An ai conversation follow-up conflict interception and redirection method and device

Abstract

The application discloses an AI conversation follow-up question conflict interception and redirection method and device, and relates to the online conversation technical field based on artificial intelligence. The method comprises the following steps: a conversation-level semantic fingerprint library is constructed to store the semantic fingerprints of the follow-up questions which have been issued; the fingerprints are clustered to form historical semantic regions with regional centers and boundaries; before the candidate follow-up question is issued, the comprehensive hit scores of the candidate follow-up question and each historical semantic region are calculated to determine the semantic conflict; if the hit score is high, a structured prohibited semantic region mark is generated, which is then converted into follow-up question redirection guide information and input into a follow-up question generation module to drive the module to generate a redirected candidate follow-up question; the redirection result is rechecked, if the result is passed, the fingerprint library is updated, and if the result is not passed, iterative retries are performed until the number of times reaches the upper limit. The application realizes efficient and accurate semantic conflict interception and follow-up question redirection.

Description

Technical Field

[0001] This invention relates to the field of online conversation technology based on artificial intelligence, and in particular to an AI conversation questioning conflict interception and redirection method and apparatus. Background Technology

[0002] In real-time human-computer dialogue scenarios such as AI interviews, automated customer service, and intelligent teaching, systems typically need to dynamically generate follow-up questions based on user responses for in-depth exploration or clarification. Currently, mainstream solutions rely on the contextual understanding and generation capabilities of Large Language Models (LLMs) to generate candidate follow-up questions. However, existing follow-up question generation and management mechanisms have significant technical shortcomings, especially in multi-turn, long conversation contexts, which easily lead to semantically repetitive questioning, affecting interaction efficiency and user experience. Specifically, this manifests in the following aspects: First, existing technologies mostly rely on string matching or keyword filtering for deduplication. This method cannot identify follow-up questions that have the same meaning but different expressions. For example, "What communication actions did you take at the time?" and "How did you communicate with the other party?" are different in string format, but their core meaning is highly consistent. String deduplication methods will misjudge them as new questions, causing the conversation to go in circles.

[0003] Second, existing solutions typically have a localized detection scope, limited to the current question or conversation round. This fails to address semantic repetition across questions and conversation phases. For example, in an interview, given two different questions about a candidate's "teamwork" and "conflict handling," the system might generate semantically similar follow-up questions to understand the details of the communication. However, existing technologies, lacking a conversation-level global perspective, cannot identify such cross-question repetitions.

[0004] Third, when the system detects potential duplication, it typically only attempts to rewrite the code by adding natural language prompts such as "do not repeat the question" to the LLM. This approach is weakly binding; the LLM may ignore the prompts or return to the semantic region that was already being asked through paraphrasing, lacking a mechanism for forced avoidance.

[0005] Fourth, existing technologies lack structured definitions, measurements, and closed-loop verification for semantic repetition. The system cannot accurately answer questions such as "which / which follow-up questions in the history are repeated by the current follow-up question" or "what type of repetition does it belong to," and after attempting rewriting, it cannot ensure that newly generated follow-up questions actually avoid prohibited semantic regions. The entire process is like a black box, with poor interpretability and auditability.

[0006] Therefore, there is an urgent need for a technology that can perform fine-grained semantic conflict identification and provide structured constraints at the entire conversation level before follow-up questions are issued, so as to drive the follow-up question generation model to effectively avoid topics or dialogue directions that have already been asked, thereby improving the coherence of AI dialogue and enhancing the efficiency and experience of AI dialogue. Summary of the Invention

[0007] In view of the above-mentioned defects or deficiencies in the prior art, the present invention provides an AI conversation follow-up question conflict interception and redirection method and apparatus, which aims to solve the technical problem of semantic repetition in multi-round AI dialogue. By constructing a conversation-level semantic fingerprint database, clustering to form historical semantic regions, performing multi-dimensional conflict detection, generating structured prohibition tags and closed-loop re-examination, efficient and accurate semantic conflict interception and follow-up question redirection are achieved.

[0008] One aspect of the present invention provides a method for intercepting and redirecting AI conversation follow-up questions, comprising the following steps: Map the follow-up questions that have been issued to semantic fingerprints, and build a semantic fingerprint library to store the semantics of the follow-up questions. The semantic fingerprints in the semantic fingerprint database are clustered to form multiple historical semantic regions, each of which is used to represent a semantic direction that has been queried. Obtain candidate follow-up questions in the current session, map the candidate follow-up questions to candidate semantic fingerprints, determine whether the candidate semantic fingerprints hit any of the historical semantic regions, if they hit, generate a prohibited semantic region marker, the prohibited semantic region marker is used to structurally describe the hit historical semantic region, and provide constraint parameters for redirecting candidate follow-up questions; The prohibited semantic region markers are converted into follow-up question redirection guidance information, and the follow-up question redirection guidance information is input into the follow-up question generation module to generate candidate follow-up questions that avoid the hit historical semantic regions; The newly generated candidate follow-up questions are re-evaluated for hit detection. If no historical semantic region is hit, the newly generated candidate follow-up question is allowed to be issued, and the candidate semantic fingerprint of the newly generated candidate follow-up question is written into the semantic fingerprint database. If at least one historical semantic region is hit, the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven to regenerate candidate follow-up questions again.

[0009] In another aspect, the present invention provides an AI conversation follow-up conflict interception and redirection device, comprising: The semantic fingerprint database construction module is used to map the follow-up question association data corresponding to the issued follow-up questions to semantic fingerprints, and to build a semantic fingerprint database for storing the semantics of the follow-up questions. The historical semantic region management module is used to cluster the semantic fingerprints in the semantic fingerprint database to form multiple historical semantic regions, each of which is used to represent a semantic direction that has been queried. The conflict detection module is used to obtain candidate follow-up questions in the current session, map the candidate follow-up questions to candidate semantic fingerprints, determine whether the candidate semantic fingerprints hit any of the historical semantic regions, and if they hit, generate a prohibited semantic region marker. The prohibited semantic region marker is used to structurally describe the hit historical semantic region and provide constraint parameters for redirecting candidate follow-up questions. The redirection guidance module is used to convert the prohibited semantic region marker into follow-up redirection guidance information, and input the follow-up redirection guidance information into the follow-up question generation module to generate candidate follow-up questions that avoid the hit historical semantic region; The re-examination and distribution module is used to re-determine the hit rate of the regenerated candidate follow-up questions. If no historical semantic region is hit, the regenerated candidate follow-up question is allowed to be distributed, and the candidate semantic fingerprint of the regenerated candidate follow-up question is written into the semantic fingerprint database. If at least one historical semantic region is hit, the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven again to regenerate the candidate follow-up question.

[0010] The AI ​​conversation follow-up questioning conflict interception and redirection method and apparatus provided by the present invention have the following beneficial effects: (1) By mapping follow-up questions to semantic fingerprints and performing clustering, it is possible to identify complex forms of semantic repetition such as synonym rewriting and near-synonym expansion, which breaks through the limitations of string matching and significantly improves the accuracy and depth of conflict identification.

[0011] (2) By constructing and maintaining a session-level semantic fingerprint database and historical semantic regions, the scope of conflict detection covers the entire dialogue lifecycle, effectively discovering and intercepting semantic recurrence across issues and stages, and ensuring the global consistency of dialogue progress.

[0012] (3) By transforming the detected conflicts into structured prohibited semantic region markers and follow-up question redirection guidance information, the follow-up question generation module is provided with clear, machine-readable avoidance targets and parameters, rather than vague natural language prompts, which greatly improves the effectiveness and stability of redirection generation.

[0013] (4) By introducing a conflict detection mechanism for new follow-up questions after redirection, a closed loop of detection, redirection and verification is formed, thereby ensuring that the final follow-up questions actually meet the avoidance conditions, preventing the follow-up question generation module from looping back to the prohibited area in another way, and ensuring the reliability of the interception effect. Attached Figure Description

[0014] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1This is a flowchart illustrating an AI conversation follow-up conflict interception and redirection method provided in one embodiment of this application; Figure 2 This is a schematic diagram of the structure of an AI conversation follow-up conflict interception and redirection device provided in another embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] To enable those skilled in the art to accurately understand the technical solution of this invention, key technical terms are first defined: Follow-up question related data: refers to the set of information related to a follow-up question that has been issued. It includes at least the follow-up question text itself, and may also include one or more of the following: the system's intention when generating the follow-up question, the user's answer summary to the follow-up question, the topic identifier to which the follow-up question belongs, the conversation round, and the role label.

[0017] Semantic fingerprint: refers to a digital identifier that represents the core semantics of a follow-up question in a vector space or other mathematical space, obtained by normalizing and semantically encoding the related data. Follow-up questions with similar semantics should have similar semantic fingerprints.

[0018] Historical semantic region: refers to a semantic set formed by clustering multiple semantic fingerprints with similar semantics in a semantic fingerprint database. In vector space, this region can be represented as a spherical region with a certain semantic radius around the region's central vector, used to abstractly represent a specific semantic direction or topic subspace that has been explored in depth during the conversation.

[0019] Prohibited semantic region marker: In this invention, it is a structured data object used to precisely describe one or more historical semantic regions that need to be avoided in the current follow-up question generation. It includes machine-readable constraint parameters such as the identifier of the prohibited region, representative follow-up question text, prohibition strength, and minimum semantic distance to be maintained.

[0020] Follow-up redirection guidance information: Control instructions generated based on prohibited semantic region markers and directly input to the follow-up question generation module. It transforms the prohibited markers into guidance conditions that the follow-up question generation module can understand and execute, such as: a list of semantic expressions that must be avoided, a minimum semantic distance that must be achieved, optional follow-up question directions, etc.

[0021] Follow-up question generation module: In this invention, it refers to the functional unit used to generate candidate follow-up questions. It receives contextual information and follow-up question redirection guidance information from the system input, and outputs one or more follow-up question texts in natural language form accordingly. Specific implementations of the follow-up question generation module include, but are not limited to, large language models, template-based follow-up question generators, retrieval-based follow-up question generators, or rule engines.

[0022] The following is combined with Figures 1 to 3 The technical solution of the present invention will be described in detail below.

[0023] This embodiment details a method for intercepting and redirecting AI conversation follow-up questions. For example... Figure 1 As shown, the method mainly includes the following steps: Step S101: Construct a session-level semantic fingerprint database.

[0024] This step aims to establish a continuously evolving semantic memory foundation for the entire session, addressing the lack of a global semantic perspective in existing technologies. Its core principle is as follows: whenever a follow-up question is actually issued (i.e., presented to the respondent), the system immediately captures the associated data of that question, transforms it into a point in a high-dimensional vector space—a semantic fingerprint—through a series of normalization and encoding operations, and persistently stores it according to session identifiers. This session-organized fingerprint set forms the data foundation for all subsequent semantic comparisons and conflict identification.

[0025] Specifically, semantic fingerprint generation is a multi-step digital process designed to create a unique, computable semantic identifier for each follow-up question. This process begins by normalizing multi-source information related to the follow-up question. This information includes the system's intent when generating the follow-up question, the text content of the follow-up question, a summary of the user's corresponding answer, and metadata (such as question identifier, dialogue turn, and speaker role). Normalization involves text cleaning, synonym normalization, role label concatenation, and sensitive information desensitization, aiming to eliminate surface-level discrepancies and transform the input information into a standard format.

[0026] Subsequently, the normalized composite information is fed into a semantic encoding function. The core task of this function is to map the text and its contextual semantics to a point in a high-dimensional vector space. The model implementing this function can be a pre-trained word vector model (such as Word2Vec, GloVe), a dedicated sentence encoding model (such as Sentence-BERT, SimCSE), or directly utilize the embedding layers of a large language model. Through semantic encoding, the semantic similarity between different follow-up questions is transformed into a distance relationship between points in the vector space; the more similar the semantics, the closer their corresponding vector positions.

[0027] The normalization and semantic encoding steps work together to ensure that follow-up questions, even if their literal expressions differ but their core semantics are similar, can ultimately be mapped to positions close to each other in the vector space. The vector generated through this process is the semantic fingerprint of that follow-up question. All semantic fingerprints of follow-up questions issued in the same session are collected to form the semantic fingerprint database for that session. This database not only records "what was asked," but more importantly, it records "the semantic meaning of the questions" in a structured way, thus laying the foundation for subsequent semantic conflict detection.

[0028] Step S102: Forming a historical semantic region.

[0029] This step aims to abstract and integrate discrete semantic fingerprints to form a more robust and representative semantic concept region, addressing the issues of noise interference and missed detections when comparing with single historical records. The technical principle is as follows: in the vector space of the semantic fingerprint, fingerprint points with adjacent locations are clustered, with each cluster forming a historical semantic region. This region is no longer an isolated point, but a geometric and semantic concept with a region center, semantic radius, and representative samples. This simulates the human cognitive process of classifying a series of similar questions into the same topic, enabling subsequent conflict detection to be compared with a robust semantic region, rather than a fragile single point.

[0030] Specifically, this step includes the following sub-steps: Step S1021: Calculate the center vector of the region. For each semantic fingerprint set formed by clustering algorithms (such as DBSCAN, hierarchical clustering, etc.), i.e. a historical semantic region, the system calculates the region center vector. The calculation of the region center vector is not a simple arithmetic mean, but a weighted average.

[0031] Specifically, the system extracts all semantic fingerprint vectors belonging to the historical semantic region. Before summing, a weight coefficient is assigned to each semantic fingerprint. The weight coefficient can be integrated into the business logic. For example, it can be determined based on the time of the follow-up question (newer follow-up questions have higher weights) or the level of detail of the corresponding answer (follow-up questions with richer content have higher weights). Subsequently, all semantic fingerprint vectors are multiplied by their corresponding weight coefficients, summed, and then divided by the sum of all weight coefficients to obtain an average vector. Finally, this average vector is normalized (e.g., using L2 normalization) to transform it into a unit vector of length 1. This final unit vector is defined as the region center vector of the historical semantic region. Mathematically, it represents the core position of the region in the vector space, reflects the overall average direction of the group of queried semantics, and is the centroid or prototype of the region's semantics.

[0032] Step S1022: Determine the semantic radius.

[0033] Defining a clear boundary for each historical semantic region is achieved by calculating a value called the semantic radius. This radius defines the extent of the region within the vector space.

[0034] When calculating the semantic radius, it is first necessary to measure the distance from each semantic fingerprint vector within the historical semantic region to the region's center vector. This distance can be calculated using metrics such as cosine distance or Euclidean distance, and the result reflects the degree of deviation of each fingerprint from the region's core.

[0035] Next, after obtaining the distance values ​​between all semantic fingerprints and the region center vector, the system selects a representative value as the final semantic radius. There are several selection strategies, such as directly taking the maximum value among all distances, which will ensure the region boundary just includes the outermost fingerprint; or taking a higher quantile, such as the 90th percentile. This means ignoring a few outliers that are particularly far away (e.g., 10%) to obtain a more stable and representative radius, avoiding unreasonable expansion of the region due to a few outlier fingerprints.

[0036] Then, the average of all distance values ​​is calculated, and a factor (e.g., one to two times) of the standard deviation is added. This method, based on the distribution characteristics of the data itself, can dynamically determine a reasonable range that covers most fingerprints.

[0037] The final determined semantic radius marks the spherical boundary with the region's center vector as its center and the radius as its length. The vector space within this boundary is defined as the range of the historical semantic region. Any new semantic fingerprint falling within this range will be considered as entering this previously queried semantic region.

[0038] Step S1023: Select representative follow-up text.

[0039] Identifying a representative follow-up text for each historical semantic region is a crucial step in making its semantic meaning more concrete and understandable. The system iterates through all the original follow-up texts that have actually been issued for that region and calculates the semantic vector of each text using a semantic encoding model. Subsequently, the system compares the similarity of these text semantic vectors with the region's center vector and selects the text with the highest similarity as the representative. This selected text, because its semantics most closely match the core direction of the region, becomes the most typical example for interpreting what was asked in that region.

[0040] Therefore, a complete historical semantic region can be summarized and represented by three core elements: the region center vector defining its semantic core, the semantic radius delineating its coverage, and representative follow-up texts serving as specific examples. All such historical semantic regions in a conversation together constitute a dynamic set, used to characterize the entire semantic landscape explored throughout the dialogue.

[0041] This semantic landscape is not static but continuously evolving. Whenever a new follow-up question is issued by the system, its corresponding semantic fingerprint is immediately added to the session's semantic fingerprint database. Next, the region management mechanism evaluates the distance between this new semantic fingerprint and existing historical semantic regions. If it is sufficiently close to an existing region, it will be assigned to that historical semantic region, potentially triggering minor adjustments to the center vector and semantic radius of that region; if it is far from all existing regions, it may serve as the starting point for forming a new, independent historical semantic region.

[0042] In this way, the system's understanding of the semantics of the questions asked can be automatically updated and improved as the dialogue progresses.

[0043] Step S103: Conflict detection and prohibition marker generation.

[0044] This step aims to accurately determine whether candidate follow-up questions are semantically redundant before they are issued, and to generate executable avoidance instructions to address the problems of coarse detection and weak constraints in existing technologies. The technical principle is as follows: candidate follow-up questions are also converted into semantic fingerprints, and then their comprehensive matching degree (region hit score) with each historical semantic region is calculated. This score integrates multiple dimensions of features, rather than simple cosine similarity. If the score exceeds a threshold, it is considered a hit, and a structured prohibition marker is generated based on the hit region and features, providing precise guidance for the next step of redirection.

[0045] Specifically, this step includes the following sub-steps: Step S1031: Obtain and encode candidate follow-up questions. When the follow-up question generation module generates a candidate follow-up question, the system obtains its related data, which is then normalized and semantically encoded to generate a candidate semantic fingerprint. Simultaneously, the system can parse or receive the candidate follow-up question intent associated with the candidate follow-up question.

[0046] Step S1032: Calculate the region hit score. This is the core of conflict detection in this embodiment. The system calculates the region hit score between the candidate semantic fingerprint and each historical semantic region. The scoring employs a multi-feature weighted model to comprehensively capture different types of semantic conflicts: The meanings of the parameters in the formula are as follows: Indicates the candidate semantic fingerprint and the first The semantic similarity of the region center vectors of a historical semantic region. It is usually calculated using cosine similarity and normalized to [0,1]. It reflects the degree of proximity between the candidate follow-up question and the core semantics of the region.

[0047] Indicates the candidate semantic fingerprint relative to the first The distance relationship between the semantic radii of historical semantic regions is used to quantify the degree to which a candidate semantic fingerprint falls within the boundary of a historical semantic region in the vector space. It is one of the core indicators for judging whether a candidate follow-up question has entered the queried semantic region. The calculation logic is based on a comparison between the actual distance from the candidate semantic fingerprint to the center vector of the target historical semantic region and the preset semantic radius of that region. Specifically, the system first calculates the actual distance from the candidate semantic fingerprint to the center vector of the target historical semantic region. If this distance is less than or equal to the semantic radius of the region, it is directly determined that the candidate fingerprint is completely within the region boundary. It is assigned a maximum value of 1. If this distance is greater than the semantic radius, the candidate fingerprint is determined to be outside the region boundary. In this case, The value doesn't drop directly to 0; instead, it decreases from 1 following a decay function as the distance from the candidate fingerprint to the region boundary increases. For example, an exponential decay model can be used, causing the score to decrease non-linearly with increasing distance beyond the boundary. The decay rate can be adjusted using a configurable decay coefficient. This design allows the score to make continuous and nuanced quantifications of suspected entry states near the boundary, rather than a binary judgment of either 0 or 1.

[0048] The text similarity between the candidate follow-up text and the representative follow-up text of the historical semantic region can be calculated using BLEU, ROUGE, or sentence encoding models. This provides evidence at the surface text level.

[0049] This is a paraphrasing feature score, specifically used to identify whether a candidate follow-up question is a repetition of an existing follow-up question in a specific historical semantic region, differing only in its form while retaining the core meaning. The aim is to penetrate the surface differences of the text and capture semantic repetitions that are essentially the same despite only a change in wording. In practice, this score can be calculated using a lightweight natural language processing model or a set of predefined semantic equivalence rules. The model or rules use representative follow-up texts from the current historical semantic region and other typical follow-up texts within that region as benchmarks to analyze the semantic equivalence between the candidate and the candidate follow-up texts. The focus is not on literal overlap, but on whether both express the same intent, inquire about the same fact, or request the same type of detail. The final score output is a quantified value. If determined to be a paraphrase, the system assigns a higher score (e.g., close to 1.0) to clearly identify this conflict. If determined to be a non-paraphrase, a lower score (e.g., close to 0) is output. This score serves as a key input, used in conjunction with other dimension scores to calculate the region hit score. If the synonym rewriting feature score is very high, even if the literal similarity is not high, it may lead to an overall conflict determination, thereby triggering the redirection mechanism.

[0050] This is a stage regression feature score. It is obtained by comparing the metadata of the current dialogue (such as the current round and stage label) with the metadata of the corresponding follow-up questions in the historical semantic regions. If the semantic content of the candidate follow-up question corresponds to an earlier dialogue stage that should have ended, a high score is output. For example, a fact-clarification question might be generated during the behavior review stage.

[0051] This is a role misalignment feature scoring. It determines whether the content of the candidate follow-up question is a restatement of what the AI ​​interviewer (e.g., an AI interviewer) said before (e.g., "As I just said...") or a confirmation of system functionality (e.g., "Do you understand?"), which is usually a meaningless repetition. This scoring is global and not specific to any particular historical semantic region.

[0052] , , , , , These are preset weight coefficients, and their sum equals 1. The weight configuration determines the system's sensitivity to various conflicts. For example, in a rigorous interview... (Synonymous rewriting) and The weight of (phase rollback) can be set relatively high.

[0053] Step S1033: Conflict determination and type identification. Each historical semantic region has a preset hit threshold. .if ≥ Then it is determined that the candidate question hit the first one. There are several historical semantic regions. Of course, the system may hit multiple historical semantic regions simultaneously. In this case, the system will not only output a Boolean hit result, but also identify the conflict type based on the sub-score item that contributes the most.

[0054] For example: Synonym rewriting conflict types, mainly caused by and High-scoring contributions trigger this.

[0055] Detailed reproduction of conflict types, by and The high score contribution is triggered and matches a specific detail area.

[0056] Phase rollback conflict types, by High score contribution triggers.

[0057] Cross-topic reproducibility of conflict types, by The high score contribution is triggered, but the current question ID is different from the question ID in the historical semantic region.

[0058] Role misalignment conflict type, by High score contribution triggers.

[0059] Step S1034: Generate prohibited semantic region markers. When a conflict detection candidate query hits a historical semantic region, the system does not simply provide a vague signal prohibiting repetition, but generates a structured prohibited semantic region marker. This marker is a machine-readable and parsable data object, whose core purpose is to transform the abstract instruction of prohibition into a series of specific and quantifiable generation constraint parameters.

[0060] Prohibited semantic region markers typically contain the following key information to collectively define a clear avoidance target: Forbidden area identifier: Clearly indicates which one or more historical semantic areas (identified by area ID) must not be entered again in this round of generation.

[0061] Representative text for the region: Provides the most representative follow-up text examples for the prohibited region, enabling the follow-up question generation module to intuitively understand the specific semantic content that needs to be avoided.

[0062] Conflict type and score: Record the specific reasons for the hit being judged (such as paraphrasing, detailed reproduction, etc.) and the quantified conflict severity score.

[0063] And, constraint strength parameters, which are the core of driving effective redirection, mainly include: Disallowance Strength: A dynamically adjusted value, typically positively correlated with the conflict score. The higher the value, the further the system requires newly generated follow-up questions to be semantically distant from the prohibited region.

[0064] Minimum semantic distance: A hard threshold that mandates that newly generated candidate fingerprints must be at least as far as the prototype vector of the forbidden region. This ensures discrepancies in the vector space.

[0065] Scope and rules: This flag is limited to a specific scope (such as the current interview question or the current stage of the conversation) and specifies that in subsequent review stages, the pass criterion must be to avoid hitting these prohibited areas.

[0066] By integrating all the above elements, prohibiting semantic region tagging deconstructs a complex semantic avoidance problem into a set of explicit, computable conditions, thereby providing the follow-up question generation module with clear, stable, and powerful redirection guidance, rather than relying on simple natural language prompts that are prone to failure.

[0067] Step S104: Generate follow-up question redirection guidance information and redirect.

[0068] This step aims to transform prohibited semantic region markers into prompting signals that the prompting generation model can directly utilize, driving it to generate new, non-repeating prompts, thus addressing the problem of insufficient constraint from natural language prompts alone. The technical principle is to translate structured prohibited semantic region markers into input constraints for the prompting generation model. These input constraints are more specific and measurable than vague input instructions like "don't repeat."

[0069] Specifically, this step includes the following sub-steps: Step S1041: Convert the boot information. The system will convert prohibited semantic region markers into follow-up redirection guidance information. This follow-up redirection guidance information includes the following: List of expressions to avoid: This is a list of key phrases extracted from representative follow-up texts and related follow-up texts of all prohibited areas in the prohibited semantic region markers. Examples include: "communication actions," "how to communicate," and "what specific measures to take."

[0070] Minimum semantic distance requirement: Take the maximum value of all minimum semantic distances among the prohibited semantic region tags as a global hard constraint, requiring that the semantic distance between the regenerated candidate follow-up question and the prohibited historical semantic region must not be less than this value.

[0071] Redirection seeds: Extracting potential directions or keywords for follow-up questions from the current dialogue context that have not yet been covered by any historical semantic regions. For example, if the historical semantic region focuses on communication, but the current answer mentions decision-making, then redirection seeds could be: decision-making basis, evaluation criteria, etc., to provide positive guidance to the follow-up question generation model on "which direction to take in the follow-up questions".

[0072] Maximum number of retries: The maximum number of times the follow-up question generation module is allowed to attempt to generate new candidate follow-up questions based on the current guidance information.

[0073] Through this step, the follow-up question redirection guidance information provides precise constraints and guidance to the follow-up question generation module at both structural and semantic levels. On the one hand, negative constraints are formed using the list of expressions to be avoided and the minimum semantic distance requirement, ensuring that newly generated candidate follow-up questions do not fall into the semantic region that has already been asked. On the other hand, positive guidance is formed using redirection prompts, helping the follow-up question generation module quickly locate effective topic directions that have not yet been covered. At the same time, the maximum retry mechanism avoids meaningless loop generation. Overall, the follow-up question redirection guidance information quantifies the originally vague instructions to avoid repeated generation of follow-up questions into machine-executable constraint parameters, significantly improving the success rate and efficiency of follow-up question redirection generation.

[0074] Step S1042, drive redirection generation. The complete prompt, including the follow-up question redirection guidance information, is input into the follow-up question generation module. The complete prompt includes system instructions, the current dialogue history, and the follow-up question redirection guidance information. The follow-up question generation module (typically a large language model) understands and executes the following structured constraints: First, you must avoid expressions from the list of expressions to be avoided; Second, the semantic distance between the newly generated candidate follow-up semantic vector and the prohibited historical semantic region must be greater than the preset minimum semantic distance requirement; Third, you can ask questions based on the direction indicated by the turning prompts.

[0075] Based on the above constraints, the follow-up question generation module will generate a new candidate follow-up question.

[0076] Step S105: Follow-up review, issuance and iteration.

[0077] This step aims to verify the redirection results and complete the loop, addressing the issue that a single change to a follow-up question might lead to repeated conflicts, thus ensuring the final effect. The technical principle is as follows: for new candidate follow-up questions generated by the redirection, the conflict detection (i.e., re-check) in step S103 is re-executed. This is a feedback loop to ensure that the output satisfies the constraints. Simultaneously, successful follow-up questions are fingerprinted and archived, enabling the system to self-update.

[0078] Specifically, this step includes the following sub-steps: Step S1051: Re-inspection. For the new candidate follow-up questions generated in step S1042, steps S1031 to S1033 are re-executed to calculate a new region hit score relative to all regions listed in the prohibited semantic region markers.

[0079] Step S1052: Judgment and distribution. The system re-evaluates the hit rate of the regenerated candidate follow-up questions. If the hit rate of the candidate follow-up question for all prohibited historical semantic regions is lower than the preset re-examination threshold (which is usually stricter than the initial hit threshold to allow for a safety margin), the re-examination is deemed successful. At this point, the system allows the candidate follow-up question to be formally issued, for example, by speech synthesis or by displaying it to the respondent in text form.

[0080] After successful delivery, the system immediately processes the related data of the query according to the semantic fingerprint generation rules, generates a new semantic fingerprint, and adds it to the session-level semantic fingerprint database. Subsequently, the historical semantic region management module can synchronously or asynchronously determine whether to assign the new fingerprint to an existing historical semantic region or to create a new historical semantic region based on the distance between the new fingerprint and the existing historical semantic regions.

[0081] Step S1053: Iterative retry. If the re-examination fails—that is, if the regenerated candidate follow-up questions still hit at least one prohibited historical semantic region—the system updates the currently used prohibited semantic region markers. Updates may include increasing the prohibition strength, increasing the minimum semantic distance requirement, or changing the redirection prompt information. Subsequently, the system, carrying the updated prohibited semantic region markers and their transformed follow-up question redirection guidance information, returns to the redirection generation step to drive the follow-up question generation module to regenerate candidate follow-up questions. The system accumulates the number of attempts in this redirection loop.

[0082] Step S1054: Failure handling. If the number of redirection attempts reaches the preset maximum retry count and the latest generated candidate follow-up question still fails the review, the system terminates the local redirection loop and outputs a status message indicating that the follow-up question redirection failed. This message includes the hit region, conflict type, and the last candidate follow-up question attempted. This status message is then submitted to a higher-level session orchestration or workflow management service, which determines subsequent actions, such as skipping the follow-up question, switching topics, or transferring the request to a human operator.

[0083] The following example uses an AI interview scenario to demonstrate the specific application process of the method of the present invention.

[0084] Scenario 1: In an AI behavioral interview, the system is asking candidates questions about their experience in "handling team conflicts".

[0085] Step S1: Session initialization and initial follow-up question.

[0086] Interview session S001 begins. Regarding the previous experience, the system has already issued a follow-up question: "What specific communication actions were taken at that time?", which the candidate has answered. The system converts this follow-up question and its answer summary into a semantic fingerprint and stores it in the semantic fingerprint database for this session. At this point, the database contains relatively few fingerprints, and no clear historical semantic region clusters have yet formed.

[0087] Step S2: New candidate follow-up questions trigger conflict detection.

[0088] The interview begins with questions about the candidate's current experience. After the candidate answers, the follow-up question generation module first generates a candidate follow-up question: "How did you communicate with the other party?"

[0089] The system semantically encodes the candidate follow-up question and compares it with existing historical semantic regions. The results show that the candidate follow-up question is highly semantically similar to the previously issued "communication action" follow-up question (similarity score reaches 0.95) and is identified as a paraphrase (paraphrase feature score is 0.88). After considering all scores, the region hit score is 0.91, exceeding the preset hit threshold of 0.85. Therefore, the system determines that the candidate follow-up question hits a historical semantic region, and the conflict type is a paraphrase conflict.

[0090] Step S3: Generate prohibited semantic region markers.

[0091] The system then generates a prohibited semantic region marker, which contains information about prohibited historical semantic regions (pointing to the relevant historical semantic regions of "communication actions"), and sets the prohibition strength to 0.9 and the minimum semantic distance requirement to 0.2.

[0092] Step S4: Redirect generation.

[0093] The system converts prohibited semantic region markers into follow-up question redirection guidance information, which includes a list of expressions to avoid (such as "communication actions," "how to communicate," etc.) and redirection prompts extracted from candidate answers (such as "seek third-party opinions"). This guidance information is then input into the follow-up question generation module (LLM).

[0094] Step S5: Initial re-inspection and pass.

[0095] The follow-up question generation module generates a new candidate follow-up question based on the guidance information: "Did you consider or try to introduce a middleman to coordinate at that time?" The system performs a hit judgment on this new follow-up question, and calculates that its hit score with the previously prohibited area is only 0.3, which is far below the threshold of 0.85, so the re-examination passes.

[0096] Step S6: Issue and update the fingerprint database.

[0097] The system allows this new follow-up question to be formally issued via voice broadcast. Once issued, a semantic fingerprint is immediately generated for the follow-up question and added to the session-level semantic fingerprint database. Subsequently, the historical semantic region management module may categorize this fingerprint into a new historical semantic region related to "seeking external assistance," or temporarily store it independently.

[0098] Scenario 2: Reproducing conflicts across different topics.

[0099] The interview moved to the next question, this time on the theme of "driving project innovation." After the candidate answered, the follow-up question generation module initially generated the following candidate question: "In this project, how did you and your team members communicate and collaborate to advance new ideas?" The system's calculations revealed that the semantic fingerprint of this follow-up question was again highly similar to the previous "communication actions" area (hit score 0.89). Since the semantic fingerprint database and historical semantic areas are maintained across questions and conversation stages, the system successfully identified this as a "cross-question recurrence conflict"—although the question topic changed, the core semantics of the follow-up question were repeated. The system then generated new prohibited semantic area markers, guiding the follow-up question generation module to avoid the "communication actions" area and instead focus on new directions such as "risk assessment" or "resource acquisition."

[0100] Scenario 3: Re-inspection failed.

[0101] In another complex scenario, the candidate's answer range is very narrow. After multiple redirects, the newly generated follow-up questions still have a slightly higher hit score in the historical semantic region than the threshold (e.g., 0.86, threshold is 0.85). In this case, the system updates the prohibited semantic region flag (e.g., increasing the minimum semantic distance requirement from 0.2 to 0.3) and drives redirection again. If the system fails to pass the re-examination after reaching the preset maximum number of retries (e.g., 3 times), it outputs a "follow-up question redirection failed" status and hands it over to the upper-layer service for processing. The upper-layer service can decide based on this status not to follow up on this topic and directly move on to the next topic.

[0102] As can be seen from this embodiment, the method of the present invention can automatically intercept various forms of semantic repetition, such as paraphrasing and cross-question repetition, throughout the entire interview session. Through structured guidance and closed-loop verification, it can effectively drive the generation of subsequent questions with entirely new semantics, significantly improving the efficiency and quality of AI interviews.

[0103] See Figure 2 Another embodiment of the present invention provides an AI conversation follow-up conflict interception and redirection device 200, including a semantic fingerprint database construction module 201, a historical semantic region management module 202, a conflict detection module 203, a redirection guidance module 204, and a review and distribution module 205. This AI conversation follow-up conflict interception and redirection device 200 is capable of executing the AI ​​conversation follow-up conflict interception and redirection method in the method embodiment.

[0104] Specifically, the AI ​​conversational questioning conflict interception and redirection device 200 includes: The semantic fingerprint database construction module 201 is used to map the follow-up question association data corresponding to the issued follow-up questions to semantic fingerprints, and to build a semantic fingerprint database for storing the semantics of the follow-up questions. The historical semantic region management module 202 is used to cluster the semantic fingerprints in the semantic fingerprint database to form multiple historical semantic regions, each of which is used to represent a semantic direction that has been queried. The conflict detection module 203 is used to obtain candidate follow-up questions in the current session, map the candidate follow-up questions to candidate semantic fingerprints, determine whether the candidate semantic fingerprints hit any of the historical semantic regions, and if they hit, generate a prohibited semantic region marker. The prohibited semantic region marker is used to structurally describe the hit historical semantic region and provide constraint parameters for redirecting candidate follow-up questions. The redirection guidance module 204 is used to convert the prohibited semantic region marker into follow-up redirection guidance information, and input the follow-up redirection guidance information into the follow-up generation module to generate candidate follow-up questions that avoid the hit historical semantic region; The re-examination and distribution module 205 is used to re-determine the hit rate of the regenerated candidate follow-up questions. If no historical semantic region is hit, the regenerated candidate follow-up question is allowed to be distributed, and the candidate semantic fingerprint of the regenerated candidate follow-up question is written into the semantic fingerprint database. If at least one historical semantic region is hit, the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven again to regenerate the candidate follow-up question.

[0105] It should be noted that the AI ​​conversation follow-up conflict interception and redirection device 200 provided in this embodiment, which can be used to execute various method embodiments, has a similar implementation principle and technical effect to the method, and will not be described again here.

[0106] Figure 3 This is a schematic diagram of an electronic device 300 provided in another embodiment of the present invention. The electronic device 300 is used to implement the AI ​​conversation follow-up conflict interception and redirection method in the method embodiment. The electronic device 300 in the embodiments of the present invention may include, but is not limited to, smartphones, tablets, PCs, laptops, servers, etc. Figure 3 The electronic device 300 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0107] like Figure 3 As shown, the electronic device 300 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes to implement the methods of the embodiments described herein, based on a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing device 301, ROM 302, and RAM 303 are interconnected via a bus 305. An input / output (I / O) interface 304 is also connected to the bus 305.

[0108] Typically, the following devices can be connected to I / O interface 304: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0109] The above description is merely a preferred embodiment of the present invention. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to the specific combination of the above-described technical features, but should also cover other technical solutions formed by any combination of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for intercepting and redirecting AI conversational follow-up questions, characterized in that, Includes the following steps: Map the follow-up questions that have been issued to semantic fingerprints, and build a semantic fingerprint library to store the semantics of the follow-up questions. The semantic fingerprints in the semantic fingerprint database are clustered to form multiple historical semantic regions, each of which is used to represent a semantic direction that has been queried. Retrieve candidate follow-up questions in the current session, map the candidate follow-up questions to candidate semantic fingerprints, and determine whether the candidate semantic fingerprints hit any of the historical semantic regions. If they do, generate a prohibited semantic region marker. The prohibited semantic region marker is used to structurally describe the hit historical semantic region and provide constraint parameters for redirecting candidate follow-up questions. The prohibited semantic region marker includes a prohibited region identifier, region representative text, conflict type and score, prohibition strength, minimum semantic distance, and scope and rules. The prohibited semantic region markers are converted into follow-up question redirection guidance information, and the follow-up question redirection guidance information is input into the follow-up question generation module to generate candidate follow-up questions that avoid the hit historical semantic regions; The follow-up question redirection guidance information includes: representative follow-up question texts of historical semantic regions that need to be avoided, the minimum semantic distance that the newly generated candidate follow-up question needs to satisfy between the hit historical semantic region, the available question directions that have not hit any historical semantic regions extracted from the current session context, and the maximum number of retries that the follow-up question generation module is allowed to try to generate redirection candidate follow-up questions. The newly generated candidate follow-up questions are re-evaluated for hit detection. If no historical semantic region is hit, the newly generated candidate follow-up question is allowed to be issued, and the candidate semantic fingerprint of the newly generated candidate follow-up question is written into the semantic fingerprint database. If at least one historical semantic region is hit, the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven to regenerate candidate follow-up questions again.

2. The AI ​​conversation follow-up question conflict interception and redirection method according to claim 1, characterized in that, The step of forming multiple historical semantic regions further includes: The weighted average of all semantic fingerprints belonging to each historical semantic region is used as the region center vector of that historical semantic region. The maximum distance from each semantic fingerprint within each historical semantic region to the center vector of the region is taken as the semantic radius of that historical semantic region. The follow-up text that is closest to the center vector of each historical semantic region is taken as the representative follow-up text of that historical semantic region.

3. The AI ​​conversation follow-up question conflict interception and redirection method according to claim 2, characterized in that, The step of determining whether the candidate semantic fingerprint matches any of the historical semantic regions includes: The region hit score for each historical semantic region corresponding to the candidate semantic fingerprint is calculated using the following formula. ： in, Indicates the first Regional hit score for each historical semantic region; Indicates the candidate semantic fingerprint and the first The similarity of the region center vectors of each historical semantic region; Indicates the candidate semantic fingerprint relative to the first Distance relationship of semantic radii of historical semantic regions; Indicates follow-up questions to the candidate and the first The similarity of representative inquiry texts in each historical semantic region; The synonym rewriting feature score is used to characterize whether the candidate follow-up question is the first one. There are already synonymous rewritings for questioning within a historical semantic region; The stage rollback feature score represents whether the candidate follow-up question requires the respondent to roll back to the previous stage. The dialogue stages corresponding to each historical semantic region are explained; The role misalignment feature score is used to characterize whether the content of the candidate follow-up question is a restatement of the AI ​​interviewer's own words or a confirmation of the system's function. , , , , , These are preset weighting coefficients; If the region hit score is not lower than the preset hit threshold, then the candidate semantic fingerprint is determined to have hit the historical semantic region corresponding to the region hit score.

4. The AI ​​conversation follow-up question conflict interception and redirection method according to claim 3, characterized in that, Also includes: If the region hit score of all historical semantic regions corresponding to the candidate semantic fingerprint of the regenerated candidate follow-up question is lower than the preset hit threshold, then the candidate follow-up question is allowed to be issued and the candidate semantic fingerprint is written into the semantic fingerprint database. If the region hit score of the candidate semantic fingerprint corresponding to at least one historical semantic region of the regenerated candidate follow-up question is not lower than the preset hit threshold, then the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven again to regenerate candidate follow-up questions. If the number of attempts to regenerate candidate follow-up questions reaches the preset maximum number of retries, and the region hit score of the candidate semantic fingerprint of the latest generated candidate follow-up question corresponding to at least one historical semantic region is not lower than the preset hit threshold, then output the status information of follow-up question redirection failure.

5. An AI conversational questioning conflict interception and redirection device, characterized in that, include: The semantic fingerprint database construction module is used to map the follow-up question association data corresponding to the issued follow-up questions to semantic fingerprints, and to build a semantic fingerprint database for storing the semantics of the follow-up questions. The historical semantic region management module is used to cluster the semantic fingerprints in the semantic fingerprint database to form multiple historical semantic regions, each of which is used to represent a semantic direction that has been queried. The conflict detection module is used to obtain candidate follow-up questions in the current session, map the candidate follow-up questions to candidate semantic fingerprints, and determine whether the candidate semantic fingerprints hit any of the historical semantic regions. If they hit, a prohibited semantic region marker is generated. The prohibited semantic region marker is used to structurally describe the hit historical semantic region and provide constraint parameters for redirecting candidate follow-up questions. The prohibited semantic region marker includes a prohibited region identifier, region representative text, conflict type and score, prohibition strength, minimum semantic distance, and scope and rules. The redirection guidance module is used to convert the prohibited semantic region marker into follow-up redirection guidance information, and input the follow-up redirection guidance information into the follow-up question generation module to generate candidate follow-up questions that avoid the hit historical semantic region; The follow-up question redirection guidance information includes: representative follow-up question texts of historical semantic regions that need to be avoided, the minimum semantic distance that the newly generated candidate follow-up question needs to satisfy between the hit historical semantic region, the available question directions that have not hit any historical semantic regions extracted from the current session context, and the maximum number of retries that the follow-up question generation module is allowed to try to generate redirection candidate follow-up questions. The re-examination and distribution module is used to re-determine the hit rate of the regenerated candidate follow-up questions. If no historical semantic region is hit, the regenerated candidate follow-up question is allowed to be distributed, and the candidate semantic fingerprint of the regenerated candidate follow-up question is written into the semantic fingerprint database. If at least one historical semantic region is hit, the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven again to regenerate the candidate follow-up question.

6. The AI ​​conversation follow-up conflict interception and redirection device according to claim 5, characterized in that, The historical semantic region management module is further used for: The weighted average of all semantic fingerprints belonging to each historical semantic region is used as the region center vector of that historical semantic region. The maximum distance from each semantic fingerprint within each historical semantic region to the center vector of the region is taken as the semantic radius of that historical semantic region. The follow-up text that is closest to the center vector of each historical semantic region is taken as the representative follow-up text of that historical semantic region.

7. The AI ​​conversation follow-up conflict interception and redirection device according to claim 6, characterized in that, The collision detection module is further used for: The region hit score corresponding to each historical semantic region for the candidate semantic fingerprint is calculated using the following formula. ： in, Indicates the first Regional hit score for each historical semantic region; Indicates the candidate semantic fingerprint and the first The similarity of the region center vectors of each historical semantic region; Indicates the candidate semantic fingerprint relative to the first Distance relationship of semantic radii of historical semantic regions; Indicates follow-up questions to the candidate and the first The similarity of representative inquiry texts in each historical semantic region; The synonym rewriting feature score is used to characterize whether the candidate follow-up question is the first one. There are already synonymous rewritings for questioning within a historical semantic region; The stage rollback feature score represents whether the candidate follow-up question requires the respondent to roll back to the previous stage. The dialogue stages corresponding to each historical semantic region are explained; The role misalignment feature score is used to characterize whether the content of the candidate follow-up question is a restatement of the AI ​​interviewer's own words or a confirmation of the system's function. , , , , , These are preset weighting coefficients; If the region hit score is not lower than the preset hit threshold, then the candidate semantic fingerprint is determined to have hit the historical semantic region corresponding to the region hit score.

8. The AI ​​conversation follow-up conflict interception and redirection device according to claim 7, characterized in that, The re-inspection and distribution module is further used for: If the region hit score of all historical semantic regions corresponding to the candidate semantic fingerprint of the regenerated candidate follow-up question is lower than the preset hit threshold, then the candidate follow-up question is allowed to be issued and the candidate semantic fingerprint is written into the semantic fingerprint database. If the region hit score of the candidate semantic fingerprint corresponding to at least one historical semantic region of the regenerated candidate follow-up question is not lower than the preset hit threshold, then the constraint parameters in the prohibited semantic region marker are updated, and the follow-up question generation module is driven again to regenerate candidate follow-up questions. If the number of attempts to regenerate candidate follow-up questions reaches the preset maximum number of retries, and the region hit score of the candidate semantic fingerprint of the latest generated candidate follow-up question corresponding to at least one historical semantic region is not lower than the preset hit threshold, then output the status information of follow-up question redirection failure.

Images