Interaction guidance and process feature evaluation method and system based on data semantic shunting

Abstract

The application relates to the technical field of a computer system based on a specific calculation model, and discloses an interactive guidance and process feature evaluation method and system based on data semantic shunting, which comprises the following steps: a front-end task routing module performs semantic recognition on user input, asynchronously records original interaction data and generates a guidance response, and records process state feature values at the end of an interaction stage; a back-end judgment module analyzes a final scheme submitted by a user to obtain a terminal state result feature value, and identifies abnormal data flow in which a process and a result do not match by searching for historical process state feature values and performing bivariate cross judgment on the terminal state result feature value. The system adopts a single-table heterogeneous storage structure, realizes data semantic shunting and isolation through an action type field, and supports rapid expansion and deployment through a motherboard architecture. The application realizes high-fidelity process data acquisition and automatic abnormality identification.

Description

Technical Field

[0001] This invention relates to the technical field of computer systems based on specific computational models, and in particular to a method and system for interactive guidance and process feature evaluation based on data semantic decomposition. Background Technology

[0002] With the popularization of Large Language Models (LLM) and artificial intelligence technologies, agent-based interactive assistance systems have been widely used in the guidance and evaluation of various complex tasks. Existing interactive evaluation systems are mainly divided into general large-model dialogue frameworks, customized agents, and traditional online process monitoring systems. However, a review and analysis of existing technologies reveals the following systemic defects in practical complex task interaction and end-to-end evaluation scenarios: 1. The data acquisition mechanism has technical defects such as synchronous blocking and feedback pollution.

[0003] Existing systems typically rely on explicit evaluation forms, intuitive scoring nodes, or simple text log exports when collecting user interaction behavior and cognitive state data. This explicit, intuitive recording generates strong observational feedback within the interaction workflow, leading to adaptive shifts in user input characteristics (i.e., input behavior distortion). Existing systems struggle to achieve asynchronous, silent offloading and storage of underlying data based on semantic intent without interfering with the main front-end interaction thread.

[0004] 2. The evaluation model is based on a single final state result, which has a blind spot for detecting abnormal generated flows.

[0005] Existing large-scale model evaluation systems generally employ a one-dimensional judgment model of result feature extraction and standard comparison during quality acceptance. The system only quantifies the quality of the user's final submitted text or solution, without incorporating the user's time-series state during the interaction process (such as exploration depth, feature retrieval frequency, and other state characteristics) into the cross-validation mechanism. Therefore, when the system receives "high-quality final-state text" directly generated by external large-scale model tools and lacking intermediate interaction chains, existing systems cannot identify abnormal features in the input stream, making it difficult to achieve automated identification for forgery and cheating prevention.

[0006] 3. The system architecture has strong coupling between the running logic and specific tasks, and weak dynamic derivation capabilities.

[0007] In complex workflow systems with multiple stages and nodes, existing agent architectures typically tightly bind knowledge graphs, decision rule sets, and task-specific interaction prompts. This highly coupled architecture necessitates the refactoring of the entire underlying logic code when facing cross-task or cross-domain migrations. It also prevents the low-cost derivation of submodules through simple parameter or variable substitutions, severely limiting the system's scalability and deployment. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and to provide an interactive guidance and process feature evaluation method and system based on data semantic diversion.

[0009] The objective of this invention is achieved through the following technical solution: The interactive guidance and process feature evaluation method based on data semantic splitting includes the following steps: S1. The front-end task routing module receives the input feature stream from the user terminal; S2. Based on the preset semantic filtering rules, the input feature stream is analyzed and identified in real time. When the input feature stream matches the exploratory interaction features, step S3 is triggered. When the input feature stream matches the structured scheme text features, step S3 is blocked. S3. Based on the asynchronous silent recording mechanism, call the data persistence interface to write the first type of action data record into the shared storage module. The first type of action data record includes the original text of the input feature and the label of the current interaction stage. S4. The front-end task routing module generates a guidance response and returns it to the user terminal based on the strategy configuration of the current interaction node and the local knowledge base. S5. When it is determined that the current interaction node meets the settlement conditions of the process stage, based on the asynchronous silent recording mechanism, the second type of action data record is written to the shared storage module through the data persistence interface. The second type of action data record contains the quantitatively generated process state feature value and qualitative evaluation label. S6. After the user submits the final solution text, the backend judgment module uses the rule parsing engine to parse the final solution text and generate the final state result feature value. At the same time, based on the user identifier, it retrieves the process state feature value corresponding to the user from the shared storage module and counts the number of historical data records of the user's existing action type as the third type of action. Based on the number of historical data records, it automatically generates the tag of the currently submitted iteration version. S7. Substitute the process state feature value of step S5 and the final state result feature value of step S6 into the preset bivariate threshold matrix for logical calculation and cross-judgment to obtain the judgment result. S8. Write a third type of action data record into the shared storage module. The third type of action data record contains the final state result feature value and the iteration version label. In the data field structure, force the corresponding field used for the stored procedure state feature value to be empty. S9. Based on the judgment result of step S7, return the corresponding final qualitative evaluation report to the user terminal.

[0010] Preferably, the bivariate threshold matrix in step S7 includes a four-quadrant determination rule, and the first and second thresholds can be dynamically adjusted and configured by the administrator according to the specific task attributes without changing the bivariate feature coupling logic of the four-quadrant determination rule itself.

[0011] Preferably, the logical calculation and cross-determination process in step S7 is as follows: when the process state feature value of step S5 is lower than the first threshold and the final state result feature value of step S6 is higher than or equal to the second threshold, it is identified as an external generation flow anomaly and an early warning action is triggered.

[0012] Preferably, the asynchronous silent recording mechanism of step S3 or step S5 is as follows: the data recording writing action precedes the system-generated guidance response or dialogue output in terms of timing, and no front-end output content of the system contains any prompt information about the executed record or the score.

[0013] Preferably, step S5 specifically includes the following steps: S501. Settlement conditions for process stages include, but are not limited to: completing a preset sub-task, reaching a threshold number of interaction rounds, or the user input containing a stage end keyword. S502. When the settlement conditions of the process stage are met, the user's interactive performance in that stage is quantitatively evaluated, and a process state characteristic value is generated. S503. Write the process status characteristic value of step S502 into the shared storage module as a second type of action type data record.

[0014] Preferably, step S2 includes the following steps: S201. Perform real-time analysis on the received input feature stream to identify semantic intent; S202. Match the identified semantic intent with the preset exploratory interaction features to obtain the first matching result, and match the identified semantic intent with the preset structured scheme text features to obtain the second matching result. S203. Perform traffic splitting based on the first and second matching results: If the first matching result is a successful match, then the asynchronous data recording mechanism in step S3 is triggered; If the second matching result is a successful match, then the asynchronous data recording mechanism in step S3 is blocked.

[0015] Preferably, the shared storage module includes a unified behavior chain shared data table; the first type of action type data record only activates the input feature original text and the current interaction stage label; the second type of action type data record only activates the process state feature value and the qualitative evaluation label; the third type of action type data record only activates the final state result feature value and the iteration version label; the first type of action type data... RecordThe second type of action type data record and the third type of action type data record are dynamically switched through the action type field under the unified structure of the behavior chain shared data table.

[0016] The interactive guidance and process feature evaluation system based on data semantic triage includes a user terminal, a front-end task routing module configured to execute steps S1 to S5, a shared storage module, and a back-end judgment module configured to execute steps S6 to S9. The output of the user terminal is bidirectionally connected to the front-end task routing module, the output of the front-end task routing module is unidirectionally connected to the input of the shared storage module, the shared storage module is bidirectionally connected to the back-end judgment module, and the output of the back-end judgment module is unidirectionally connected to the input of the user terminal. The user terminal is used for user interaction with the system; The front-end task routing module is used to collect user interaction data while guiding the user to complete the task; The shared storage module is used for heterogeneous data storage, behavior chain association, and data service provision. The backend determination module is used to receive the final solution text submitted by the user. It is configured with a conditional retrieval engine based on the behavior chain shared data table. Through the user identifier and action type field, it extracts the user's historical process state feature values ​​from the unified underlying data table.

[0017] Preferably, it also includes an interactive web page module, the output of which is connected to the input of the shared storage module, and the interactive web page module is bidirectionally connected to the user terminal.

[0018] Preferably, after the backend judgment module outputs the final result feature value in each round, the system keeps the multi-round iterative tracking channel open and does not forcibly terminate the interaction; the system triggers dynamic branch routing according to the received autonomous settlement instruction or the externally configured settlement conditions to execute the advanced refinement path or output the final qualitative evaluation report.

[0019] Preferably, the front-end task routing module and the back-end judgment module are built on a logical master board. The system derives functional sub-entities suitable for different task scenarios by injecting task parameters into the logical master board and mounting a local knowledge base.

[0020] The present invention has the following advantages and beneficial effects compared with the prior art: The system architecture of this invention adopts a decoupled design, packaging the underlying operating logic into a basic template. For cross-task or cross-disciplinary application scenarios, new subsystem matrices can be derived at low cost simply by injecting new task parameters through data interfaces, mounting local knowledge base documents of the corresponding discipline, and replacing the whitelist search sources.

[0021] This system overcomes the limitations of traditional one-dimensional result evaluation by constructing a bivariate judgment matrix (four-quadrant rule) by extracting process data and final state data. The system can accurately filter out machine-generated content that lacks process features but has excellent result features from the underlying data feature level, providing a reliable anti-cheating technology barrier for complex online training tasks.

[0022] 3. This system decouples system logic code from specific subject content through master template design and dynamic parameter injection. When reused across tasks, there is no need to rewrite the underlying workflow, significantly reducing the cost and time required for secondary development. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the interactive guidance and process feature evaluation method based on data semantic diversion of the present invention; Figure 2 This is a schematic diagram of the interactive guidance and process feature evaluation system based on data semantic diversion of the present invention (excluding interactive web page modules); Figure 3 This is a schematic diagram of the interactive guidance and process feature evaluation system based on data semantic diversion (including an interactive web page module) of the present invention. Detailed Implementation

[0024] The invention's objective will be further described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described in detail here, but the implementation of the invention is not limited to the following embodiments.

[0025] Example 1 like Figure 1 As shown, the interactive guidance and process feature evaluation method based on data semantic decomposition includes the following steps: S1. The front-end task routing module receives the input feature stream from the user terminal; S2. Based on preset semantic filtering rules, the input feature stream is analyzed and identified in real time. When the input feature stream matches exploratory interactive features containing interrogative words, expressions of confusion, or requests for detailed evaluation, the asynchronous data recording mechanism in step S3 is triggered. When the input feature stream matches structured solution text features that conform to a preset format, exceed a threshold in length, or contain key solution elements, the asynchronous data recording mechanism in step S3 is blocked. Step S2 includes the following steps: S201. Perform real-time analysis on the received input feature stream to identify its syntactic structure and semantic intent; S202. The identified semantic intent is matched with the preset exploratory interaction features to obtain the first matching result, and the identified semantic intent is matched with the structured solution text features to obtain the second matching result; wherein, the exploratory interaction features include, but are not limited to, statements containing interrogative words, expressing confusion or requesting detailed evaluation; the structured solution text features include, but are not limited to, text that conforms to the preset format, exceeds the threshold in length, or contains key solution elements; S203. Perform traffic splitting based on the first and second matching results: If the first matching result is a successful match, then the asynchronous data recording mechanism in step S3 is triggered; If the second matching result is a successful match, then the asynchronous data recording mechanism in step S3 is blocked.

[0026] S3. Based on the asynchronous silent recording mechanism, the data persistence interface is called to write the first type of action data record to the shared storage module. The first type of action data record contains the original text of the input characteristics and the label of the current interaction stage. The writing of the first type of action data record in step S3 precedes the system-generated guidance response or dialogue output in terms of timing, and no front-end output content of the system contains any prompts about executed records or scores. This step uses an asynchronous silent data writing method in conjunction with filtering rules to avoid the observability effect caused by explicit form records, ensuring that user interaction behavior data remains true and objective.

[0027] S4. The front-end task routing module generates a guided response and returns it to the user terminal based on the strategy configuration of the current interaction node and the local knowledge base. The strategy configuration refers to a set of parameters preset for different interaction stages, including but not limited to guided response styles (such as encouraging or challenging), knowledge base query weights, and state machine switching conditions. The strategy configuration is stored locally on the system as a structured configuration file (such as JSON or XML format). The front-end task routing module loads this configuration at startup and dynamically calls the corresponding strategy parameters based on the current interaction stage tag. Administrators can modify this file through the configuration interface provided by the system without modifying the underlying code. The local knowledge base refers to a pre-imported collection of structured knowledge documents related to the current task domain, including but not limited to textbook chapters, standard operating procedures, and case studies. The knowledge base is stored using vectorized indexes or keyword indexes. When generating a guided response, the front-end task routing module retrieves the knowledge fragments most semantically relevant to the user input through the knowledge base mapping interface, using this as the basis for the response content. The knowledge base supports dynamic mounting and unmounting by task type to achieve cross-domain reuse.

[0028] S5. When the current interaction node is determined to meet the process stage settlement conditions, based on the asynchronous silent recording mechanism, a second type of action type data record (evaluation) is written to the shared storage module through the data persistence interface. The second type of action type data record contains quantitatively generated process state feature values ​​and qualitative evaluation labels. The process stage settlement conditions refer to the conditions determined by the system according to preset rules (such as interaction rounds, task node completion status, user intent changes, etc.) for the end of the current continuous interaction process. When this condition is met, the system performs a quantitative evaluation of the user's interaction performance in the previous stage. The writing of the second type of action type data record in step S5 precedes the system-generated guidance response or dialogue output in terms of timing, and no front-end output content of the system contains any prompt information about executed records or scores. This step uses an asynchronous silent data writing method in conjunction with filtering rules to avoid the observability effect caused by explicit form records, ensuring that user interaction behavior data remains true and objective. For example, in this task, the system can be configured to automatically trigger the settlement of the current interaction stage and generate the process status characteristic value of that stage when it detects that the user has made three consecutive interactions regarding "agenda time allocation" or "issue sorting logic," or when the user's input contains keywords such as "Is this arrangement reasonable?" or "Please provide feedback on the current draft," which contain the intention to confirm the stage. Step S5 specifically includes the following steps: S501. Settlement conditions for process stages include, but are not limited to: completing a preset sub-task, reaching a threshold number of interaction rounds, or the user input containing a stage end keyword. S502. When the settlement conditions of the process stage are met, the user's interactive performance in that stage is quantitatively evaluated, and a process state characteristic value is generated. S503. Write the process status characteristic value of step S502 into the shared storage module as a second type of action type data record.

[0029] S6. After the user submits the final solution text, the backend judgment module uses the rule parsing engine to parse the final solution text and generate the final state result feature value. At the same time, based on the user identifier (ID), the backend module retrieves the process state feature value corresponding to the user from the shared storage module, and counts the number of historical data records of the user's existing action type as the third type of action. Based on the number of historical data records, the current submission iteration version tag (submission_version) is automatically generated.

[0030] S7. Substitute the process state feature value of step S5 and the final state result feature value of step S6 into the preset bivariate threshold matrix for logical calculation and cross-judgment to obtain the judgment result.

[0031] The bivariate threshold matrix in step S7 includes four-quadrant judgment rules, and the first and second thresholds can be dynamically adjusted and configured by administrators (i.e., teachers) according to specific task attributes without changing the bivariate feature coupling logic of the four-quadrant judgment rules themselves. The bivariate threshold matrix is ​​an algorithmic model (PODA model, Process-Outcome Dual-variable Analysis) with specific executable logic built into the backend judgment module; dynamic threshold adjustment is accomplished through the management configuration interface / interface provided by the system; the early warning action is executed by the algorithmic model automatically calling the system's early warning message push interface.

[0032] The logical calculation and cross-determination process of step S7 is as follows: when the process state characteristic value of step S5 is lower than the first threshold and the final state result characteristic value of step S6 is higher than or equal to the second threshold, it is identified as an external generation flow anomaly and an early warning action is triggered.

[0033] S8. Write a third type of action data record (qc_review) to the shared storage module. The third type of action data record contains the final state result feature value and the iteration version label (submission_version). In the data field structure, force the corresponding field used for the stored procedure state feature value to be empty.

[0034] S9. Based on the judgment result of step S7, return the corresponding final qualitative evaluation report to the user terminal.

[0035] like Figure 2 As shown, the interactive guidance and process feature evaluation system based on data semantic triage includes a user terminal, a front-end task routing module, a shared storage module, and a back-end judgment module. The user terminal continuously sends input feature streams (such as questions or drafts) to the front-end task routing module; simultaneously, the user terminal receives guidance responses generated by the front-end task routing module. The front-end task routing module, through its internal data persistence interface, unidirectionally and asynchronously writes first-type action type data records (records) and second-type action type data records (evaluations) to the behavior chain shared data table (learning_log) of the shared storage module. The back-end judgment module, based on the user identifier, retrieves process state feature values ​​from the shared storage module and writes the final state result feature value and iteration version label as a third-type action type data record (qc_review) back to the shared storage module. After completing the review and judgment, the back-end judgment module unidirectionally returns the judgment results (judgment conclusions, feedback suggestions, warnings, etc.) to the user terminal.

[0036] The user terminal is used for user interaction with the system. The front-end task routing module includes an intent recognition hub, an external information source capture interface, a knowledge base mapping interface, a state machine simulation interface, and a data persistence interface.

[0037] The front-end task routing module executes steps S1-S5. While guiding users to complete tasks, it seamlessly and efficiently collects user interaction data. Based on a state machine, it sets pre-defined role parameters and automatically switches the system's dialogue strategy weights when specific interaction conditions are triggered, such as dynamically switching from knowledge tutoring mode to stress testing mode. Internally, the front-end task routing module is configured with a central hub and four functional extension interfaces: an external information source capture interface, a knowledge base mapping interface, a state machine simulation interface, and a data persistence interface. The central control logic, based on a pre-defined task instruction set, deeply analyzes the technical requirements and judgment boundaries of the current teaching task. The external information source capture interface retrieves industry news in real time as interactive material through a configured whitelist of authoritative media. The local knowledge base mapping interface loads local teaching materials to ensure the system response strictly conforms to professional knowledge standards. The state machine simulation interface handles stress testing logic; when a specific task node is triggered, the system automatically adjusts the response weights. The data persistence interface performs asynchronous silent recording. The front-end task routing module features an asynchronous silent recording mechanism based on filtering rules, an asynchronous data recording mechanism, and a dynamic switching of interaction strategies based on a state machine.

[0038] 1. Features an asynchronous silent recording mechanism based on filtering rules: In addition to the main interaction thread, the system is configured with a set of background monitoring rules based on semantic filtering. When the user's input feature stream is determined to be asking a question, expressing confusion, or requesting evaluation details, the system will, before generating a response, first call the data persistence interface to asynchronously write the raw text of the input feature (user_raw_input) and the current interaction phase label (current_phase) to the shared storage module, and the writing action will not generate any feedback prompts on the front-end interface.

[0039] 2. Asynchronous data recording mechanism: If the input feature stream is a structured scheme text feature (determined to be a direct submission scheme), the system will automatically block write requests for the first type of action data record, thereby ensuring that the shared storage module retains only the real user cognitive evolution trajectory and eliminates redundant and interfering data.

[0040] 3. Dynamic switching of interaction strategies based on state machines: Based on the progress of the interaction, this system uses a three-stage state machine: (1) Guiding phase: The system adopts a parameter configuration of low-intensity thorn picking and high-frequency support guidance, focusing on the analysis of knowledge points and task decomposition.

[0041] (2) Real-world simulation phase: Once the system determines that the user has started to attempt to build a solution, the system will trigger a state switching command and dynamically replace the system prompt word template. At this time, the system will increase the weight of logical conflict detection and switch to a result-oriented interaction strategy, verifying the rigor of the solution through pressure-based questioning.

[0042] (3) Settlement stage: After the task is completed, the system performs the last asynchronous data record writing, writes the process state feature value (score, 1-5 points) based on the dialogue quality quantification into the shared storage module, and completes the intermediate state storage of this node.

[0043] The backend judgment module executes steps S6-S9. It receives the final solution text submitted by the user and is equipped with a conditional retrieval engine based on a shared data table of behavioral chains. Using the user identifier (ID) and action type field (action_type), it extracts the user's historical process state feature values ​​from a unified underlying data table. This breaks away from single-dimensional text result evaluation, performing bidirectional joint calculations by extracting historical process state feature values ​​and final result feature values ​​through targeted retrieval. After outputting the final result feature value in each round, the system keeps the multi-round iterative tracking channel open and does not forcibly terminate the interaction. The system triggers dynamic branch routing based on received autonomous settlement instructions or externally configured settlement conditions to execute advanced refinement paths or output the final qualitative evaluation report.

[0044] The execution of this backend decision module includes the following consecutive workflow nodes: 1. Text parsing and benchmark comparison node The system receives the final solution text submitted by the user, calls the rule parsing engine, extracts key parameters (such as time nodes, budget values, etc.) from the solution, and performs a strong logical comparison with the preset standard benchmark library in the system's global variables to generate preliminary final result feature values ​​(0-100 points).

[0045] 2. Single-table targeted feature retrieval node At this node, the system triggers a database read instruction. Instead of using a multi-table join query, the system directly performs a single-table retrieval on the behavior chain shared data table (learning_log) through the user identifier. At the same time, using the semantic shunt rule, the system accurately extracts the process status feature values (1-5 scale) in the record rows with action_type = "evaluation". In addition, the system automatically calculates and generates an iterative version label (submission_version) by counting the number of historical records with action_type = "qc_review".

[0046] 3. Abnormal Feature Recognition Algorithm Based on Bivariate Threshold Matrix The system substitutes both the pre-acquired process status feature values and the final state result feature values into a preset bivariate four-quadrant threshold matrix for cross-calculation. The process status feature values can be obtained by weighted calculation through dimensions such as the interaction rounds in this stage, the depth of user questions, and the number of active retrievals. The final state result feature values can be obtained by comparing and scoring the final solution with the preset standard answer in dimensions such as integrity, accuracy, and innovation. The first threshold T1 and the second threshold T2 can be preset by the administrator according to historical data statistics or task difficulty. Let the threshold of the process status feature be T1 and the threshold of the final state result feature value be T2 (this threshold supports dynamic configuration according to task attributes), and the system executes the following logical branches: (1) Branch 1 (the first quadrant, normal and excellent): If the process status feature ≥ T1 and the final state result feature value ≥ T2, it is determined as a normally excellent output with logical consistency, and a pass instruction is generated.

[0047] (2) Branch 2 (the second quadrant, abnormal feature recognition): If the process status feature < T1 and the final state result feature value ≥ T2, this is the core defense mechanism of the present invention. This data shows abnormal features that violate the conventional cognitive evolution law of lacking an interaction process but having an extremely excellent final state result. The system automatically captures this abnormality, marks it as a suspected non-artificially generated external abnormal input stream (i.e., intercepting external large model ghostwriting), and triggers an audit warning channel.

[0048] (3) Branch 3 (the third quadrant, high interaction and low efficiency): If the process status feature ≥ T1 and the final state result feature value < T2, it is determined that the interaction is sufficient but the output quality does not meet the standard, and a targeted optimization suggestion is triggered.

[0049] (4) Branch 4 (the fourth quadrant, both low): If the process status feature < T1 and the final state result feature value < T2, it is determined as an invalid output, and a rejection instruction is triggered.

[0050] 4. Heterogeneous Data Silent Writing Node After the judgment is completed, the system calls the persistence interface again to write the calculation results of this acceptance back to the same behavior chain shared data table. To prevent data pollution, the system strictly enforces field isolation writes: it forcibly assigns the action_type="qc_review" to the record, only writes the final result feature value (result_score) and the iteration version label (submission_version), and forcibly sets the corresponding field (score) used for stored procedure status features to NULL. This ensures the absolute independence of the 1-5 point process scale and the 0-100 point result scale in the single-table structure from the bottom layer.

[0051] 5. Dynamic feedback routing nodes The system triggers dynamic feedback routing based on the current judgment results and interaction status. By default, the system outputs a structured diagnostic report and rectification instructions for the current version immediately after each round of interaction, and keeps the multi-round iteration channel open without forcibly terminating the interaction, in order to support users in continuously refining their solutions. The system only executes the branch path, calls the historical behavior chain data in the shared storage module, and renders the final qualitative evaluation report or multi-dimensional visualization panel to the user interface when and only when it receives a settlement instruction from the user terminal (such as the user manually confirming the end of optimization) or meets the externally configured stage settlement conditions.

[0052] The backend decision module in this embodiment has the function of multi-round iterative tracking and threshold-configurable decision-making power transfer, as detailed below: 1. Multi-round Iteration Tracking Logic: The backend judgment module supports students submitting the same task multiple times. The system automatically generates an iteration version label (submission_version) by calculating the historical frequency of the action type field (action_type) = "qc_review". The backend judgment module arranges the feature values ​​(result_score) of the final state results of the historical multiple versions in chronological order to generate the user's optimization iteration trajectory, which serves as one of the reference weights for the final evaluation.

[0053] 2. Real-time evaluation and differentiated routing: After each round of submission, the system will immediately generate the final result feature value (result_score) and rectification instructions for that version.

[0054] 3. Autonomous settlement control: (1) Student decision-making power: The system does not force the end of optimization. Students can decide whether to continue to improve the quality of the solution through iteration based on their current score.

[0055] (2) Teacher allocation rights: Teaching administrators can flexibly configure settlement thresholds according to students' learning situation. For example, the option to activate the final evaluation will be set after a score of 90 is reached.

[0056] (3) Qualitative evaluation trigger: Only when the preset threshold is met or the student actively chooses to finish perfectly will the system call historical data from the shared storage module based on the four-quadrant judgment results of the PODA (Process-Outcome Dual-variable Analysis) model to generate the final qualitative evaluation report (including a workplace competency map with symbolic scales).

[0057] The heterogeneous data records in the shared storage module include user identifiers, task types, action types, process state feature values, final result feature values, original input feature text, current interaction stage labels, and iteration version labels (submission_version). These records are used for storing heterogeneous data records, associating behavior chains, and providing data services. The shared storage module contains a behavior chain shared data table (learning_log). The front-end task routing module and the back-end decision module achieve semantic data separation and sharing by assigning different values ​​to and recognizing the action type field in this data table. The system achieves semantic data separation at the underlying level by configuring action types. When recording the original input feature text, it is written to the first type of action type data record (record); when recording process state feature values, it is written to the second type of action type data record (evaluation); and when the back-end decision module outputs its decision, it is written to the third type of action type data record (qc_review). The shared storage module contains a unified behavior chain shared data table (learning_log); the first type of action type data record (record) only activates the input feature original text and the current interaction stage label; the second type of action type data record (evaluation) only activates the process state feature value and qualitative evaluation label; the third type of action type data record (qc_review) only activates the final state result feature value and iteration version label; the first type of action type data, the second type of action type data record, and the third type of action type data record are all dynamically switched through the action type field under the unified structure of the behavior chain shared data table.

[0058] The front-end task routing module and the back-end decision-making module adopt a modular architecture that decouples logic from content. The system configures a unified logic template and dynamically injects task variables, mounts corresponding professional knowledge base documents, and configures search whitelist sources through data interfaces when facing different tasks, so as to quickly derive a matrix of functional sub-modules with the same data diversion and decision-making logic.

[0059] To enable large-scale deployment and rapid migration of the system across different task nodes or multidisciplinary application scenarios, this system adopts a modular, scalable architecture that decouples logic and content.

[0060] 1. General Logic Master Board: The logic master board encapsulates the core underlying control code and data processing logic of the system, specifically including: unified central control logic and multi-function interfaces, a two-layer decoupled architecture between the front end and back end, an asynchronous silent recording mechanism based on filtering rules, and a two-variable feature coupling judgment algorithm (PODA model). As a standardized technical foundation, the logic master board is not strongly bound to any specific business knowledge of any particular discipline.

[0061] 2. Dynamic parameter injection and child body derivation By injecting specific task variables into the logic master board through standardized interfaces, mounting corresponding local professional knowledge base documents, and configuring specific retrieval whitelist sources, a corresponding number of functional sub-intelligent agents (sub-agents) can be dynamically derived according to the needs of different application scenarios without modifying the underlying running logic and flow control code.

[0062] For example, in a specific training course implementation of this system (taking the "Meeting Planning and Organization" course as an example), the system has successfully and quickly derived more than 40 functional sub-entities by injecting specific task variables such as meeting agenda design and meeting budget preparation, covering the complete task chain, which significantly improves the efficiency of secondary development and cross-scenario migration of the system.

[0063] This system has the following beneficial effects: 1. This invention achieves seamless capture of the user's true cognitive evolution trajectory through a combination of preset semantic filtering rules and an asynchronous silent recording mechanism. The system can automatically filter out redundant and interfering data, such as large sections of structured text, without interfering with the main front-end interaction thread. This fundamentally avoids the observer effect caused by explicit scoring, significantly improving the authenticity and objectivity of process evaluation data.

[0064] 2. This invention breaks through the limitations of traditional systems that rely solely on final deliverables for single-dimensional evaluation, and innovatively proposes a feature coupling algorithm for bivariate processes and results. The system can accurately capture abnormal data pairs where "process state feature values ​​are missing, but final result feature values ​​are extremely high," thereby achieving automated identification of deceptive behaviors such as direct copying of external large models and falsification of schemes at the system level.

[0065] 3. The system of this invention flexibly decouples the evaluation termination right from the quality control threshold, introducing configurable autonomous settlement control. While providing real-time, multi-round version iteration tracking, the system empowers users to decide whether to stop optimization and reserves an interface for configuring learning thresholds for administrators. This human-machine collaborative control method significantly enhances the system's application flexibility and practical value in complex training guidance scenarios.

[0066] 4. This invention achieves deep decoupling between the logic master board and specific task content, eliminating the need to reconstruct the underlying workflow and prompt word framework when reusing the system. Low-cost derivation of the subsystem matrix can be achieved simply through parameter injection and knowledge base mounting (stable operation of 40 functional sub-sub-body in a single course has been empirically verified), giving the system extremely high industrial migration capability and large-scale application prospects.

[0067] Example 2 Except for the following technical features, the other technical features of this embodiment are the same as those of Embodiment 1: like Figure 3 As shown, this embodiment also includes an interactive webpage module. The output of the interactive webpage module is connected to the input of the shared storage module, and the interactive webpage module is bidirectionally connected to the user terminal. The interactive webpage module's function is to strengthen the control and data collection of the user process from another interactive dimension through two mechanisms: pre-assessment unlocking and implicit heterogeneous data feedback. This complements the core conversational system, jointly constructing a more comprehensive and rigorous automated interactive guidance and evaluation environment.

[0068] The interactive web page module executes autonomous settlement logic: after outputting the final result feature value (result_score) in each round, the system keeps the multi-round iterative tracking channel open and does not forcibly terminate the interaction; the system triggers dynamic branch routing according to the received autonomous settlement instruction or the externally configured settlement conditions to execute the advanced refinement path or output the final qualitative evaluation report.

[0069] The interactive webpage module in this embodiment has a knowledge access control unlocking mechanism and an implicit write-back of the interaction results.

[0070] 1. Knowledge Access Control Unlocking Mechanism: The tool modules on the webpage (such as the intelligent agenda planner) are locked by default. Students must first complete the corresponding knowledge point assessment. Only after achieving the required accuracy will the system trigger a CSS state switch to unlock the subsequent operation tools.

[0071] 2. Implicit backhaul of heterogeneous data: The interactive webpage module has a reserved data write-back interface. The results of students' interactions, such as logical sorting and crisis decision-making, completed within the webpage can be automatically written to the behavior chain shared data table (learning_log) through implicit data channels, forming a complete learning behavior chain together with the dialogue records.

[0072] The above-described specific embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any other changes or equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.

Claims

1. A method for interactive guidance and process feature evaluation based on data semantic decomposition, characterized in that: Includes the following steps: S1. The front-end task routing module receives the input feature stream from the user terminal; S2. Based on the preset semantic filtering rules, the input feature stream is analyzed and identified in real time. When the input feature stream matches the exploratory interaction features, step S3 is triggered. When the input feature stream matches the structured scheme text features, step S3 is blocked. S3. Based on the asynchronous silent recording mechanism, call the data persistence interface to write the first type of action data record into the shared storage module. The first type of action data record includes the original text of the input feature and the label of the current interaction stage. S4. The front-end task routing module generates a guidance response and returns it to the user terminal based on the strategy configuration of the current interaction node and the local knowledge base. S5. When it is determined that the current interaction node meets the settlement conditions of the process stage, based on the asynchronous silent recording mechanism, the second type of action data record is written to the shared storage module through the data persistence interface. The second type of action data record contains the quantitatively generated process state feature value and qualitative evaluation label. S6. After the user submits the final solution text, the backend judgment module uses the rule parsing engine to parse the final solution text and generate the final state result feature value. At the same time, based on the user identifier, it retrieves the process state feature value corresponding to the user from the shared storage module and counts the number of historical data records of the user's existing action type as the third type of action. Based on the number of historical data records, it automatically generates the tag of the currently submitted iteration version. S7. Substitute the process state feature value of step S5 and the final state result feature value of step S6 into the preset bivariate threshold matrix for logical calculation and cross-judgment to obtain the judgment result. S8. Write a third type of action data record into the shared storage module. The third type of action data record contains the final state result feature value and the iteration version label. In the data field structure, force the corresponding field used for the stored procedure state feature value to be empty. S9. Based on the judgment result of step S7, return the corresponding final qualitative evaluation report to the user terminal.

2. The interactive guidance and process feature evaluation method based on data semantic triage according to claim 1, characterized in that: The bivariate threshold matrix in step S7 includes four-quadrant decision rules, and the first and second thresholds can be dynamically adjusted and configured by administrators according to specific task attributes without changing the bivariate feature coupling logic of the four-quadrant decision rules themselves.

3. The interactive guidance and process feature evaluation method based on data semantic splitting according to claim 1, characterized in that: The logical calculation and cross-determination process of step S7 is as follows: when the process state characteristic value of step S5 is lower than the first threshold and the final state result characteristic value of step S6 is higher than or equal to the second threshold, it is identified as an external generation flow anomaly and an early warning action is triggered.

4. The interactive guidance and process feature evaluation method based on data semantic triage according to claim 1, characterized in that: The asynchronous silent recording mechanism of step S3 or step S5 is as follows: the data recording writing action precedes the system-generated guidance response or dialogue output in terms of timing, and no front-end output content of the system contains any prompt information about the executed record or the score.

5. The interactive guidance and process feature evaluation method based on data semantic triage according to claim 1, characterized in that: Step S5 specifically includes the following steps: S501. Settlement conditions for process stages include, but are not limited to: completing a preset sub-task, reaching a threshold number of interaction rounds, or the user input containing a stage end keyword. S502. When the settlement conditions of the process stage are met, the user's interactive performance in that stage is quantitatively evaluated, and a process state characteristic value is generated. S503. Write the process status characteristic value of step S502 into the shared storage module as a second type of action type data record.

6. The interactive guidance and process feature evaluation method based on data semantic triage according to claim 1, characterized in that: Step S2 includes the following steps: S201. Perform real-time analysis on the received input feature stream to identify semantic intent; S202. Match the identified semantic intent with the preset exploratory interaction features to obtain the first matching result, and match the identified semantic intent with the preset structured scheme text features to obtain the second matching result. S203. Perform traffic splitting based on the first and second matching results: If the first matching result is a successful match, then the asynchronous data recording mechanism in step S3 is triggered; If the second matching result is a successful match, then the asynchronous data recording mechanism in step S3 is blocked.

7. The interactive guidance and process feature evaluation method based on data semantic triage according to claim 1, characterized in that: The shared storage module contains a unified behavior chain shared data table; the first type of action type data record only activates the input feature original text and the current interaction stage label; the second type of action type data record only activates the process state feature value and qualitative evaluation label; the third type of action type data record only activates the final state result feature value and iterative version label; the first type of action type data... Record The second type of action type data record and the third type of action type data record are dynamically switched through the action type field under the unified structure of the behavior chain shared data table.

8. An interactive guidance and process feature evaluation system based on data semantic decomposition, characterized in that: The system includes a user terminal, a front-end task routing module configured to perform steps S1-S5 of claim 1, a shared storage module, and a back-end determination module configured to perform steps S6-S9 of claim 1. The output of the user terminal is bidirectionally connected to the front-end task routing module, the output of the front-end task routing module is unidirectionally connected to the input of the shared storage module, the shared storage module is bidirectionally connected to the back-end determination module, and the output of the back-end determination module is unidirectionally connected to the input of the user terminal. The user terminal is used for user interaction with the system; The front-end task routing module is used to collect user interaction data while guiding the user to complete the task; The shared storage module is used for heterogeneous data storage, behavior chain association, and data service provision. The backend determination module is used to receive the final solution text submitted by the user. It is configured with a conditional retrieval engine based on the behavior chain shared data table. Through the user identifier and action type field, it extracts the user's historical process state feature values ​​from the unified underlying data table.

9. The interactive guidance and process feature evaluation system based on data semantic decomposition according to claim 8, characterized in that: It also includes an interactive web page module, the output of which is connected to the input of the shared storage module, and the interactive web page module is bidirectionally connected to the user terminal.

10. The interactive guidance and process feature evaluation system based on data semantic decomposition according to claim 8, characterized in that: After the backend judgment module outputs the final result feature value in each round, the system keeps the multi-round iterative tracking channel open and does not forcibly terminate the interaction; the system triggers dynamic branch routing according to the received autonomous settlement instruction or the externally configured settlement conditions to execute the advanced refinement path or output the final qualitative evaluation report.

11. The interactive guidance and process feature evaluation system based on data semantic decomposition according to claim 8, characterized in that: The front-end task routing module and the back-end judgment module are built on a logical master board. The system derives functional sub-entities suitable for different task scenarios by injecting task parameters and mounting a local knowledge base into the logical master board.

Images