifies an abnormal user, it executes the routing constraint policy tied to that identification result. This policy includes a reversible constraint mechanism to provide the target user with an opportunity to correct their behavior while maintaining community order. The system automatically routes the target user's subsequent posts to an isolated channel, a restricted communication environment where their posts do not directly enter the public discussion area.

[0046] Upon entering the isolation channel, the system outputs a verification recovery prompt to the target user. The prompt clearly lists the verification elements that need to be supplemented and reiterates the current discussion boundary requirements. A preset time limit is set in the prompt to define the timeframe within which the target user must complete the corrective action. The target user must supplement all verification elements within the preset time limit and restate their request within the discussion boundaries.

[0047] After receiving a message from a target user in the isolated channel, the system verifies the supplementary content. The verification process includes checking the completeness of elements and reviewing the compliance with discussion boundaries. When the target user completes the verification of all required elements within a preset time limit and meets the discussion boundary requirements, the system removes the isolation channel restriction, restores the normal dialogue route, and allows the user's message to re-enter the public discussion area.

[0048] If the target user fails to complete the verification process within the preset time limit, or if they provide supplementary information but still fail to express their demands within the discussion boundaries, the system maintains the isolation constraint. The system can also escalate the isolation constraint based on the degree of continued non-compliance, including extending the isolation period or further restricting speaking privileges. By setting up a reversible constraint mechanism, the anomaly handling process has clear recovery conditions and escalation rules, forming a controllable and traceable community governance path.

[0049] The verification recovery prompt generates a recovery receipt identifier corresponding to each isolation channel during output. This recovery receipt identifier is a unique identifier generated by the system, used to confirm the corrective actions taken by the target user in response to this isolation process. The recovery receipt identifier is bound to the current anomaly determination record and isolation status to distinguish recovery requests at different stages.

[0050] When the system outputs the verification recovery prompt, it explicitly requires the target user to submit a restated version of the statement with a recovery receipt identifier within a preset time limit. The restated statement must include the completed verification elements and express the original request in a standardized manner within the established discussion boundaries. After the target user submits their statement in the isolated channel, the system first verifies whether it contains the recovery receipt identifier. If it does not contain this identifier, the system determines that the statement does not constitute a valid recovery request, does not enter the verification process, and directly determines that the verification recovery prompt requirements are not met.

[0051] When restated content carries a recovery receipt flag, the system performs a dual review of its element completeness and discussion boundary compliance. If all verification elements are completed and the content meets the discussion boundary requirements, the system removes the isolation status, restores the target user's normal dialogue route, and allows their speech to re-enter the public discussion area. If the content carries the flag but still lacks verification elements or fails to express its demands within the discussion boundaries, the system maintains the current isolation constraints and, if necessary, implements escalated isolation measures based on the persistence of the violation.

[0052] The reversible constraint mechanism further sets recovery bifurcation conditions to improve the operability and verifiability of the anomaly handling process. When the target user enters the isolation channel, the system provides at least two preset recovery paths while outputting a verification recovery prompt. Each recovery path corresponds to a different priority correction direction: one path prioritizes completing the verification elements, and the other prioritizes restating the request within the discussion boundaries. Both paths clearly define their respective completion criteria and time limits.

[0053] After a target user selects a recovery path within the isolation channel, the system records their path selection and sets corresponding priority conditions and deadlines accordingly. If the target user selects the verification element priority path, all verification elements must be completed within the preset time limit; after completing the verification element completion, the request must be restated in accordance with the discussion boundaries. If the target user selects the discussion boundary restatement priority path, the request must be restated in accordance with the discussion boundaries within the preset time limit; after completing the restatement, the relevant verification elements must be completed.

[0054] The system verifies the content submitted by the target user in real time. If the target user fails to complete the priority conditions corresponding to the selected path within the preset time limit, the system escalates the isolation constraints, including extending the isolation period or adding speaking restrictions. If the target user completes the priority conditions of the selected path within the preset time limit and subsequently fulfills another condition, the system removes the isolation status and restores their normal chat routing permissions.

[0055] When providing preset recovery paths, the verification and recovery prompts simultaneously generate a corresponding path receipt identifier for each recovery path. The path receipt identifier is a unique identifier assigned by the system, corresponding one-to-one with a specific recovery path, used to confirm the correction direction selected by the target user. When the target user selects a recovery path within the isolation channel, they must return the corresponding path receipt identifier. The system records their path selection result accordingly and establishes a correspondence between the path and subsequent submissions.

[0056] When a target user submits corrective content in the isolation channel, the system verifies whether it carries a path receipt identifier. Only if the submitted content carries a path receipt identifier that matches the selected path will the system include the content in the priority condition completion determination for the selected path. If the submitted content does not carry a path receipt identifier, or carries an identifier that is inconsistent with the selected path, the system will not include it in the priority condition completion statistics and will determine that it has not triggered the release from isolation conditions.

[0057] Upgraded isolation constraints include a one-way convergence constraint mechanism bound to the selected recovery path. If the target user selects the verification element priority path, the isolation channel only allows submissions of content corresponding to the verification element fields until the priority condition for verification element completion is met. During this stage, the system automatically blocks other content related to discussion boundary restatements. If the target user selects the discussion boundary restatement priority path, the isolation channel only allows submissions of restatement content that meets the discussion boundary requirements until the restatement priority condition is met. Before the priority condition is met, the system blocks submissions related to paths not selected. If the priority condition for the selected path is not met by the preset time limit, the system executes upgraded isolation constraints, including extending the isolation period or further restricting speaking permissions.

[0058] Example 1:

[0059] This embodiment provides a method for routing dialogue and identifying abnormal users in a community chatbot for e-commerce after-sales complaint groups, supporting a group of approximately 100,000 users seeking after-sales support. This community is deployed within a unified complaint portal on an e-commerce platform, where users provide centralized feedback on issues such as order fulfillment, logistics delays, and after-sales service. In practice, many posts contain only emotional accusations and lack crucial information such as order numbers or timestamps, and some users continue to expand their accusations even after being reminded. This embodiment uses a community chatbot to collaboratively guide and identify such posts through structured verification and discussion boundary constraints.

[0060] In terms of overall system architecture, three community robots—the first, second, and third—are deployed in parallel on the platform's message bus. The message bus receives real-time messages from target users and distributes them to the corresponding robots according to the routing strategy. The first community robot is primarily responsible for collecting verification elements and performing field-level validation; the second community robot is mainly responsible for setting discussion boundaries and restating the constraints of the rules; and the third community robot is used to output governance consequence prompts and serve as a final reminder before judging abnormal users. By drawing a multi-layered routing structure diagram of the e-commerce after-sales complaint community robot in a simulation environment, the following results are obtained: Figure 2 The structure shown includes nodes such as user U_101, a message bus, three-stage community robots, and resource nodes such as an order database and a community rule base. Directed edges represent speech and dependency relationships between nodes. This indicates that the shortest path length from user U_101 to the first, second, and third community robots is 2 hops, meaning that a user's speech can reach any community robot via the message bus. This demonstrates a simple and clear sequential structure of verification, standardization, and governance, facilitating a traceable segmented description of the processing logic at each stage in subsequent embodiments. Figure 2 This demonstrates the multi-layered routing structure.

[0061] The system predefines a set of verification elements adapted to e-commerce after-sales complaint scenarios. These include the order number and user account in the event subject information, the order time and complaint time in the occurrence time, the customer service employee number and store identifier in the involved object information, and links to chat logs, logistics tracking numbers, and screenshots of vouchers in the factual evidence or information sources. After receiving a natural language statement from a target user, the First Community Robot uses intent recognition and named entity recognition modules to parse the statement into structured fragments and compares them with the above set of verification elements to generate a list of missing elements. For each element in the list, the system constructs a field template. The field template includes two parts: a field identifier and a value. For example, the value corresponding to the order number is a string, the value corresponding to the event time is a timestamp or date string, and the value corresponding to the complained object is a user identifier or store identifier. The field template is sent to the user in a structured form, forming the first round of clarification prompts.

[0062] To control the convergence of the clarification process, each field template also maintains a clarification count identifier, denoted as count index i, to record the rounds in which the field was asked to supplement or correct. After the user returns the first response, the first community robot verifies the values ​​of each field one by one, including non-emptiness verification, format rule verification, and basic consistency verification with the platform's order database. When a field is found to have a missing or empty value, or a format that clearly does not conform to the rules, or a transaction record matching the order number and time cannot be found in the order database, the system considers the field to have failed the preset verification rules, increments the clarification count identifier for the corresponding field by one, and only re-outputs the field template for the field that failed the verification; other fields that have passed the verification are not prompted again. When the number of clarifications for a field reaches the preset limit, such as 3 times and still fails the verification, the system determines that the user has failed to provide verifiable information that meets the verification requirements within a reasonable interaction cost, triggering subsequent speech routing to the second community robot.

[0063] In addition to performing field-level validation, the first community robot also introduces a cross-field consistency validation mechanism to identify situations where a single field seems reasonable but a combination of fields results in logical conflicts. The system uses the order time field and the complained store identifier field as key association fields, corresponding to transaction records in the platform's order database. For user u's first response at dialogue time step t, the system constructs a verification state vector, denoted as:

[0064]

[0065] Each component This indicates the pass / fail status of a verification field, with a value of 1 for pass and 0 for fail; d represents the total number of fields. Based on this, a verification element completeness score is defined:

[0066]

[0067] in For the first The weighting coefficients for each field are used to reflect the higher importance of order number and time information compared to fields such as supplementary description text. To characterize cross-field consistency, the system defines a consistency penalty term:

[0068]

[0069] in The value to be taken for the event time field. The function is used to determine the value of the "Identification" field for the store being complained about. This indicates that a matching condition does not exist in the order database. and Transaction records, indicator functions The value is 1 when a conflict occurs, and 0 otherwise. The system uses the following inequality for judgment:

[0070]

[0071] in This is the conflict penalty coefficient. The threshold for verification is set. When the above inequality holds, even if some fields pass the verification individually, the system still determines that the user has not provided verifiable information that meets the preset verification elements due to cross-field conflicts or a low overall score, and routes subsequent messages to the second community robot for processing. To further demonstrate the impact of this judgment mechanism on different users, this embodiment selects users U_101, U_203, and U_350 as representative samples. Field pass vectors and conflict indices are constructed for dialogue rounds T3, T4, T5, and T8, respectively. The verification score after penalty is calculated according to the above formula, resulting in the following... Figure 4 The curves showing the comparison of verification element completeness scores after multi-user penalties are presented. Figure 4 It can be seen that user U_203 can stably reach the threshold after round T4. The above indicates that users quickly filled in key elements such as event time and the complained party after receiving the field template prompts; user U_101's score was significantly lower than the threshold after being penalized in round T5 due to cross-field consistency conflicts, and it was not until round T8 that the score exceeded the threshold after all fields were filled in and the conflicts were resolved; user U_350 remained in the low score range for a long time in the first few rounds, and even in round T8, the score was only close to the threshold due to weak factual evidence. This set of curves intuitively depicts the differences in convergence speed among different users under the influence of clarification number control and cross-field penalties.

[0072] During the overall system operation, routing results can be statistically analyzed to guide the design of thresholds and weights. For example, in a simulated dataset, within a certain time period, 100 users sent 500 complaint messages to a rights protection group. Approximately 320 messages, after initially entering the first community bot, passed field-level verification and cross-field consistency checks, and were not subsequently routed to other bots. About 130 messages, accounting for 26% of the total messages, were deemed incomplete due to missing key information or field conflicts and were routed to the second community bot. The remaining approximately 50 messages, accounting for 10% of the total messages, were not effectively standardized in the second and third community bots and were ultimately marked as potentially abnormal users. The system can also estimate the probability of a message entering the second community bot.

[0073]

[0074] in The number of messages routed to the second community bot. This represents the total number of messages. By monitoring this ratio in real time, the threshold can be dynamically adjusted. With penalty coefficient This allows the verification mechanism to strike a balance between recalling abnormal behavior and controlling the burden on users. To further refine the above statistics, a simulation of the daily distribution and phased routing of 500 complaint messages was conducted within a one-week time window. The distribution of messages on each day among the three categories of "directly passed by the first community bot," "routed to the second community bot," and "entered the third community bot or marked as abnormal" was plotted in the form of a stacked bar chart, such as... Figure 3 As shown. By Figure 3 As can be seen, the total number of messages within a week was 500. Messages directly verified by the first community robot accounted for over 60% of the total, messages routed to the second community robot accounted for approximately 26%, and messages entering the third community robot or being marked as potentially suspicious users accounted for approximately 10%. The daily stacking results also show that the proportion of messages entering the second community robot was roughly distributed between 24% and 29% each day. This indicates that, with fixed thresholds and penalty coefficients, the first community robot had filtered out most complaints lacking verification elements at the source, while the second and third community robots mainly focused on the regulation and management of the remaining long-tail cases. On the other hand, to observe the overall verification score... Cross-field conflict penalty item To understand the relationship between the original verification scores, conflict indices, and post-penalty verification scores, this embodiment also randomly generates simulated samples of several users at different time steps to construct a two-dimensional scatter distribution, resulting in the following: Figure 5 The diagram shows a two-dimensional distribution of verification scores and cross-field conflicts in anomaly detection. Figure 5 The middle horizontal axis represents the original verification score. The vertical axis represents the score after the penalty. Based on whether there are conflicts and whether the sample points have passed verification, the data is grouped and plotted. It can be clearly seen that the samples with conflicts are generally plotted along a slope of approximately 1 and an intercept of approximately [missing information]. The direction of the "" is shifted downwards relative to conflict-free samples. Even if some samples have high original scores, they may still fall below the threshold after applying cross-field penalties. If the user is deemed to have failed to provide verifiable information that meets the preset verification criteria, it needs to be routed to the second community robot to execute boundary constraints. This planar diagram intuitively illustrates the compression effect of the penalty term on the decision boundary, facilitating the explanation of the quantification mechanism for abnormal user identification in the accompanying diagram.

[0075] When the first community bot determines that a user has failed to provide verifiable information in the current round, the second community bot takes over the user's subsequent statements. Based on the e-commerce platform's community rule base and the current conversation's topic context, the second community bot automatically generates the boundaries of the current discussion round. For example, it limits the discussion to a specific order's logistics delay issue, prohibits extending accusations to the entire platform, and prohibits adding other merchants or historical disputes unrelated to the current order. The second community bot then generates a boundary token, denoted as `boundary_token`, that corresponds one-to-one with the current specification prompt. This token can be a random string or an encrypted digest, and the specification prompt explicitly requires users to include this token when restating their demands, while also expressing it according to the structured format of the problem description module and the expected processing result module.

[0076] After receiving the user's second response, the second community bot first identifies the message content and checks if it contains the correct boundary_token. If no identifier is detected, the system assumes the user has not effectively acknowledged the specification prompt, determines the user has not restated their request within the discussion boundaries, and maintains the routing state unchanged, meaning the second community bot continues to handle subsequent prompts. If a correct boundary_token is detected, the system further performs boundary compliance checks, matching the text in the second response with the current discussion boundaries to check if the user has limited their statement to the specified order and issue type, removed irrelevant extended accusations, and provided a problem description and expected handling method in the specified structured format. If the second response, although carrying a boundary_token, still spans multiple orders or indiscriminately attacks the platform as a whole, the system determines that the user has still not restated their request within the discussion boundaries and, if necessary, escalates the routing of the user's subsequent statements to the third community bot.

[0077] In this embodiment, the third community robot acts as a governance consequence notification module. Upon receiving a route from the second community robot, it clearly explains to the user the potential governance consequences of continuing to disregard verification elements and discussion boundaries, including restrictions on speaking frequency and temporary suspension of public discussion privileges. If the user still fails to complete the verification information or fails to express their demands within the discussion boundaries in the third response, the system can mark the user as an abnormal user and pass the user mark to the downstream reversible constraint and isolation channel module. The specific constraint strategies will be detailed in other embodiments and will not be elaborated upon in this embodiment.

[0078] To illustrate the above method's operation more intuitively, a specific user case is provided below. User U_101 first posts in a rights protection group, in dialogue round T1, stating that the store is ripping people off without providing an order number, time, or any evidence. The message bus routes this message to the first community robot. After parsing, the first community robot finds that all verification elements are missing, generates a list of missing elements including order number, user account, event time, the complained party, and links to factual evidence, constructs corresponding field templates, and sends a clarification prompt to U_101, in round T2. In the first response in T3, U_101 only fills in the order number as ORD-20250101, leaves the event time field blank, fills in a vague name for the complained party, and does not provide any screenshots or logistics information. After the system performs field-level validation on this response, it finds that multiple fields fail, corresponding to the vector... Only a few components are 1, and the calculated verification element completeness score C is significantly lower than the threshold. Simultaneously, the clarification count for failed fields is incremented by one, and field templates are reissued only for these fields. Figure 4 In the multi-user scoring trajectory, U_101's post-penalty verification score in round T3 was significantly below the threshold. The area.

[0079] In dialogue rounds T4 and T5, U_101 attempted to supplement information twice. The first time, they added an incorrectly formatted time string; the second time, they added a date that did not match the actual transaction date of the order in the platform's order database. After format validation in T4, the system prompted for correction again and incremented the count value. In T5, by accessing the order database, it was discovered that the combination of the user-provided time and store identifier did not have a corresponding transaction record in the system, triggering a cross-field consistency conflict, and incurring the corresponding conflict penalty. The score becomes 1, resulting in a penalized overall score. Continue to be below the threshold At this point, the clarification count in the event time field has reached the preset limit of 3. The system determines that U_101 has failed to provide verifiable information at this stage and routes subsequent messages to the second community bot. Figure 4 As shown in the curve, the inflection point of the score for round T5, U_101, clearly reflects the penalty effect caused by cross-field conflict, and its score after penalty is still below the threshold line.

[0080] After the route switch, in dialogue round T6, the second community bot sent a specification prompt to U_101. The prompt clearly stated that this round of discussion only revolved around the logistics delay issue of order ORD-20250101, requiring the user to provide a structured statement according to the problem description module and the expected processing result module, and not to include any extended accusations related to other orders or the overall reputation of the platform. Simultaneously, the system generated a boundary token (boundary_token) and displayed it at the end of the message, requiring the user to include this token in the next restatement. When U_101 responded in T7, although providing a more detailed complaint text, it did not include the boundary token in the message and again included generalized accusations such as all shops on the platform being scammers. The second community bot first noticed the lack of a boundary token, determining that U_101 had not effectively confirmed the discussion boundary, and therefore directly concluded that the request was not restated within the discussion boundary, and remained in the current routing stage.

[0081] After a brief silence, U_101 responded again at T8, this time copying the boundary_token at the beginning of the message. Following the specifications, it described the problem in one sentence, stating that order ORD-20250101 was signed for three days after its expected delivery time, and listed its desired solutions, including a refund of shipping fees and a coupon. In this round, the second community bot detected the correct boundary_token, confirming that the user had acknowledged the discussion boundary. The topic matching algorithm further confirmed that the text content was largely confined to the logistics issue of that order, without any new, extended accusations. Based on this, the system determined that U_101 had completed a structured restatement within the discussion boundary and marked the response as successfully verified. At this point, the user was no longer routed to the third community bot but returned to the regular after-sales process, where their request was handled by human customer service or the automated ticketing system.

[0082] Example 2:

[0083] In this embodiment, a city public service complaint community platform with over 500,000 users is used. This platform provides centralized feedback from citizens on matters such as municipal facility maintenance, public transportation operation, and community governance. In a high-concurrency discussion environment, abnormal behaviors such as unfounded accusations, malicious spamming, and incitement of emotions can significantly reduce the usability and efficiency of public discussions. Therefore, it is necessary to establish traceable, reversible, and quantifiable constraint paths while ensuring the smooth operation of the complaint channel.

[0084] This embodiment uses the three-stage community robot mechanism of Embodiment 1 as a pre-process, whereby the target user's speech in the public discussion area undergoes hierarchical guidance through verification and clarification, discussion boundary standardization, and governance consequence prompts. If the target user still fails to complete the verification elements or fails to restate their request in a standardized manner within the discussion boundaries in the third stage, the system marks them as an abnormal user and binds this abnormal mark to the routing strategy for subsequent speeches. Figure 6 Presenting the scale changes from suspected to isolated and then to recovered using group statistics, this illustrates the immediate isolation effect of routing strategies on public area order after anomaly marking takes effect.

[0085] Once the system determines that a target user is an anomalous user, its subsequent posts are automatically routed to an isolated channel. The isolated channel is a restricted communication environment; its messages do not directly enter the public discussion area, but still retain the ability to interact with the governance module, allowing users to make corrections within a controlled scope. The isolated channel and the public discussion area employ different message visibility strategies. Isolated channel messages are only visible to the target user and the governance system by default, and are recorded as evidence of subsequent restrictions being lifted or escalation measures. This isolation strategy manifests at the group level as a significant reduction in the spillover of anomalous posts from the public area. In the example statistics, this can be approximated as the visible anomalous posts of the anomalous user in the public area approaching zero after isolation, thus providing a basic quantitative condition for the reduction in anomalous posts in the subsequent cost function.

[0086] When a user enters the isolation channel, the system outputs a verification recovery prompt, which includes the verification elements to be completed, the current discussion boundaries, and a preset time limit. Verification elements are configured based on the municipal complaint scenario and include at least the facility or road segment number, geographical coordinates, time of occurrence, description of the scope of impact, verifiable evidence links such as photos or videos, and optional work order numbers or contract numbers. The discussion boundaries limit the discussion to a specific public service matter, such as the frequency of garbage collection in a certain community or the repair timeframe for a damaged road segment, prohibiting generalized accusations against unrelated entities. The preset time limit can be configured as 24 hours or 48 hours, defining the window within which the user can complete the corrective action. This preset time limit can be mapped to the window length of discrete time steps in subsequent state transition analysis; for example, a 6-hour step corresponds to 4 steps in 24 hours, thus characterizing the recovery action as a time-constrained reversible transition process.

[0087] When the verification recovery prompt is issued, the system generates and binds a recovery receipt identifier, `recovery_token`, which corresponds one-to-one with the isolation channel session and is used to confirm the validity of the recovery request under this isolation process. The system requires users to include the `recovery_token` when submitting restatements in the isolation channel. Messages without this identifier will not enter the verification process and will be directly judged as invalid recovery requests, and will not be counted in the statistics for meeting the release conditions. This mechanism provides a clear trigger boundary for the recovery process, preventing users from interfering with governance decisions by spamming the isolation channel without an identifier. In terms of operations and maintenance statistics, submissions without a `recovery_token` can be counted as invalid submissions and used as one of the inputs for calculating the upgrade isolation constraint strength, together with the duration of non-compliance, to determine the subsequent restriction level.

[0088] To improve the operability and verifiability of corrections, the verification recovery prompt provides at least two recovery paths and forms recovery bifurcation conditions. Path A is the verification element priority path, requiring users to first complete all verification element fields, and then perform a structured restatement that conforms to the discussion boundary after satisfying the verification element completeness requirement. Path B is the discussion boundary restatement priority path, requiring users to first complete the specification restatement within the discussion boundary, and then complete the verification elements after satisfying the restatement requirements. The system configures path receipt identifiers path_token_A and path_token_B for each path. When users select a path within the isolated channel, they need to return the corresponding identifier. The system records the path selection result accordingly and establishes the correspondence between the path and subsequent submissions. The group-level path bifurcation results can be viewed... Figure 6 It is visualized in the middle. Figure 6 The system uses a waterfall chart to sequentially display the total number of suspicious users, the pre-stage correction, and the entry into isolation. After isolation, it is split into path A and path B, and further split into two categories: 24-hour recovery and no recovery. This makes the scale convergence chain of the governance process readable and verifiable.

[0089] When a user submits corrective actions within the isolation channel, the system performs dual-identity verification and path consistency verification. Dual-identity verification requires the submitted content to carry both a `recovery_token` and a `path_token` corresponding to the selected path; otherwise, the submission will not be prioritized for completion. Path consistency verification requires that the `path_token` carried in the submitted content matches the path selection recorded by the system. If the user carries a `path_token` that does not match the selected path, the system determines that the conditions for lifting isolation have not been triggered and may prompt the user to resubmit using the selected path. Through the combination of the receipt identifier and the path identifier, the system achieves traceable recording of recovery actions and reduces the risk of recovery requests being forged or mistakenly recorded. This mechanism is as follows: Figure 6In the path bifurcation statistics, only when both the path identifier and the recovery identifier are correct is the submission counted in the path recovery statistics, thus ensuring that the path recovery rate matches the actual effective recovery requests.

[0090] To achieve unidirectional convergence constraints, the system imposes directional restrictions on the content that can be submitted through the isolation channel after the user selects different paths, until the priority conditions of the selected path are met. If the user selects path A, before the priority conditions of the verification elements are met, the isolation channel only allows submissions of content related to the verification fields, including facility number, time, location, evidence links, etc. The system automatically blocks long comments or emotional texts unrelated to the boundary restatement. If the user selects path B, before the priority conditions of the discussion boundary restatement are met, the isolation channel only allows submissions of text that meets the structured restatement template, such as a three-part structure of problem description, scope of impact, and expected solution. The system blocks submissions that only supplement evidence but do not follow the structured restatement. Once the priority conditions are met, the system removes the unidirectional restrictions and allows the user to complete the other condition to achieve a dual satisfaction of the verification elements and the discussion boundary. This unidirectional convergence is represented in the state model as the transition probability from abnormal isolation to recovery being constrained by both the path and the time limit, and can be achieved through... Figure 7 The discrete-time evolution curve shown represents a gradually increasing recovered probability.

[0091] The system sets independent priority conditions and completion deadlines for each path, such as completing the element completion for path A within 12 hours, or completing the structured restatement for path B within 12 hours. If the completion is not completed by the deadline, the system implements upgraded isolation constraints, including extending the isolation period, further reducing the upper limit of speaking frequency, and temporarily freezing the user's historical speaking history visible in the public area. The strength of the upgraded isolation constraints can be correlated with the duration of non-compliance and the number of invalid submissions to achieve progressively stronger control over persistent abnormal behavior. When a user completes the priority conditions of the selected path within the preset time limit and subsequently completes another condition, the system removes the isolation state and restores their normal dialogue route, allowing their speech to re-enter the public discussion area. This upgrade strategy corresponds to the constraint boundary between the isolation duration term and the probability of false positives in the cost trade-off analysis, and in... Figure 8 The optimal strategy frontier is reflected in the selection of the optimal isolation duration under a given probability of accidental injury.

[0092] To quantify the governance effectiveness of this reversible mechanism, this embodiment provides a set of example statistics. In a certain week, the platform detected 800 potentially abnormal users. Of these, 200 were identified as abnormal users and entered the isolation channel after completing the initial three-stage guidance. The remaining 600 completed corrections before the third stage and did not enter isolation. Of the 200 abnormal users who entered the isolation channel, 120 chose path A, and 80 chose path B. The example results show that among users on path A, 84 recovered and were released from isolation within 24 hours, a recovery rate of 70%, with an average isolation duration of 16 hours. Among users on path B, 48 recovered and were released from isolation within 24 hours, a recovery rate of 60%, with an average isolation duration of 14 hours. The remaining unrecovered users were subject to upgraded isolation; some recovered within an extended window, while others were transferred to stricter restrictions due to continued non-compliance. The above statistics are presented in... Figure 6 The corresponding total number of suspected cases is 800. After a preliminary adjustment reducing the number by 600, 200 enter quarantine. After quarantine, these are further divided into 120 choosing path A and 80 choosing path B, and then further subdivided into 84 recovering and 36 not recovering on path A, and 48 recovering and 32 not recovering on path B. Based on this statistic, the percentage of cases entering quarantine can also be calculated as follows: That is, 25%, the proportion of path A selection is That is, 60%, the proportion of path B selection is That is, 40%, and the overall recovery rate within the 24-hour window is... That is, 66%.

[0093] To characterize the reversible transition process from an abnormal state to a normal state, the user state set is defined as follows:

[0094] S = {Normal, Suspicious, Abnormal (Isolated), Recovered, Permanently Restricted}

[0095] With a discrete time step of 6 hours, using row vectors Indicates the first The probability distribution of the user's state during each step is calculated using the state transition matrix. Describe the transfer relationships under the current governance strategy, and satisfy...

[0096]

[0097] in The Line number Column elements represent states Transition to state The probability of a state transition during anomalous isolation can be characterized by path hybrid weights, which represent the effective transition probability from anomalous isolation to recovery. It is obtained by weighting the path selection ratio and the recovery probability within the path.

[0098]

[0099] In the example data above, below , ,thereby This indicates that the expected proportion of individuals transferred from abnormal isolation to recovered individuals within a 24-hour window is approximately 66%. If... Let the selected unit column vector represent the recovered state. Then, the expected recovery ratio after (n) time steps can be written as:

[0100]

[0101] in This refers to matrix exponentiation. Let be the initial distribution vector when entering isolation. For individuals entering isolation, the probability of an abnormal isolation state can be 1. Figure 7 Based on the aforementioned discrete time setting, a stepped evolution curve is plotted with 6-hour intervals to represent the probabilities of recovery, remaining in isolation, and permanent confinement. A horizontal target line is used to indicate example statistics at the 24-hour mark. Under the simulation parameters corresponding to the attached figure, Figure 7 exist The recovery probability given at the hour is approximately consistent with the target value of 66%, which illustrates that the state transition matrix parameters can be calibrated using monitoring data from the operation and maintenance side to align the recovery ratio within the window with the actual governance effect.

[0102] To weigh the governance benefits of upgrading isolation constraints against the potential costs of collateral damage, the cost function is constructed as follows:

[0103]

[0104] in This indicates a reduction in abnormal speech. Weighting for order gains. This represents the probability that a normal user is mistakenly identified and placed in quarantine. Weighting of accidental injury penalties This represents the expected duration of the quarantine. Weighting the interaction costs associated with isolation. Using example data, the isolation mechanism can reduce the visible speech of abnormal users in public areas to approximately zero, making... This significantly increased the number of users, while simultaneously enabling 66% of the isolated users to return to normal discussions within 24 hours through a reversible recovery path, thereby controlling the spread of the virus. Within an acceptable range. If the platform keeps the probability of false positives low, the cost function can remain positive under reasonable parameters, indicating that the reversible constraint strategy has an adjustable optimal range between order benefits and user experience. Figure 8Frontier computation of the policy space based on the cost function, given Traversal in the case of Get the value and select to make Maximum optimal isolation time At the same time, the optimal cost function value is given. Thus forming The frontier curve of the changing optimal strategy. Figure 8 Example points in the middle are adopted Compared with weighted average isolation time Calculated in hours value, and the same Compare the optimal points below, and mark the upgrade isolation points. Hours The change is used to illustrate that when the probability of false positives is fixed, excessively extending the isolation period will increase the interaction cost term and make... The decrease provides a quantitative reference for setting up increased isolation levels.

[0105] Citizen user U_205 repeatedly posted accusations in the municipal road maintenance group, claiming that a certain leader was taking kickbacks, leading to annual road damage. Despite repeated requests to provide specific road sections, times, work orders, or contract numbers, U_205 refused to provide these details and continued to broaden the scope of their accusations. After the initial three-stage guidance, U_205 still failed to provide the required verification elements in the third stage and did not express a clear demand within the discussion boundaries. The system marked U_205 as an abnormal user and switched their route. In dialogue round T1 corresponding to t=0 hours, U_205's message was automatically redirected to an isolation channel. The system immediately issued a verification recovery prompt, requiring U_205 to provide the road section number, geographic coordinates, time of occurrence, photo link, and optional work order number. The discussion was limited to requests only regarding the repair timeframe for a specific road section, and a recovery token was issued for subsequent recovery requests. This moment can be considered the initial condition of the state vector. In the abnormal isolation, the component takes a value of 1, while in other states it takes a value of 0, thus entering the subsequent discrete-time evolution.

[0106] In dialogue round T2 Hours later, U_205 sent another emotionally charged accusatory text message in the isolation channel without a recovery token. The system determined this to be an invalid recovery request and did not include it in the statistics. It also prompted the user to include a recovery token and select a recovery path. This corresponds to dialogue round T3. At hour 10:00, U_205 selects path B and returns path_token_B. The system records its path selection and enables one-way convergence constraints, allowing only three-part structured restatements to be submitted before the priority conditions of path B are met. Dialogue round T4 corresponds to... Hourly, U_205 submitted a text containing recovery_token and path_token_B, but it was still a generalized accusation and not expressed according to the three-part template. The system recognized that it did not meet the content constraints of path B and rejected it, returning a structured template prompt. Invalid submissions in this process can be counted in the invalid interaction count on the operations side, and after reaching a certain threshold or continuously timeout, the strengthening parameters of the isolation constraints will be triggered.

[0107] Dialogue round T5 corresponds to Within 24 hours, U_205 submitted a structured restatement according to the template. The problem description included multiple potholes on a road section affecting traffic, impacting nearby residents' vehicles and bus stop access. The desired solution was to complete temporary repairs within 48 hours and develop a formal repair plan within 7 days, including a recovery_token and path_token_B. The system determined that it had fulfilled the priority condition for path B, included the submission in the priority condition statistics, and removed the one-way restriction on path B, allowing it to enter the stage of supplementing verification elements. At this point, the system still required it to supplement verification elements such as road section number, location coordinates, occurrence time, and photo link within the total 24-hour time limit; otherwise, it would trigger an escalation isolation constraint. This behavior corresponds to the intermediate process in the anomaly isolation where an individual meets some conditions within the window but has not yet reached the two-condition release from isolation, explaining why there are still unrecovered individuals entering the escalation isolation stage within the 24-hour window.

[0108] Dialogue round T6 corresponds to Hourly, U_205 submitted a photo link but omitted the road segment number and coordinate information. The system determined that the verification elements were still incomplete and prompted a missing field. Dialogue round T7 corresponds to... After one hour, U_205 still hadn't completed the crucial fields. The system then implemented an upgrade to the isolation constraints, extending the isolation period by 24 hours and reducing the maximum speaking frequency to one message per hour. Simultaneously, it froze some of U_205's historical messages visible in the public area to minimize spillover effects. Dialogue round T8 corresponds to... Within hours, U_205 completed the road segment number and coordinates and submitted a photo link with a timestamp. The system completed the element integrity verification and confirmed that its restated content was still within the discussion boundary. Based on this, the isolation was lifted and its normal route was restored, allowing its subsequent posts to re-enter the public discussion area and be transferred to the work order processing flow.

[0109] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for routing dialogues and identifying abnormal users in a social media chatbot, characterized in that... include: Receive messages from target users in community conversations; The message will be routed to the first community bot to provide clarification prompts and obtain the first response; If the first response does not provide verifiable information that meets the preset verification elements, subsequent messages will be routed to the second community robot to output specification prompts and obtain the second response; If the second response does not restate the request within the discussion boundaries as prompted by the specifications, the subsequent speech will be routed to the third community robot to output the governance consequences and obtain the third response; If the third response still fails to meet the verification elements or fails to restate the request within the discussion boundaries, the target user is determined to be an abnormal user; and a routing constraint policy bound to the determination is applied to subsequent statements.

2. The method for routing dialogues and identifying abnormal users in a social robot according to claim 1, characterized in that... The first community robot generates a list of missing elements based on the preset verification elements and outputs a field template for each missing element. The field template includes a field identifier and a value position. If the first response is missing any field, has an empty value, or fails to meet the preset verification rules corresponding to that field, it is determined that no verifiable information that meets the preset verification elements has been provided, and routing to the second community robot is triggered.

3. The method for routing dialogues and identifying abnormal users in a social robot according to claim 1, characterized in that... The second community robot issues a boundary receipt identifier bound to the discussion boundary in the specification prompt, and requires the target user to include the identifier when restating the request and satisfy the discussion boundary; when the second response does not include the identifier or includes it but does not restate it according to the specification, it is determined that the request has not been restated within the discussion boundary and the routing to the third community robot is triggered.

4. The method for routing dialogues and identifying abnormal users in a social robot according to claim 1, characterized in that... The routing constraint strategy includes a reversible constraint mechanism: the target user's subsequent statements are routed to the isolation channel and a verification recovery prompt is output, requiring the verification elements to be supplemented within a preset time limit and the request to be restated within the discussion boundary; if the condition is met, the isolation is lifted and the normal routing is restored; if the condition is not met, the isolation constraint is maintained or upgraded.

5. The method for routing dialogues and identifying abnormal users in a social robot according to claim 2, characterized in that... The preset verification rules include cross-field consistency verification: when the missing element list contains both a time field and an object identifier field, the time field value and the object identifier field value in the first response are required to be true simultaneously under the preset coexistence condition; when either field passes the verification individually but the two do not meet the coexistence condition, it is determined that the value has not passed the preset verification rules, and routing to the second community robot is triggered.

6. The method for routing dialogues and identifying abnormal users in a social robot according to claim 2, characterized in that... The field template includes a clarification count identifier corresponding to each field. When the first community robot fails the preset verification rule in the first response, it only outputs the field template again for the field that failed the verification and updates the clarification count identifier. When the number of clarifications for the same field reaches the preset upper limit and still fails the preset verification rule, it is directly determined that no verifiable information that meets the preset verification elements has been provided, and routing to the second community robot is triggered.

7. The method for routing dialogues and identifying abnormal users in a social robot according to claim 4, characterized in that... The verification recovery prompt includes a recovery receipt identifier bound to the isolation channel, requiring the target user to submit restatement content carrying the identifier within a preset time limit; when the restatement content completes the verification elements and restates the request within the discussion boundary, the isolation is lifted and the normal route is restored. If the aforementioned identifier is not carried, it is determined that the verification recovery prompt has not been met and the isolation constraints will be maintained or upgraded.

8. The method for routing dialogues and identifying abnormal users in a social robot according to claim 4, characterized in that... The reversible constraint mechanism sets recovery bifurcation conditions: the verification recovery prompt provides at least two preset recovery paths, corresponding to the priority of verification elements and the priority of discussion boundary restatement, respectively; if the target user fails to complete the corresponding priority condition within the preset time limit after selecting a path, the isolation constraint is upgraded; if the corresponding priority condition is completed and another condition is fulfilled, the isolation is lifted and the normal route is restored.

9. The method for routing dialogues and identifying abnormal users in a social robot according to claim 8, characterized in that... The verification recovery prompt is configured with a path receipt identifier for each preset recovery path; When the target user selects a path, a corresponding receipt identifier is returned. Subsequent submissions carrying this receipt identifier are included in the priority criteria for the selected path to complete the determination. Submissions without a receipt will not trigger the conditions for lifting quarantine.

10. The method for routing dialogues and identifying abnormal users in a social robot according to claim 8, characterized in that... The upgraded isolation constraints include unidirectional convergence constraints bound to the selected recovery path: when selecting the verification element priority path, the isolation channel only allows submission of content corresponding to the verification element field until the priority condition is met; when selecting the discussion boundary restatement priority path, only submission of restatement content that meets the discussion boundary is allowed until the priority condition is met; if content from non-selected paths is blocked before completion, and the preset time limit has expired without completion, the upgraded isolation constraints will be implemented.

Images