Real-time online conversation ai tracking and feedback system and method based on emotion recognition model

Abstract

This invention proposes a real-time online dialogue AI tracking and feedback system and method based on an emotion recognition model. It belongs to the interdisciplinary fields of artificial intelligence, affective computing, and human-computer interaction. The method includes: synchronously collecting speech and text content, speech prosody features, and micro-expression dynamic data from user dialogues using a lightweight multimodal perception module to generate a multimodal emotion perception data stream; and constructing a dynamic emotion and intent graph based on the multimodal emotion perception data stream, where nodes represent emotional states and intent types, and edge weights reflect the strength of the correlation between the two. Constructing the dynamic emotion and intent graph using lightweight multimodal perception can comprehensively and accurately capture user emotions and intents, improving the accuracy of emotion recognition and the depth of intent understanding, and avoiding misjudgments caused by the limitations of a single modality.

Description

Technical Field

[0001] This invention proposes a real-time online dialogue AI tracking and feedback system and method based on an emotion recognition model, belonging to the interdisciplinary fields of artificial intelligence, affective computing and human-computer interaction. Background Technology

[0002] In the field of human-computer interaction, while real-time online dialogue AI has made some progress, existing technologies have many limitations. Most systems focus only on text sentiment analysis, paying insufficient attention to key non-verbal cues such as micro-expressions and speech rhythm, resulting in incomplete emotion perception and difficulty in accurately grasping the user's true emotional state.

[0003] At the same time, the system's handling of the connection between emotion and intent is weak, often resulting in a disconnect between emotion labels and the user's true intentions. For example, a user may still want to cooperate when they are angry, but the system cannot recognize this complex relationship, leading to misaligned feedback and affecting the effectiveness of the conversation.

[0004] Feedback mechanisms are mostly passive and static, relying on a pre-set script library. They cannot dynamically correct their own mistakes or guide the user's emotions based on the dialogue context, making them appear rigid and inflexible.

[0005] Furthermore, existing technologies lack causal thinking, failing to trace the triggers of users' negative emotions. They can only describe phenomena and cannot answer crucial questions such as "which sentence caused the user to get angry?" Moreover, most published patents describe highly homogenized functions, failing to deeply integrate emotion-intention dynamic graphs, counterfactual reasoning, and ethical constraint reinforcement learning to construct an intelligent dialogue architecture with a cognitive closed loop of "perception-attribution-correction-empathy." Therefore, there is an urgent need for a novel dialogue tracking and feedback method that breaks with convention and deeply integrates causal reasoning, dynamic graphs, and ethical AI to achieve advanced human-machine empathy. Summary of the Invention

[0006] This invention provides a real-time online dialogue AI tracking and feedback system and method based on an emotion recognition model, to solve the problems mentioned in the background art above: The present invention proposes a real-time online dialogue AI tracking and feedback method based on an emotion recognition model, the method comprising: S1. Simultaneously collect voice and text content, voice prosody features and micro-expression dynamic data in user dialogue through a lightweight multimodal perception module to generate a multimodal emotion perception data stream; construct an emotion and intention dynamic graph based on the multimodal emotion perception data stream, where nodes represent emotional state and intention type, and edge weights reflect the correlation strength between the two. S2. Use the counterfactual reasoning engine to perform causal tracing analysis on the dynamic graph of emotions and intentions. By comparing the differences between the actual dialogue path and the virtual correction path, locate the key dialogue nodes that trigger users' negative emotions and generate emotional trigger causal chain data. S3. Generate differentiated response strategies based on the causal chain data of emotional triggers, and finally output multi-type response strategy data; S4. Input multi-type response strategy data into the ethical constraint reinforcement learning model, and perform security verification and value alignment on the response content through a preset ethical rule base, filter out statements with bias, discrimination or privacy leakage risks, and generate optimized response data. S5. Based on the closed-loop interaction of optimized response data and real-time user feedback, dynamically adjust the edge weight parameters of the emotion-intent dynamic graph, iteratively optimize the causal judgment accuracy of the counterfactual reasoning engine, and update the rule base of the ethical constraint reinforcement learning model, ultimately generating a continuously evolving dialogue AI tracking and feedback system.

[0007] The real-time online dialogue AI tracking and feedback system based on an emotion recognition model proposed in this invention includes: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0008] The beneficial effects of this invention are as follows: By constructing dynamic graphs of emotions and intentions through lightweight multimodal perception, it can comprehensively and accurately capture user emotions and intentions, improving the accuracy of emotion recognition and the depth of intention understanding, and avoiding misjudgments caused by the limitations of a single modality. Utilizing counterfactual reasoning to trace the triggers of emotions can accurately locate key dialogue nodes that cause negative user emotions, reducing the probability of user dissatisfaction due to emotional misunderstandings and enhancing the relevance of dialogue. Generating empathic / corrective / guided responses can not only meet users' emotional needs but also promptly correct system errors, guide the direction of dialogue, reduce ineffective communication, and improve dialogue efficiency. Through ethically constrained reinforcement learning filtering, it can ensure that the response content complies with ethical norms, avoiding biased, discriminatory, or other inappropriate remarks, and preventing user resentment or social controversy. This method can achieve high-level human-computer empathy, providing users with a considerate dialogue experience, while also ensuring the security and credibility of dialogue, promoting the development of real-time online dialogue AI towards a more intelligent and human-centered direction. Attached Figure Description

[0009] Figure 1 This is a diagram illustrating the steps of the method described in this invention. Detailed Implementation

[0010] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0011] One embodiment of the present invention, such as Figure 1As shown, a real-time online dialogue AI tracking and feedback method based on an emotion recognition model is described, the method comprising: S1. Simultaneously collect voice and text content, voice prosody features and micro-expression dynamic data in user dialogue through a lightweight multimodal perception module to generate a multimodal emotion perception data stream; construct an emotion and intention dynamic graph based on the multimodal emotion perception data stream, where nodes represent emotional state and intention type, and edge weights reflect the correlation strength between the two. S2. Use the counterfactual reasoning engine to perform causal tracing analysis on the dynamic graph of emotions and intentions. By comparing the differences between the actual dialogue path and the virtual correction path, locate the key dialogue nodes that trigger users' negative emotions and generate emotional trigger causal chain data. S3. Generate differentiated response strategies based on emotional trigger causal chain data: When abnormal user emotions are detected, activate the empathy response mode to generate emotionally resonant words; when the system's own error is identified, trigger the correction response mode to generate targeted apology statements; when the user's intention tendency is predicted, activate the guidance response mode to generate suggestive dialogue content, and finally output multi-type response strategy data. S4. Input multi-type response strategy data into the ethical constraint reinforcement learning model, and perform security verification and value alignment on the response content through a preset ethical rule base, filter out statements with bias, discrimination or privacy leakage risks, and generate optimized response data that conforms to ethical norms. S5. Based on the closed-loop interaction of optimized response data and real-time user feedback, dynamically adjust the edge weight parameters of the emotion-intent dynamic graph, iteratively optimize the causal judgment accuracy of the counterfactual reasoning engine, and update the rule base of the ethical constraint reinforcement learning model, ultimately generating a continuously evolving dialogue AI tracking and feedback system.

[0012] The working principle and effects of the above technical solution are as follows: Through multimodal data acquisition and dynamic graph construction, the accuracy of dialogue AI in recognizing user emotions and intentions is improved, enhancing the relevance of response content. Counterfactual reasoning is used to locate key nodes of negative emotions, reducing invalid responses and lowering the probability of escalating negative emotions. Differentiated response modes balance empathy, correction, and guidance, alleviating negative emotions while accurately matching user needs and improving the dialogue experience. Ethical constraints are used to filter risky statements, avoiding bias, discrimination, and privacy leaks, thus enhancing dialogue security. Closed-loop iterative optimization continuously improves system performance, preventing issues such as decreased recognition accuracy and response lag after long-term use, further improving system stability and enabling the continuous evolution of dialogue AI.

[0013] In one embodiment of the present invention, S1 includes: S11. The lightweight multimodal perception module synchronously collects the voice and text content, voice prosody features and micro-expression dynamic data in the user's dialogue to generate a multimodal emotion perception data stream. S12. Perform time-series synchronization calibration on the multimodal emotion perception data stream to eliminate the acquisition delay difference of text, voice and micro-expression data, and generate time-series synchronized emotion perception data. S13. Perform noise filtering and feature purification on the time-synchronized emotion perception data, remove invalid data caused by environmental interference and collection errors, retain core emotion-related features, and generate time-normalized emotion perception data. S14. Construct a dynamic graph of emotion and intention based on time-series normalized emotion perception data, where nodes represent emotional states and intention types, and edge weights reflect the strength of their correlation. S15. Capture changes in emotion and intent during user dialogue in real time, dynamically update node information and edge weight parameters of the emotion and intent dynamic graph, and generate real-time updated emotion and intent association data.

[0014] The working principle and effects of the above technical solution are as follows: A lightweight multimodal perception module synchronously collects text, speech, and micro-expression data, improving the perception dimension of user emotions and intentions by the dialogue AI and enhancing the comprehensiveness of recognition. Time-series synchronous calibration eliminates acquisition latency differences, reduces matching errors between different modalities, and improves the accuracy of emotion perception data. Noise filtering and feature purification remove invalid data, retaining core emotion-related features, reducing the interference of redundant information on analysis results, and improving the purity of emotion perception data. A dynamic emotion and intention graph is constructed based on time-series normalized data, intuitively presenting the correlation between emotional states and intention types, avoiding recognition bias caused by disordered data timelines. Real-time capture of emotion and intention changes during the dialogue process and dynamic updates of the graph parameters improve the real-time performance and effectiveness of emotion and intention correlation data, avoiding response mismatch problems caused by information update lag.

[0015] In one embodiment of the present invention, S15 includes: Real-time capture of changes in user emotions and intentions during conversations, extraction of emotional fluctuation characteristics and intention shift information during conversations, and generation of dynamic emotional and intention change data; Based on the dynamic changes in emotion and intention data, update the node information of the dynamic graph of emotion and intention, add new emotion states and intention types, delete invalid nodes, and generate a dynamic graph with updated nodes. Calculate the magnitude of the correlation between emotional state and intention type, adjust the edge weight parameters of the dynamic graph, strengthen strong correlations, weaken weak correlations, and generate a dynamic graph with updated weights. By integrating the dynamic graph information after node updates and weight updates, a comprehensive dynamic update of the emotion and intention dynamic graph is completed, generating real-time updated emotion and intention association data. The real-time updated emotion and intention association data is time-stamped, and the data update time and corresponding dialogue node are recorded to generate emotion and intention association data with time stamps.

[0016] The working principle and effects of the above technical solution are as follows: Real-time capture of changes in user emotions and intentions during dialogues; extraction of relevant feature information; improvement of the completeness of dynamic emotion and intention change data; reduction of omissions of key change information. Update of dynamic graph node information, supplementing newly added states and deleting invalid nodes, avoiding recognition bias caused by redundancy or missing graph nodes, and enhancing the applicability of the graph. Adjustment of edge weight parameters, strengthening strong correlations and weakening weak correlations, improving the accuracy of emotion and intention association relationships and reducing the probability of misjudgment. Integration of node and weight update information to complete a comprehensive graph update, improving the real-time nature of emotion and intention association data and avoiding untimely response problems caused by information lag. Time-series labeling of associated data, recording update time and corresponding dialogue nodes, facilitating subsequent traceability analysis, reducing inconvenience caused by data chaos, and further improving the efficiency of subsequent causal tracing and response optimization.

[0017] In one embodiment of the present invention, S2 includes: S21. Utilize the counterfactual reasoning engine to retrieve real-time updated emotion and intention correlation data, and conduct a comprehensive causal tracing analysis on the dynamic graph of emotions and intentions. S22. Extract dialogue flow feature information from the emotion and intention association data, sort out the evolution trajectory of emotional state and intention type during the dialogue, and generate dialogue emotion evolution time series data. S23. Construct a virtual corrected dialogue path, and by comparing the differences in emotional changes between the actual dialogue path and the virtual corrected path, quantify the degree of influence of each dialogue node on emotional fluctuations and generate quantitative data on emotional impact. S24. Based on quantitative data on the impact of emotions, locate the key dialogue nodes that trigger negative emotions in users, sort out the transmission relationship between key nodes and negative emotions, and generate emotional trigger causal chain data. S25. Perform hierarchical analysis of the causal chain data of emotional triggers, distinguish between core and secondary triggers, clarify the scope of influence and transmission intensity of each trigger, and generate structured emotional trigger source data.

[0018] The working principle and effects of the above technical solution are as follows: It utilizes a counterfactual reasoning engine to conduct comprehensive causal source analysis, combined with real-time updated emotional intent correlation data, to improve the comprehensiveness of emotional trigger analysis and reduce source omissions. It extracts dialogue flow characteristics, sorts out the trajectory of emotional evolution, generates time-series data, clearly presents the patterns of emotional changes, and avoids misjudgments of triggers due to chaotic trajectories. It constructs virtual correction paths and compares differences, quantifies the degree of emotional influence at each node, enhances the accuracy of trigger analysis, and reduces subjective judgment errors. It locates key nodes of negative emotions, sorts out transmission relationships, generates causal chain data, quickly identifies core issues, and reduces ineffective investigation time. It performs layered processing of causal chain data, distinguishes between core and secondary triggers, generates structured source data, avoids trigger confusion, facilitates subsequent targeted response strategy development, improves response efficiency, and further optimizes the dialogue interaction experience.

[0019] In one embodiment of the present invention, step S23 includes: By combining the time-series data of dialogue emotion evolution, a virtual corrected dialogue path is constructed to simulate the emotion evolution process after adjustment at different dialogue nodes, and virtual dialogue emotion time-series data is generated. Extract emotion change data corresponding to the actual dialogue path, compare it node by node with the virtual dialogue emotion time series data, and generate emotion change difference data. Based on the data on differences in emotional changes, the magnitude of the impact of each dialogue node on emotional fluctuations is calculated, the degree of impact is quantified, and a quantitative value of node impact is generated. Integrate the quantitative values ​​of the impact of all dialogue nodes, label the emotional impact weight of each node, and form a complete quantitative dataset of emotional impact. Redundancy is removed from the quantitative dataset of emotion impact, and core quantitative data is retained to finally generate quantitative data of emotion impact.

[0020] The working principle and effects of the above technical solution are as follows: A virtual, corrected dialogue path is constructed by combining time-series data of dialogue emotion evolution. This simulates the emotion evolution after node adjustments, generating virtual time-series data. This avoids the one-sidedness of analyzing a single actual path and enhances the comprehensiveness of emotion impact analysis. Actual dialogue emotion change data is extracted and compared node-by-node with the virtual data to generate difference data. This accurately captures the differences in the impact of each node on emotion, reducing missed or false judgments. Based on the difference data, the node impact amplitude is calculated, completing the quantification process and improving the accuracy of emotion impact analysis while reducing subjective assessment errors. All node quantification values ​​are integrated and weighted to form a complete dataset, avoiding analytical chaos caused by fragmented data and improving data usability. Redundant data is removed, and core information is retained to reduce interference from invalid data in subsequent analysis. This ensures the simplicity of the quantified data and improves the efficiency of locating key nodes, providing reliable support for accurately tracing the causes of negative emotions.

[0021] In one embodiment of the present invention, S3 includes: S31. Extract the core trigger feature information from the structured emotional trigger source data, and combine it with the dialogue scenario and user historical interaction data to complete the comprehensive judgment of the dialogue scenario and user state. S32. When an abnormal user emotion is detected, the empathy response mode is activated, and emotionally resonant words that fit the user's current emotional state and the dialogue scenario are generated. S33. When a system error is detected, a correction response mode is triggered to accurately match the error type and its impact, and generate a targeted apology statement and error correction instructions. S34. When the user's intention tendency is predicted, activate the guided response mode and generate suggestive dialogue content that is instructive and practical, based on the user's intention type and needs preferences. S35. Integrate the dialogue output content of the three modes of empathy, correction, and guidance, optimize the fluency and adaptability of sentence expression, and finally output multi-type response strategy data.

[0022] The working principle and effects of the above technical solution are as follows: Extracting core triggering features and combining them with scenario and historical data to make a comprehensive judgment improves the targeting of dialogue responses and reduces the sense of alienation caused by indiscriminate responses. Activating the empathic response mode generates emotionally appropriate language to alleviate users' negative emotions and prevent the continued deterioration of emotions from leading to communication conflicts. Triggering the corrective response mode outputs corresponding apologies and explanations, reducing user dissatisfaction caused by system errors and improving the credibility of the interaction. Activating the guided response mode produces practical suggestions to meet users' real needs and enhance the practical value of the dialogue. Integrating the output content of multiple modes and optimizing the fluency of expression makes the response more appropriate to the scene and atmosphere, avoiding stiff sentences or logical gaps. This not only adapts to different dialogue situations but also improves the overall dialogue experience, allowing users to feel more humanized interactive feedback.

[0023] In one embodiment of the present invention, step S4 includes: S41. Input multi-type response strategy data into the ethical constraint reinforcement learning model, retrieve the built-in ethical rule library of the model, and perform security and compliance verification on each response content; S42. Based on the core requirements of the ethical rule base, conduct value alignment and adaptation processing on response statements to filter out risky statements that contain bias, discrimination, privacy leaks or do not conform to public order and good morals. S43. Optimize the expression and adjust the tone of the retained compliant statements to make them fit the user's current emotional state and dialogue scenario, thereby improving the naturalness and comfort of the dialogue. S44. Perform a second compliance check on the optimized response data to ensure that no risky statements are omitted and generate optimized response data that complies with ethical standards. S45. Sort the output time sequence of the optimized response data, and generate orderly output ethical compliance response data by combining the rhythm of user dialogue and changes in emotion.

[0024] The working principle and effects of the above technical solution are as follows: A reinforcement learning model based on ethical constraints is used to perform security and compliance checks on response content, improving the security level of dialogue output and preventing risky content from entering the interaction process. Value alignment and adaptation are applied to response statements to filter out biased, discriminatory, and privacy-leaking content, reducing the negative risks of harmful information dissemination and enhancing the positive guidance of the interaction process. The tone and expression of compliant statements are optimized to match user emotions and dialogue scenarios, improving the naturalness of the dialogue and avoiding stiff expressions that reduce user experience. The optimized response data undergoes secondary verification to further reduce the probability of missing risky statements, ensuring that the output content continuously complies with ethical norms. The output content is sorted according to the rhythm of the dialogue and changes in emotion, ensuring both logical coherence and smoothness in the response, and making the interaction rhythm more in line with user experience, improving the overall comfort and credibility of the dialogue.

[0025] In one embodiment of the present invention, S45 includes: Extract sentence features from the optimized response data, match them with user dialogue rhythm features, and generate dialogue rhythm matching data. Based on the real-time emotional changes of users, the output sequence of response data is divided into hierarchical levels, and response hierarchy data is generated. Based on the response hierarchy, the data is arranged in the order of response statements to complete the initial arrangement of the output timing and generate the initial timing response data; Perform connection verification on the initial timing response data to eliminate jump conflicts between statements and generate seamless timing response data; The validated time-series response data is integrated to form a complete output sequence, ultimately generating ordered ethical compliance response data.

[0026] The working principle and effects of the above technical solution are as follows: Extracting sentence features and matching them to the user's dialogue rhythm improves the fit between the response output and the user's interaction rhythm, avoiding discomfort caused by content being pushed too quickly or too slowly. Combining real-time emotion to classify output levels enhances the relevance of the response content and reduces interference from irrelevant information on core communication. Arranging the sentence order according to hierarchy makes the output logic more aligned with the dialogue process, reducing comprehension barriers caused by content confusion. Performing connection checks on time-series data eliminates sentence jump conflicts, avoids abrupt breaks in the dialogue, and improves the smoothness of interaction. Integrating to form a complete output sequence ensures both the orderly delivery of ethically compliant content and a natural and coherent overall dialogue flow, further enhancing the user's interactive experience and communication feel.

[0027] In one embodiment of the present invention, step S5 includes: S51. Push the orderly output of ethical compliance response data to the user according to the preset time sequence, and simultaneously collect the user's real-time interactive feedback information, including dialogue reply content, emotional changes and operation behavior. S52. Combine real-time user feedback information with optimized response data to form closed-loop interactive data, analyze the matching degree between feedback results and response content, and dynamically adjust the edge weight parameters of the emotional intent dynamic graph. S53. Using the updated edge weight parameters, iteratively optimize the causal judgment algorithm of the counterfactual reasoning engine to improve the accuracy of locating key emotional triggers and the precision of causal analysis. S54. Based on the feedback information in the closed-loop interaction data, the rule base and judgment threshold of the ethical constraint reinforcement learning model are updated synchronously to improve the ethical verification standards. S55. Repeat the above closed-loop optimization process to continuously iterate and optimize the overall emotion recognition, causal judgment and response output capabilities of the system, and finally generate a continuously evolving dialogue AI tracking and feedback system.

[0028] The working principle and effects of the above technical solution are as follows: It pushes compliant response data and simultaneously collects real-time user feedback, improving the system's ability to perceive interactive effects and reducing one-way outputs that are detached from actual user experience. By combining feedback to form closed-loop data and adjusting graph weight parameters, it enhances the accuracy of emotion and intent correlation judgments, reducing the probability of continuous misjudgments. It optimizes the causal judgment algorithm by updating parameters, improving the accuracy of cause localization and analysis, and avoiding source tracing bias caused by model rigidity. It updates ethical verification rules and thresholds based on feedback, improving the adaptability of content security control and reducing risks and oversights caused by rule lag. By repeating the closed-loop iteration process, it continuously optimizes various system capabilities, ensuring that the dialogue AI remains in a dynamic evolutionary state while avoiding performance degradation after long-term operation, thus maintaining a stable and efficient level of interactive service.

[0029] In one embodiment of the present invention, S52 includes: By integrating real-time user feedback and optimized response data, extracting core elements of feedback content and key features of response statements, aligning and mapping data dimensions, and generating structured closed-loop interactive data; Similarity calculation is performed on structured closed-loop interaction data to quantify the degree of matching between feedback results and response content in three dimensions: semantics, emotional fit, and demand satisfaction, and to generate multidimensional matching score data. Extract the original edge weight parameters of the corresponding nodes in the dynamic graph of emotion intention, and record the relationship between the emotion state and intention type corresponding to each edge weight; By combining multidimensional matching score data, the original edge weight parameters are gradient adjusted to strengthen the weight values ​​of high matching relationships and weaken the weight values ​​of low matching relationships, generating intermediate weight adjustment data. The intermediate data for weight adjustment is normalized to ensure that the edge weight parameters are within a reasonable range, thereby completing the dynamic update of the edge weight parameters of the emotional intent dynamic graph and generating updated graph weight data.

[0030] The working principle and effects of the above technical solution are as follows: Integrating user feedback and optimized response data, extracting core elements and aligning dimensions, generating structured closed-loop data avoids analytical biases caused by disorganized data and improves data utilization efficiency. Similarity calculation is performed on the closed-loop data to quantify the degree of matching from multiple dimensions, making the fit between feedback and response more intuitive and reducing errors caused by subjective judgment. Original edge weight parameters are extracted and their relationships recorded to ensure that parameter adjustments are based on evidence and avoid graph chaos caused by blind adjustments. Weights are adjusted using multi-dimensional scoring gradients to strengthen high-matching associations and weaken low-matching associations, improving the accuracy of emotion and intent association judgments and reducing the probability of misjudgment. The adjusted data is normalized to ensure that parameters are within a reasonable range, avoiding abnormal weights from affecting graph performance. This ensures that the dynamic graph always reflects actual user interactions and provides reliable data support for subsequent system optimization.

[0031] In one embodiment of the present invention, S53 includes: Extract the updated edge weight parameters and map them into the causal judgment model of the counterfactual reasoning engine to generate a parameter-adapted causal reasoning model. Retrieve historical data on the causal chain of emotional triggers, input parameters to a causal inference model, conduct batch causal tracing verification, and generate a set of verification inference results; By comparing and verifying the inference result set with the real historical cause data, the difference in the accuracy of causal judgment is calculated, and the accuracy difference data is generated. Based on the accuracy difference data, the core algorithm parameters of the causal inference model were iteratively adjusted to optimize the logic of node influence quantification and path association determination. After completing the algorithm parameter iteration, test samples are input again to verify the causal judgment effect and generate optimized causal inference engine data.

[0032] The working principle and effects of the above technical solution are as follows: The updated edge weight parameters are mapped to the causal judgment model, improving the model's adaptability to real-time interactive scenarios and avoiding biased causal judgment caused by parameter misalignment. Historical data is retrieved for batch source tracing verification, generating a complete verification result set, enhancing the comprehensiveness of model optimization references and reducing optimization limitations caused by single samples. The inference results are compared with real data, and the accuracy difference is calculated, intuitively reflecting the model performance gap and reducing ineffective adjustments caused by ambiguous optimization directions. Based on the difference, core algorithm parameters are adjusted to optimize the node and path judgment logic, improving the accuracy of causal judgment and reducing the occurrence of erroneous source tracing. Test samples are input again to complete the effect verification, generating optimized engine data. This ensures continuous improvement in causal analysis capabilities while avoiding a decline in recognition accuracy due to model solidification, allowing the system to maintain stable and reliable source tracing performance.

[0033] One embodiment of the present invention provides a real-time online dialogue AI tracking and feedback system based on an emotion recognition model, comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0034] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time online dialogue AI tracking and feedback method based on an emotion recognition model, characterized in that, The method includes: S1. Simultaneously collect voice and text content, voice prosody features and micro-expression dynamic data in user dialogue through a lightweight multimodal perception module to generate a multimodal emotion perception data stream; construct an emotion and intention dynamic graph based on the multimodal emotion perception data stream, where nodes represent emotional state and intention type, and edge weights reflect the correlation strength between the two. S2. Use the counterfactual reasoning engine to perform causal tracing analysis on the dynamic graph of emotions and intentions. By comparing the differences between the actual dialogue path and the virtual correction path, locate the key dialogue nodes that trigger users' negative emotions and generate emotional trigger causal chain data. S3. Generate differentiated response strategies based on the causal chain data of emotional triggers, and finally output multi-type response strategy data; S4. Input multi-type response strategy data into the ethical constraint reinforcement learning model, and perform security verification and value alignment on the response content through a preset ethical rule base, filter out statements with bias, discrimination or privacy leakage risks, and generate optimized response data. S5. Based on the closed-loop interaction of optimized response data and real-time user feedback, dynamically adjust the edge weight parameters of the emotion-intent dynamic graph, iteratively optimize the causal judgment accuracy of the counterfactual reasoning engine, and update the rule base of the ethical constraint reinforcement learning model, ultimately generating a continuously evolving dialogue AI tracking and feedback system.

2. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 1, characterized in that, S1 includes: S11. The lightweight multimodal perception module synchronously collects the voice and text content, voice prosody features and micro-expression dynamic data in the user's dialogue to generate a multimodal emotion perception data stream. S12. Perform time-series synchronization calibration on the multimodal emotion perception data stream to eliminate the acquisition delay difference of text, voice and micro-expression data, and generate time-series synchronized emotion perception data. S13. Perform noise filtering and feature purification on the time-synchronized emotion perception data, remove invalid data caused by environmental interference and collection errors, retain core emotion-related features, and generate time-normalized emotion perception data. S14. Construct a dynamic graph of emotion and intention based on time-series normalized emotion perception data, where nodes represent emotional states and intention types, and edge weights reflect the strength of their correlation. S15. Capture changes in emotion and intent during user dialogue in real time, dynamically update node information and edge weight parameters of the emotion and intent dynamic graph, and generate real-time updated emotion and intent association data.

3. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 2, characterized in that, S15 includes: Real-time capture of changes in user emotions and intentions during conversations, extraction of emotional fluctuation characteristics and intention shift information during conversations, and generation of dynamic emotional and intention change data; Based on the dynamic changes in emotion and intention data, update the node information of the dynamic graph of emotion and intention, add new emotion states and intention types, delete invalid nodes, and generate a dynamic graph with updated nodes. Calculate the magnitude of the correlation between emotional state and intention type, adjust the edge weight parameters of the dynamic graph, strengthen strong correlations, weaken weak correlations, and generate a dynamic graph with updated weights. By integrating the dynamic graph information after node updates and weight updates, a comprehensive dynamic update of the emotion and intention dynamic graph is completed, generating real-time updated emotion and intention association data. The real-time updated emotion and intention association data is time-stamped, and the data update time and corresponding dialogue node are recorded to generate emotion and intention association data with time stamps.

4. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 1, characterized in that, The S2 includes: S21. Utilize the counterfactual reasoning engine to retrieve real-time updated emotion and intention correlation data, and conduct a comprehensive causal tracing analysis on the dynamic graph of emotions and intentions. S22. Extract dialogue flow feature information from the emotion and intention association data, sort out the evolution trajectory of emotional state and intention type during the dialogue, and generate dialogue emotion evolution time series data. S23. Construct a virtual corrected dialogue path, and by comparing the differences in emotional changes between the actual dialogue path and the virtual corrected path, quantify the degree of influence of each dialogue node on emotional fluctuations and generate quantitative data on emotional impact. S24. Based on quantitative data on the impact of emotions, locate the key dialogue nodes that trigger negative emotions in users, sort out the transmission relationship between key nodes and negative emotions, and generate emotional trigger causal chain data. S25. Perform hierarchical analysis of the causal chain data of emotional triggers, distinguish between core and secondary triggers, clarify the scope of influence and transmission intensity of each trigger, and generate structured emotional trigger source data.

5. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 1, characterized in that, The S3 includes: S31. Extract the core trigger feature information from the structured emotional trigger source data, and combine it with the dialogue scenario and user historical interaction data to complete the comprehensive judgment of the dialogue scenario and user state. S32. When an abnormal user emotion is detected, the empathy response mode is activated, and emotionally resonant words that fit the user's current emotional state and the dialogue scenario are generated. S33. When a system error is detected, a correction response mode is triggered to accurately match the error type and its impact, and generate a targeted apology statement and error correction instructions. S34. When the user's intention tendency is predicted, activate the guided response mode and generate suggestive dialogue content that is instructive and practical, based on the user's intention type and needs preferences. S35. Integrate the dialogue output content of the three modes of empathy, correction, and guidance, optimize the fluency and adaptability of sentence expression, and finally output multi-type response strategy data.

6. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 1, characterized in that, The S4 includes: S41. Input multi-type response strategy data into the ethical constraint reinforcement learning model, retrieve the built-in ethical rule library of the model, and perform security and compliance verification on each response content; S42. Based on the core requirements of the ethical rule base, perform value alignment and adaptation processing on the response statements; S43. Optimize the expression and adjust the tone of the retained compliant statements to make them fit the user's current emotional state and dialogue scenario; S44. Perform a second compliance check on the optimized response data to ensure that no risky statements are omitted and generate optimized response data that complies with ethical standards. S45. Sort the output time sequence of the optimized response data, and generate orderly output ethical compliance response data by combining the rhythm of user dialogue and changes in emotion.

7. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 1, characterized in that, The S5 includes: S51. Push the orderly output of ethical compliance response data to the user according to the preset time sequence, and collect the user's real-time interactive feedback information simultaneously. S52. Combine real-time user feedback information with optimized response data to form closed-loop interactive data, analyze the matching degree between feedback results and response content, and dynamically adjust the edge weight parameters of the emotional intent dynamic graph. S53. Using the updated edge weight parameters, iteratively optimize the causal judgment algorithm of the counterfactual reasoning engine to improve the accuracy of locating key emotional triggers and the precision of causal analysis. S54. Based on the feedback information in the closed-loop interaction data, the rule base and judgment threshold of the ethical constraint reinforcement learning model are updated synchronously to improve the ethical verification standards. S55. Repeat the above closed-loop optimization process to continuously iterate and optimize the overall emotion recognition, causal judgment and response output capabilities of the system, and finally generate a continuously evolving dialogue AI tracking and feedback system.

8. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 7, characterized in that, S52 includes: By integrating real-time user feedback and optimized response data, extracting core elements of feedback content and key features of response statements, aligning and mapping data dimensions, and generating structured closed-loop interactive data; Similarity calculation is performed on structured closed-loop interaction data to quantify the degree of matching between feedback results and response content in three dimensions: semantics, emotional fit, and demand satisfaction, and to generate multidimensional matching score data. Extract the original edge weight parameters of the corresponding nodes in the dynamic graph of emotion intention, and record the relationship between the emotion state and intention type corresponding to each edge weight; By combining multidimensional matching score data, the original edge weight parameters are gradient adjusted to strengthen the weight values ​​of high matching relationships and weaken the weight values ​​of low matching relationships, generating intermediate weight adjustment data. The intermediate data for weight adjustment is normalized to ensure that the edge weight parameters are within a reasonable range, thereby completing the dynamic update of the edge weight parameters of the emotional intent dynamic graph and generating updated graph weight data.

9. The real-time online dialogue AI tracking and feedback method based on an emotion recognition model according to claim 7, characterized in that, S53 includes: Extract the updated edge weight parameters and map them into the causal judgment model of the counterfactual reasoning engine to generate a parameter-adapted causal reasoning model. Retrieve historical data on the causal chain of emotional triggers, input parameters to a causal inference model, conduct batch causal tracing verification, and generate a set of verification inference results; By comparing and verifying the inference result set with the real historical cause data, the difference in the accuracy of causal judgment is calculated, and the accuracy difference data is generated. Based on the accuracy difference data, the core algorithm parameters of the causal inference model were iteratively adjusted to optimize the logic of node influence quantification and path association determination. After completing the algorithm parameter iteration, test samples are input again to verify the causal judgment effect and generate optimized causal inference engine data.

10. A real-time online dialogue AI tracking and feedback system based on an emotion recognition model, including: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

Images