Library reference consultation question and answer method based on generative model for education scenario

Abstract

This application discloses a generative model-based question-answering method for library reference services in educational settings. The method includes: acquiring and parsing reference request text to obtain structured parsing results containing clues such as task type and integrity intent; constructing a learner profile by combining a set of course policy parameters, calculating a comprehensive risk score, and determining the risk level; subsequently determining the highest permissible output level based on the risk level, course policy, and task completion rate; constructing a constraint-based inference input and generating candidate responses; finally, obtaining a score through a submitability discrimination model; if the score meets the standard, rewriting the candidate responses to generate compliant responses. This method achieves a balance between compliance and pedagogical relevance, effectively mitigating the risk of academic misconduct and improving the quality of reference services.

Description

Technical Field

[0001] This application relates to the fields of educational information services and natural language processing technology, and more specifically, to a generative model-based library reference consultation question-and-answer method for educational scenarios. Background Technology

[0002] In educational settings such as university and primary / secondary school libraries and subject-based service platforms, online reference and consultation services have become an important means of supporting learners' self-directed learning. Their core requirement is to provide academic support while mitigating the risk of academic misconduct. Currently, with the widespread application of generative models in natural language processing, related technologies are gradually being introduced into reference and consultation services to improve response efficiency and content richness.

[0003] However, existing technologies have significant shortcomings in practical applications in education. On the one hand, the boundary between reference consultation and academic misconduct is blurred. Users may gradually approach a draft assignment that can be directly submitted through multiple rounds of dialogue, while existing models often adopt simple rejection or unconstrained free generation methods, either outputting a reduced-duplication text or a complete draft that can be directly submitted, resulting in insufficient compliance; or excessive rejection to avoid violations, failing to provide effective heuristic guidance and limiting its teaching support role. On the other hand, existing technologies lack effective verification of the substitutability of output text, cannot automatically determine whether the text meets the submission requirements and downgrade it, and are difficult to dynamically adjust the output granularity according to the policy requirements of different courses and the actual status of learners. The process has poor consistency and low reproducibility, seriously affecting the quality of reference consultation services and the implementation of academic integrity norms. Summary of the Invention

[0004] The application provides a generative model-based library reference consultation question-and-answer method for educational scenarios, which aims to at least solve the technical problems existing in the related technologies described above.

[0005] According to a first aspect of the embodiments of this application, a generative model-based library reference consultation question-and-answer method for educational scenarios is provided, including:

[0006] Obtain the reference consultation request text, parse the request text to obtain a structured parsing result, wherein the structured parsing result includes task type, deliverable type, completion status clues, and integrity intent clues; the integrity intent clues are obtained by keyword matching or template matching of the request text;

[0007] Based on the structured analysis results and the preset set of course policy parameters, a learner profile is constructed;

[0008] Based on the learner profile and the course policy parameter set, a comprehensive risk score is calculated, and the risk level is determined according to the comprehensive risk score and the preset risk threshold.

[0009] The highest allowed output level is determined based on the risk level, the course policy parameter set, and the task completion rate in the learner profile.

[0010] The inference input is provided to the generative model to generate candidate responses, wherein the inference input includes constraints extracted from the structured parsing results, the learner profile, and the risk level;

[0011] The candidate responses are processed using a submitability discrimination model to obtain submitability scores;

[0012] If the submitability score is greater than or equal to the preset submitability threshold, then the candidate response is downgraded and rewritten to generate a compliant response and output it.

[0013] As an optional approach, the learner profile includes an integrity risk propensity score and a task completion rate; the integrity risk propensity score is calculated based on integrity intention cues in the structured parsing results; and the task completion rate is calculated based on completion rate cues in the structured parsing results.

[0014] As an optional approach, the comprehensive risk score is obtained by weighted summation of the integrity risk tendency score, the task completion score, and the approximation index; the approximation index is calculated based on the number and magnitude of user request type transitions in multiple rounds of reference consultation sessions.

[0015] As an optional approach, the step of determining the maximum allowed output level includes: determining an initial upper limit based on the preset maximum output level in the course policy parameter set; constraining the initial upper limit based on the risk level; wherein, if the risk level is high risk, the maximum allowed output level is not higher than the retrieval strategy level.

[0016] As an alternative, the generative model inference input includes a set of constraint fields, which at least includes the highest allowed output level and a list of prohibited output items derived from the highest allowed output level.

[0017] As an optional approach, the step of processing the candidate responses using a submitability discrimination model to obtain a submitability score specifically includes: extracting text features from the candidate responses, including coherence indicators, argument completeness indicators, and relevance to the request text indicators; fusing the text features with the risk level and the maximum allowed output level, and inputting them into a classifier to obtain the submitability score.

[0018] As an optional approach, the downgrade rewriting operation includes: locating highly submitable segments in the candidate responses based on the triggering cause vector; rewriting the complete argument sentences in the highly submitable segments into paragraph function descriptions and placeholders, or retaining the structural framework of the segments and removing the complete expressions therein.

[0019] As an optional approach, the downgrade rewrite operation further includes: after performing the rewrite, temporarily lowering the allowed highest output level and outputting a compliant response containing the required user input field template.

[0020] As an optional approach, the method further includes the step of generating a learning step chain: before generating the candidate response, a learning step chain containing multiple step nodes is generated based on the task type in the structured parsing result and the allowed highest output level, each step node containing the required input fields and corresponding validation rules; and the current node information of the learning step chain is written into the generative model inference input.

[0021] According to a second aspect of the embodiments of this application, a generative model-based library reference consultation question-and-answer system for educational scenarios is also provided, comprising:

[0022] The request parsing module is used to obtain reference consultation request text, parse the request text, and obtain structured parsing results. The structured parsing results include task type, deliverable type, completion clues, and integrity intent clues. The integrity intent clues are obtained by keyword matching or template matching of the request text.

[0023] The profile building module is used to build learner profiles based on the structured parsing results and a preset set of course policy parameters.

[0024] The risk assessment module is used to calculate a comprehensive risk score based on the learner profile and the course policy parameter set, and to determine the risk level based on the comprehensive risk score and a preset risk threshold.

[0025] The hierarchical control module is used to determine the highest allowed output level based on the risk level, the course policy parameter set, and the task completion rate in the learner profile.

[0026] The candidate generation module is used to provide inference input to the generative model and generate candidate responses. The inference input includes constraints extracted from the structured parsing results, the learner profile, and the risk level.

[0027] The submitability discrimination module is used to process the candidate responses through a submitability discrimination model to obtain a submitability score;

[0028] The output rewriting module is used to perform a downgrade rewriting operation on the candidate response if the submitability score is greater than or equal to a preset submitability threshold, generate a compliant response, and output it.

[0029] According to a third aspect of the embodiments of this application, an electronic device is provided, including: a processor; a memory for storing a computer program executable by the processor; wherein the processor is configured to execute the computer program in the memory to implement the method described in the first aspect.

[0030] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, which, when an executable computer program in the storage medium is executed by a processor, enables the implementation of the method described in the first aspect.

[0031] This application achieves precise characterization of users' abilities and integrity tendencies by constructing learner profiles, quantitatively assesses academic misconduct risks by combining comprehensive risk scores, and dynamically adjusts the output level. Utilizing a constrained generative model and a submitability judgment mechanism, it effectively avoids outputting ready-to-submit drafts or reduced-duplication text, ensuring service compliance. Simultaneously, it provides step-by-step heuristic guidance through a learning process chain, coupled with the downgrading and rewriting of highly submitable text. This not only solves the problem of insufficient teaching support caused by excessive rejection of answers in existing technologies but also achieves control over output content and standardization of processes. It ensures that the reference and consultation service not only complies with academic integrity norms but also adapts to different course policies and learner statuses, significantly improving the professionalism, relevance, and reliability of reference and consultation services in educational scenarios.

[0032] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. Furthermore, no embodiment in this application needs to achieve all the effects described above. Attached Figure Description

[0033] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0034] Figure 1 This is a schematic diagram of a generative model-based library reference consultation question-and-answer method for educational scenarios, provided as an embodiment of this application.

[0035] Figure 2 This is a schematic diagram illustrating the process for determining the highest permissible output level provided in an embodiment of this application.

[0036] Figure 3 This is a schematic diagram of the structure of the generative model provided in the embodiments of this application.

[0037] Figure 4 This is a schematic diagram of the downgrade rewriting operation process provided in the embodiments of this application.

[0038] Figure 5 This is a schematic diagram of the structure of a generative model-based library reference consultation question-and-answer system for educational scenarios provided in an embodiment of this application.

[0039] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0040] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0041] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, collected data, used data, generated data, processed data, etc.) and the data (including but not limited to data used for analysis, stored data, displayed data, collected information, used information, generated information, processed information, etc.) are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0042] This method is suitable for online reference and consultation services in educational settings such as libraries and learning support centers. The execution entities include user terminals, reference and consultation servers, and model inference nodes. These three entities can be deployed on the same computing device or in a distributed manner. The user terminal can handle information input, result display, and conversation interaction. The reference and consultation server is responsible for overall process orchestration, data processing, and status management. The model inference nodes specifically run generative models and submitability discrimination models. The execution entities interact with each other through standardized data interfaces, meeting the real-time and high-efficiency data processing requirements of online reference and consultation services.

[0043] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0044] Figure 1 This is a flowchart of a generative model-based library reference consultation question-and-answer method for educational scenarios according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes steps S1-S7:

[0045] In step S1, the reference consultation request text is obtained, and the request text is parsed to obtain a structured parsing result. The structured parsing result includes task type, deliverable type, completion clues, and integrity intent clues. The integrity intent clues are obtained by keyword matching or template matching of the request text.

[0046] Specifically, when a user terminal submits a reference request text to the reference server, the reference server receives the request text and performs input normalization processing. This involves removing redundant whitespace characters from the request text, retaining citation symbols and the original text of the question, and avoiding deviations in subsequent parsing due to text formatting issues. For example, if the request text contains consecutive spaces or line breaks, they are uniformly processed into single spaces; if the text contains citation symbols such as square brackets or quotation marks, their original position and form are preserved.

[0047] After input normalization, the reference and consultation server performs parsing on the request text to obtain structured parsing results. These results include task type, deliverable type, completion cues, and integrity intent cues. All information is extracted from the request text, ensuring a high degree of match between the parsing results and the user's actual consultation needs. Specifically, the task type refers to the category of the specific consultation task obtained after parsing the user's reference and consultation request, including various types such as concept explanation, retrieval assignments, paper writing, book reports, citation formatting, and polishing / rewriting. The deliverable type is the specific form of the consultation result that the user expects, extracted from the user's reference and consultation request text. The completion cues are relevant information extracted from the user's reference and consultation request that reflects the user's current progress and existing achievements in the corresponding learning task.

[0048] Integrity intent clues are obtained by keyword matching or template matching of the request text. The reference consultation server has a pre-set keyword dictionary and matching templates related to academic misconduct. The keyword dictionary includes expressions related to directly requesting a draft, reducing plagiarism and rewriting, and avoiding citations. The matching template is a common sentence structure of the above expressions. When the request text contains keywords from the dictionary or sentence structures from the matching template, it is marked as relevant content of integrity intent clues.

[0049] Specifically, the parsing operation employs a combination of rule-based features and a classification model. The rule layer, based on a pre-defined keyword dictionary and template matching algorithm, extracts deliverable type and integrity intent clues from the request text. The classification model performs multi-classification processing of task types, including common reference and consultation task types such as concept explanation, retrieval assignments, paper writing, book reports, citation formatting, and editing / rewriting. This classification model can utilize convolutional neural networks, recurrent neural networks, or pre-trained language models. Before application, the model is trained using corpora labeled with educational scenario reference and consultation task types. After training, the corpora are deployed to the model's inference nodes to ensure the accuracy of task type recognition.

[0050] In step S2, a learner profile is constructed based on the structured parsing results and the preset course policy parameter set.

[0051] The reference service server has a pre-configured set of course policy parameters. These parameters are pre-configured by librarians or course instructors based on course requirements and academic integrity guidelines. Stored in a structured record format, they include at least the following configuration items: preset maximum output level, whether polishing is allowed, whether example sentences are allowed, whether direct rewriting is allowed, allowed citation formats, and priority resource types. If no corresponding course configuration parameters are detected, the system's default course policy parameter set is used, with the preset maximum output level in the default set limited to no higher than the writing scaffolding level. The writing scaffolding provides learners with a framework for argumentation, paragraph function descriptions, and explanations of replaceable connecting sentence categories for their writing tasks. It does not contain complete, ready-to-submit paragraphs and serves as both instructional and compliance support for writing.

[0052] Based on the above structured analysis results and the set of curriculum policy parameters, a learner profile is constructed. This learner profile is a structured set of features, including core features such as integrity risk tendency score and task completion rate, and may also include auxiliary features such as grade level, writing proficiency score, retrieval ability score, and applicable curriculum policy identifier. The values ​​of all features are calculated based on actual data, and the quantification results of each feature are mapped to a numerical range of 0 to 1.

[0053] The integrity risk propensity score is calculated based on integrity intent clues from the structured parsing results, and is further refined by incorporating user behavior statistics from historical session logs. The specific calculation formula is as follows: In the formula The credit risk propensity score ranges from 0 to 1, with the value closer to 1 indicating a higher academic integrity risk propensity. The binary feature represents whether there is a direct request to obtain the final draft; if yes, it is set to 1, otherwise it is set to 0. This is a binary feature indicating whether there is a request for deduplication rewriting; if yes, it is set to 1, otherwise it is set to 0. To determine whether there are multiple rounds of dialogue that gradually approach the final draft, take 1 if there are, and 0 if there are none. The binary feature indicates whether no self-created content has been provided; if yes, it is set to 1, otherwise it is set to 0. , , , The weight coefficients for each feature are obtained by offline statistical analysis of the contribution of academic misconduct cases in educational scenarios. The values ​​range from 0 to 0 and can be adjusted according to the academic integrity requirements of different educational scenarios.

[0054] The task completion rate is calculated based on completion clues from the structured parsing results, and is also verified by combining actual materials submitted by users, such as drafts, bibliographies, and notes. The specific calculation formula is as follows: In the formula This represents the task completion rate, with a value ranging from 0 to 1. The closer the value is to 1, the higher the completion rate of the user's reference and consultation-related tasks. The binary feature indicating whether there is an outline is set to 1 if it is, and 0 if it is not. The binary feature indicating whether there is a draft is set to 1 if it is, and 0 if it is not. The binary feature indicating whether there is a record is used; if so, it is set to 1, otherwise it is set to 0. The binary feature indicates whether there is a retrieval log; if yes, it is set to 1, otherwise it is set to 0. , , , Here are the weighting coefficients for each feature. These coefficients are pre-configured by the course policy parameter set, and their values ​​range from 0 to 1, and satisfy the following conditions: It can be customized according to the emphasis placed on each stage of the task in different courses.

[0055] Understandably, during multiple rounds of reference consultation sessions, the learner profile will be incrementally updated based on new structured analysis results and supplementary materials submitted by the user. The updates follow the principle of prioritizing new evidence; that is, when new evidence differs from historical data, the feature values ​​of the learner profile are corrected based on the new evidence. Simultaneously, the data source for each update is recorded to ensure the real-time nature and accuracy of the learner profile. By constructing a multi-dimensional and updatable learner profile, accurate user identification is achieved, avoiding the technical problem in existing technologies where the inability to accurately depict learner abilities and task status leads to a mismatch between the granularity of reference consultation output and user needs.

[0056] In step S3, based on the learner profile and the course policy parameter set, a comprehensive risk score is calculated, and the risk level is determined according to the comprehensive risk score and a preset risk threshold.

[0057] Based on the learner profile and course policy parameter set constructed above, the reference consultation server calculates a comprehensive risk score. This comprehensive risk score is obtained by weighted summation of the integrity risk tendency score, the complement of task completion degree, and the approximation index. The specific calculation formula is as follows: In the formula The comprehensive risk score ranges from 0 to 1. The closer the value is to 1, the higher the comprehensive risk of academic misconduct of the user. Scoring the integrity risk propensity in learner profiles; This is a complement to the task completion score, representing the degree to which the user's task is not completed, and its value ranges from 0 to 1; To approximate the index, it is calculated based on the number and magnitude of user request type shifts in multiple rounds of reference consultation sessions, with a value range of 0 to 1. The closer the value is to 1, the closer the user's request is to the final draft.

[0058] , , Here are the weighting coefficients for each feature. These coefficients are pre-configured by the course policy parameter set, and their values ​​range from 0 to 1, and satisfy the following conditions: Its value is set according to the degree of importance attached to each risk factor in different educational scenarios. For example, for postgraduate courses with high requirements for academic integrity, it can be... The value is configured to be relatively large to highlight the impact of integrity risk tendency on overall risk; for basic courses, it can be... The value is configured to be relatively large to focus on the risk arising from task incompleteness. This weighting coefficient can be adjusted by librarians or course instructors according to actual teaching needs. After adjustment, it can be tested in a small-scale scenario to ensure the rationality of the comprehensive risk score calculation.

[0059] The approximation index is calculated as follows: The reference service server, based on preset round statistics in the course policy parameter set, counts the number and magnitude of user reference service requests shifting from strategy outlines to sentence-by-sentence rewriting of drafts in the most recent rounds of conversations. The magnitude is represented by the difference in request type sequence numbers. The reference service server assigns ordered sequence numbers to different reference service request types; the closer the request type is to the draft, the larger the sequence number. The shift magnitude is the difference in sequence numbers between two consecutive rounds of request types. The approximation index is obtained by normalizing the weighted sum of the number of shifts and the magnitude. The more shifts and the larger the magnitude, the higher the approximation index. The weighting coefficients used in the calculation are obtained from offline system statistics and range from 0 to 1.

[0060] As an example, requests are first ordered from low to high risk by type, such as L1 clarification interview as 1, L2 retrieval strategy as 2, L3 reading comprehension as 3, L4 writing scaffolding as 4, and L5 draft level as 5. Then, the most recent 3 rounds of conversations are counted according to the k-value preset in the course policy parameter set. If a user request changes from retrieval strategy 2 to reading comprehension 3, and then to writing scaffolding 4, the magnitudes of the two changes are 1 and 1 respectively, and the number of changes is 2. The total magnitude 2 is normalized to 0-1, and then weighted by the number of changes to obtain the approximation index. .

[0061] Meanwhile, the course policy parameter set has preset risk thresholds, including medium-risk and high-risk thresholds. If the course policy parameter set does not have a preset threshold, the system default value is used. The default value can be obtained by offline statistical analysis of academic misconduct risk cases in educational scenarios and can be updated based on actual application results. The reference consultation server compares the calculated comprehensive risk score with the preset risk thresholds to determine the risk level. The risk level is divided into three levels: low risk, medium risk, and high risk. Specifically, if the comprehensive risk score is greater than or equal to the high-risk threshold, it is judged as high risk; if the comprehensive risk score is greater than or equal to the medium-risk threshold but less than the high-risk threshold, it is judged as medium risk; if the comprehensive risk score is less than the medium-risk threshold, it is judged as low risk. The above thresholds can be determined by performing five-fold cross-validation on the labeled dataset, using the parameter group with the highest F1 score and a false positive rate lower than the preset upper limit. In one embodiment, T1 = 0.40, T2 = 0.70, T3 = 0.50, and T4 = 0.60.

[0062] After the risk level is determined, the reference consultation server will store the comprehensive risk score and risk level in a structured manner and associate them with the corresponding learner profile. By constructing a multi-dimensional comprehensive risk score calculation model, the server can achieve quantitative assessment and classification of academic misconduct risk.

[0063] In step S4, the highest allowed output level is determined based on the risk level, the course policy parameter set, and the task completion rate in the learner profile.

[0064] Based on the risk level, course policy parameter set, and task completion rate in the learner profile, the highest allowed output level is determined. In this embodiment, the output level is divided into five levels, from low to high: clarification interview level, retrieval strategy level, reading comprehension level, writing scaffolding level, and final draft text level. Each level corresponds to different reference consultation output granularity and content type. The final draft text level is the highest output level and is not open by default.

[0065] In one embodiment, the output levels are defined as follows: Clarification Interview Level (L1): The system only outputs a set of clarification questions and task boundary descriptions to help users define their research questions without providing any substantive content output; Retrieval Strategy Level (L2): The system provides users with guidance on retrieval strategies such as keyword expansion methods, retrieval structure, screening criteria, and database access paths to help users learn how to find literature resources, but does not directly provide the literature content itself; Reading Comprehension Level (L3): The system outputs relevant guidance on reading and note-taking methods to help users master how to effectively read and extract from retrieved literature, but does not replace the user's reading and summarizing; Writing Scaffolding Level (L4): The system provides structural writing aids such as argumentation frameworks, paragraph function descriptions, and functional categories of replaceable connecting sentences to guide users to fill in the writing within the framework, but does not output complete paragraphs that can be directly submitted; Final Draft Text Level (L5): The system outputs a nearly complete final draft text that can be directly submitted. This level is the highest output level and is not open by default, and may only be temporarily allowed in very rare cases where there is low risk and the user provides sufficient verifiable materials.

[0066] The determination of the maximum allowable output level follows the principle of first setting an initial upper limit, then constraining according to risk level, and fine-tuning in conjunction with task completion rate. Figure 2 A schematic diagram of the process for determining the highest allowed output level provided in an embodiment of this application is shown.

[0067] like Figure 2 As shown, in step S201, the initial upper limit is determined based on the preset highest output level in the course policy parameter set. The specific calculation formula is as follows: In the formula As the initial upper limit, This is the preset highest output level in the course policy parameter set. As a writing scaffolding level, the purpose of this calculation rule is to avoid directly opening the final draft level of text by default, thereby avoiding the risk of academic misconduct from the source.

[0068] In step S202, the initial upper limit is constrained based on the risk level, and fine-tuned according to the task completion rate to obtain the highest allowed output level. If the risk level is high, the highest allowed output level cannot exceed the retrieval strategy level, and the specific calculation formula is as follows: In the formula To allow the highest output level, To ensure that only basic search strategy guidance is provided to users in high-risk situations, the output granularity is strictly limited. If the risk level is medium, the highest allowed output level is the smaller of the initial upper limit and the writing scaffolding level. This is contingent on the task completion rate in the learner profile being greater than or equal to the progressive disclosure threshold preset in the course policy parameter set. If the task completion rate is less than the progressive disclosure threshold, the highest allowed output level will be lowered to the reading comprehension level. If the risk level is low, the highest allowed output level is equal to the initial upper limit. Only when the user meets specific conditions can the allowed highest output level be temporarily raised to the draft text level. These specific conditions include providing the user with no less than a preset number of verifiable bibliographic records and location-based reading notes, and the task completion rate being greater than or equal to a preset high task completion threshold. The preset number and threshold are configured by the course policy parameter set.

[0069] The progressive disclosure threshold is used to determine whether a user has the conditions to obtain higher-level output content. Its value is pre-configured by the course policy parameter set and ranges from 0 to 1. It can be adjusted according to the teaching progress and requirements of different courses. After adjustment, it needs to be verified in combination with the actual teaching scenario to ensure its rationality. This parameter can be adjusted by librarians or course teachers when the course teaching requirements change.

[0070] After determining the highest allowed output level, the reference and consultation server associates and stores it with the risk level and learner profile. Through the output level constraint mechanism based on the risk level, it achieves a precise match between the output granularity and the user's risk level, thereby preventing academic misconduct from the source and taking into account the reference and consultation needs of different users.

[0071] In step S5, the inference input is provided to the generative model to generate candidate responses. The inference input includes constraints extracted from the structured parsing results, the learner profile, and the risk level. Here, a candidate response refers to the initial output generated by the generative model based on the inference input; the final response refers to the output text sent to the user terminal; and a compliant response refers to the final response that passes the submitability test or, after being downgraded and rewritten, meets the course policy constraints and academic integrity requirements. If the submitability score of a candidate response is lower than a preset threshold, the candidate response can be directly output as a compliant response.

[0072] Optionally, before generating candidate responses, the reference consultation server generates a learning step chain based on the task type in the structured parsing results and the maximum allowed output level determined above. This learning step chain contains multiple step nodes, each containing the required input fields and corresponding validation rules, providing users with step-by-step learning guidance to guide them through completing the reference consultation-related tasks step by step, avoiding obtaining the final draft content in one step.

[0073] The rules for generating the learning step chain are customized based on the task type. For example, if the task type is thesis writing, the learning step chain includes steps such as defining the research question, retrieving the search plan, submitting bibliographic data, reading notes, and establishing an argumentation framework. The required input fields for the research question definition step are one research question and three keywords. The corresponding validation rule is that the keywords are not empty and the semantic similarity between the keywords and the research question is greater than a preset threshold. The required input fields for the search plan step are database selection or resource type. The corresponding validation rule is that the resource type belongs to the set of resource types allowed by the course policy parameter set. If the highest allowed output level is no higher than the search strategy level, then the learning step chain must contain at least two step nodes: defining the research question and retrieving the search plan. Users are required to submit search results and pass the validation rules before proceeding to the next step.

[0074] The reference consultation server writes the current node information of the learning step chain into the generative model inference input. Then, based on the structured parsing results, learner profile, risk level, maximum allowed output level, and the current node information of the learning step chain, it constructs the generative model inference input. This inference input includes constraints extracted from the structured parsing results, learner profile, and risk level, as well as a set of constraint fields. This set of constraint fields includes at least the maximum allowed output level and a list of prohibited output items derived from the maximum allowed output level.

[0075] Specifically, the input to the generative model's inference is a structured dataset comprising four parts: a dialogue context summary, a set of constraint fields, a scaffolding template, and verifiable materials provided by the user. The dialogue context summary contains only task-type-related information and is obtained by the reference server summarizing relevant data from the current and historical sessions. The set of constraint fields includes the highest allowed output level and a list of prohibited outputs, as well as constraints extracted from the course policy parameter set, such as citation format requirements and whether polishing is allowed. The list of prohibited outputs is derived from the highest allowed output level; for example, if the highest allowed output level is no higher than the retrieval strategy level, the list includes complete paragraphs and directly submittable text. The scaffolding template is a predefined output structure template for different output levels. The reference server selects the corresponding scaffolding template based on the highest allowed output level to ensure the generative model's output structure meets the requirements. Verifiable materials provided by the user include bibliographical entries, notes, retrieval logs, and draft fragments submitted by the user during the session, providing practical material support for the generative model's inference.

[0076] During the construction of the input for generative model inference, if the highest output level is allowed to be no higher than the retrieval strategy level, the scaffolding template is fixed as a template combining clarification questions and retrieval strategies, and constraints are explicitly added to the constraint field set to prohibit the output of complete paragraphs and allow direct text submission; if the highest output level is allowed to be the writing scaffolding level, the scaffolding template is a template combining argumentation framework, paragraph function descriptions and placeholder hints, and constraints are added to the constraint field set to prohibit the output of complete draft paragraphs that are highly relevant to the topic.

[0077] In one embodiment, a generative model is deployed in the model inference node. This generative model employs a combined structure of a generative master model and a constraint template layer. Figure 3 A schematic diagram of the structure of the generative model provided in an embodiment of this disclosure is shown. For example... Figure 3 As shown, the generative model includes an input encoding module, a constraint fusion module, a decoding generation module, and a template rendering module. These modules are connected in series in sequence, with the output of the previous module serving as the input of the next module, forming a complete model inference process. The input of this generative model is the inference input of the generative model constructed above, and the output is a standardized candidate response.

[0078] Specifically, the input encoding module receives the inference input from the generative model, performs structured feature extraction and sequence concatenation on the dialogue context summary, constraint field set, scaffolding template, and user-provided verifiable materials, transforming the structured inference input into a vector sequence recognizable by the generative main model, and outputs an encoded representation. This encoded representation is a high-dimensional vector sequence, the dimension of which is preset by the model structure of the generative main model. The input encoding module can be implemented using the encoding layer of a pre-trained language model, such as the BERT-based pre-trained encoding algorithm, which segments and embeds each field of the structured inference input and then concatenates them to output a multi-dimensional encoded representation. It can be understood that before applying the model, it can be pre-trained using a large-scale educational text corpus to ensure the accuracy of feature extraction.

[0079] The constraint fusion module encodes the highest allowed output level and prohibited output items into constraint vectors. The dimension of these constraint vectors matches the dimension of the encoded representation output by the input encoding module. The constraint vectors and encoded representations are then fused to output a fused representation. The fusion processing algorithm can be either a vector concatenation algorithm or an attention fusion algorithm. Preferably, when using the attention fusion algorithm, the constraint vectors are used as attention weights to weight the encoded representations, allowing the generative model to focus on information related to the constraints during inference, thereby improving the matching degree between the generated results and the constraints.

[0080] The decoding and generation module generates the original text of candidate responses based on the fusion representation output by the constraint fusion module, according to the paragraph structure specified in the scaffolding template. This original text includes three parts: the main body, the next node in the learning step chain, and the required user input instructions. The decoding and generation module is implemented using the decoding layer of the generative master model, following autoregressive generation rules, such as the GPT-2 autoregressive decoding algorithm. It can be understood that during model training, fine-tuning can be performed using corpora labeled with educational scenario reference consultation content and corresponding output levels. The training loss function adopts the cross-entropy loss function to ensure that the generated results meet the output level requirements and teaching support needs.

[0081] The template rendering module standardizes the raw text output by the decoding and generation module. It uses preset rule templates in the course policy parameter set to render content such as citation format, numbering, and placeholders in the text. It can replace the output with standardized candidate responses through regular expression matching. The citation format follows the allowed citation format configured in the course policy parameter set, the numbering adopts a standardized hierarchical numbering method, and the placeholders are replaced with prompt text that meets the teaching support requirements, ensuring the standardization and uniformity of the candidate response format.

[0082] When generating candidate responses, the generative model follows a hierarchical generation rule, generating only content corresponding to the highest allowed output level and not generating higher-granularity content across levels. If the highest allowed output level is the clarification interview level, only the set of clarification questions and task boundary descriptions are output; if it is the retrieval strategy level, only content related to retrieval strategies such as keyword expansion methods, retrieval structure, screening criteria, and database access paths are output; if it is the reading comprehension level, only content related to reading and note-taking methods is output; if it is the writing scaffolding level, only content such as the argumentation framework, paragraph function descriptions, and functional categories of replaceable connecting sentences are output, without providing complete paragraphs that can be directly submitted. By constructing a constrained generative model structure and hierarchical generation rules, precise matching between generated content and output levels is achieved.

[0083] In step S6, the candidate responses are processed using a submitability discrimination model to obtain a submitability score.

[0084] The model inference node also deploys a submitability discrimination model, which is a deep learning model containing a text feature extraction module and a fusion classification module, connected in sequence. This model is used to evaluate the submitability of the candidate responses generated above, and outputs a submitability score and a triggering reason vector. The submitability score is used to quantitatively represent whether the candidate response can be directly submitted as an assignment, and the triggering reason vector is used to represent the contribution of each feature to the submitability score.

[0085] For example, the text feature extraction module uses the RoBERTa-base pre-trained algorithm. After inputting candidate responses for word segmentation and vector embedding, it extracts multi-dimensional feature vectors through pooling layers, which can cover core indicators such as coherence and argument completeness. The fusion classification module first concatenates the feature vector with risk level, allowable level, and task completion degree numerically to obtain a fused feature vector. Then, it inputs the fused feature vector into a three-layer fully connected regression network. After processing with ReLU activation function and Dropout regularization, it obtains the submitability score in the 0-1 interval through the Sigmoid output layer. At the same time, it generates the triggering reason vector of each feature through the gradient contribution analysis algorithm to quantify the contribution of each feature to the score.

[0086] Specifically, the text feature extraction module receives candidate responses, extracts multi-dimensional text features related to submitability from them, and outputs a feature set, which is a high-dimensional feature vector. It includes at least six core indicators: coherence, completeness of argumentation, fit with the request text, citation density and standardization, presence of a final draft structure, and lack of learning process traces. Each indicator is quantified and has a value range of 0 to 1.

[0087] The coherence index uses a weighted sum of inter-sentence and inter-paragraph similarity; the argument completeness index uses the completeness of the argument chain; the fit with the request text index uses the semantic similarity between the candidate response and the reference consultation request text; the citation density and standardization index uses a weighted sum of the density and standardization of citation marks in the text; the presence of a final draft structure index is a binary feature, which is 1 if the candidate response has a final draft structure of introduction-method-discussion-conclusion, and 0 otherwise; the lack of learning process traces index is a binary feature, which is 1 if the candidate response lacks prompts for users to supplement content themselves, and 0 otherwise.

[0088] The fusion classification module concatenates the feature set output by the text feature extraction module with the risk level and the highest allowed output level to obtain a fusion feature vector. This fusion feature vector integrates text features and conversation features, and more comprehensively reflects the submitability of candidate responses.

[0089] The fused feature vector is then input into the classifier, which outputs a submitability score and a triggering reason vector. The submitability score ranges from 0 to 1, with values ​​closer to 1 indicating higher submitability of the candidate response. The dimension of the triggering reason vector matches the dimension of the feature set, and each element in the vector represents the contribution of the corresponding feature to the submitability score; larger element values ​​indicate a greater impact of the feature on submitability.

[0090] As an example, the generation of the trigger cause vector through the gradient contribution analysis algorithm is as follows: First, taking the submitability score output by the model as the target value, the partial derivatives of the score with respect to each dimension of the feature set output by the text feature extraction module are calculated to obtain the gradient value of each feature. This gradient value represents the degree of influence of a small change in a single feature on the score. Then, the absolute values ​​of all gradient values ​​are processed to eliminate the sign difference between positive and negative influences. Next, the absolute value gradient values ​​are normalized by 0-1 so that the sum of the contributions of each feature is 1. The normalized value is the contribution of each feature to the submitability score. Finally, the contributions of all features are arranged in order of the dimensions of the feature set to form the trigger cause vector. The larger the value of the element in the vector, the higher the contribution of the corresponding feature to the excessively high submitability of the candidate response.

[0091] Understandably, before applying a submitability discrimination model, it needs to undergo complete training. First, a pre-trained language model is used to initialize the text feature extraction module. Then, reference consultation texts in educational scenarios are collected, their submitability scores and the contribution of each feature are labeled, and a training dataset is constructed. Finally, the entire discrimination model is trained end-to-end using the training dataset. The training objective is to minimize the error between the model's predicted submitability scores and the true labeled values, while accurately outputting the contribution of each feature.

[0092] The reference consultation server has a preset submitability threshold, which is configured by the course policy parameter set. If not configured, the system default value is used. The default value is obtained by labeling the submitability features of assignment texts in offline statistical education scenarios, and the value ranges from 0 to 1. The reference consultation server compares the submitability score output by the submitability discrimination model with the preset submitability threshold to determine whether the submitability of the candidate response is too high. By constructing a dedicated submitability discrimination model, a quantitative assessment of submitability is achieved.

[0093] In step S7, if the submitability score is greater than or equal to the preset submitability threshold, then the candidate response is downgraded and rewritten to generate a compliant response and output it.

[0094] If the submitability score is less than the preset submitability threshold, it means that the submitability of the candidate response meets the requirements. The reference consultation server will encapsulate the candidate response as the final response and output it to the user terminal. If the submitability score is greater than or equal to the preset submitability threshold, it means that the submitability of the candidate response is too high. The reference consultation server will perform a downgrade rewriting operation on the candidate response, generate a compliant response, and output it.

[0095] Figure 4 A schematic diagram of the downgrade rewriting operation process provided in an embodiment of this disclosure is shown. For example... Figure 4 As shown, the downgrade rewriting operation includes steps such as locating highly submitable segments, rewriting the highly submitable segments, temporarily lowering the allowed maximum output level, and outputting a compliant response. Specifically, in step S401, highly submitable segments in the candidate response are identified and located based on the triggering cause vector output by the submitability discrimination model and the paragraph boundaries of the candidate response. Specifically, firstly, the feature contributing the most to the submitability score is determined based on the triggering cause vector; then, paragraphs containing this feature are located in the candidate response. Simultaneously, combined with the paragraph boundary division of the text, paragraphs with complete argument chains, direct copyability, and high relevance to the question are identified as highly submitable segments. For example, if the argument completeness index has the highest contribution in the triggering cause vector, then paragraphs in the candidate response containing complete argument chains are identified as highly submitable segments.

[0096] In step S402, the highly submitable fragments obtained from the location are rewritten. As an example, the rewriting method involves either rewriting complete argument sentences in the highly submitable fragment into paragraph function descriptions and placeholders, or retaining the fragment's structural framework while removing complete statements. These two rewriting methods can be used individually or in combination. Rewriting complete argument sentences into paragraph function descriptions and placeholders specifically involves removing specific assertions and complete statements from the fragment and replacing them with clear paragraph function descriptions and placeholders prompting the user to fill in the content. For example, a specific argument statement might be rewritten as "Fill in your claim here," citing relevant literature research conclusions as evidence. The function of this paragraph is to explain the practical value of the research. Retaining the fragment's structural framework while removing complete statements specifically involves retaining the title, key point numbers, hierarchical structure, and other framework content of the highly submitable fragment, while removing coherent complete sentence groups and specific statements from the fragment. Only a structural framework is provided to guide the user in filling in the content independently based on the framework.

[0097] In step S403, the maximum output level will be temporarily lowered. After rewriting the highly submitable segment, the reference server will allow the maximum output level to be temporarily lowered. The lowering rule is to adjust it to the smaller value between the original maximum output level and the reading comprehension level, or to lower it according to the downgrading rules preset in the course policy parameter set. The temporarily lowered maximum output level will be applied to subsequent session rounds to prevent the generation of overly submitable content in the same round or subsequent rounds. This temporarily lowered level is only valid within the current session period. When the user ends the session and initiates a new session, the maximum output level will be recalculated.

[0098] In one embodiment, the downgraded and rewritten text is input again into the submitability discrimination model for review; if the review score is still greater than or equal to the preset submitability threshold, downgraded rewriting continues until the review score is less than the threshold and then output.

[0099] In step S404, the rewritten content is integrated to generate a compliant response. The reference consultation server integrates the rewritten content to generate a compliant response, which includes three core parts: a compliance boundary statement, the next learning steps, and a template for required user input fields. The compliance boundary statement clarifies the scope of assistance and academic integrity requirements of this reference consultation to the user. The next learning steps are subsequent nodes in the learning step chain, and the template for required user input fields is a standardized input template corresponding to the next learning step, providing clear input guidance for the user. By constructing targeted downgrading rewriting rules, highly submitable text is transformed into compliant teaching support content.

[0100] Finally, the reference consultation server encapsulates the compliance response into a terminal display data packet. This data packet contains the response text, next learning steps, required input instructions, compliance prompts, etc., and is transmitted to the user terminal through the data interface. The user terminal parses and displays the data, providing the user with a compliance reference consultation response and clear follow-up learning guidance.

[0101] Optionally, if the user submits supplementary materials as required, the reference consultation server will re-execute all the above steps, update the learner profile based on the new materials, recalculate the risk level and the maximum allowed output level, and form a complete reference consultation process. If the user continues to make high-risk requests, causing the risk level to remain high for more than the preset number of rounds, the reference consultation server will fix the learning step chain at the clarification interview level and the retrieval strategy level, and only output clarification questions and retrieval strategy guidance content until the user provides verifiable materials or the session ends.

[0102] Therefore, this method, through the coordinated execution of the above steps, achieves the dual goals of compliance and pedagogical relevance in library reference consultation and Q&A within educational scenarios, effectively addressing technical issues such as insufficient compliance, inadequate pedagogical support, and lack of consistent processes in existing technologies. It achieves precise user profiling by constructing learner profiles and dynamically adjusts output levels based on risk levels to prevent academic misconduct at its source; it achieves precise control and submitability assessment of generated content by constructing constrained generative models and submitability discrimination models; it automatically corrects highly submitable text through downgrading and rewriting operations, forming compliance control; and it provides step-by-step pedagogical support to users by generating learning step chains, enhancing the pedagogical value of reference consultation.

[0103] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a generative model-based library reference consultation question-and-answer system for educational scenarios, provided in an embodiment of this application. Figure 5 As shown, the system includes:

[0104] The request parsing module 501 is used to obtain the reference consultation request text, parse the request text, and obtain a structured parsing result. The structured parsing result includes task type, deliverable type, completion clues, and integrity intent clues. The integrity intent clues are obtained by keyword matching or template matching of the request text.

[0105] The profile building module 502 is used to build learner profiles based on the structured parsing results and the preset course policy parameter set;

[0106] The risk assessment module 503 is used to calculate a comprehensive risk score based on the learner profile and the course policy parameter set, and to determine the risk level based on the comprehensive risk score and a preset risk threshold.

[0107] The hierarchical control module 504 is used to determine the highest allowed output level based on the risk level, the course policy parameter set, and the task completion rate in the learner profile.

[0108] The candidate generation module 505 is used to provide inference input to the generative model to generate candidate responses. The inference input includes constraints extracted from the structured parsing results, the learner profile, and the risk level.

[0109] The submitability discrimination module 506 is used to process the candidate response through the submitability discrimination model to obtain a submitability score;

[0110] The output rewriting module 507 is used to perform a downgrade rewriting operation on the candidate response if the submitability score is greater than or equal to a preset submitability threshold, generate a compliant response and output it.

[0111] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0112] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0113] Please see Figure 6 It shows a schematic diagram of the structure of an electronic device according to an embodiment of this application, which can be used to implement... Figure 1 The method in the illustrated embodiment. (As shown) Figure 6 As shown, the electronic device may include:

[0114] The system includes at least one processor 601, at least one network interface 604, a user interface 603, a memory 605, and at least one communication bus 602. The communication bus 602 is used to enable connection and communication between the components. The user interface 603 may include buttons, and optionally include a standard wired or wireless interface. The network interface 604 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0115] The processor 601 may include one or more processing cores and connect to various parts within the electronic device through various interfaces and lines. It implements various functions and data processing of the electronic device by running or executing instructions, programs, code sets, or instruction sets stored in the memory 605, and by accessing data in the memory 605. Optionally, the processor 601 may be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor 601 may also integrate one or more combinations of CPU, GPU, and modem.

[0116] Memory 605 may include random access memory (RAM) or read-only memory (ROM). Optionally, memory 605 includes a non-transitory computer-readable medium for storing instructions, programs, code, code sets, or instruction sets. Memory 605 may be divided into a program storage area and a data storage area, wherein the program storage area can be used to store instructions for implementing an operating system and instructions for implementing the foregoing method embodiments; the data storage area can be used to store data related to the relevant method embodiments. Memory 605 may also be at least one storage device located remotely from processor 601. Figure 6 As shown, the memory 605, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and program instructions.

[0117] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by processor 601, it performs the functions defined in the methods of this application.

[0118] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0119] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A generative model-based question-and-answer method for library reference services in educational settings, characterized in that: Includes the following steps: Obtain the reference consultation request text, parse the request text to obtain a structured parsing result, wherein the structured parsing result includes task type, deliverable type, completion status clues, and integrity intent clues; the integrity intent clues are obtained by keyword matching or template matching of the request text; Based on the structured analysis results and the preset set of course policy parameters, a learner profile is constructed; Based on the learner profile and the course policy parameter set, a comprehensive risk score is calculated, and the risk level is determined according to the comprehensive risk score and the preset risk threshold. The highest allowed output level is determined based on the risk level, the course policy parameter set, and the task completion rate in the learner profile. The inference input is provided to the generative model to generate candidate responses. The inference input includes constraints extracted from the structured parsing results, the learner profile, and the risk level, and also includes a set of constraint fields, which at least includes the highest allowed output level and a list of prohibited output items derived from the highest allowed output level. The candidate responses are processed using a submitability discrimination model to obtain submitability scores; If the submitability score is greater than or equal to the preset submitability threshold, a downgrade rewriting operation is performed on the candidate response to generate a compliant response and output it. The downgrade rewriting operation includes: locating the highly submitable segments in the candidate response based on the triggering cause vector; rewriting the complete argument sentences in the highly submitable segments into paragraph function descriptions and placeholders, or retaining the structural framework of the segment and removing the complete expressions therein.

2. The method according to claim 1, characterized in that, The learner profile includes a trust risk propensity score and task completion rate; the trust risk propensity score is calculated based on the trust intention cues in the structured parsing results. The task completion rate is calculated based on the completion rate clues in the structured parsing results.

3. The method according to claim 2, characterized in that, The comprehensive risk score is obtained by weighted summation of the integrity risk tendency score, the task completion score, and the approximation index; the approximation index is calculated based on the number and magnitude of user request type transitions in multiple rounds of reference consultation sessions.

4. The method according to claim 1, characterized in that, The steps for determining the maximum allowed output level include: determining an initial upper limit based on the preset maximum output level in the course policy parameter set; constraining the initial upper limit based on the risk level; wherein, if the risk level is high risk, the maximum allowed output level is no higher than the retrieval strategy level, wherein the output levels are ordered from low to high according to output priority as follows: clarification interview level, retrieval strategy level, reading comprehension level, writing scaffolding level, and final draft text level; and the final draft text level is not open by default.

5. The method according to claim 1, characterized in that, The step of processing the candidate response using a submitability discrimination model to obtain a submitability score specifically includes: extracting text features from the candidate response, including coherence indicators, argument completeness indicators, and relevance to the request text indicators; fusing the text features with the risk level and the highest allowed output level, and inputting them into a classifier to obtain the submitability score.

6. The method according to claim 5, characterized in that, The downgrade rewrite operation further includes: after performing the rewrite, temporarily lowering the allowed highest output level and outputting a compliant response containing the required user input field template.

7. The method according to claim 1, characterized in that, It also includes the step of generating a learning step chain: before generating the candidate response, a learning step chain containing multiple step nodes is generated based on the task type in the structured parsing result and the allowed highest output level, each step node containing the required input fields and the corresponding verification rules; and the current node information of the learning step chain is written into the generative model inference input.

8. A generative model-based library reference consultation question-and-answer system for educational scenarios, characterized in that, include: The request parsing module is used to obtain reference consultation request text, parse the request text, and obtain structured parsing results. The structured parsing results include task type, deliverable type, completion clues, and integrity intent clues. The integrity intent clues are obtained by keyword matching or template matching of the request text. The profile building module is used to build learner profiles based on the structured parsing results and a preset set of course policy parameters. The risk assessment module is used to calculate a comprehensive risk score based on the learner profile and the course policy parameter set, and to determine the risk level based on the comprehensive risk score and a preset risk threshold. The hierarchical control module is used to determine the highest allowed output level based on the risk level, the course policy parameter set, and the task completion rate in the learner profile. The candidate generation module is used to provide inference input to the generative model and generate candidate responses. The inference input includes constraints extracted from the structured parsing results, the learner profile and the risk level, and also includes a set of constraint fields. The set of constraint fields includes at least the highest allowed output level and a list of prohibited output items derived from the highest allowed output level. The submitability discrimination module is used to process the candidate responses through a submitability discrimination model to obtain a submitability score; The output rewriting module is used to perform a downgrade rewriting operation on the candidate response if the submitability score is greater than or equal to a preset submitability threshold, generate a compliant response and output it; the downgrade rewriting operation includes: locating the highly submitable fragments in the candidate response according to the triggering cause vector; rewriting the complete argument sentences in the highly submitable fragments into paragraph function descriptions and placeholders, or retaining the structural framework of the fragments and removing the complete expressions therein.

Images