Social interaction adaptive question generating and hierarchical scheduling method and system based on AI

Abstract

The application discloses a social interaction adaptive question generating method and system based on AI generation and hierarchical scheduling, and belongs to the technical field of social interaction content generation and scheduling. The method comprises the following steps: generating candidate topics according to a theme dimension and evaluating ice-breaking adaptation, and marking core topics and non-core topics according to scores in a hierarchical manner; determining round rules according to a question number in a conversation; collecting user interaction behavior signals in real time and determining an interaction state; covering the round rule determination result according to the interaction state, switching to a core pool when being tired, switching to a non-core pool when being active, and obtaining a final target pool; and selecting topics from the target pool and pushing the topics. The system can dynamically adjust a question generating strategy based on real-time user behavior feedback, realize interaction fatigue intervention and adaptive control of a question rhythm.

Description

Technical Field

[0001] This invention belongs to the field of social interactive content generation and scheduling technology, and in particular relates to a social interactive adaptive question generation method and system based on AI generation and hierarchical scheduling. Background Technology

[0002] In real-time social interaction scenarios, especially during two-person voice or video calls, fun quiz games are a common way to break the ice and promote communication. In this scenario, two users take turns playing rock-paper-scissors, revealing topics, exchanging answers, and moving on to the next question, completing multiple rounds of interaction. The quality of the topics and the pace of the questions directly affect the interactive experience.

[0003] In existing technologies, solutions related to question setting mainly focus on the following areas: I. Item Selection Methods in Computerized Adaptive Testing (CAT). This type of approach is driven by the ability parameter estimation based on Item Response Theory (IRT). It uses a weighted bias model to achieve optimal ability estimation accuracy while considering content constraints, item overlap constraints, and item group constraints, and selects items heuristically step by step. Alternatively, it uses computer simulation to generate an exposure control parameter set to control the probability of each item in the item bank being administered. This approach requires pre-calibrating item parameters and relies on extensive historical data for statistical exposure control. It is suitable for standardized tests in educational assessment but not for lightweight social scenarios with dynamically generated AI content. It does not support dynamic scheduling based on real-time user interaction behavior, nor does it differentiate between the needs of social icebreakers and in-depth interactions.

[0004] II. AI-Assisted Intelligent Test Paper Generation and Question Conversion Solutions. These solutions propose grouping questions into hierarchical question pools based on content domains, automatically selecting a subset of questions based on domain coverage constraints, and automatically generating test papers for professional qualification certification exams. These solutions are static test paper generation, meaning they generate a complete test paper at once, do not support real-time selection of questions individually, do not involve quality grading of AI-generated content, and lack a dynamic scheduling mechanism based on user behavior.

[0005] III. Question Generation Methods in Game Scenarios. This type of solution dynamically adjusts the difficulty of questions based on user performance, determining the next question by matching the difficulty level with the user's ability. It is applied to single-player attention training games. This type of solution is driven by an individual ability model, does not involve AI content generation, is not suitable for two-player social interaction scenarios, lacks topic diversity, and does not have conversation-level deduplication and multi-level degradation mechanisms. It does not distinguish between icebreaker questions and in-depth questions.

[0006] In summary, the existing technology has the following shortcomings: First, the question bank content is all manually compiled or pre-fixed, lacking the ability of AI large-scale models to generate topics by theme dimension and automatically assess icebreaking suitability for classification; Second, the question generation strategy is based on preset fixed rules or pure statistical models, unable to dynamically adjust the question generation strategy according to real-time behavioral signals of users during interaction, and unable to detect user interaction fatigue and intervene in a timely manner; Third, the selection of questions only considers ability matching or randomness, without controlling the diversity of theme classification for consecutive questions, which may lead to the continuous appearance of the same theme, affecting the experience; Fourth, it usually terminates or reports an error when the question bank subset is exhausted, lacking conversation-level deduplication and multi-level degradation mechanisms; Fifth, content generation, content scheduling, and effect perception are completely separated, lacking an end-to-end closed loop from AI generation to behavior perception scheduling.

[0007] Based on the above, there is a need for an AI-generated and hierarchical scheduling-based adaptive question generation method and system for social interaction that can dynamically adjust the question generation strategy according to the real-time behavioral signals of users during the interaction process, sense user interaction fatigue and intervene in a timely manner, and maintain the current rhythm without interference when the interaction is in a good state. Summary of the Invention

[0008] To address the shortcomings of existing question-generating schemes that rely on preset fixed rules or purely statistical models, failing to dynamically adjust question-generating strategies based on real-time user behavior signals during interaction, lack of awareness of user fatigue and timely intervention, and inability to maintain the current pace without interference when interaction is in good condition, this invention aims to provide a social interaction adaptive question-generating method and system based on AI content generation and hierarchical scheduling. Through a two-factor adaptive scheduling approach using round rule determination and real-time behavior status determination, it automatically switches to the core question bank for high-quality topic intervention when user interaction fatigue is detected, and maintains the non-core question bank question generation without interrupting the good pace when high user interaction activity is detected. This achieves an automated pipeline from AI generation to automatic hierarchical labeling. When user interaction fatigue is detected, core topic intervention is automatically triggered; when high interaction activity is detected, the current question-generating pace is maintained without interference. This ensures that the same topic category is not repeated within several consecutive questions, enhancing topic freshness and increasing the probability of selecting topics with higher icebreaker relevance. It achieves a quality gradient within the core pool, ensuring high availability with a continuous stream of questions. Different question-generating strategy parameters are applied to users engaging in first-time and repeat interactions.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides a social interaction adaptive question generation method based on AI generation and hierarchical scheduling, comprising the following steps: Candidate topics are generated in batches according to predefined social interaction theme dimensions. The icebreaking suitability of the candidate topics is evaluated to obtain an icebreaking suitability score for each topic. Based on the comparison result of the icebreaking suitability score and the stratification threshold, the topics are marked as core topics or non-core topics and stored in the core question bank and non-core question bank respectively. In a social interaction conversation, the current question number is obtained, and the round rule is determined based on the question number to determine whether the topic should be selected from the core question bank or the non-core question bank in this round. Real-time acquisition of interactive behavior signals from participating users; determination of the current interactive state based on the interactive behavior signals; the interactive state includes fatigue state, active state, and normal state. The final target pool decision is obtained by overwriting the results of the round rule determination based on the interaction state. Specifically, when the interaction state is fatigued, the final target pool is covered by the core question bank. When the interaction state is active and the round rule determination result is the core question bank, the final target pool is covered by the non-core question bank. When the interaction state is normal, the result of the round rule determination remains unchanged. Topics are selected from the question bank corresponding to the final target pool decision, and the selected topics are pushed to users who participate in the interaction.

[0010] As a further aspect of the present invention, the icebreaking suitability assessment of the candidate topics includes: Each candidate topic is scored from multiple dimensions, including topic openness, safety, interest, and information exchange value, to obtain a score value for each dimension. The score value for each dimension is obtained by a large language model performing a structured score on the topic according to preset scoring prompts, or by a lightweight classifier outputting according to preset judgment rules. The score values ​​of each dimension are weighted and summed according to the corresponding dimension weights to obtain the icebreaking suitability score, where the sum of the weights of each dimension is 1.

[0011] As a further aspect of the present invention, the method also includes online feedback correction of the ice-breaking adaptability score: Collect performance data for each topic in actual interactions. The performance data includes the historical average dwell time for the topic, the number of rounds of continued interaction between the two parties after the topic appears, and the proportion of topics that are quickly skipped. The effectiveness return of the topic is calculated based on the effect data, and the icebreaking suitability score is corrected online using an exponential moving average method. The corrected score is written back and triggers the dynamic re-stratification of the topic between the core question bank and the non-core question bank.

[0012] As a further aspect of the present invention, the stratification threshold is a dynamic threshold that is adaptively adjusted according to the deviation between the actual proportion of the core question bank and the target proportion: when the actual proportion of the core question bank is higher than the target proportion, the stratification threshold is increased; when the actual proportion of the core question bank is lower than the target proportion, the stratification threshold is decreased.

[0013] As a further aspect of the present invention, determining the round rule based on the question number includes: When the question number is not greater than the first preset value, it is determined that the question should be selected from the core question bank in this round; When the question number is not less than the second preset value and the difference between the question number and the second preset value is an integer multiple of the third preset value, it is determined that the question should be selected from the core question bank in this round. In other cases, the question bank should be selected from non-core questions in this round; wherein the second preset value is greater than the first preset value.

[0014] As a further aspect of the present invention, determining the current interaction state based on the interaction behavior signal includes at least one of the following methods: The first method involves collecting the single-round dwell time for each round of interaction, comparing the single-round dwell time with a fast skip threshold, and recording a dwell time lower than the fast skip threshold as a fast skip. The number of consecutive fast skips is counted, and when the number of consecutive fast skips reaches a preset number, it is determined to be a fatigue state. When the single-round dwell time is higher than the activity determination threshold, it is determined to be an active state. After triggering the fatigue state overwrite, the number of consecutive fast skips is reset to zero. The second method involves collecting multi-channel behavioral signals for each round of interaction. These signals include single-round dwell time, text reply length, voice speech duration, and emotional polarity score. The emotional polarity score is obtained by performing voice emotion recognition on the user's voice content in this round of interaction. After normalizing each behavioral signal, the signals are fused according to their corresponding weights to obtain the single-round interaction engagement. Within a preset review window, the interaction engagement of each round is weighted by time decay, with signals closer to the current round having a larger weight. Based on this, fatigue level and activity level are calculated. When the fatigue level reaches the fatigue threshold, it is determined to be a fatigued state. When the activity level reaches the activity threshold and is not determined to be a fatigued state, it is determined to be an active state. Other situations are determined to be normal states.

[0015] As a further aspect of the present invention, the social interaction is a two-person interaction, and the behavior sequence and interaction engagement of the two users are maintained separately for each relationship pair, and the fatigue level and activity level of each user are calculated separately; the fatigue level of the relationship pair is taken as the larger value of the fatigue levels of both users, and the relationship pair is determined to be in a fatigue state when either user reaches a fatigue state; the activity level of the relationship pair is taken as the smaller value of the activity levels of both users, and the relationship pair is determined to be in an active state only when both users reach an active state.

[0016] As a further aspect of the present invention, selecting topics from the question bank corresponding to the final target pool decision includes: maintaining a topic sliding window with a preset window value and recording the topic categories of the most recent questions; excluding topics whose topic categories are the same as those already present in the topic sliding window from the available topics in the target question bank, thereby obtaining a topic-filtered candidate set; and clearing the topic sliding window and canceling topic constraints when the topic-filtered candidate set is empty.

[0017] As a further aspect of the present invention, the selection of topics from the question bank corresponding to the final target pool decision includes: when the final target pool is a core question bank, a weighted random selection is performed based on the icebreaking suitability score of each topic, and the topic with the higher the icebreaking suitability score is more likely to be selected; when the final target pool is a non-core question bank, a uniform random selection is performed.

[0018] As a further aspect of the present invention, weighted random selection based on the icebreaking suitability score of each topic includes: calculating a selection priority score for each candidate topic, wherein the selection priority score is equal to the icebreaking suitability score of the topic plus an exploration bonus, wherein the exploration bonus is negatively correlated with the number of times the topic has been selected in the past; and performing softmax temperature sampling based on the selection priority score to determine the final selected topic.

[0019] As a further aspect of the present invention, the method also includes session-level deduplication and multi-level degradation: Maintain a set of used topics for the current session, and filter used topics from the candidate set; When the available set in the target question bank is empty after deducting used topics, it will be downgraded to selecting from the entire question bank. When the available set of the entire question bank is empty after deducting the used topics, the set of used topics is cleared and a new question generation cycle begins.

[0020] As a further aspect of the present invention, the method also includes cold start and hot start parameter adaptation: Query the historical call records of both users involved in the interaction; When both parties are interacting for the first time, a cold start parameter set is used to increase the number of questions that are forcibly selected from the core question bank for the first few questions. When the interaction between the two parties is repetitive, a hot start parameter set is used to reduce the number of questions that are forcibly selected from the core question bank for the first few questions.

[0021] As a further aspect of the present invention, the method also includes personalized adaptation based on user profiles: Obtain the user profile tags and interest tags of both parties involved in the interaction; Based on the topic dimensions corresponding to the shared interest tags of both parties, the selection weight of topics in the icebreaking stage is biased, so that the topic dimensions corresponding to the shared interest tags are given priority in the questions.

[0022] Secondly, this invention provides an AI-generated and hierarchically scheduled social interaction adaptive question-generating system, comprising: The AI ​​question bank generation and hierarchical labeling module is used to generate candidate topics in batches according to predefined social interaction theme dimensions, evaluate the icebreaking suitability of candidate topics to obtain icebreaking suitability scores, and label topics as core topics or non-core topics according to the comparison results of the icebreaking suitability scores and hierarchical thresholds, and store them in the core question bank and non-core question bank respectively. The round rule determination module is used to obtain the current question number in the social interaction conversation, and determine whether the topic should be selected from the core question bank or the non-core question bank in this round based on the question number. The behavior signal acquisition and state determination module is used to acquire the interactive behavior signals of users participating in the interaction in real time, and determine the current interaction state based on the interactive behavior signals. The interaction state includes fatigue state, active state and normal state. The behavior signal overlay module is used to overlay the judgment result of the round rule judgment module according to the interaction state to obtain the final target pool decision. When the interaction state is fatigued, the final target pool is covered as the core question bank. When the interaction state is active and the round rule judgment result is the core question bank, the final target pool is covered as the non-core question bank. When the interaction state is normal, the round rule judgment result remains unchanged. The topic selection and push module is used to select topics from the question bank corresponding to the final target pool decision and push the selected topics to users who participate in the interaction.

[0023] Compared with existing technologies, the AI-generated and hierarchical scheduling-based adaptive question generation method and system for social interaction provided in this application have the following advantages: This invention employs a two-factor adaptive scheduling approach, combining round-based rule-based judgment and real-time behavior status judgment. When user interaction fatigue is detected, it automatically switches to the core question bank for high-quality topic intervention. Conversely, when high user interaction activity is detected, it maintains a smooth rhythm by continuously generating questions from non-core question banks, effectively improving topic adaptability and user retention in social interactions. Through an AI-powered large-scale model, topics are generated in batches according to thematic dimensions. After multi-dimensional weighted evaluation of icebreaking adaptability, they are automatically stratified and labeled as core and non-core topics. This achieves a fully automated question bank construction pipeline encompassing generation, risk control, evaluation, and labeling. Furthermore, by using a sliding window constraint for thematic classification, topics identical to those of the most recent questions are excluded during selection, ensuring... To ensure topic freshness and multi-dimensional information exchange, and avoid aesthetic fatigue caused by the continuous appearance of similar topics; to randomly select topics from the core question bank based on icebreaking suitability scores, topics with higher icebreaking suitability have a greater probability of being selected, thus optimizing the quality gradient within the core question bank; through a differentiated parameter strategy for cold start and warm start, the parameter set is automatically switched based on the historical interaction records of both parties, strengthening the icebreaking density in the first interaction and quickly entering in-depth topics in repeated interactions; through conversation granular deduplication, three-level degradation and automatic loop mechanism, when the target question bank is exhausted, it is downgraded to the full question bank, and when the full question bank is exhausted, the used set is cleared and a new cycle begins, ensuring continuous question generation without interruption in any scenario. Attached Figure Description

[0024] Figure 1 This is a system block diagram of a social interactive adaptive question generation system based on AI generation and hierarchical scheduling provided by the present invention.

[0025] Figure 2 This is a flowchart of the AI ​​question bank generation and hierarchical labeling process in an AI-based adaptive question generation and hierarchical scheduling method for social interaction provided by the present invention.

[0026] Figure 3 This is a flowchart of the real-time adaptive question generation process in a social interaction adaptive question generation method based on AI generation and hierarchical scheduling provided by the present invention.

[0027] Figure 4 This is a flowchart of the effect feedback and online scoring update process in a social interaction adaptive question generation method based on AI generation and hierarchical scheduling provided by the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0029] To address the shortcomings of existing question-generating schemes that rely on preset fixed rules or purely statistical models, failing to dynamically adjust question strategies based on real-time user behavior signals during interaction, lack of awareness of user fatigue and timely intervention, and inability to maintain the current pace without interference when interaction is in good condition, this invention provides an AI-generated and hierarchical scheduling-based adaptive question-generating method and system for social interaction. This system is applied to real-time social interaction scenarios, specifically a fun question-and-answer game scenario during two-person voice or video calls. Two users sequentially complete multiple rounds of interaction by playing rock-paper-scissors, presenting topics, exchanging answers, and moving on to the next question. This invention constructs a complete self-learning closed-loop system encompassing AI generation, hierarchical labeling, behavior perception, adaptive scheduling, effect feedback, and online scoring correction. The overall architecture is as follows: Figure 1 As shown, the overall system architecture of this invention is divided into five layers: AI content generation and evaluation layer, data storage layer, memory caching layer, adaptive hierarchical scheduling engine, social interaction business layer, and user behavior perception layer. This AI-generated and hierarchically scheduled social interaction adaptive question generation system is functionally divided into the following modules: AI question bank generation and hierarchical marking module, round rule determination module, behavior signal acquisition and status determination module, behavior signal coverage module, and topic selection and push module. In this embodiment, the AI ​​question bank generation and hierarchical marking module is used to generate candidate topics in batches according to predefined social interaction theme dimensions, evaluate the icebreaking suitability of candidate topics to obtain an icebreaking suitability score, and mark topics as core topics or non-core topics based on the comparison result of the icebreaking suitability score and the hierarchical threshold, storing them in the core question bank and non-core question bank respectively; the round rule determination module is used to obtain the current question generation sequence number in the social interaction session, and determine whether the topic should be selected from the core question bank or the non-core question bank in this round based on the question generation sequence number; the behavior signal acquisition and status determination module is used to collect the interactive behavior signals of users participating in the interaction in real time, and determine the topic selection and push based on the interactive behavior signals. The system determines the current interaction state, which includes fatigue, activity, and normal states. A behavior signal overlay module is used to overlay the judgment result of the round rule judgment module based on the interaction state to obtain the final target pool decision. Specifically, when the interaction state is fatigue, the final target pool is overlaid with the core question bank; when the interaction state is activity and the round rule judgment result is the core question bank, the final target pool is overlaid with a non-core question bank; and when the interaction state is normal, the round rule judgment result remains unchanged. A topic selection and push module is used to select topics from the question bank corresponding to the final target pool decision and push the selected topics to users participating in the interaction.

[0030] The AI-generated and hierarchically scheduled social interaction adaptive question generation system includes modules for AI question bank generation and hierarchical labeling, round rule determination, behavior signal acquisition and status determination, behavior signal coverage, and topic selection and push, which respectively perform the functions of the corresponding links in the AI-generated and hierarchically scheduled social interaction adaptive question generation method.

[0031] See Figures 2 to 4 As shown, this embodiment of the invention provides a social interaction adaptive question generation method based on AI generation and hierarchical scheduling, including the following steps: Candidate topics are generated in batches according to predefined social interaction theme dimensions. The icebreaking suitability of the candidate topics is evaluated to obtain an icebreaking suitability score for each topic. Based on the comparison result of the icebreaking suitability score and the stratification threshold, the topics are marked as core topics or non-core topics and stored in the core question bank and non-core question bank respectively. In a social interaction conversation, the current question number is obtained, and the round rule is determined based on the question number to determine whether the topic should be selected from the core question bank or the non-core question bank in this round. Real-time acquisition of interactive behavior signals from participating users; determination of the current interactive state based on the interactive behavior signals; the interactive state includes fatigue state, active state, and normal state. The final target pool decision is obtained by overwriting the results of the round rule determination based on the interaction state. Specifically, when the interaction state is fatigued, the final target pool is covered by the core question bank. When the interaction state is active and the round rule determination result is the core question bank, the final target pool is covered by the non-core question bank. When the interaction state is normal, the result of the round rule determination remains unchanged. Topics are selected from the question bank corresponding to the final target pool decision, and the selected topics are pushed to users who participate in the interaction.

[0032] In this embodiment, when generating candidate topics in batches according to predefined social interaction theme dimensions, the large language model is used to generate candidate topics according to the predefined social interaction theme dimension set. The system generates candidate topics in batches, covering categories such as personality, family, views on love, work, interests, and plans. Prompt words control the tone, length, and openness of the topics. The generated candidate topics undergo AI-powered risk control filtering, including sensitive word detection, violation content identification, and sentiment analysis, filtering out content unsuitable for social scenarios.

[0033] In this embodiment, the icebreaking suitability assessment of the candidate topics includes: Each candidate topic is scored across multiple dimensions, including openness, safety, interest, and information exchange value, yielding a score for each dimension. These scores are obtained either through a large language model using pre-defined scoring prompts or through a lightweight classifier using pre-defined rules. The scores for each dimension are then weighted and summed to obtain the icebreaking suitability score, where the sum of all weights is 1. In the multi-dimensional weighted evaluation of icebreaking suitability, for candidate topics that pass risk control, an initial icebreaking suitability score is calculated using the icebreaking suitability evaluation model. This score measures the topic's potential to spark effective dialogue in initial interactions with strangers. The evaluation employs a multi-dimensional weighted approach, as shown in the following formula:

[0034] In the formula, For the topic The initial icebreaking suitability score ranges from [0,1]. A higher score indicates a stronger ability to break the ice. To evaluate the dimension indices, iterate through the set {openness, security, fun, information exchange value}; For the first The weights of each evaluation dimension satisfy the following: The default weight for each dimension is 0.25; For the topic In the The ratings are for each dimension, with values ​​ranging from [0,1].

[0035] The meanings of the above evaluation dimensions are as follows: Topic openness refers to the size of the answer space and the degree of subjective development of the topic, that is, whether the topic allows respondents to give personalized, non-unique, and freely elaborating answers; Topic safety refers to the compliance and comfort level of the topic content in a stranger social context, that is, whether the topic may cause offense, privacy violations, or touch on sensitive areas; Topic interest refers to the ability of the topic to stimulate respondents' desire to express themselves and their enthusiasm for participation, which is reflected in the novelty, topicality, and emotional arousal potential of the topic; Information exchange value refers to the degree to which the topic facilitates the exchange of effective personal information between the two parties and deepens their mutual understanding, that is, how much incremental information can both parties obtain after answering the topic that helps to deepen their understanding.

[0036] The scores for each dimension are obtained by a large language model performing structured scoring on the topic according to preset scoring prompts, or by a lightweight classifier outputting scores according to preset judgment rules. The score range is [0,1]. Specifically: For open-ended topic scoring: based on preset... Level (default) An open-ended rating scale, where a large language model assigns a score of 1 to 1 based on the scale's anchor points for each topic. Integer rating The dimension scores are obtained after linear normalization:

[0037] Topic openness scoring Linear normalization; for example If the topic is rated as level 4, then... .

[0038] For topic safety scoring: the content safety classification model outputs the overall risk confidence level of the topic. Safety scores and risk values ​​are inversely correlated.

[0039] Topic safety rating Calculations show that higher risk generally equates to lower security. When... Topics exceeding the risk control interception threshold are discarded directly; topics not discarded are handled accordingly. Calculate a safety score, for example, for the topic "What kind of personality do you find most interesting?" hour, .

[0040] For topic interest rating: based on preset... The level of interest rating scale, based on the large language model, assesses interest from 1 to 1 on each of the three sub-dimensions: novelty, emotional arousal potential, and personalization. The arithmetic mean is taken after the grade assessment. Linear normalization yields the dimension scores:

[0041] Topic interest rating Linear normalization. For example, when the three sub-dimensions are rated 4, 5, and 4 respectively, , .

[0042] For information exchange value scoring: pre-set personal experiences, personal preferences, values, and future plans. Personal information dimensions (default) The large language model extracts structured information dimensions from the expected answers to a topic, and outputs the number of information dimensions that are matched. The dimension rating is as follows:

[0043] Information exchange value rating Calculation, where This represents the maximum score for a preset information dimension, with a default value of 4. For example, the expected answers to the topic "What did you most want to be when you were a child? Why?" would fall into three categories: personal experience, personal preferences, and values. ,but .

[0044] The scores of the above four dimensions are weighted and summed according to the corresponding dimension weights, where let the score be the first... The weights of each evaluation dimension are: Satisfying weights and constraints (By default, all dimensions have a weight of 0.25), initial icebreaking fit score:

[0045] Obtain the initial icebreaking fitness score The default weight for each dimension is 0.25.

[0046] For example, the open-ended rating for the topic "What kind of personality do you find most interesting?" is 5 ( ), Security , The overall fun factor rating is 4. The value of information exchange falls into three categories: personal experience, personal preferences, and values. ;but A topic that is not lower than the tier threshold (default 0.7) is marked as a core topic.

[0047] In this embodiment, the method further includes online feedback correction of the ice-breaking suitability score: Collect performance data for each topic in actual interactions. The performance data includes the historical average dwell time for the topic, the number of rounds of continued interaction between the two parties after the topic appears, and the proportion of topics that are quickly skipped. The effectiveness return of the topic is calculated based on the effect data, and the icebreaking suitability score is corrected online using an exponential moving average method. The corrected score is written back and triggers the dynamic re-stratification of the topic between the core question bank and the non-core question bank.

[0048] In this embodiment, the stratification threshold is a dynamic threshold, adaptively adjusted based on the deviation between the actual and target proportions of the core question bank: the stratification threshold is increased when the actual proportion of the core question bank is higher than the target proportion, and decreased when the actual proportion of the core question bank is lower than the target proportion. Specifically, during automatic stratification labeling, topics are automatically labeled as core topics or non-core topics based on the comparison between the icebreaking suitability score and the stratification threshold, according to the following labeling rules:

[0049] In the formula, For the topic The core marker, with a value of 1, represents a core topic and is stored in the core question bank. A value of 0 indicates a non-core topic, which is then stored in the non-core question bank. ; For the topic The icebreaking fit score, the initial state is It is dynamically updated based on online feedback during operation; The threshold for icebreaking adaptability is set to a value range of [0,1], with a baseline default value of 0.7; the entire question bank set satisfies... .

[0050] Among them, the stratification threshold A dynamic adaptive mechanism is adopted to automatically adjust the core question pool size based on the deviation between the actual and target proportions. The formula is as follows:

[0051] In the formula, This is a dynamic stratification threshold, with a value range of [0,1]. Actual stratification marking and scheduling are based on this threshold. For the truncation function, Limited to Within the interval, if Then output ,like Then output Otherwise output ; This is the baseline threshold, with a value range of [0,1] and a default value of 0.7. Adjust the intensity for the threshold; a non-negative real number; the default value is 0.5. The larger the core pool size, the more sensitive the adjustment of the core pool proportion; It degenerates to a fixed threshold; This represents the actual proportion of the current core question bank, that is, the ratio of the number of core topics to the total number of topics, with a value range of [0,1]. This represents the percentage of the target core question bank, with a value range of [0,1] and a default value of 0.1. This is the lower limit of the threshold, with a value range of [0,1] and a default value of 0.5; This is the upper limit of the threshold, with a value range of [0,1] and a default value of 0.9.

[0052] When the actual proportion of core topics is higher than the target proportion, the threshold is raised to tighten the core standards; when the actual proportion is lower than the target proportion, the threshold is lowered to relax the core standards, so as to ensure that the size of the core pool remains stable within a reasonable range.

[0053] After a topic is launched, the system continuously collects real-world interaction data and uses an exponential moving average to adjust the icebreaking suitability score online. This allows the score to self-calibrate based on real social feedback, forming a self-learning loop. First, the system calculates the return on real-world interaction for the topic:

[0054] In the formula, For the topic The actual interactive effect reward, with a value range of [0,1]; The Sigmoid activation function compresses the linear combination of inputs to the [0,1] interval; , , The weights of each component of the effect reward are all non-negative real numbers, corresponding to the influence weights of dwell time, subsequent rounds, and skip ratio, respectively; For the topic The historical average dwell time, after normalization, ranges from [0,1]. To generate topics The number of rounds of continued interaction between the two parties, after normalization, takes a value in the range of [0,1]. For the topic The percentage of items that are quickly skipped, with a value range of [0,1].

[0055] Then, the score is corrected online using an exponential moving average:

[0056] In the formula, For the first The topic after the second revision Fit score; For the first The topic after the second revision Fit score; This represents the online learning rate, with a value range of [0.05, 0.3] and a default value of 0.1. The larger the value, the faster the rating follows recent feedback; The smaller the value, the more stable the score. For the topic The real interactive effect returns.

[0057] The corrected score is written back to the database and triggers a cache refresh when the score crosses a dynamic threshold. In this case, the core or non-core layer markers of the topic are automatically switched to achieve dynamic re-layering.

[0058] In this embodiment, determining the round rule based on the question number includes: When the question number is not greater than the first preset value, it is determined that the question should be selected from the core question bank in this round; When the question number is not less than the second preset value and the difference between the question number and the second preset value is an integer multiple of the third preset value, it is determined that the question should be selected from the core question bank in this round. In other cases, the question bank should be selected from non-core questions in this round; wherein the second preset value is greater than the first preset value.

[0059] In the basic determination of the round rules, the current question number is used. and configuration parameters The formula for determining whether to select from the core pool or the non-core pool in this round is as follows:

[0060] In the formula, For the first The basic judgment result of the question round, This indicates that questions should be selected from the core question bank. This indicates that questions should be selected from non-core question banks; The current question number is a positive integer starting from 1 and incrementing. For the front The problem requires a core pool size that is a positive integer; the default value for a cold start is 5, and the default value for a warm start is 2. The starting index of the core question is inserted periodically, a positive integer, with a default value of 6, satisfying the following condition: ; The interval for inserting core questions periodically, a positive integer, with a default value of 4; This indicates the modulo operation.

[0061] Before the rules of this round are implemented The basic rhythm of forcibly breaking the ice with questions and then periodically inserting core questions, such as the default configuration. , , When there is no behavior coverage, the question sequence is: core, core, core, non-core, non-core, core, non-core, non-core, non-core, core...

[0062] When collecting interactive behavior signals from participating users in real time, multiple behavioral signals from each round of interaction are collected in real time, and the behavioral sequences of users A and B are maintained separately for each "relationship pair": Single-wheel dwell time The difference between the timestamp of the user clicking "Next Question" and the timestamp of the last question, in seconds; Text reply length : The number of text characters entered by both parties in this round of the conversation (optional); Voice message duration : Effective voice duration for both parties in this round (optional, voice scenario); Emotional polarity The social interaction scenario in this application is a two-person voice call, in which both parties communicate and answer each other through voice, without involving text input. The emotional polarity score is entirely based on voice signal acquisition, that is, the polarity score obtained by voice emotion recognition of the voice content in the user's current interaction (optional).

[0063] In each round of interaction, the system records the effective speech segments of both parties, that is, the speech data from the presentation of the topic to the user clicking "next question". The collected speech segments are input into the speech emotion recognition model. The model determines the emotional tendency based on the acoustic features of the speech (including fundamental frequency / pitch variation, speech rate, energy / loudness, intonation fluctuation pattern, pause distribution and other prosodic features), and outputs the original speech emotion polarity value. The range of values ​​is [ [1,1], values ​​close to 1 represent positive emotions, close to 0 represent neutral emotions, and close to 1 represent positive emotions. 1 indicates negative emotion; when the effective voice duration is below the minimum effective threshold (e.g., 2 seconds), the voice signal for that round is considered missing, and the original polarity is set to the default value of 0; since this scenario only has a voice signal source, the original emotional polarity for that round... The original emotional polarity in a single round equals the emotional polarity of the speech; the original polarity is shifted from [ A linear mapping from [1,1] to [0,1] yields a normalized sentiment polarity score that can be directly incorporated into multiple signals. , as the value of the sentiment polarity score in multi-channel signal fusion; the mapping relationship is: Strongly negative: hour, ; Neutral mood: hour, ; Strongly positive: hour, ; Emotional polarity The value range is [0,1]. The closer the value is to 1, the more positive the user's emotions are in that round of interaction.

[0064] When determining the interaction state, the current interaction state is determined based on the behavioral signal sequence, including three states: fatigued, active, and normal. This invention provides two determination methods, which can be used selectively or in combination. In this embodiment, determining the current interaction state based on the interactive behavioral signals includes at least one of the following methods: The first method involves collecting the single-round dwell time for each round of interaction, comparing the single-round dwell time with a quick-skip threshold, and counting dwell times lower than the quick-skip threshold as a quick skip. The number of consecutive quick skips is counted, and when the number of consecutive quick skips reaches a preset number, it is determined to be a fatigue state. When the single-round dwell time is higher than the activity determination threshold, it is determined to be an active state. After triggering the fatigue state overwrite, the number of consecutive quick skips is reset to zero. The active state is updated in real time with the current round of behavior and does not require additional reset.

[0065] Among them, during the continuous and rapid skipping of hard counting judgment, the dwell time of a single round is collected. , and fast skip threshold Compare: like One quick jump is recorded as a single quick jump, and consecutive quick jumps are counted. Add 1; like This indicates a state of fatigue. If the duration of a single round It is determined to be in an active state; All other conditions are normal.

[0066] in, To quickly skip the judgment threshold, the default value is 10 seconds; The number of consecutive quick jumps required to trigger fatigue detection; a positive integer, with a default value of 3. The activity threshold is set to 60 seconds by default. This is the current consecutive fast skip count, a non-negative integer, which is reset to 0 after fatigue coverage is triggered.

[0067] The second method involves collecting multi-channel behavioral signals for each round of interaction. These signals include single-round dwell time, text reply length, voice speech duration, and emotional polarity score. The emotional polarity score is obtained by performing voice emotion recognition on the user's voice content in this round of interaction. After normalizing each behavioral signal, the signals are fused according to their corresponding weights to obtain the single-round interaction engagement. Within a preset review window, the interaction engagement of each round is weighted by time decay, with signals closer to the current round having a larger weight. Based on this, fatigue level and activity level are calculated. When the fatigue level reaches the fatigue threshold, it is determined to be a fatigued state. When the activity level reaches the activity threshold and is not determined to be a fatigued state, it is determined to be an active state. Other situations are determined to be normal states.

[0068] In the multi-signal time decay fusion judgment, multiple behavioral signals are fused and time decay weighting is introduced. Signals closer to the current round have greater weights, resulting in more accurate judgment. First, the interaction engagement level in a single round is calculated:

[0069] In the formula, For the first The level of interaction engagement in the round, with a value range of [0,1]; , , , The fusion weights of the various signals satisfy the following conditions: The default values ​​are 0.5, 0.2, 0.2, and 0.1, respectively. This is a normalization function that maps the input to the interval [0,1]. For the first The duration of a single wheel's stay; For the first The length of the text reply in the wheel; For the first The duration of each round of voice messages; For the first The emotional polarity score of the round, with a value range of [0,1], is the original emotional polarity value output by the speech emotion recognition model. according to Obtained through linear mapping.

[0070] The emotional polarity score is obtained as follows: In a voice scenario, voice emotion recognition is performed on the voice content, and the voice emotional polarity score is output through an emotion recognition model based on acoustic features; when both text content and voice content exist in the same round of interaction, the text emotional polarity score and the voice emotional polarity score are normalized separately and then weighted and fused according to a preset ratio to obtain the emotional polarity score for that round; the emotional polarity score ranges from [0,1], and the closer the value is to 1, the more positive the user's emotion is in that round of interaction.

[0071] Then, fatigue and activity levels are calculated using a time-decay weighted average within the review window:

[0072]

[0073] In the formula, The fatigue level is a non-negative real number. Activity level, a non-negative real number; This is the current question number; The number of window rotations for status determination review, a positive integer, with a default value of 5; For time decay weight, It is a natural constant. Time decay coefficient; rounds The closer to the current wheel The greater the weight; This is the time decay coefficient, a non-negative real number, with a value range of [0.1, 1] and a default value of 0.5; The larger the value, the more important recent signals become. This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. This is the threshold for engagement level in quick jumps, with a value range of [0.2, 0.5] and a default value of 0.3; engagement levels below this value are considered quick jumps.

[0074] Finally, the state is determined:

[0075] In the formula, This represents the current interaction status; The fatigue trigger threshold is a positive real number. The active trigger threshold is a positive real number.

[0076] In this embodiment, the social interaction is a two-person interaction. The behavior sequences and interaction engagement of both users are maintained separately for each relationship pair, and the fatigue and activity levels of each user are calculated separately. The fatigue level of a relationship pair is the larger of the fatigue levels of both users; if either user reaches a fatigue state, the relationship pair is considered fatigued. The activity level of a relationship pair is the smaller of the activity levels of both users; if both users reach an active state, the relationship pair is considered active. Specifically, for the two-person interaction scenario, the states of both users are merged at the "relationship pair" level to achieve collaborative scheduling between the two users.

[0077]

[0078] In the formula, For the fatigue degree of the relationship level; For relationship-level activity; The fatigue levels of user A and user B respectively; The activity levels of user A and user B respectively; To find the maximum value of the function, fatigue on either side is used to determine the fatigue relationship. To achieve the minimum value function, both parties must be active for the relationship to be considered active; the final interaction state is determined by... , Respectively with threshold , The comparison yields the results.

[0079] Based on the fundamental rules of each round, dynamic overlay is performed according to the interaction status to obtain the final target pool decision:

[0080] In the formula, For the first The final source of the questions is determined, i.e., the final target pool decision; This represents the current interaction status; This is the result of the round-based determination.

[0081] Among them, fatigue coverage: when the state is fatigued, regardless of the original round's rule judgment result, it will be covered to the core pool, using high-quality topics to interrupt fatigue and increase interaction engagement; after triggering coverage, the continuous fast skip counter will be reset. Active Coverage: When the status is active and the original rule determines it to be in the core pool, the coverage is changed to the non-core pool to avoid disrupting the good interaction rhythm, while retaining the core question quota; Normal Status: The original round rule determination remains unchanged.

[0082] In this embodiment, selecting topics from the question bank corresponding to the final target pool decision includes: maintaining a topic sliding window of a preset length, recording the topic categories of the most recent questions; excluding topics whose topic categories are the same as those already present in the topic sliding window from the available topics in the target question bank, thus obtaining a topic-filtered candidate set; and clearing the topic sliding window and canceling topic constraints when the topic-filtered candidate set is empty. The maintenance length is... Theme sliding window Record the latest The topics are categorized by theme, and topics with the same theme are excluded first to ensure freshness. The filtered candidate set is as follows:

[0083] In the formula, A set of candidate topics filtered by theme; The set of available topics corresponding to the final target pool; For the topic The topic classification belongs to the topic dimension set. ; For the most recent The list of topics should not exceed a certain length. An ordered list; The length of the sliding window is a positive integer, with a default value of 3. This represents the length of the filtered candidate set. If empty, relax the constraints: clear the topic window. Remove topic restrictions. .

[0084] In this embodiment, selecting topics from the question bank corresponding to the final target pool decision includes: when the final target pool is a core question bank, selecting topics randomly with weighted average based on their icebreaking suitability scores, with topics having a higher icebreaking suitability score having a greater probability of being selected; when the final target pool is a non-core question bank, selecting topics randomly and evenly.

[0085] In this embodiment, the weighted random selection based on the icebreaking suitability score of each topic includes: calculating a selection priority score for each candidate topic, wherein the selection priority score is equal to the icebreaking suitability score of the topic plus an exploration bonus, wherein the exploration bonus is negatively correlated with the number of times the topic has been selected in the past; and performing softmax temperature sampling based on the selection priority score to determine the final selected topic.

[0086] The final topics are selected from the candidate set, with a differentiated selection strategy used between the core pool and non-core pool. The core topic bank is randomly selected based on a weighted average of icebreaker suitability scores; higher scores increase the probability of selection.

[0087] In the formula, For the topic The probability of being selected; For the topic Icebreaking compatibility score; As a candidate set for the core question bank; This is the first topic in the candidate set.

[0088] Non-core question banks are selected using uniform random selection:

[0089] In the formula, This represents the number of elements in the non-core question bank candidate set.

[0090] In addition to the weighted allocation in the core pool, an exploration mechanism is introduced to balance high-scoring utilization with low-exposure exploration, allowing long-tail topics to gain exposure and accumulate feedback data. First, the selection priority score for each candidate topic is calculated:

[0091] In the formula, For the topic Priority selection; The utilization item is the icebreaking fit score; To explore the coefficient, the value range is [0,2], a non-negative real number, and the default value is 0.5; It degenerates into a pure fraction weighted average; For the current session The cumulative number of question rounds within the period; For the topic The number of times it has been selected in history, a non-negative integer; Avoid having a denominator of zero.

[0092] Then, softmax temperature sampling is performed according to priority to obtain the final selection probability:

[0093] In the formula, It is a natural exponential function; This is the softmax temperature coefficient, a positive real number, with a value range of [0.5, 2] and a default value of 1.0; The smaller the value, the more concentrated the selection of high-scoring topics; The larger the value, the more uniform the distribution; alternatively, Thompson Sampling, sampling by topic effect posterior distribution, can be used as an equivalent alternative.

[0094] In this embodiment, the method further includes session-level deduplication and multi-level degradation: Maintain a set of used topics for the current session, and filter used topics from the candidate set; When the available set in the target question bank is empty after deducting used topics, it will be downgraded to selecting from the entire question bank. When the available set of the entire question bank is empty after deducting the used topics, the set of used topics is cleared and a new question generation cycle begins.

[0095] This includes maintaining the set of topics already used in the current session. Before selecting a topic, filter out topics that have already been used:

[0096] In the formula, For the target question bank In the conversation Available sets in; For the target question bank set; For the session The set of used topics is then used; then a set difference operation is performed, and a three-level degradation mechanism is executed to ensure that topics are never broken: If the target question bank's available set is empty, downgrade to the full question bank. Select; If the available set for the entire question bank is also empty, clear the used sets. Entering a new question-setting cycle: like ,but

[0097] After selection, the topic content and metadata are pushed to both terminals in real time; at the same time, the timestamp of this round of question generation is recorded, the topic sliding window is updated, the topic is marked as used, and written to the distributed cache.

[0098] In this embodiment, the method further includes cold start and hot start parameter adaptation: Query the historical call records of both users involved in the interaction; When both parties are interacting for the first time, a cold start parameter set is used, increasing the number of questions that must be selected from the core question bank for the first few questions. During a cold start, when both parties are interacting for the first time and have no historical call records, a cold start parameter set is used, employing a larger number of questions selected from the core question bank for the first few questions. The question forces a maximum number of cores (default). ), to increase ice-breaking density; When both parties are engaging in repetitive interactions, a hot-start parameter set is used to reduce the number of questions that must be selected from the core question bank for the first few questions. During a hot start, when both parties are engaging in repetitive interactions and have historical call records, a hot-start parameter set is used, employing a smaller initial... The question forces a maximum number of cores (default). ), quickly get into in-depth topics.

[0099] Optionally, cross-conversation topic filtering can be enabled to prevent repeated topics from being discussed in multiple calls. Optionally, personalized adaptation can be performed by combining the user profiles and interest tags of both parties: based on the topic dimensions corresponding to the common interest tags of both parties, the selection weight of topics in the icebreaking stage is biased, so that topics of common interest are given priority, thereby improving the efficiency of icebreaking.

[0100] In this embodiment, the method further includes personalized adaptation based on user profiles: Obtain the user profile tags and interest tags of both parties involved in the interaction; Based on the topic dimensions corresponding to the shared interest tags of both parties, the selection weight of topics in the icebreaking stage is biased, so that the topic dimensions corresponding to the shared interest tags are given priority in the questions.

[0101] After each round of interaction, the system collects the current round's performance data for that topic (duration of stay, whether it skipped quickly, number of subsequent interaction rounds, etc.), asynchronously aggregating it into historical statistical data at the topic level, and then... Calculate the return on investment , and according to The ice-breaking fit score is adjusted using an exponential moving average. The topic exposure count is also maintained. This is used for the exploration items in the weighted selection module. When the corrected score crosses the dynamic threshold... At that time, the system automatically switches between core and non-core topic hierarchical tags, writes back to the database, and triggers a cache refresh. This process forms a self-learning closed loop of selection, feedback, correction, and re-selection, making the system more accurate with use and continuously improving the actual interactive quality of the core question bank.

[0102] The following table, Table 1, provides a complete example of question generation and execution, using a set of typical parameter configurations: , , , , , , .

[0103] Table 1. Examples of Question Generation and Execution

[0104] Through a two-factor adaptive scheduling system combining round-based rule-based judgment and real-time behavior status judgment, the system automatically switches to the core question bank for high-quality topic intervention when user interaction fatigue is detected, and maintains a smooth rhythm by generating questions from non-core question banks when high user interaction activity is detected, effectively improving topic adaptability and user retention in social interactions. The system uses an AI large-scale model to generate topics in batches according to thematic dimensions, and automatically categorizes and marks them as core and non-core topics after multi-dimensional weighted evaluation of icebreaking adaptability. This achieves a fully automated question bank construction pipeline encompassing generation, risk control, evaluation, and marking. A sliding window constraint based on topic classification excludes topics with the same themes as the most recent few questions during selection, ensuring the smooth flow of questions. To ensure freshness and multi-dimensional information exchange, the system avoids aesthetic fatigue caused by the continuous appearance of similar topics. Within the core question bank, topics are randomly selected based on icebreaker suitability scores, increasing the probability of selection for topics with higher icebreaker suitability and optimizing the quality gradient within the core question bank. A differentiated parameter strategy for cold and warm starts automatically switches parameter sets based on historical interaction records, strengthening icebreaker density during the first interaction and quickly leading to deeper topics during repeated interactions. Through conversation-level deduplication, a three-level degradation mechanism, and an automatic loop mechanism, the system downgrades to the full question bank when the target question bank is exhausted, and clears the used set and enters a new cycle when the full question bank is exhausted, ensuring uninterrupted question generation in any scenario.

[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A social interaction adaptive question generation method based on AI generation and hierarchical scheduling, characterized in that, Includes the following steps: Candidate topics are generated in batches according to predefined social interaction theme dimensions. The icebreaking suitability of the candidate topics is evaluated to obtain an icebreaking suitability score for each topic. Based on the comparison result of the icebreaking suitability score and the stratification threshold, the topics are marked as core topics or non-core topics and stored in the core question bank and non-core question bank respectively. In a social interaction conversation, the current question number is obtained, and the round rule is determined based on the question number to determine whether the topic should be selected from the core question bank or the non-core question bank in this round. Real-time acquisition of interactive behavior signals from participating users; determination of the current interactive state based on the interactive behavior signals; the interactive state includes fatigue state, active state, and normal state. The final target pool decision is obtained by overwriting the results of the round rule determination based on the interaction state. Specifically, when the interaction state is fatigued, the final target pool is covered by the core question bank. When the interaction state is active and the round rule determination result is the core question bank, the final target pool is covered by the non-core question bank. When the interaction state is normal, the result of the round rule determination remains unchanged. Topics are selected from the question bank corresponding to the final target pool decision, and the selected topics are pushed to users who participate in the interaction.

2. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, The icebreaking suitability assessment of the candidate topics includes: Each candidate topic is scored from multiple dimensions, including topic openness, safety, interest, and information exchange value, to obtain a score value for each dimension. The score values ​​of each dimension are then weighted and summed according to their respective dimension weights to obtain the icebreaking suitability score, where the sum of the weights of each dimension is 1.

3. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 2, characterized in that, The method also includes online feedback correction of the icebreaking fitness score: Collect performance data for each topic in actual interactions. The performance data includes the historical average dwell time for the topic, the number of rounds of continued interaction between the two parties after the topic appears, and the proportion of topics that are quickly skipped. The effectiveness return of the topic is calculated based on the effect data, and the icebreaking suitability score is corrected online using an exponential moving average method. The corrected score is written back and triggers the dynamic re-stratification of the topic between the core question bank and the non-core question bank.

4. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, The stratification threshold is a dynamic threshold that is adaptively adjusted based on the deviation between the actual proportion of the core question bank and the target proportion: when the actual proportion of the core question bank is higher than the target proportion, the stratification threshold is increased; when the actual proportion of the core question bank is lower than the target proportion, the stratification threshold is decreased.

5. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, The round-based rule determination based on the question number includes: When the question number is not greater than the first preset value, it is determined that the question should be selected from the core question bank in this round; When the question number is not less than the second preset value and the difference between the question number and the second preset value is an integer multiple of the third preset value, it is determined that the question should be selected from the core question bank in this round. In other cases, the question bank should be selected from non-core questions in this round; wherein the second preset value is greater than the first preset value.

6. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, Determining the current interaction state based on the interaction behavior signals includes at least one of the following methods: The first method involves collecting the single-round dwell time for each round of interaction, comparing the single-round dwell time with a fast skip threshold, and recording a dwell time lower than the fast skip threshold as a fast skip. The number of consecutive fast skips is counted, and when the number of consecutive fast skips reaches a preset number, it is determined to be a fatigue state. When the single-round dwell time is higher than the activity determination threshold, it is determined to be an active state. After triggering the fatigue state overwrite, the number of consecutive fast skips is reset to zero. The second method involves collecting multi-channel behavioral signals for each round of interaction. These signals include single-round dwell time, text reply length, voice speech duration, and emotional polarity score. The emotional polarity score is obtained by performing voice emotion recognition on the user's voice content in this round of interaction. After normalizing each behavioral signal, the signals are fused according to their corresponding weights to obtain the single-round interaction engagement. Within a preset review window, the interaction engagement of each round is weighted by time decay, with signals closer to the current round having a larger weight. Based on this, fatigue level and activity level are calculated. When the fatigue level reaches the fatigue threshold, it is determined to be a fatigued state. When the activity level reaches the activity threshold and is not determined to be a fatigued state, it is determined to be an active state. Other situations are determined to be normal states.

7. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 6, characterized in that, The social interaction is a two-person interaction. The behavior sequence and interaction engagement of the two users are maintained separately for each relationship pair. The fatigue level and activity level of each user are calculated separately. The fatigue level of the relationship pair is the larger value between the fatigue levels of the two users. The relationship pair is determined to be in a fatigue state when either user reaches a fatigue state. The activity level of the relationship pair is the smaller value between the activity levels of the two users. The relationship pair is determined to be in an active state only when both users reach an active state.

8. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, Selecting topics from the question bank corresponding to the final target pool decision includes: maintaining a topic sliding window with a preset window value and recording the topic categories of the most recent questions; excluding topics whose topic categories are the same as those already present in the topic sliding window from the available topics in the target question bank, thus obtaining a topic-filtered candidate set; and clearing the topic sliding window and canceling topic constraints when the topic-filtered candidate set is empty.

9. The adaptive question generation method for social interaction based on AI generation and hierarchical scheduling as described in claim 1, characterized in that, The selection of topics from the question bank corresponding to the final target pool decision includes: when the final target pool is a core question bank, topics are selected randomly with weighted average based on their icebreaking suitability scores, and topics with higher icebreaking suitability scores have a greater probability of being selected; when the final target pool is a non-core question bank, topics are selected randomly and evenly.

10. A social interactive adaptive question generation system based on AI generation and hierarchical scheduling, characterized in that, The method for implementing the AI-generated and hierarchical scheduling-based adaptive question generation method for social interaction as described in any one of claims 1-9 includes: The AI ​​question bank generation and hierarchical labeling module is used to generate candidate topics in batches according to predefined social interaction theme dimensions, evaluate the icebreaking suitability of candidate topics to obtain icebreaking suitability scores, and label topics as core topics or non-core topics according to the comparison results of the icebreaking suitability scores and hierarchical thresholds, and store them in the core question bank and non-core question bank respectively. The round rule determination module is used to obtain the current question number in the social interaction conversation, and determine whether the topic should be selected from the core question bank or the non-core question bank in this round based on the question number. The behavior signal acquisition and state determination module is used to acquire the interactive behavior signals of users participating in the interaction in real time, and determine the current interaction state based on the interactive behavior signals. The interaction state includes fatigue state, active state and normal state. The behavior signal overlay module is used to overlay the judgment result of the round rule judgment module according to the interaction state to obtain the final target pool decision. When the interaction state is fatigued, the final target pool is covered as the core question bank. When the interaction state is active and the round rule judgment result is the core question bank, the final target pool is covered as the non-core question bank. When the interaction state is normal, the round rule judgment result remains unchanged. The topic selection and push module is used to select topics from the question bank corresponding to the final target pool decision and push the selected topics to users who participate in the interaction.

Images