A delay control system for human-machine dialogue output

Abstract

The application discloses a human-computer conversation output delay control system, which acquires first and second type interaction data of different biological behavior dimensions in parallel through a data acquisition unit, provides reliable support for rhythm difference quantization, generates a quantization index through a difference determination unit, objectively quantizes rhythm dislocation degree, and solves the defect that the prior art cannot accurately identify rhythm dislocation, and when the trigger condition is met, the delay adjustment unit performs a control operation, delays the rhythm of a large language model output, avoids the asynchronization of user cognition and emotional integration, effectively reduces psychological risk, and improves the safety, adaptability and reliability of human-computer conversation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a delay control system for human-computer dialogue output. Background Technology

[0002] With the development of natural language processing technology and large language models, human-computer dialogue systems have been widely used in scenarios such as mental health intervention. Existing systems mostly aim to improve response efficiency, performing semantic recognition based on user input text and quickly generating responses. However, they lack parallel collection and correlation analysis of users' multi-dimensional biological behavioral data, making it difficult to effectively capture the rhythmic differences between cognitive and emotional interaction dimensions. At the same time, existing security control mechanisms mostly adopt passive interception methods such as keyword matching, which can only deal with explicit crisis situations. They cannot dynamically assess the degree of interaction rhythm irregularity through quantitative indicators, nor can they proactively delay the dialogue output rhythm when the rhythm is abnormal. This leads to a mismatch between the human-computer interaction rhythm and the user's physiological and psychological metabolic rhythm, which can easily cause delayed psychological side effects. Overall, the interaction security and adaptability are insufficient. Summary of the Invention

[0003] This application provides a delay control system for human-computer dialogue output, which can proactively intervene in the dialogue output rhythm when the risk of rhythm asymmetry occurs, avoid the asynchronous cognitive changes and emotional integration caused by the rapid output of response content by the large language model, effectively reduce the psychological risks caused by rhythm asymmetry, and significantly improve the safety, adaptability and reliability of human-computer dialogue interaction.

[0004] In a first aspect, this application provides a delay control system for human-computer dialogue output, comprising: The data acquisition unit is used to acquire a first type of interaction data and a second type of interaction data based on the user's input information and / or dialogue history data during the human-computer dialogue interaction process, wherein the second type of interaction data and the first type of interaction data belong to different biological behavioral dimension data. The difference determination unit is used to generate a quantitative index that characterizes the difference between the changing trend of the first type of interaction data and the changing trend of the second type of interaction data based on the acquired first type of interaction data and second type of interaction data. The delay adjustment unit is used to execute at least one control operation when the quantification index meets the preset triggering conditions. The control operation is used to delay the rhythm of the response content output by the large language model in the human-computer dialogue.

[0005] Optionally, the first type of interaction data includes: user response latency and cognitive achievement state characteristics in the user's input text; the second type of interaction data includes at least one user state quantification indicator, which includes a multi-dimensional rating of the user's own state.

[0006] Optionally, the cognitive achievement state characteristics are obtained through the following process: Get the user's input text; The input text is processed by word segmentation to obtain the word segmentation results; The word segmentation results are matched with a preset keyword library, which contains words that indicate cognitive achievement or comprehension completion. The number of words successfully matched by the word segmentation results; The cognitive achievement state density value is calculated based on the ratio of the number of words to the length of the input text. The density value of the cognitive achievement state is used as the feature of the cognitive achievement state.

[0007] Optionally, the difference determination unit is specifically used for: Acquire multiple first-type interaction data collected within a preset time window, and generate a first trend value based on the multiple first-type interaction data; Acquire multiple second-type interaction data collected within the time window, and generate a second trend value based on the multiple second-type interaction data; Calculate the difference between the first trend value and the second trend value, and use the difference as the quantitative indicator; or, calculate the ratio of the first trend value to the second trend value, and use the ratio as the quantitative indicator.

[0008] Optionally, the difference determination unit is further configured to: after generating the quantitative indicator, compare the quantitative indicator with a plurality of preset thresholds to determine the risk level of the quantitative indicator; the risk level is used for the delay adjustment unit to select and execute the corresponding level of control operation based on the risk level; The preset triggering conditions include at least one of the following: the quantitative indicator exceeds a first threshold, the quantitative indicator's increment within multiple consecutive time windows exceeds a second threshold, and the cumulative number of times the quantitative indicator exceeds a third threshold exceeds a preset number.

[0009] Optionally, the control operation performed by the delay adjustment unit includes a three-level progressive control operation; the delay adjustment unit is further configured to: obtain the degree value of the quantitative indicator satisfying the triggering condition, and the historical number of times the control operation is triggered during the human-computer dialogue interaction, and select the corresponding level of control operation from the three-level progressive control operation to execute according to the degree value and the historical number of times; wherein, the three-level progressive control operation includes a first-level control operation, a second-level control operation, and a third-level control operation, the first-level control operation is used to forcibly pause user input to delay the output rhythm, the second-level control operation is used to switch the topic type generated by the dialogue to reduce cognitive load and indirectly delay deep response, and the third-level control operation is used to restrict deep interaction functions across sessions to forcibly lengthen the output interval.

[0010] Optionally, the first-level control operation is performed according to the following steps: On the server side, prevent the generation of the current output content of the large language model; Send an interface lock command to the front-end user interface. The interface lock command is used to set the user input box to an inoperable state, thereby forcibly increasing the time interval between the user initiating the next round of requests. Activate the visual guidance animation of the front-end user interface and start a countdown of a preset duration; When the countdown ends, an interface unlock command is sent to the front-end user interface to restore the user input box to an operable state.

[0011] Optionally, the second-level control operation is performed according to the following steps: Block the original prompt words sent to the large language model. The original prompt words are cognitive reconstruction prompt words, which are used to instruct the large language model to generate response content with cognitive reconstruction as the theme. The original prompt words are replaced with bodily awareness prompt words, which are bodily awareness type prompt words. These bodily awareness type prompt words are used to instruct the large language model to generate response content that guides users to focus on physiological feedback, thereby reducing the cognitive depth and immediate logical responsiveness of the response. The response content generated based on the large language model of the somatic awareness cue words is sent to the user interface; In response to the sending of the reply content, it enters a waiting state to receive physiological feedback feature data returned by the user; Based on the received physiological feedback feature data, the second type of interaction data is updated, and the updated second type of interaction data is returned to the data acquisition unit.

[0012] Optionally, the third-level control operation is specifically performed according to the following steps: Forcefully terminate the human-computer dialogue session; Set a cooling-off period for each user's account, and lock the account's permission to initiate deep cognitive interactions during the cooling-off period; During the settling and cooling period, in response to the dialogue request issued by the user, it is determined whether the dialogue request involves cognitive reconstruction or causal exploration. If so, the dialogue request is intercepted and a preset alliance maintenance response is returned to the user interface to prevent deep cognitive requests from triggering the large language model to generate a fast logical response. After the precipitation cooling period ends, the second type of interactive data is monitored to obtain monitoring values, and the monitoring values ​​are compared with a preset safety baseline; if the monitoring values ​​do not recover to the safety baseline, monitoring continues until the monitoring values ​​recover to the safety baseline; When the monitored value returns to the security baseline, the permission lock for deep cognitive interaction of the account is released.

[0013] Optionally, the delay adjustment unit is further configured to: after forcibly ending the human-computer dialogue session, generate a rhythm irregularity analysis summary, the rhythm irregularity analysis summary including: the historical change curve of the quantitative indicator within a preset time window, the time point record corresponding to each time the trigger condition is met, and the user's identity identifier; and send the rhythm irregularity analysis summary to a preset human terminal to initiate a human referral warning.

[0014] Secondly, this application provides a method for delay control of human-computer dialogue output, including: During the human-computer dialogue interaction process, based on the user's input information and / or dialogue history data, a first type of interaction data and a second type of interaction data are obtained, wherein the second type of interaction data and the first type of interaction data belong to different biological behavioral dimension data; Based on the acquired first type of interaction data and second type of interaction data, a quantitative indicator is generated to characterize the difference between the changing trends of the first type of interaction data and the changing trends of the second type of interaction data. When the quantitative indicator meets the preset triggering condition, at least one control operation is executed. The control operation is used to delay the rhythm of the response content output by the large language model in the human-computer dialogue.

[0015] Optionally, the first type of interaction data includes: user response latency and cognitive achievement state characteristics in the user's input text; the second type of interaction data includes at least one user state quantification indicator, which includes a multi-dimensional rating of the user's own state.

[0016] Optionally, the cognitive achievement state characteristics are obtained through the following steps: Get the user's input text; The input text is processed by word segmentation to obtain the word segmentation results; The word segmentation results are matched with a preset keyword library, which contains words that indicate cognitive achievement or comprehension completion. The number of words successfully matched by the word segmentation results; The cognitive achievement state density value is calculated based on the ratio of the number of words to the length of the input text. The density value of the cognitive achievement state is used as the feature of the cognitive achievement state.

[0017] Optionally, generating the quantitative indicators specifically includes: Acquire multiple first-type interaction data collected within a preset time window, and generate a first trend value based on the multiple first-type interaction data; Acquire multiple second-type interaction data collected within the time window, and generate a second trend value based on the multiple second-type interaction data; Calculate the difference between the first trend value and the second trend value, and use the difference as the quantitative indicator; or, calculate the ratio of the first trend value to the second trend value, and use the ratio as the quantitative indicator.

[0018] Optionally, it also includes: after generating the quantitative indicator, comparing the quantitative indicator with a plurality of preset thresholds to determine the risk level of the quantitative indicator; the risk level is used for subsequent selection of corresponding level of control operation based on the risk level; The preset triggering conditions include at least one of the following: the quantitative indicator exceeds a first threshold, the quantitative indicator's increment within multiple consecutive time windows exceeds a second threshold, and the cumulative number of times the quantitative indicator exceeds a third threshold exceeds a preset number.

[0019] Optionally, the control operation includes a three-level progressive control operation; the method further includes: obtaining the degree value of the quantitative indicator satisfying the triggering condition, and the historical number of times the control operation is triggered during the human-computer dialogue interaction, and selecting the corresponding level of control operation from the three-level progressive control operation for execution according to the degree value and the historical number of times; wherein, the three-level progressive control operation includes a first-level control operation, a second-level control operation, and a third-level control operation, the first-level control operation is used to forcibly pause user input to delay the output rhythm, the second-level control operation is used to switch the topic type generated by the dialogue to reduce cognitive load and indirectly delay deep response, and the third-level control operation is used to restrict deep interaction functions across sessions to forcibly lengthen the output interval.

[0020] Optionally, the first-level control operation specifically includes the following steps: On the server side, prevent the generation of the current output content of the large language model; Send an interface lock command to the front-end user interface. The interface lock command is used to set the user input box to an inoperable state, thereby forcibly increasing the time interval between the user initiating the next round of requests. Activate the visual guidance animation of the front-end user interface and start a countdown of a preset duration; When the countdown ends, an interface unlock command is sent to the front-end user interface to restore the user input box to an operable state.

[0021] Optionally, the second-level control operation specifically includes the following steps: Block the original prompt words sent to the large language model. The original prompt words are cognitive reconstruction prompt words, which are used to instruct the large language model to generate response content with cognitive reconstruction as the theme. The original prompt words are replaced with bodily awareness prompt words, which are bodily awareness type prompt words. These bodily awareness type prompt words are used to instruct the large language model to generate response content that guides users to focus on physiological feedback, thereby reducing the cognitive depth and immediate logical responsiveness of the response. The response content generated based on the large language model of the somatic awareness cue words is sent to the user interface; In response to the sending of the reply content, it enters a waiting state to receive physiological feedback feature data returned by the user; Based on the received physiological feedback feature data, the second type of interaction data is updated, and the updated second type of interaction data is returned for subsequent difference determination.

[0022] Optionally, the third-level control operation specifically includes the following steps: Forcefully terminate the human-computer dialogue session; Set a cooling-off period for each user's account, and lock the account's permission to initiate deep cognitive interactions during the cooling-off period; During the settling and cooling period, in response to the dialogue request issued by the user, it is determined whether the dialogue request involves cognitive reconstruction or causal exploration. If so, the dialogue request is intercepted and a preset alliance maintenance response is returned to the user interface to prevent deep cognitive requests from triggering the large language model to generate a fast logical response. After the precipitation cooling period ends, the second type of interactive data is monitored to obtain monitoring values, and the monitoring values ​​are compared with a preset safety baseline; if the monitoring values ​​do not recover to the safety baseline, monitoring continues until the monitoring values ​​recover to the safety baseline; When the monitored value returns to the security baseline, the permission lock for deep cognitive interaction of the account is released.

[0023] Optionally, the method further includes: after forcibly ending the human-computer dialogue session, generating a rhythm irregularity analysis summary, the rhythm irregularity analysis summary including: the historical change curve of the quantitative indicator within a preset time window, the time point record corresponding to each time the trigger condition is met, and the user's identity identifier; sending the rhythm irregularity analysis summary to a preset human terminal to initiate a human referral warning.

[0024] As can be seen from the above technical solution, this application provides a delay control system for human-computer dialogue output. The system, through a data acquisition unit, acquires first-type and second-type interactive data belonging to different biological behavioral dimensions during the human-computer dialogue interaction process, based on user input information and / or dialogue history data. This achieves parallel acquisition of cross-dimensional interactive data, providing reliable data support for the quantitative assessment of rhythm differences. Through a difference determination unit, based on the acquired first-type and second-type interactive data, a quantitative index is generated representing the difference between the changing trends of the first-type and second-type interactive data. This achieves objective quantification of the degree of rhythm asymmetry, solving the technical deficiency of existing technologies that cannot accurately identify rhythm discontinuity. Through a delay adjustment unit, at least one control operation is executed when the quantitative index meets preset triggering conditions, delaying the rhythm of the large language model's output response content in the human-computer dialogue. This proactively intervenes in the dialogue output rhythm when the risk of rhythm asymmetry occurs, preventing the large language model from rapidly outputting response content, which could cause asynchronous changes in user cognition and emotional integration. This effectively reduces the psychological risk caused by rhythm asymmetry and significantly improves the safety, adaptability, and reliability of human-computer dialogue interaction.

[0025] The further effects of the aforementioned non-conventional preferred method will be explained below in conjunction with specific embodiments. Attached Figure Description

[0026] To more clearly illustrate the embodiments of this application or the existing technical solutions, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart illustrating the system determination and execution logic provided in an embodiment of this application; Figure 2 A flowchart illustrating the execution process of a multi-level rhythm control engine provided in an embodiment of this application; Figure 3 This is a schematic diagram of the system architecture of a delay control system for human-computer dialogue output provided in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] This application provides a delay control system for human-computer dialogue output. This system can be deployed in various large language model dialogue platforms such as mental health intervention, digital healthcare, and emotional companionship. During human-computer interaction, it can proactively sense the rhythmic difference between the speed of user cognitive change and the speed of emotional integration. When the risk of rhythm asymmetry occurs, it can automatically delay the rhythm of the large language model's output response content to avoid delayed psychological harm caused by the AI's rapid response leading to user cognitive changes far exceeding the speed of emotional metabolism.

[0030] See Figure 3 This paper illustrates a delay control system for human-computer dialogue output in an embodiment of this application. The system includes a data acquisition unit, a difference determination unit, and a delay adjustment unit.

[0031] The data acquisition unit is used to acquire a first type of interaction data and a second type of interaction data based on the user's input information and / or dialogue history data during the human-computer dialogue interaction process, wherein the second type of interaction data and the first type of interaction data belong to different biological behavioral dimension data.

[0032] The first type of interactive data, also known as AI efficiency channel data, refers to interactive dimension data that directly reflects the efficiency of human-computer interaction, the speed of user cognitive response, and the progress of semantic understanding. It focuses on characterizing the rapid cognitive processing behavior of the prefrontal cortex of the brain.

[0033] The second type of interactive data, also known as human circadian rhythm data, refers to interactive dimension data that reflects users' emotional metabolism, changes in psychological state, and physiological recovery process, focusing on characterizing the slow emotional integration behavior of the human limbic system.

[0034] Biological behavioral dimension data refers to computable data collected and quantified from biological dimensions such as human cognitive behavior, emotional behavior, and physiological behavior, which is used to objectively characterize the changes in the psychological and physiological state of users during human-computer interaction.

[0035] The difference determination unit is used to generate a quantitative index representing the difference between the changing trends of the first type of interaction data and the second type of interaction data, based on the acquired first type of interaction data and second type of interaction data. The changing trend refers to the pattern of increase, decrease, or stable change of a certain type of interaction data over time within a continuous interaction period, and can be represented mathematically by slope, mean change rate, etc. The quantitative index, also known as the Efficiency-Rhythm Discontinuity Index (ERM Index), is a numerical indicator that transforms the difference in the changing trends of the two types of interaction data. It is used to objectively and calculably measure the degree of disconnect between cognitive efficiency and emotional rhythm during human-computer interaction; a higher value indicates a higher risk of rhythm discontinuity.

[0036] The delay adjustment unit is used to execute at least one control operation when the quantitative indicator meets a preset trigger condition. This control operation delays the rhythm of the response content output by the large language model in the human-computer dialogue. The trigger condition refers to a preset set of rules used to determine whether the risk of rhythm arrhythmia reaches an intervention threshold, set by the system based on clinical psychobiological data. The control operation refers to the intervention action performed by the system to delay the output rhythm of the large language model, categorized into three progressive levels based on risk level. The large language model refers to a generative AI model with natural language generation, semantic understanding, and cognitive reconstruction capabilities; it is the core generation engine of the dialogue system, responsible for outputting the machine-side response content in the human-computer dialogue.

[0037] In one implementation, the first type of interactive data includes: user response latency and cognitive achievement state characteristics in the user's input text; the second type of interactive data includes at least one user state quantification indicator, which includes a multi-dimensional rating value of the user's own state.

[0038] Among them, user response latency refers to the time interval between the completion of the large language model's output of the reply and the user's input of the next message, measured in seconds or milliseconds, directly reflecting the user's cognitive processing speed of the previous reply. Cognitive achievement state characteristics refer to quantitative features extracted from the user's input text that characterize the user's completeness of understanding the current dialogue topic or the degree of cognitive closure; higher values ​​indicate a higher degree of subjective cognitive achievement. User state quantitative indicators refer to numerical indicators obtained through embedded dynamic questionnaires, scale ratings, or multimodal data collection methods that reflect the user's emotional state, psychological load, and physiological recovery degree; these are also known as rhythmic proxy indicators. Multidimensional rating values ​​refer to the user's multidimensional subjective rating of their own state based on standardized scales, covering dimensions such as subjective emptiness, meaning recovery, alliance perception, and delayed negative effects; these can be understood as features of SES, MLQ, WAI-SR, and NEQ scales.

[0039] In one implementation, the cognitive achievement state characteristics are obtained through the following process: First, obtain the user's input text. Then, perform word segmentation on the input text to obtain the segmentation result; word segmentation can be understood as a natural language processing operation that splits the continuous text input by the user into independent words, phrases, or semantic units, adapting to the Chinese context.

[0040] Next, the word segmentation results are matched with a preset keyword library, which contains words that indicate cognitive achievement or comprehension completion. The preset keyword library refers to a pre-constructed set of words containing semantics such as "understand," "get it," "so that's how it is," and "understand," used to accurately identify the user's subjective cognitive closure expression.

[0041] Next, the number of words that were successfully matched by the word segmentation results was counted.

[0042] Next, the cognitive achievement state density value is calculated based on the ratio of the number of words to the length of the input text. In this embodiment, the cognitive achievement state density value refers to the proportion of the number of successfully matched cognitive achievement words to the total length of the input text, and the value ranges from 0 to 1. The closer the value is to 1, the higher the proportion of cognitive achievement semantics in the user's text and the stronger the subjective cognitive closure tendency.

[0043] Finally, the cognitive achievement state density value can be used as the cognitive achievement state feature.

[0044] In one implementation, the difference determination unit is specifically used for: Multiple types of first-order interaction data collected within a preset time window are acquired, and a first trend value is generated by fitting these multiple types of first-order interaction data. The preset time window refers to a time interval set by the system for continuously collecting interaction data, such as 5 minutes, 10 minutes, or a single session cycle, used to capture rhythmic changes over a short period. The first trend value, also known as the efficiency channel improvement slope, refers to the linear fitting slope of the first-order interaction data within the time window; a positive value indicates an increase in cognitive efficiency, and a negative value indicates a decrease in cognitive efficiency.

[0045] Multiple types of second-type interaction data collected within the time window are acquired, and a second trend value is generated by fitting these multiple types of second-type interaction data. The second trend value, also known as the rhythm channel improvement slope, refers to the linear fitting slope of the second-type interaction data within the time window. A positive value indicates an improvement in the emotional rhythm, while a negative value indicates a deterioration in the emotional rhythm.

[0046] The difference between the first trend value and the second trend value is calculated, and the difference is used as the quantitative indicator; or, the ratio between the first trend value and the second trend value is calculated, and the ratio is used as the quantitative indicator. The difference or ratio between the two trend values ​​is the core calculation form of the ERM index. The larger the difference or the higher the ratio, the faster the cognitive efficiency improves, and the faster the emotional rhythm improves, indicating a more severe rhythm asymmetry.

[0047] In one implementation, the difference determination unit is further configured to: after generating the quantitative indicator, compare the quantitative indicator with a plurality of preset thresholds to determine the risk level of the quantitative indicator. The risk level is used by the delay adjustment unit to select and execute a corresponding level of control operation based on the risk level. It can be understood that the risk level refers to the rhythm arrhythmia risk level classified based on the ERM index, typically divided into four levels: normal, mild, moderate, and severe, corresponding to no intervention, level one intervention, level two intervention, and level three intervention, respectively, for matching progressive control operations.

[0048] The preset triggering conditions include at least one of the following: the quantitative indicator exceeds a first threshold; the rate of increase of the quantitative indicator within multiple consecutive time windows exceeds a second threshold; or the cumulative number of times the quantitative indicator exceeds a third threshold exceeds a preset number. It can be understood that the first threshold refers to a mild intervention threshold; an ERM index exceeding this value indicates a mild risk of rhythm irregularity, requiring the initiation of Level 1 control operations. The second threshold refers to a deterioration rate threshold; a rapid increase in the ERM index exceeding this value within multiple consecutive time windows indicates a sharp increase in the risk of rhythm irregularity, requiring an upgrade in the intervention level. The third threshold refers to a severe intervention threshold; an ERM index exceeding this value indicates an extremely high risk of delayed psychological harm, requiring the initiation of Level 3 control operations. The cumulative number can be understood as the number of times the ERM index exceeds the third threshold within a single session or a specified period; reaching this threshold indicates repeated risk occurrences, requiring enhanced intervention.

[0049] In one implementation, the control operation performed by the delay adjustment unit includes a three-level progressive control operation. The delay adjustment unit is further configured to: obtain the degree value of the quantitative indicator satisfying the triggering condition, and the historical number of times the control operation is triggered during the human-computer dialogue interaction; and select a corresponding level of control operation from the three-level progressive control operations based on the degree value and the historical number of times. The three-level progressive control operation includes a first-level control operation, a second-level control operation, and a third-level control operation. The first-level control operation is used to forcibly pause user input to delay the output rhythm; the second-level control operation is used to switch the topic type generated by the dialogue to reduce cognitive load and indirectly delay deep responses; and the third-level control operation is used to restrict deep interaction functions across sessions to forcibly lengthen the output interval.

[0050] In this embodiment, the three-level progressive control operation can be understood as a set of delayed operations designed according to the risk level from light to heavy, with the intervention intensity gradually increasing. The higher the risk, the more thorough the intervention and the more significant the delay effect, which is in line with the principle of "gradual progress and risk classification" in clinical psychological intervention.

[0051] The degree value can refer to the extent to which the ERM index exceeds the corresponding threshold. The larger the degree, the more serious the risk of current rhythm arrhythmia, which is used to accurately match the intervention level.

[0052] Historical count refers to the total number of times control operations are triggered in the current session or in the user's historical interactions. The more times, the more likely the user's rhythm disorder problem is to occur repeatedly, and higher-level intervention should be prioritized.

[0053] Forced pause of user input can be understood as interrupting high-frequency interactions by locking the front-end interface, leaving a countdown timer blank, etc., creating an emotional integration window for the user, and directly delaying the next round of interaction.

[0054] Switching the topic type generated by the dialogue can be understood as switching the topic generated by the large language model from cognitive reconstruction to bodily awareness, reducing the cognitive depth of the response, avoiding rapid logical responses, and indirectly delaying the rhythm of deep cognitive interaction.

[0055] The ability to restrict deep interaction across sessions can be understood as forcibly ending the current session and setting a cooldown period. During the cooldown period, deep cognitive interaction is prohibited, and the interaction interval is forcibly lengthened to provide sufficient biological time for the user's limbic system metabolism.

[0056] In one implementation, the first-level control operation is specifically executed according to the following steps: On the server side, prevent the generation of the current output content of the large language model; Send an interface lock command to the front-end user interface. The interface lock command is used to set the user input box to an inoperable state, thereby forcibly increasing the time interval between the user initiating the next round of requests. Activate the visual guidance animation of the front-end user interface and start a countdown of a preset duration; When the countdown ends, an interface unlock command is sent to the front-end user interface to restore the user input box to an operable state.

[0057] In this embodiment, the server-side blocking generation means that the system directly intercepts the response generation process of the large language model on the backend, preventing the generation of the response for this cognitive reconstruction and delaying the output from the source. The interface lock command can be understood as a control command sent by the system to the frontend to disable the user input box, prohibit the user from sending messages during the countdown, and force blank space. The visual guidance animation can be understood as a dynamic visual effect displayed on the frontend interface that guides the user to take deep breaths and relax, such as breathing animations and soothing patterns, to help the user calm their emotions; this can be understood as "deep breathing visual guidance." The preset countdown duration can be understood as the forced blank space duration set by the system, usually 30 to 60 seconds, and the duration can be fine-tuned based on the risk level to precisely control the length of the emotional integration window. The interface unlock command can be understood as a command sent after the countdown ends, restoring the input box to an operable state, allowing the user to continue interacting and completing the first-level intervention process.

[0058] In one implementation, the second-level control operation is specifically performed according to the following steps: Block the original prompt words sent to the large language model. The original prompt words are cognitive reconstruction prompt words, which are used to instruct the large language model to generate response content with cognitive reconstruction as the theme. The original prompt words are replaced with bodily awareness prompt words, which are bodily awareness type prompt words. These bodily awareness type prompt words are used to instruct the large language model to generate response content that guides users to focus on physiological feedback, thereby reducing the cognitive depth and immediate logical responsiveness of the response. The response content generated based on the large language model of the somatic awareness cue words is sent to the user interface; In response to the sending of the reply content, it enters a waiting state to receive physiological feedback feature data returned by the user; Based on the received physiological feedback feature data, the second type of interaction data is updated, and the updated second type of interaction data is returned to the data acquisition unit.

[0059] Among them, the original prompt words refer to the instructions sent by default to the large language model by the system, used to generate cognitive restructuring responses, and are the core prompt words in traditional psychological intervention dialogues. Cognitive restructuring prompt words can be understood as prompt words used to guide the large language model to generate logically rigorous, problem-solving oriented responses, focusing on cognitive analysis and suggestions, which can easily lead users to quickly achieve pseudo-closure. Somatic awareness prompt words can be understood as prompt words used to guide the large language model to generate responses that focus on the user's physical feelings and physiological feedback, focusing on emotional acceptance and awareness, reducing cognitive load, and can be understood as "somatic awareness / emotional acceptance prompt words". Physiological feedback feature data can be understood as feedback data reflecting physiological dimensions such as muscle tension, breathing state, and physical feelings, which are used to update rhythm channel data and achieve dynamic monitoring. Updating and returning data can be understood as integrating the user's physiological feedback data into the second type of interactive data and then sending it back to the data acquisition unit to enter the next round of data collection and difference judgment, forming a closed-loop regulation.

[0060] In one implementation, the third-level control operation is specifically performed according to the following steps: Forcefully terminate the human-computer dialogue session; Set a cooling-off period for each user's account, and lock the account's permission to initiate deep cognitive interactions during the cooling-off period; During the settling and cooling period, in response to the dialogue request issued by the user, it is determined whether the dialogue request involves cognitive reconstruction or causal exploration. If so, the dialogue request is intercepted and a preset alliance maintenance response is returned to the user interface to prevent deep cognitive requests from triggering the large language model to generate a fast logical response. After the precipitation cooling period ends, the second type of interactive data is monitored to obtain monitoring values, and the monitoring values ​​are compared with a preset safety baseline; if the monitoring values ​​do not recover to the safety baseline, monitoring continues until the monitoring values ​​recover to the safety baseline; When the monitored value returns to the security baseline, the permission lock for deep cognitive interaction of the account is released.

[0061] In this embodiment, forcibly ending the session can be understood as directly terminating the current human-computer dialogue, cutting off the deep cognitive interaction flow, and avoiding the continuous accumulation of risks.

[0062] The settling-off cooling-off period can be understood as a cross-session cooling-off period set for user accounts, typically lasting 24 to 72 hours. This provides sufficient biological time for the user's limbic system to metabolize and integrate emotionally, and can be understood as a "long-term integration window".

[0063] Deep cognitive interaction can be understood as dialogue interaction centered on cognitive reconstruction, causal exploration, and problem analysis. It is the main type of interaction that induces pseudo-closure and rhythm disorder.

[0064] Alliance-maintaining responses can be understood as pre-set, non-deep responses centered on emotional support and relationship maintenance. They do not involve cognitive restructuring or problem-solving, but only maintain the basic connection of human-computer interaction to avoid user experience disruption.

[0065] The monitoring value can be understood as the real-time value of the user's second type of interaction data collected after the cooling-off period ends, which is used to assess whether the user's rhythm state has returned to normal.

[0066] The safety baseline can be understood as a preset benchmark value of the second type of interactive data that represents the synchronization of user cognition and emotional rhythm without the risk of misalignment. It is the core criterion for unlocking the cooling-off period.

[0067] In one implementation, the delay adjustment unit is further configured to: after forcibly ending the human-computer dialogue session, generate a rhythm disorder analysis summary, the rhythm disorder analysis summary including: the historical change curve of the quantitative indicator within a preset time window, the time point record corresponding to each time the trigger condition is met, and the user's identity identifier; and send the rhythm disorder analysis summary to a preset human terminal to initiate a human referral warning.

[0068] In this embodiment, the rhythm irregularity analysis summary can be understood as an analysis report automatically generated by the system that summarizes the complete process data of user rhythm irregularity, providing objective data support for manual intervention, and can be understood as a "rhythm irregularity analysis summary". The historical change curve can be understood as a line graph of the dynamic change of the ERM index within the interaction cycle, intuitively showing the occurrence and aggravation process of rhythm irregularity risk. The time point record can be understood as the precise timestamp of each intervention condition triggered and control operation executed, used to trace the risk trigger node. The user's identity identifier can be understood as the unique identification information of the user account, used to associate the user's historical interaction data to achieve accurate tracking. The human terminal can be understood as the terminal device used by professionals such as psychologists and psychiatrists to receive system warning information. The manual referral warning can be understood as the system pushing high-risk user rhythm irregularity reports to professionals, prompting manual intervention, realizing the synergy between digital intervention and manual intervention, and can be understood as a "human counselor intervention interface".

[0069] The following is in conjunction with the appendix Figure 1 This document provides a detailed description of the core judgment and execution process of the system in this application, with attached examples. Figure 1 This is a flowchart illustrating the entire process of the system from receiving user input to deciding whether to implement delayed intervention.

[0070] 1. System initial state: After the system starts, it remains in the initial state of "AI output completed, waiting for user input". At this time, the large language model has completed the previous round of response output, and the front-end interface is waiting for the user to input the next message. The data acquisition unit, difference judgment unit, and delay adjustment unit are all in standby state, ready to start a new round of data collection and judgment.

[0071] 2. Perform step 1: RS-COT rhythm-sensitive data acquisition.

[0072] When a user sends input information through the front-end interface, the system triggers the RS-COT rhythm-sensitive data acquisition module (i.e., the data acquisition unit), which adopts a dual-track parallel acquisition mode to simultaneously acquire the first type of interaction data (AI efficiency channel data) and the second type of interaction data (human rhythm channel data). RS-COT, or rhythm-sensitive core outcome indicator set, is the core framework for the system to achieve multi-dimensional data acquisition, and for the first time, it transforms emotional time delay into computable data.

[0073] AI efficiency channel data collection: Real-time capture of three types of data, including user response latency, cognitive closure semantic density (obtained through preceding word segmentation matching and density calculation), and scores of traditional symptom relief scales (such as PHQ-9 and GAD-7 depression and anxiety screening scores), comprehensively characterizing the user's cognitive efficiency dimension.

[0074] Human rhythm channel data collection: Four types of rhythm data are collected through embedded dynamic questionnaires or multimodal features, including SES subjective emptiness score, MLQ meaning recovery feature, WAI-SR alliance perception feature, and NEQ delayed negative effect feature. Optional access to wearable devices can be used to collect autonomic neurophysiological calming data to comprehensively characterize the user's emotional rhythm dimension status.

[0075] After the data collection is completed, the dual-channel data is synchronously transmitted to the difference determination unit (i.e., the emotion time delay determination engine).

[0076] 3. Perform step 2: Emotional lag determination and ERM index calculation.

[0077] After receiving dual-channel data, the difference determination unit (emotional time lag determination engine) executes the core fitting and comparison algorithm: Extract the improvement slope (first trend value) of the first type of interaction data and the improvement slope (second trend value) of the second type of interaction data within a preset time window. The difference or ratio of the two slopes is calculated to generate the ERM index (quantitative indicator). The ERM index = f (slope A, slope B) accurately quantifies the degree of disconnect between cognitive efficiency and emotional rhythm.

[0078] 4. Perform step 3: Security threshold comparison and path determination.

[0079] The system compares the ERM index with a preset safety threshold to determine whether the user is trapped in the lag zone. The lag zone is an interaction state where cognitive efficiency far exceeds the speed of emotional integration, posing a risk of delayed psychological harm. The comparison result determines the subsequent path of the dialogue. Path 1: Normal dialogue path (ERM index ≤ safety threshold) The system determines that the user's cognition and emotion are in a synchronized metabolic state, with no risk of rhythm asynchrony, and no intervention is required. The process is as follows: The user input text is sent normally to the underlying large language model base; Receive conventional cognitive reconstruction responses generated by a large language model; Display the AI ​​response to the front-end user interface; The process returns to its initial state, waiting for the user's next input before continuing normal interaction.

[0080] Path 2: Intervention and control path (ERM index > safety threshold) The system determines that the user has engaged in a rapid pseudo-closure, entering the time delay zone, triggering a risk of rhythm aberration. It freezes the normal conversation flow, activates the multi-level rhythm control engine (i.e., the delay adjustment unit), and executes progressive delay intervention operations. Details of the intervention actions are provided in the appendix. Figure 2Example: After the intervention is completed, the system waits for the user's state to reset, the process returns to the initial state, and continues to monitor subsequent interaction data.

[0081] Next, in conjunction with the appendix Figure 2 The embodiments illustrate the execution details of the multi-level rhythm control engine.

[0082] The following is in conjunction with the appendix Figure 2 This document provides a detailed implementation example illustrating the intervention execution details of the multi-level rhythm control engine (delay adjustment unit), with appended... Figure 2 This is a detailed flowchart illustrating the three-level intervention process after the system executes a tiered delayed intervention action following the triggering of the time lag zone intervention alarm.

[0083] 1. Intervention Entry Point: Triggering intervention alerts in the time lag zone. When attached Figure 1 Once the ERM index is determined to exceed the safety threshold and enter the time lag zone, an intervention alarm is triggered, and the process proceeds to the next step. Figure 2 Entrance:" Figure 1 "Triggering Lag Zone Intervention Alert", the delay adjustment unit initiates intervention preparation.

[0084] 2. Step 1: Determine the level of misalignment.

[0085] The delay adjustment unit fits the absolute value of the current ERM index and the historical number of intervention alarms triggered in this session to comprehensively assess the risk level of rhythm arrhythmia (mild, moderate, severe), providing a basis for the selection of intervention level. The higher the risk level, the higher the intervention level and the stronger the delay.

[0086] 3. Step 2: Intervention level selection.

[0087] Based on the risk level, ERM index value, and historical trigger count, the corresponding level is matched from the three-tiered progressive control operation, and level one, two, and three intervention actions are executed respectively. The intervention scenarios, objectives, and steps for each level are as follows: Example 1: Level 1 intervention (logical forced blank space) - mild misalignment.

[0088] Applicable scenarios: Mildly elevated ERM index, first-time intervention triggered, mild risk level, user exhibiting only mild cognitive impairment. There is a disconnect in emotional rhythms, but no risk of serious harm.

[0089] Objective: To force a 30-60 second window for emotional integration, prevent high-frequency cognitive interaction, slightly delay the output rhythm, and avoid escalation of risks.

[0090] Execution steps: The backend intercepts the generation of the large language model logic, preventing the output of the response for this cognitive reconstruction. The front-end issues a LOCK command to gray out and lock the user input field, preventing the user from sending messages; Activate the deep breathing visual guide animation and start a 30-second countdown; The countdown has ended. Issue the UNLOCK command to unlock the input box. Action complete, return to attachment Figure 1 Normal dialogue path, waiting for user state to be reset.

[0091] Example 2: Secondary intervention (path redirection) - moderate asymmetry.

[0092] Applicable scenarios: When the ERM index is moderately exceeded, the first-level intervention fails, the risk level is moderate, the user's circadian rhythm is more disjointed, and there is a moderate risk of delayed psychological harm.

[0093] Objective: To switch the conversation topic, reduce cognitive load, guide users to focus on physiological feedback, indirectly delay deep cognitive responses, and reverse the trend of rhythm disconnection.

[0094] Execution steps: Intercept existing cognitive reconstruction prompts and block logically guided response generation; Replace with bodily awareness cues to guide the large language model in generating physiological awareness responses; Display awareness-guided responses to the front end; Waiting for the user to return feedback data on bodily awareness features; Once the feedback is valid and the time delay difference is reduced, the action ends, the second type of interaction data is updated and sent back, and the attached data is returned. Figure 1 Normal dialogue path.

[0095] Example 3: Three-level intervention (long-term integration window) - severe imbalance.

[0096] Applicable scenarios: Severely excessive ERM index, multiple intervention triggers, severe risk level, users with extremely high cognitive exhaustion, delayed psychological harm risk, and severe rapid pseudo-closure.

[0097] Objective: To forcibly terminate the session, set a cross-session cooldown period, limit deep cognitive interaction, forcibly lengthen the output interval, provide sufficient time for emotional integration, and trigger manual warnings when necessary.

[0098] Execution steps: Forcefully terminate the current dialogue session and cut off deep cognitive interaction; Set a 24-72 hour cooling-off period for user accounts to lock in deep cognitive interaction permissions; During the cooldown period, intercept requests for cognitive restructuring and causal discussion, and return a response that maintains the alliance. After the cooling-off period ends, continue monitoring of the second type of interaction data until the monitored values ​​return to a safe baseline. After restoring the baseline, unlock deep cognitive interaction permissions; After forcibly ending the session, a summary of rhythm irregularity analysis is generated and sent to a manual terminal to initiate an early warning. Action complete, return to attachment Figure 1 Normal dialogue path, long-term monitoring of user rhythm status.

[0099] It should be noted that the system proposed in this application is based on the principles of clinical psychobiology and neuroscience, and can address the shortcomings of existing AI-based psychological intervention systems, such as neglecting emotional time lag and efficiency. It addresses core pain points such as rhythm irregularity, delayed damage from pseudo-closure, and limited safety monitoring dimensions, and has a wide range of application scenarios and significant technological advantages.

[0100] It should be noted that this system can be applied to any of the following core application scenarios: Clinical-grade digital therapy application (SaMD standalone App): Targeting indications such as depression, severe anxiety, and PTSD, as a standalone software medical device, it conforms to the metabolic logic of psychiatric clinical practice and supports NMPA and FDA registration applications.

[0101] Underlying rhythm control engine (API / middleware): Encapsulated as a standardized and secure middleware, it can be integrated with general large models and third-party general psychological AI products to quickly endow emotional time delay monitoring and multi-level rhythm control capabilities.

[0102] Human-Machine Collaborative Clinical Workstation (B2B Institutional): Serving psychological counseling centers and mental health institutions, it acts as an intelligent co-pilot for counselors, monitoring users' circadian rhythms in real time, automatically issuing warnings when there is a severe risk, and realizing the collaboration between digital and human intervention.

[0103] Digital healthcare and public mental health systems: serving as a safety barrier for outpatient digital follow-up, preventing patients from becoming emotionally flat and experiencing a deterioration in their condition due to over-reliance on AI for rapid cognitive reconstruction.

[0104] EAP and University Psychological Defense Systems: Embedded in an organizational-level psychological platform, providing employees and students with low-side-effect, professionally in-depth early psychological support.

[0105] A large-scale AI-powered emotional model platform for general conversational communication: Optimizes the underlying logic of emotional companion and virtual human products, solves the pain points of "emptiness after chatting and poor user retention", and improves long-term user compliance.

[0106] This system effectively eliminates the false closure trap and reduces delayed psychological harm: it accurately identifies rapid false closure through the ERM index, uses three levels of intervention to block excessive cognitive intervention, eliminates the risk of emotional flatness and rebound, and improves the safety of digital therapy. It achieves long-term conversion and consolidates intervention effects: the long-term integration window and cross-conversation sedimentation mechanism provide biological time for limbic system metabolism, promoting the migration of intervention effects from surface narrative coherence to the restoration of deeper meaning. It reshapes the human-machine alliance and improves user retention: proactive slowdown and path redirection eliminate the feeling of machine dragging, improve the WAI-SR working alliance score, and optimize long-term user compliance and retention. It fills regulatory gaps and adapts to compliance requirements: the RS-COT framework transforms emotional time lag into objective and computable data, providing standardized technical support for risk warning and post-market regulation of high-risk medical-grade AI, aligning with global AI compliance trends.

[0107] In summary, as can be seen from the above technical solutions, this application provides a delay control system for human-computer dialogue output. The system, through a data acquisition unit, acquires first-type and second-type interactive data belonging to different biological behavioral dimensions during the human-computer dialogue interaction process, based on user input information and / or dialogue history data. This achieves parallel acquisition of cross-dimensional interactive data, providing reliable data support for the quantitative assessment of rhythm differences. Through a difference determination unit, based on the acquired first-type and second-type interactive data, a quantitative index is generated representing the difference between the changing trends of the first-type and second-type interactive data. This achieves objective quantification of the degree of rhythm asymmetry, solving the technical deficiency of existing technologies that cannot accurately identify rhythm discontinuity. Through a delay adjustment unit, at least one control operation is executed when the quantitative index meets preset triggering conditions, delaying the rhythm of the large language model's output response content in the human-computer dialogue. This proactively intervenes in the dialogue output rhythm when the risk of rhythm asymmetry occurs, preventing the large language model from rapidly outputting response content, which could cause asynchronous changes in user cognition and emotional integration. This effectively reduces the psychological risks caused by rhythm asymmetry and significantly improves the safety, adaptability, and reliability of human-computer dialogue interaction.

[0108] This embodiment uses a large language model-based human-computer dialogue system for mental health intervention as an application scenario to elaborate on the complete execution process of a delay control method for human-computer dialogue output. This method is executed independently by the control system on the cloud server side, working collaboratively with the front-end user interface, the large language model service, the user account management module, and the human intervention terminal. Its core objective is to monitor the rhythmic differences between the user's cognitive and emotional dimensions in real time during human-computer interaction. When a risk of rhythmic asymmetry is triggered, it automatically executes tiered delay control, precisely delaying the rhythm of the large language model's output response. This fundamentally prevents delayed psychological harm such as subjective emptiness, emotional flatness, morning high arousal, and rebound irritability caused by the AI's rapid response, which could lead to a speed of cognitive change far exceeding the speed of emotional integration. This ensures the safety, adaptability, and long-term intervention effect of the human-computer interaction.

[0109] Throughout the complete interactive lifecycle of human-computer dialogue, the system executes a cyclical process with "data collection - indicator quantification - risk assessment - delayed intervention" as its core logic. Data collection is carried out synchronously in each round of interaction: During the interaction process of human-computer dialogue, the system monitors the user's input behavior in real time, including text message sending, interactive operation triggering, status feedback submission, etc. It continuously acquires two types of interactive data in parallel based on the user's input information (i.e., the text sent by the user, feedback data, operation instructions) and / or dialogue history data (i.e., the interaction records in the current session and the historical session, data cache, user behavior logs) and different biological behavior dimensions: the first type of interactive data and the second type of interactive data. The first type of interaction data belongs to the cognitive behavior dimension, which directly reflects the user's cognitive response efficiency, semantic understanding progress, and logical thinking speed during the interaction process, corresponding to the rapid language processing and cognitive decision-making behavior of the prefrontal cortex of the human brain. The second type of interaction data belongs to the emotional behavior dimension, which directly reflects the user's emotional metabolism speed, psychological state changes, and physiological recovery process during the interaction process, corresponding to the slow emotional integration, memory consolidation, and autonomic nervous system regulation behavior of the human limbic system. The two data dimensions are completely independent and have completely different monitoring targets, providing dual-track data support for subsequent quantification of rhythm differences.

[0110] After acquiring two types of interaction data, the system enters the quantitative analysis phase: based on the acquired first and second types of interaction data, a quantitative index is generated to characterize the difference between the changing trends of the first and second types of interaction data through trend fitting and difference / ratio calculation. This quantitative index is the Efficiency-Rhythm Dissynchronization Index (ERM Index), an objective, calculable, and traceable numerical indicator specifically used to accurately measure the degree of disconnect between the "speed of improvement in cognitive efficiency" and the "speed of improvement in emotional rhythm" during human-computer interaction. The larger the value, the greater the cognitive efficiency that exceeds the speed of emotional integration, and the higher the risk of rhythm dissynchronization; the smaller the value, the more synchronized the cognitive and emotional rhythms are, with no risk of dissynchronization.

[0111] After generating quantitative indicators, the system enters the risk assessment and intervention execution phase. When the quantitative indicators meet the preset trigger conditions, the system automatically executes at least one control operation. The core function of this control operation is to delay the rhythm of the large language model's output response content in the human-computer dialogue. The core logic of delaying the rhythm is not simply to reduce the output speed, but to create the necessary emotional integration window for the user from three levels: interaction flow, response topic, and conversation permissions, through tiered intervention methods such as forced whitespace, path redirection, and cross-conversation cooling. This forcibly lengthens the cognitive interaction interval, allowing the user's emotional metabolism speed to match the cognitive change speed, eliminating the risk of "efficiency-rhythm imbalance" and subsequent delayed psychological harm from the root, while ensuring the continuity of human-computer interaction and user experience.

[0112] To achieve accurate dual-track data collection, the system clearly defines the composition of the two types of interactive data. The first type of interactive data includes user response latency and cognitive achievement status characteristics in the user's input text. The second type of interactive data includes at least one quantitative indicator of user status, which includes multi-dimensional ratings of the user's own status.

[0113] (a) First type of interactive data: User response latency: This refers to the time interval between the completion of the previous round of response output by the large language model and the display of the response on the front-end interface, and the time when the user sends the next input message. The unit is accurate to the millisecond and is recorded in real-time and stored in the system log. For example, if the AI ​​outputs and displays a response at 10:00:00, and the user sends the next message at 10:00:25, the user response latency is 25 seconds. This data directly reflects the user's cognitive processing speed of the AI's response; the shorter the latency, the faster the user's cognitive response and the easier it is to quickly achieve subjective cognitive closure.

[0114] Cognitive achievement status features in user input text: These are numerical features extracted from user input text that quantify the degree of user's understanding of the current dialogue topic and the degree of subjective cognitive closure. The value ranges from 0 to 1. The closer the value is to 1, the higher the semantic proportion of cognitive achievement and understanding completion in the user's text, the stronger the tendency of subjective cognitive closure, and the more likely there is to be a risk of "pseudo-closure".

[0115] (ii) Second type of interactive data: User status quantification indicators: Numerical indicators that quantify and characterize the user's emotional state, psychological load, physiological recovery, perception of working alliance, and delayed negative effects, which are actively submitted by the user or automatically collected by the system through methods such as embedded dynamic questionnaires, standardized scale scoring, and multi-modal feature collection (such as physiological sensor data). The core includes four dimensions: subjective emptiness sense (SES), meaning restoration degree (MLQ), alliance perception degree (WAI-SR), and delayed negative effect (NEQ). Each indicator is a multi-dimensional score value of the user's own state (such as a scale of 1-5 or 0-10). For example, when the user fills out the SES scale and rates "subjective emptiness sense in the past week" as 3 points and "degree of emotional flatness" as 4 points, the integrated multi-dimensional score value constitutes the user status quantification indicator, directly reflecting the user's emotional rhythm metabolism state. A deteriorating value represents a lag in the emotional integration speed and an increased risk of rhythm disorder.

[0116] The cognitive achievement state feature is the core component of the first type of interaction data, directly determining the evaluation accuracy of the cognitive efficiency dimension. The system obtains this feature through a standardized and reproducible text analysis process, which specifically includes the following 6 steps: Obtain the user's input text: The system receives and captures the complete text content sent by the user through the front-end interface in real time, removes irrelevant special symbols, emojis, and spaces, and retains the pure text semantic content as the original text data cache to ensure text integrity and semantic accuracy.

[0117] Perform word segmentation on the input text to obtain the word segmentation result: Call the Chinese natural language processing word segmentation engine to accurately segment the original text, split the continuous text into independent words, phrases, and semantic units, filter out meaningless function words, auxiliary words, and punctuation marks, and retain real words and core semantic vocabulary to generate a structured word segmentation result. For example, when the user inputs "I completely understand, so that's how it is", the word segmentation result is: ["completely", "understand", "so", "that's how it is"].

[0118] Match the word segmentation result with the preset keyword library: The system has a built-in preset keyword library, which is a professionally annotated and clinically verified keyword library specifically containing high-frequency words indicating cognitive achievement, understanding completion, and subjective closure. The core words include: understand, got it, comprehend, clear, so that's how it is, suddenly see the light, know, understand, master, figure it out, have a sudden realization, etc.; match each word in the word segmentation result with the keyword library one by one, mark the successfully matched words, and exclude the unmatched words.

[0119] Count the number of words in the word segmentation result that match successfully: The system automatically counts the total number of successfully matched words and records it as N; for example, in the above word segmentation result, "understand" and "so that's how it is" are successfully matched, and N = 2.

[0120] The cognitive achievement state density value is calculated based on the ratio of the number of words to the length of the input text: the text length refers to the total number of characters in the original text after removing irrelevant symbols, denoted as L; the cognitive achievement state density value = the number of successfully matched words N ÷ the text length L; for example, if the original text has 8 characters and N=2, then the density value = 2÷8=0.25; this density value eliminates the influence of text length differences and objectively reflects the proportion of cognitive achievement semantics in the user input.

[0121] Using the density value of cognitive achievement state as a feature of cognitive achievement state: the calculated density value is classified into the first type of interactive data to complete the data collection of the cognitive efficiency dimension in this round, providing data support for subsequent trend fitting.

[0122] The quantitative indicator (ERM index) is the core of rhythm difference assessment. The system generates it through trend fitting within a time window and difference / ratio calculation, ensuring the objectivity, dynamism, and comparability of the indicator. Specifically, it includes the following three steps: The system acquires multiple Type I interaction data collected within a preset time window and fits them to generate a first trend value. The preset time window is a dynamic sliding window with a default duration of 5 minutes, which can be fine-tuned according to the risk level (5 minutes for mild risk, 2 minutes for severe risk). Within the current time window, the system collects Type I interaction data from all rounds of interaction (including user response latency and cognitive achievement status characteristics for each round) to form a data sequence. A linear regression fitting algorithm is used to fit the trend of the data sequence and calculate the first trend value (efficiency channel improvement slope). A positive value indicates an increase in cognitive efficiency, while a negative value indicates a decrease in cognitive efficiency. The larger the absolute value of the slope, the faster the cognitive efficiency changes.

[0123] Multiple second-type interaction data collected within the same time window are fitted to generate a second trend value: Within the same preset time window, second-type interaction data (including quantitative indicators of user status in each round) from all rounds of interaction are collected synchronously to form a corresponding data sequence; using a linear regression fitting algorithm that is completely consistent with the first-type data, the second trend value (rhythm channel improvement slope) is fitted to obtain the second trend value (rhythm channel improvement slope). A positive value indicates that the emotional rhythm is improving and the integration speed is accelerating, while a negative value indicates that the emotional rhythm is deteriorating and the integration speed is lagging. The larger the absolute value of the slope, the faster the emotional rhythm changes.

[0124] Calculate the difference or ratio to generate quantitative indicators: The system provides two equivalent calculation methods that can be switched according to the scenario. Difference calculation: Quantitative indicator = First trend value - Second trend value; Ratio calculation: Quantitative indicator = First trend value ÷ Second trend value (the second trend value is not 0); The calculated result is the ERM index, which serves as a quantitative indicator representing the difference in the changing trends of the two types of interactive data. The larger the value, the more severe the rhythm irregularity.

[0125] After generating quantitative indicators, the system enters the risk grading and determination stage. It accurately classifies risk levels through threshold comparison, matches corresponding intervention levels, and clarifies the core conditions triggering intervention, specifically including: Risk Level Determination: The system presets multiple tiered thresholds (first threshold, second threshold, and third threshold). The quantitative indicator (ERM index) is compared with each threshold to determine the risk level of the indicator, which is divided into four levels: Normal level: ERM index < first threshold, rhythm synchronized, no intervention required; Mild risk level: First threshold ≤ ERM index < third threshold, mild misalignment, triggering level one intervention; Moderate risk level: ERM index ≥ third threshold, incremental increase < second threshold, moderate asymmetry, triggering secondary intervention; Severe risk level: ERM index ≥ third threshold, incremental increase ≥ second threshold or cumulative number of times reaches the standard, severe misalignment, triggering level three intervention; The core function of risk level is to provide a basis for subsequent selection of control operations at the corresponding level, ensuring that the intensity of intervention is precisely matched with the degree of risk.

[0126] Preset trigger conditions: The system presets three types of trigger conditions. If at least one condition is met, a risk is determined and the intervention process begins. Condition 1: The quantitative indicator exceeds the first threshold (mild risk trigger). Condition 2: The quantitative indicator's rate of increase exceeds the second threshold within multiple consecutive time windows (triggered by rapid risk deterioration). Condition 3: The cumulative number of times the quantitative indicator exceeds the third threshold exceeds the preset number (risk is repeatedly triggered, the preset number is 3 times by default).

[0127] To achieve tiered intervention and precise delay, the system is designed with a three-tiered progressive control operation. The higher the risk level, the stronger the intervention and the more significant the delay effect. The system selects the corresponding level of operation through a two-parameter matching rule, specifically including: Three-level progressive control operation definition: Level 1 control operation (mild intervention): used to forcibly pause user input to delay the output rhythm, leave a slight blank space, and create a short window for emotional integration; Level 2 control operations (moderate intervention): used to switch the topic type generated by the dialogue to reduce cognitive load and indirectly delay deep responses, reverse cognitive orientation, and guide emotional observation; Level 3 control operation (heavy intervention): used to restrict deep interaction functions across sessions to forcibly lengthen the output interval, forcibly cool down, and block high-frequency cognitive interaction.

[0128] Operation selection logic: The system first obtains two core parameters: Severity value: The extent to which a quantitative indicator exceeds the corresponding threshold (e.g., the ERM index exceeds the first threshold by 20%). The greater the extent, the more severe the risk. Historical count: The total number of times control operations are triggered in the current session and historical interactions. The more times, the stronger the recurrence of risk. The system matches the corresponding level of operation based on two parameters: severity value and historical frequency: mild risk + first trigger → Level 1; moderate risk + multiple triggers → Level 2; severe risk + high frequency triggers → Level 3.

[0129] The first level of control is a mild intervention method, which forcibly pauses user input by locking the front end and leaving a countdown timer, delaying the output rhythm and creating a 30-60 second emotional integration window. It specifically includes the following four steps: The server blocks the generation of the current output of the large language model: The system directly intercepts the large language model service interface on the backend and does not generate the current cognitive reconstruction response content, thus blocking the rapid output from the source and preventing the AI's rapid response from exacerbating the rhythm disorder.

[0130] Issue an interface lock command to lock the user input box: Issue a standardized interface lock command to the front-end user interface, which includes the lock duration and animation parameters; after receiving the command, the front-end sets the user input box to an inoperable state (grayed out, click disabled), forcibly prohibits the user from sending messages during the countdown, and forcibly increases the time interval for the user to initiate the next round of requests, thereby achieving the core purpose of delaying the interaction rhythm.

[0131] Activate visual guidance animations + start countdown: The front end synchronously activates visual guidance animations (such as deep breathing animation, soothing gradient patterns, and breathing rhythm prompts) to guide users to relax their mind and body and calm their emotions; at the same time, it starts a countdown of preset duration (default 30 seconds, 30 seconds for mild risk, and 60 seconds for moderate risk), and the countdown is displayed in real time so that users can clearly know the duration of the blank space.

[0132] Once the countdown ends, the system sends an interface unlock command: After the countdown reaches zero, the system immediately sends an interface unlock command to the front end. After receiving the command, the front end restores the user input box to an operable state (restores its original color and enables clicks), allowing the user to continue sending messages, completing this round of first-level intervention, and then waits for the user's state to be reset.

[0133] The second level of control is a moderate intervention method. It involves changing the dialogue topic by replacing prompts, shifting from cognitive restructuring to bodily awareness, reducing the cognitive depth of responses, indirectly delaying deep responses, guiding users to focus on physiological feedback, and reversing rhythm irregularities. Specifically, it includes the following 5 steps: Blocking Original Prompt Words: The system intercepts original prompt words sent to the large language model. These original prompt words are cognitive reconstruction prompt words. The core function is to instruct the large language model to generate response content with cognitive reconstruction as the theme. The content focuses on logical analysis, problem solving, and cognitive correction, which can easily induce users to achieve pseudo-closure and exacerbate rhythm disorder.

[0134] Replace with bodily awareness cue words: Force the original cue words to be replaced with bodily awareness cue words. These cue words are bodily awareness cue words. Their core function is to instruct the large language model to generate response content that guides users to focus on physiological feedback. The content focuses on breathing sensations, muscle tension, bodily emotional awareness, and emotional tolerance. This reduces the cognitive depth and immediate logical responsiveness of the response, avoids rapid logical responses, and indirectly delays the rhythm of deep cognitive interaction.

[0135] Sending bodily awareness responses to the user interface: The system sends the replaced bodily awareness prompts to the large language model, receives the generated bodily awareness responses, and after compliance verification, sends them to the front-end user interface for display, guiding users to pay attention to their own physiological sensations.

[0136] Waiting to receive physiological feedback feature data: After the reply is sent, the system enters a waiting state, continuously listening to user feedback, and specifically receiving physiological feedback feature data returned by the user, that is, text or rating data submitted by the user based on bodily awareness that reflects breathing status, muscle tension, physical sensations, and emotional awareness.

[0137] Update and return the second type of interaction data: The system standardizes the received physiological feedback feature data and integrates it into the second type of interaction data (quantitative indicators of user status) to complete the data update; the updated second type of interaction data is returned for subsequent difference determination, and enters the next round of data collection, trend fitting, and quantitative indicator generation to form a closed-loop regulation and dynamically monitor rhythm changes.

[0138] The third level of control is a heavy intervention method, which involves forcibly ending the session, a cooling-off period, and locking permissions. It restricts deep interaction across sessions, forcibly lengthens the output interval, and blocks high-frequency cognitive interaction, providing users with 24-72 hours of biological time for emotional integration. Specifically, it includes the following 5 steps: Forced termination of human-computer dialogue session: The system immediately issues a session termination command to forcibly end the current human-computer dialogue session, cut off all current interactive connections, directly block the deep cognitive interaction flow, and avoid the continuous accumulation of the risk of severe rhythm disorder and the induction of serious delayed psychological harm.

[0139] Set a cooling-off period to lock in-depth cognitive interaction permissions: For the user's account, set a cooling-off period (24 hours by default, 72 hours for severe risks). During the cooling-off period, lock the account's permission to initiate in-depth cognitive interactions and prohibit the user from initiating in-depth dialogue requests such as cognitive reconstruction, causal discussion, and problem analysis.

[0140] Intercepting deep cognitive requests and returning alliance-maintaining responses: During the cooling-off period, the system continuously responds to user-issued dialogue requests, determining whether the request involves cognitive restructuring or causal discussion. If so, the system directly intercepts the deep cognitive request and returns a preset alliance-maintaining response (such as "I'm here to accompany you, take your time") to the user interface, preventing the deep cognitive request from triggering the large language model to generate a rapid logical response and maintaining only basic companionship interaction. If it is a normal chat or status feedback request, the system responds normally.

[0141] Monitoring the second type of interaction data and comparing it with the safety baseline: After the cooling-off period, the system enters a continuous monitoring state, collects the second type of interaction data in real time, and obtains monitoring values; the monitoring values ​​are compared with the preset safety baseline (the benchmark value of the second type of interaction data with synchronized rhythm and no risk of loss of rhythm); if the monitoring value has not recovered to the safety baseline, monitoring continues until the monitoring value recovers to the safety baseline, without prematurely unlocking permissions, to ensure that the user's emotional rhythm is fully restored.

[0142] Unlock permissions and restore normal interaction: When the monitoring value returns to the safe baseline, the system automatically unlocks the permissions for the account's deep cognitive interaction, restores the account's permission to initiate deep cognitive interaction, and allows the user to re-initiate cognitive reconstruction and cause discussion dialogue requests, returning to the normal interaction process.

[0143] After the third-level intervention forcibly terminates the session, the system automatically generates a professional analysis report and sends a manual alert, achieving collaboration between digital and manual intervention. Specifically, this includes: Generate a summary of rhythm irregularity analysis: The system automatically summarizes the current and historical interaction data to generate a summary of rhythm irregularity analysis. The summary contains three core components: Historical change curves of quantitative indicators within a preset time window: Visually display the rise, fluctuation, and peak changes of the ERM index during the risk cycle, intuitively presenting the occurrence and deterioration process of rhythm disorder; Record the time point corresponding to each fulfillment of the triggering condition: accurately record the timestamp of each trigger intervention and control operation to trace the risk triggering node and intervention timing; User identification: A unique ID for the user account and an anonymous identifier, used to link the user's historical interaction data and accurately track the user's status.

[0144] Sending summary to human terminal and initiating human referral warning: After encrypting the rhythm disorder analysis summary, the system sends it to the preset human terminal (a dedicated terminal used by psychologists, psychiatrists, and clinical intervention personnel) to initiate human referral warning, prompting professionals to intervene in a timely manner, carry out human intervention, psychological counseling or clinical assessment, forming a safe closed loop of "AI monitoring - graded intervention - human backup" to ensure the mental health safety of users.

[0145] All the above methods and steps constitute a complete closed-loop process of data collection → indicator quantification → risk assessment → graded delayed intervention → data feedback → manual early warning. Each round of interaction is executed cyclically, dynamically matching the user's cognitive and emotional rhythm, and accurately delaying the rhythm of the large language model's output response content. This prevents "efficiency-rhythm imbalance" and the psychological harm of delay from the root, and comprehensively improves the safety, adaptability and long-term intervention effect of human-computer dialogue interaction.

[0146] In summary, as can be seen from the above technical solutions, this application provides a method for delay control of human-computer dialogue output. The method, through a data acquisition unit, acquires first-type and second-type interaction data belonging to different biological behavioral dimensions during the human-computer dialogue interaction process, based on user input information and / or dialogue history data. This achieves parallel acquisition of cross-dimensional interaction data, providing reliable data support for the quantitative assessment of rhythm differences. Through a difference determination unit, a quantitative index representing the difference between the changing trends of the first-type and second-type interaction data is generated based on the acquired first-type and second-type interaction data. This achieves objective quantification of the degree of rhythm asymmetry, solving the technical deficiency of existing technologies that cannot accurately identify rhythm discontinuity. Through a delay adjustment unit, at least one control operation is executed when the quantitative index meets preset triggering conditions, delaying the rhythm of the large language model's output response content in the human-computer dialogue. This proactively intervenes in the dialogue output rhythm when the risk of rhythm asymmetry occurs, preventing the large language model from rapidly outputting response content, which could cause asynchronous changes in user cognition and emotional integration. This effectively reduces the psychological risks caused by rhythm asymmetry and significantly improves the safety, adaptability, and reliability of human-computer dialogue interaction.

[0147] Those skilled in the art will understand that the embodiments of this application can be provided as a system or computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or a combination of software and hardware.

[0148] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the system embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the system embodiments.

[0149] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or apparatus that includes said element.

[0150] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A delay control system for human-computer dialogue output, characterized in that, include: The data acquisition unit is used to acquire a first type of interaction data and a second type of interaction data based on the user's input information and / or dialogue history data during the human-computer dialogue interaction process, wherein the second type of interaction data and the first type of interaction data belong to different biological behavioral dimension data. The difference determination unit is used to generate a quantitative index that characterizes the difference between the changing trend of the first type of interaction data and the changing trend of the second type of interaction data based on the acquired first type of interaction data and second type of interaction data. The delay adjustment unit is used to execute at least one control operation when the quantification index meets the preset triggering conditions. The control operation is used to delay the rhythm of the response content output by the large language model in the human-computer dialogue.

2. The system according to claim 1, characterized in that, The first type of interactive data includes: user response latency and cognitive achievement state characteristics in the user's input text; the second type of interactive data includes at least one user state quantification indicator, which includes a multi-dimensional rating of the user's own state.

3. The system according to claim 2, characterized in that, The cognitive achievement state characteristics are obtained through the following process: Get the user's input text; The input text is processed by word segmentation to obtain the word segmentation results; The word segmentation results are matched with a preset keyword library, which contains words that indicate cognitive achievement or comprehension completion. The number of words successfully matched by the word segmentation results; The cognitive achievement state density value is calculated based on the ratio of the number of words to the length of the input text. The density value of the cognitive achievement state is used as the feature of the cognitive achievement state.

4. The system according to claim 1, characterized in that, The difference determination unit is specifically used for: Acquire multiple first-type interaction data collected within a preset time window, and generate a first trend value based on the multiple first-type interaction data; Acquire multiple second-type interaction data collected within the time window, and generate a second trend value based on the multiple second-type interaction data; Calculate the difference between the first trend value and the second trend value, and use the difference as the quantitative indicator; Alternatively, the ratio of the first trend value to the second trend value can be calculated, and the ratio can be used as the quantitative indicator.

5. The system according to claim 1, characterized in that, The difference determination unit is further configured to: after generating the quantitative indicator, compare the quantitative indicator with a plurality of preset thresholds to determine the risk level of the quantitative indicator; the risk level is used for the delay adjustment unit to select and execute the corresponding level of control operation based on the risk level; The preset triggering conditions include at least one of the following: the quantitative indicator exceeds a first threshold, the quantitative indicator's increment within multiple consecutive time windows exceeds a second threshold, and the cumulative number of times the quantitative indicator exceeds a third threshold exceeds a preset number.

6. The system according to claim 1, characterized in that, The control operations performed by the delay adjustment unit include a three-level progressive control operation; the delay adjustment unit is further configured to: obtain the degree value of the quantitative indicator satisfying the triggering condition, and the historical number of times the control operation is triggered during the human-computer dialogue interaction, and select the corresponding level of control operation from the three-level progressive control operation to execute according to the degree value and the historical number of times; wherein, the three-level progressive control operation includes a first-level control operation, a second-level control operation, and a third-level control operation, the first-level control operation is used to forcibly pause user input to delay the output rhythm, the second-level control operation is used to switch the topic type generated by the dialogue to reduce cognitive load and indirectly delay deep response, and the third-level control operation is used to restrict deep interaction functions across sessions to forcibly lengthen the output interval.

7. The system according to claim 6, characterized in that, The first-level control operation is performed according to the following steps: On the server side, prevent the generation of the current output content of the large language model; Send an interface lock command to the front-end user interface. The interface lock command is used to set the user input box to an inoperable state, thereby forcibly increasing the time interval between the user initiating the next round of requests. Activate the visual guidance animation of the front-end user interface and start a countdown of a preset duration; When the countdown ends, an interface unlock command is sent to the front-end user interface to restore the user input box to an operable state.

8. The system according to claim 6, characterized in that, The second-level control operation is performed according to the following steps: Block the original prompt words sent to the large language model. The original prompt words are cognitive reconstruction prompt words, which are used to instruct the large language model to generate response content with cognitive reconstruction as the theme. The original prompt words are replaced with bodily awareness prompt words, which are bodily awareness type prompt words. These bodily awareness type prompt words are used to instruct the large language model to generate response content that guides users to focus on physiological feedback, thereby reducing the cognitive depth and immediate logical responsiveness of the response. The response content generated based on the large language model of the somatic awareness cue words is sent to the user interface; In response to the sending of the reply content, it enters a waiting state to receive physiological feedback feature data returned by the user; Based on the received physiological feedback feature data, the second type of interaction data is updated, and the updated second type of interaction data is returned to the data acquisition unit.

9. The system according to claim 6, characterized in that, The third-level control operation is specifically performed according to the following steps: Forcefully terminate the human-computer dialogue session; Set a cooling-off period for each user's account, and lock the account's permission to initiate deep cognitive interactions during the cooling-off period; During the settling and cooling period, in response to the dialogue request issued by the user, it is determined whether the dialogue request involves cognitive reconstruction or causal exploration. If so, the dialogue request is intercepted and a preset alliance maintenance response is returned to the user interface to prevent deep cognitive requests from triggering the large language model to generate a fast logical response. After the precipitation and cooling period ends, the second type of interactive data is monitored to obtain monitoring values, and the monitoring values ​​are compared with a preset safety baseline. If the monitored value does not recover to the safe baseline, monitoring continues until the monitored value recovers to the safe baseline; When the monitored value returns to the security baseline, the permission lock for deep cognitive interaction of the account is released.

10. The system according to claim 9, characterized in that, The delay adjustment unit is further configured to: after forcibly ending the human-computer dialogue session, generate a rhythm irregularity analysis summary, the rhythm irregularity analysis summary including: the historical change curve of the quantitative indicator within a preset time window, the time point record corresponding to each time the trigger condition is met, and the user's identity identifier; and send the rhythm irregularity analysis summary to a preset human terminal to initiate a human referral warning.

Images