Artificial intelligence-based emotional interaction method and storage medium

Abstract

This application discloses an artificial intelligence-based emotion interaction method and storage medium, relating to the field of artificial intelligence. The method includes: constructing a value ontology model comprising at least two different levels of elements, the elements including baseline value elements and individual preference elements, with elements at different levels being associated through semantic features; acquiring multimodal data of the target object, and based on scene semantic information, selecting the target baseline value elements activated in the current scene from the value ontology model and determining their activation intensity; inputting the multimodal data and the activation intensity of the target baseline value elements into an emotion generation model to output emotion state information; inputting the emotion state information into a deep reinforcement learning network, which calculates the reward value of each candidate emotion interaction decision based at least on value matching degree and emotion matching degree, and outputs the candidate emotion interaction decision with the highest reward value as the optimal emotion interaction decision.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based emotional interaction method and storage medium. Background Technology

[0002] With the continuous development of artificial intelligence technologies such as natural language processing and multimodal perception, various human-computer emotional interaction systems are rapidly developing, such as home companion robots, medical-assisted general intelligence (AGI), and intelligent cockpit interaction systems.

[0003] However, current emotional interaction models typically rely on predefined rule bases or pattern matching to generate emotional expressions, remaining at a mechanical response level and lacking the intrinsic value motivation to support emotional decision-making and expression. They are unable to establish a truly human-like emotional interaction relationship with users. Summary of the Invention

[0004] This application provides an artificial intelligence-based emotional interaction method and storage medium, which can significantly improve the human-like naturalness of artificial intelligence-based emotional interaction.

[0005] Firstly, this application provides an artificial intelligence-based emotion interaction method, which includes: Construct a value ontology model, wherein the value ontology model includes at least two different levels of elements, the elements including at least a baseline value element based on general principles and an individual preference element based on the historical behavior data of the target object, and the elements at different levels are associated through semantic features; Acquire multimodal data of the target object, wherein the multimodal data includes at least the scene semantic information of the target object; Based on the semantic information of the scene, the target benchmark value elements activated in the current scene are selected from the value ontology model and their activation intensity is determined. The multimodal data and the activation intensity of the target benchmark value element are input into the emotion generation model to output emotion state information, wherein the emotion state information includes at least the emotion type and the correlation between the emotion type and the target benchmark value element. The emotional state information is input into a deep reinforcement learning network, which calculates the reward value of each candidate emotional interaction decision based at least on the value matching degree and the emotional matching degree, and outputs the candidate emotional interaction decision with the highest reward value as the optimal emotional interaction decision; wherein, the value matching degree is determined at least based on the matching degree between the candidate interaction decision and the target benchmark value element and / or the individual preference element.

[0006] In one alternative implementation of the first aspect, the at least two different levels include: At the macro level, it consists of elements based on pre-set ethical rules; The meso-level consists of cultural feature elements constructed based on cultural patterns and mapped to pre-defined brain map regions; The micro-layer consists of individual preference elements constructed from the historical behavioral data of the target object and mapped to the corresponding neural activation patterns.

[0007] In one optional implementation of the first aspect, the step of filtering out the target benchmark value elements activated in the current scene from the value ontology model based on the scene semantic information and determining their activation intensity includes: Based on the scene recognition semantic results for the target object in the scene semantic information, semantic matching is performed on the elements in the value ontology model to obtain matching elements and their matching degree that meet the preset matching conditions. If the matching element is not the ethical rule element, then the associated ethical rule element is determined based on the relationship between the elements, and the associated ethical rule element is used as the target benchmark value element. If the matching element is the ethical rule element, then the matched ethical rule element will be used as the target benchmark value element. The activation intensity of the target benchmark value element is output based on the matching degree of the matching element.

[0008] In an alternative implementation of the first aspect, the emotional state information further includes emotional intensity, and the method further includes: Based on the behavior recognition results of the auxiliary object in the scene semantic information, the behavior feature vector of the auxiliary object is extracted; The behavioral feature vector is semantically matched with a preset positive behavioral feature library to obtain a behavioral value score. The emotional intensity is adjusted based on the behavioral value score.

[0009] In one optional implementation of the first aspect, the step of inputting the emotional state information into a deep reinforcement learning network, and having the deep reinforcement learning network calculate the reward value for each candidate emotional interaction decision based at least on the value matching degree and the emotional matching degree, includes: Based on the emotion type and adjusted emotion intensity in the emotional state information, the weights of value matching degree and emotion matching degree used to calculate reward value in the deep reinforcement learning network are dynamically adjusted. The reward value for each candidate emotional interaction decision is calculated based on the adjusted weights and the value matching degree and the emotional matching degree.

[0010] In an alternative embodiment of the first aspect, the method further includes: Based on a preset brain semantic map algorithm, the cortical activation mode corresponding to the candidate emotional interaction decision is determined; Calculate the first matching degree between the cortical activation pattern corresponding to the candidate emotional interaction decision and the cortical activation pattern corresponding to the individual preference element, and the second matching degree between the cortical activation pattern corresponding to the target benchmark value element. The value matching degree is determined based on the first matching degree and the second matching degree.

[0011] In an alternative embodiment of the first aspect, the method further includes: After executing the optimal emotional interaction decision, obtain the target object's feedback data on the optimal emotional interaction decision; The feedback data is associated with the optimal emotional interaction decision, and the value ontology model is updated through a reinforcement learning mechanism. The update includes at least adjusting the priority parameters and / or the weights of the neural activation patterns of the value elements associated with the optimal emotional interaction decision.

[0012] In an alternative embodiment of the first aspect, the method further includes: During the execution of the optimal emotional interaction decision, the computational resource allocation strategy for perceiving the target object and / or executing the optimal emotional interaction decision is adjusted based on the priority information of the activated target benchmark value elements.

[0013] In an alternative embodiment of the first aspect, the method further includes: During the execution of the optimal emotional interaction decision, corresponding execution behavior data is collected in real time; Calculate the behavioral matching degree between the execution behavior data and the rules associated with the target benchmark value elements; When the behavior matching degree is lower than a preset abnormal threshold, the abnormal handling mechanism is triggered.

[0014] Secondly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the artificial intelligence-based emotional interaction method provided in the first aspect.

[0015] The artificial intelligence-based emotion interaction method and storage medium provided in this application have the following beneficial effects: By constructing an interpretable and transferable ontology model of human values, a structured representation of universally applicable human values ​​and individual preferences is achieved, thereby solving the problem of the difficulty in quantifying values ​​in the cognition of artificial general intelligence (such as robots). Through a value-oriented emotion generation mechanism, activated value semantics are transformed into specific emotional representations, which guide deep reinforcement learning networks to make emotional interaction decisions with value conformity and emotional matching degree as the core factors of the reward function. This can break through the limitations of traditional deep reinforcement learning in single task optimization, promote the organic integration of values, emotions and decisions, and thus significantly improve the human-like naturalness of AI-based emotional interaction. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] Figure 1 A flowchart illustrating an artificial intelligence-based emotion interaction method provided in an embodiment of this application; Figure 2 An example diagram of the value ontology model provided in the embodiments of this application; Figure 3 This is an architecture diagram related to the AI-based emotion interaction method of this application. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0019] The current research and application of brain-like decision-making systems (such as neuromorphic path decision-making systems) and emotion interaction models have the following shortcomings: First, there is a lack of value anchoring. Existing mainstream models (such as GPT-4V and traditional spiking neural network systems) typically rely solely on task data for decision optimization, lacking an interpretable and structured human value system as a basis for decision-making. This easily leads to "ethical deviations" in actual interactions, such as prioritizing task efficiency over user safety, violating human-centered ethical principles.

[0020] Second, there are limitations in the design of the DQN reward function. The reward functions of traditional deep reinforcement learning networks and their improved models (such as DuelingDQN and RainbowDQN) are mostly limited to the two dimensions of "task completion" and "system energy consumption," without taking "value alignment" as a core optimization objective. This design limitation makes the model's decision-making lack intrinsic meaning perception and emotional resonance in human-computer interaction, making it difficult to transcend mechanical responses.

[0021] Third, the emotion-decision coordination mechanism is broken: existing emotion interaction systems (such as Microsoft Xiaoice) mostly adopt rule-based emotion generation methods, failing to form a closed-loop coordination with the decision-making module. Emotional states cannot dynamically adjust decision priorities, and decision results are difficult to provide feedback to optimize emotional expression, which is inconsistent with the natural cognitive mechanism of humans where "values ​​drive emotions and emotions guide decisions."

[0022] From a cognitive neuroscience perspective, the root of the above problems lies in the fact that existing models fail to effectively reproduce the neural circuit mechanisms of human value decision-making (such as the "orbitofrontal cortex-mirror neurons-thalamus" pathway). Human decision-making is essentially a closed-loop process of "value-based attention filtering → emotion-based value quantification → prefrontal cortex execution of behavior," but existing systems lack the following capabilities: (1) The ability of values ​​to be semantically represented at the cortical level, for example, the inability to map ethical concepts to activation patterns of specific brain regions.

[0023] (2) The neural mapping relationship between emotion and value, such as the correlation modeling of mirror neuron activity and behavioral value rating.

[0024] (3) The realization of value-weighted mechanisms in the decision-making process, such as the simulation of the influence of orbitofrontal neural activity on behavioral choices.

[0025] To address at least one of the above technical problems, this application provides an artificial intelligence-based emotional interaction method. By constructing an interpretable and transferable human value ontology model, it achieves a structured representation of universally applicable human values ​​and individual preferences, thereby solving the problem of the difficulty in quantifying values ​​in the cognition of artificial intelligence systems (such as robots). Through a value-oriented emotion generation mechanism, the activated value semantics are transformed into specific emotional representations, which guide a deep reinforcement learning network to make emotional interaction decisions using value conformity and emotional matching degree as core factors of the reward function. This method can overcome the limitations of traditional deep reinforcement learning in single-task optimization, promote the organic integration of values, emotions, and decisions, and thus significantly improve the human-like naturalness of artificial intelligence-based emotional interaction.

[0026] The specific implementation of the artificial intelligence-based emotion interaction method provided in this application will be described in detail below with reference to the accompanying drawings.

[0027] Please see Figure 1 , Figure 1 An embodiment of this application provides an artificial intelligence-based emotion interaction method, including steps S10-S50: S10. Construct a value ontology model, where the value ontology model includes at least two elements at different levels. The elements at least include benchmark value elements based on general principles and individual preference elements based on the historical behavior data of the target object. The elements at different levels are associated through semantic sememes.

[0028] S20. Obtain the multi-modal data of the target object, where the multi-modal data at least includes the scene semantic information of the target object.

[0029] S30. Based on the scene semantic information, screen out the activated target benchmark value elements in the current scene from the value ontology model and determine their activation intensities.

[0030] S40. Input the multi-modal data and the activation intensity of the target benchmark value elements into an emotion generation model to output emotion state information, where the emotion state information at least includes the emotion type and the association degree between the emotion type and the target benchmark value elements.

[0031] S50. Input the emotion state information into a deep reinforcement learning network. The deep reinforcement learning network calculates the return values of each candidate emotion interaction decision at least based on the value matching degree and the emotion matching degree, and outputs the candidate emotion interaction decision with the highest return value as the optimal emotion interaction decision.

[0032] Among them, the value matching degree is at least determined based on the matching degree between the candidate interaction decision and the target benchmark value elements and / or the individual preference elements.

[0033] Specifically, the benchmark value elements refer to the value criteria that are generally recognized by human society and applicable to intelligent systems. Their content can be formed by integrating the "Global AI Ethics Consensus" and specific regional cultural characteristics. For example, the "principle of non-harm" (robots shall not harm humans), "Eastern culture of respecting the elderly", etc. Such elements adopt an ontology-based knowledge representation method and are structurally semantically defined and logically represented through triples in the form of "subject - relationship - object". For example, it can be formally described in OWL language as <AI, should abide by, principle of non-harm>. In this triple, the subject "AI" represents the acting subject of the value constraint, the relationship "should abide by" represents the value obligation or code of conduct that the subject needs to fulfill, and the object "principle of non-harm" represents the specific content or target object of the value constraint.

[0034] In an optional implementation manner, the value ontology model constructed in S10 above includes: Macro layer, which is composed of ethical rule elements constructed based on preset ethical principles; Middle layer, which is composed of cultural feature elements constructed based on cultural patterns and mapped to preset brain map regions; The micro layer consists of individual preference elements constructed based on the historical behavior data of the target object and mapped to corresponding neural activation patterns.

[0035] Among them, the ethical rule elements in the macro layer extract general value criteria such as "non - harm" and "justice" from universal ethical ontologies such as the Global AI Ethical Consensus. Each extracted ethical rule element is structurally defined using ontology - based knowledge representation methods to form machine - interpretable semantic units.

[0036] For example, the rule "principle of non - harm" can be formally defined as <AI, should abide by, principle of non - harm>. In addition, the specific constraints of this rule are described in detail through attributes. For example, the object "principle of non - harm" in the triple <AI, should abide by, principle of non - harm> is bound with attributes such as "contact force does not exceed 5N" and "movement speed does not exceed 1m / s".

[0037] Among them, the elements in the middle layer are derived from the cultural adaptation ontology. Combining the PrAGMATiC partitioning algorithm of the Human Brainnetome Atlas or the Berkeley Brain Semantic Atlas, regional cultural characteristics are transformed into cortical semantic sememes. For example, cultural characteristics such as "respect for the elderly in the East" and "individual autonomy in the West" are mapped to clusters of activated voxels in specific cerebral cortical regions. The "respect for the elderly in the East" feature corresponds to a cluster of activated voxels in area A9 (dorsolateral prefrontal cortex) of the frontal lobe.

[0038] The attribute information of cultural characteristic elements can include, but is not limited to: the corresponding cortical region identifier, cortical region voxel coordinates, activation pattern and its functional role (such as decision - making enhancement, emotion inhibition).

[0039] The micro layer extracts personal preference features from the user's historical behavior data, which can include historical human - machine interaction data and historical physiological state data. The attribute information of each individual preference element includes, but is not limited to, its corresponding neural activation pattern and voxel coordinates. For example, the personal preference element of "dislike of noise" corresponds to an inhibitory activation pattern and associated voxel cluster in the auditory cortex area A1.

[0040] After that, based on semantic sememes, the ethical rule elements in the macro layer, the cultural characteristic elements in the middle layer and the individual preference elements in the micro layer can be associated. For example, the individual preference element of "dislike of noise" coincides with the behavior norm of "speak in a soft voice" in the "respect for the elderly in the East" culture in the middle layer, and at the same time meets the condition constraint of "avoid psychological harm" in the "principle of non - harm" in the macro layer.

[0041] Finally, the knowledge at different levels is integrated into a knowledge graph to establish clear cross - layer association relationships. The rules in the macro layer are instantiated as cultural characteristics in the middle layer, and the cultural characteristics in the middle layer are personalized as individual preferences in the micro layer.

[0042] As an example, please see Figure 2 The provided value ontology model has the following elements: the macro level includes ethical rules such as the "non-harm principle" and the "privacy protection principle"; the meso level includes cultural characteristics such as "Eastern respect for the elderly" and "Eastern autonomy"; and the micro level includes personal preference elements such as "noise avoidance" and "asking questions before making decisions". The "Eastern respect for the elderly" culture can be mapped to the activation voxel clusters of the A9 area of ​​the frontal lobe using the PrAGMATiC partitioning method. The individual preference of "noise aversion" corresponds to the inhibitory activation pattern of the A1 area of ​​the auditory cortex, and this neural activation pattern can be associated with a specific voxel cluster.

[0043] Understandably, the aforementioned mechanism of mapping value elements to specific brain regions (i.e., "neural anchoring") has a neuroscience basis. Existing research has shown that semantic maps modeled using functional magnetic resonance imaging (fMRI) data demonstrate cross-individual consistency in the representation of specific semantic categories in the human cerebral cortex. Based on this mechanism, this application, combined with the PrAGMATiC partitioning algorithm of the Berkeley Brain Semantic Map, precisely maps value elements at different levels—including macro-level ethical rules, meso-level cultural characteristics, and micro-level individual preferences—to corresponding cortical regions. It also clearly identifies the cortical markers, voxel coordinates, neural activation patterns, and functional roles associated with each element. This enables artificial general intelligence (AGI) to learn the semantic elements of values, reproduce the neural circuits of human value decision-making, and thus achieve ethical alignment and human-like interactive decision-making that conforms to human cognitive patterns.

[0044] In an optional implementation, the multimodal data in S20 above may include physiological state data, scene semantic information, and historical behavior data.

[0045] Among them, the physiological status data comes from the somatosensory module set on the target object. For example, the heart rate, blood pressure, body temperature and other signals of the elderly can be monitored in real time through millimeter-wave radar, photoelectric sensors and other devices.

[0046] The scene semantic data is generated by the vision processing module after analyzing target object tracking videos / images input from devices such as cameras. Using a target detection model (such as YOLOv11) combined with 3D point cloud reconstruction technology, the visual scene can be parsed into a structured semantic description. For example, semantic vectors representing the current environment and the state of the target object, such as "an elderly person is unsteady while leaning against a wall," can be generated.

[0047] The historical behavior data is retrieved from the working memory module, including records of the target object's past behaviors and the system's own historical responses. For example, it can retrieve "the robot's response sequence and feedback data when the elderly person was unsteady on their feet or fell in the past three times".

[0048] In an optional implementation, S30 above, based on scene semantic information, filters out the target baseline value elements activated in the current scene from the value ontology model and determines their activation intensity, including: Based on the scene recognition semantic results for the target object in the scene semantic information, the cultural feature elements and ethical rule elements in the value ontology model are semantically matched to obtain the matching elements and their matching degree that meet the preset matching conditions. If the matching element is not an ethical rule element, then the associated ethical rule element is determined based on the relationship between the elements, and the associated ethical rule element is used as the target benchmark value element. If the matching element is an ethical rule element, then the matched ethical rule element will be used as the target benchmark value element. The activation intensity of the target benchmark value element is output based on the matching degree of the matching elements.

[0049] Specifically, the semantic similarity of the scene recognition results of the target object (such as "an elderly person is unsteady while holding onto the wall") with the semantic similarity of each element library at different levels predefined in the value ontology model will be calculated.

[0050] The similarity calculation process can employ vector cosine similarity comparison based on pre-trained language models (such as BERT or Sentence-BERT) or semantic distance measurement based on knowledge graph embedding. The final output consists of one or more matching elements and their corresponding matching scores, where the matching score is a similarity value ranging from [0,1]. In actual calculations, the element with the highest similarity can be considered as the matching element that meets the preset matching conditions.

[0051] If the matching element is a cultural characteristic element or an individual preference element, the ethical rule element associated with that cultural characteristic element or individual preference element is located by querying predefined semantic relationships (such as "instantiation" or "association") in the value ontology model. Then, this ethical rule element is determined as the target benchmark value element. If the matching element is an ethical rule element, then the matched ethical rule element is directly determined as the target benchmark value element.

[0052] When determining the activation intensity of target benchmark value elements, the matching degree of the matching elements or the value after normalization and nonlinear transformation can be directly used as its activation intensity.

[0053] Through the above steps, the semantic information of the scene can be transformed into target benchmark value elements with activation intensity, providing value guidance for subsequent emotion generation and decision optimization.

[0054] In an optional implementation, in S40 above, the activation intensity of the multimodal data and the target benchmark value elements can be combined to form a multidimensional feature vector X=[X 生理 X 环境 X 行为 X 价值观 [Input into the sentiment generation model. Where X...] 生理 X represents the physiological data of the target object. 环境 Representing scene semantic information, X 行为 X represents the historical behavior data of the target object. 价值观 This represents the target baseline value elements activated in the current scenario and their activation intensity. The activation intensity of the target baseline value elements serves as a constraint condition for the emotion generation model to predict emotional states.

[0055] In an alternative implementation, the sentiment generation model employs a hybrid architecture combining Transformer-XL and Graph Convolutional Networks (GCN).

[0056] The Transformer-XL layer is responsible for X. 生理 and X 行为 The time-series data is used to capture cross-temporal dependencies such as "sustained increase in blood pressure → subsequent behavioral abnormalities" through long-sequence modeling capabilities. The GCN layer takes the scene semantic graph and the value ontology model as input graphs and calculates the correlation between the semantic scene and the baseline value ontology through graph convolution operations. For example, the matching degree between the "fall scene" node and the "safety rule" element is calculated to be 0.95.

[0057] The emotional state representation output by the emotion generation model can be represented in a structured way. This representation is usually a vector that includes at least three dimensions: emotion type, emotion intensity, and the degree of correlation between the emotion and the target benchmark value element.

[0058] Among these, emotion type can be defined based on basic emotion classification theory as categories such as "worry," "mild," and "urgent." Emotion intensity is a quantitative scalar. Correlation represents the degree of correlation between the emotion state and specific values.

[0059] In training the emotion generation model, multiple labeled multimodal emotion datasets are used as the training set. For example, the "Human Emotion-Behavior" dataset (containing over 100,000 multimodal emotion annotations) from the Beijing General Artificial Intelligence Research Institute can be used as the training set. Then, a network model based on Transformer-XL and GCN is trained using the training set and a predefined loss function. The loss function can be defined using cross-entropy loss (for the emotion type dimension) and MSE loss (for the emotion intensity and relevance dimension).

[0060] In an alternative implementation, the AI-based emotion interaction method further includes: Based on the behavior recognition results of the auxiliary object in the scene semantic information, extract the behavior feature vector of the auxiliary object; The behavioral feature vector is semantically matched with a pre-defined positive behavioral feature library to obtain a behavioral value score. Emotional intensity is adjusted based on behavioral value scores.

[0061] In this application, the helper behavior based on the mirror neuron activation model also provides emotional reference signals for the generation of the neural network model's own emotional state. Specifically, the helper is an object that is in the same scene as the target object and interacts with it; for example, the helper could be the nurse of the elderly person.

[0062] Specifically, the behavior recognition results of the auxiliary object are semantically matched with a pre-defined positive behavior feature library. This feature library stores typical positive behavior examples and their feature vectors that are labeled with a value ontology and conform to ethical and cultural norms. During semantic matching, methods such as calculating cosine similarity can be used, and the resulting similarity score is the behavior value score (e.g., ranging from 0 to 1). The higher the score, the more consistent the observed behavior is with the positive behaviors recognized in the value system.

[0063] If the score is higher than the preset positive threshold (e.g., >0.8), the emotion generation model will enhance the positive emotional components in the current emotional state accordingly (e.g., adjust "worry" to a more positive "reassurance" or "appreciation"), or increase the intensity of the output emotion according to preset rules.

[0064] For example, in the scenario of "an elderly person who is unsteady on their feet and has to lean against a wall", the observation that "a nurse helps the elderly person up" increases the emotional intensity from 0.8 to 0.9.

[0065] Furthermore, the value score of this behavior can serve as an additional positive reward signal, which can be fed into the reward function of subsequent deep reinforcement learning networks to incentivize general artificial intelligence to learn and imitate such high-value behaviors.

[0066] This mechanism can simulate the cognitive process of "observation-understanding-resonance" in the human mirror nervous system, enabling artificial intelligence systems not only to generate emotions based on their own perceptions, but also to learn values ​​and calibrate emotions by observing the behavior of others in social interactions, thereby enhancing the social adaptability and ethical guidance of emotional interactions.

[0067] In an optional implementation, the reward function in S40 above is calculated by comprehensively considering evaluation factors from three dimensions: value matching degree, emotional matching degree, and task efficiency score.

[0068] The value matching degree is determined by calculating the consistency between the neural representations of candidate emotional interaction decisions and the activated target benchmark value elements, including the following steps: Based on a pre-defined brain semantic map algorithm, the cortical activation patterns corresponding to candidate emotional interaction decisions are determined. Calculate the first degree of matching between the cortical activation patterns corresponding to candidate emotional interaction decisions and the cortical activation patterns corresponding to individual preference elements, and the second degree of matching between the cortical activation patterns corresponding to target benchmark value elements. Value matching degree is determined based on the first matching degree and the second matching degree.

[0069] Specifically, the preset brain semantic mapping algorithm can be the PrAGMATiC partitioning algorithm of the Berkeley Brain Semantic Map. By performing semantic parsing on candidate decisions, key behavioral elements are extracted, such as "moving quickly," "touching to help up," and "asking loudly."

[0070] Then, based on the established mapping relationship between "behavioral semantics and cortical regions" in the brain semantic map, the cortical regions corresponding to each behavioral element and their expected activation states (such as enhanced activation and inhibitory activation) are determined.

[0071] Then, the expected activation states of each region are integrated into a unified cortical activation pattern vector, with each dimension of the vector corresponding to the activation intensity of a specific cortical voxel or functional area.

[0072] Specifically, the similarity between the predicted cortical activation patterns of candidate emotion interaction decisions and the cortical activation patterns corresponding to the following two types of elements is calculated: The first matching degree is obtained by calculating the similarity between the cortical activation patterns of candidate emotional interaction decisions and the cortical activation patterns corresponding to individual preference elements. For example, if the individual preference element is "aversion to noise", which corresponds to the inhibitory activation pattern of the A1 area of ​​the auditory cortex, then it is calculated whether the candidate emotional interaction decision (such as "asking loudly") will cause the expected activation of the auditory cortex.

[0073] The second matching degree is obtained by calculating the similarity between the cortical activation patterns of candidate emotional interaction decisions and the cortical activation patterns corresponding to the target baseline value elements. For example, if the target baseline value element is the "non-harm principle," the activation patterns of specific areas of the prefrontal cortex corresponding to it are used to calculate whether candidate emotional interaction decisions (such as "quickly rush up and help") are consistent with the expected activation patterns of these areas.

[0074] Similarity can be calculated using one of the following methods: Cosine similarity: Directly calculate the cosine similarity between two cortical activation pattern vectors.

[0075] Weighted Euclidean distance: Considering the differences in importance of different cortical regions in value representation, the distance between each region is weighted and then converted into similarity.

[0076] Finally, based on the first and second matching degrees, different weights are assigned to the first and second matching degrees. For example, the individual preference matching degree is weighted at 0.4 and the benchmark value matching degree is weighted at 0.6. The value matching degree is obtained by weighted fusion of the first and second matching degrees.

[0077] Ultimately, the resulting comprehensive value matching score serves as the input score for the value dimension in the reward function, and together with other dimensions (emotional matching score and task efficiency score), determines the total reward value of the candidate decision. This method achieves a calculable and interpretable evaluation of value conformity by simulating the neural mechanisms of human value decision-making.

[0078] In an optional implementation, to more comprehensively assess the consistency between the decision and the value system, the value matching degree V can also be calculated by combining it with knowledge graph alignment matching. Specifically, candidate emotional interaction decisions are parsed into a structured behavioral semantic subgraph G. A And associate it with the target value subgraph G associated with the target baseline value elements activated in the current scenario. v The constraints in the code are checked for compliance. For example, it checks whether the estimated contact force meets the contact force threshold constraint in the "no harm" rule.

[0079] More specifically, the candidate emotional interaction decisions (such as "quickly rush up to help and ask loudly") are semantically parsed to generate a structured behavioral semantic subgraph G. A This subgraph describes the subjects, relationships, objects, and attributes involved in decision-making in the form of triples.

[0080] For example, if the subject is a robot, the relationship is "helping up," and the object is an elderly person, it can be represented as <robot, help up, elderly person> or <robot, ask, elderly person>. The attributes of "helping up" include estimated contact force, and the attributes of "asking" include sound, with the value of the sound attribute being equal to "softly."

[0081] Extract the target benchmark value elements activated in the current scenario from the value ontology model, along with their associated cultural characteristic elements and individual preference elements. Integrate all quantitative constraints and qualitative requirements defined in these elements to form the target value subgraph G. v .

[0082] For example, in the scenario of "an elderly person falling," the activated element is the "safety principle," the cultural characteristic associated with the "safety principle" is "respect for the elderly in Eastern cultures," and the associated individual preference element is "aversion to noise." Based on this, a target value subgraph G is constructed. v It includes three key elements: "safety principle", "respect for elders in the East", and "aversion to noise". It also includes the following constraints for "safety principle", "respect for elders in the East", and "aversion to noise", namely: maximum contact force ≤ 5N, gentle action, and maximum acceptable volume ≤ 50dB.

[0083] Then, the behavioral semantic subgraph G A Behavioral attributes and target value subgraph G v The corresponding constraints are checked item by item to determine whether they meet the constraint requirements: Contact force constraint check: If the query behavior semantic subgraph G A The estimated contact force is 8N, compared with the target value subgraph G. v The constraint that the maximum contact force is ≤5N is deemed to be violated.

[0084] Volume constraint check: If the query behavior semantic subgraph G A The obtained volume is 75dB, compared with the target value subgraph G. v If the maximum acceptable volume is ≤50dB, it is considered a violation.

[0085] The results of the inspection of all constraints are quantified and assigned values, with 1 for compliance and 0 for violation. Then, according to the preset priority of each constraint in the current scenario, the corresponding weight coefficients are assigned, and the value matching degree V, normalized to [0,1], is obtained by weighted average calculation.

[0086] In addition to value alignment, the calculation of the reward function also includes the following two dimensions: Emotional fit E: obtained by calculating the fit between emotional state information and candidate interaction decisions.

[0087] In an optional implementation, the emotional matching degree E is obtained by calculating the fit between the emotional state vector M and the candidate emotional interaction decision vector A:

[0088] in, It is the behavioral needs and preferences implied by the current emotion type, which can be directly obtained through a preset emotion-behavior tendency mapping table. For example, when the emotion type is "worry", the corresponding need preference vector is "low interference, high security". It is the emotional expression feature of the behavior corresponding to the candidate emotion interaction decision. For example, the behavior of "slowly approaching" corresponds to the "low interference" feature, which can be obtained through a preset behavior-emotion expression feature mapping table.

[0089] Among them, the task efficiency score T is used to measure the effectiveness of candidate emotional interaction decisions in achieving the given task objectives, and its calculation can be formalized as follows: ,in, These represent the estimated energy consumption and estimated time required to execute the candidate emotion interaction decision, respectively. This is an efficiency evaluation function designed to balance costs and benefits, and to reward efficient and energy-saving behaviors.

[0090] Finally, the total reward value R for each candidate emotional interaction decision is obtained by weighted summation of the above three dimensions: , These are dynamically adjustable weighting coefficients.

[0091] In an optional implementation, the training process of the deep reinforcement learning network (DQN) is configured as follows: Network infrastructure: RainbowDQN is adopted as the basic framework, which integrates several improvement strategies such as Prioritized Experience Replay, DoubleQ-Learning, and Dueling Network architecture to improve training stability and policy performance.

[0092] Value-based constraint embedding: A "value-based filtering mechanism" is introduced into the experience replay buffer. Specifically, for each behavioral sample in the buffer, its value matching degree V is calculated. If the value matching degree V of a sample is lower than a preset ethical threshold (e.g., V < -0.5), the probability of it being sampled and replayed is systematically reduced (e.g., its sampling weight is multiplied by 0.3), thereby preventing the model from overlearning behavioral patterns that violate core values.

[0093] Training objective and loss function: Training aims to minimize a joint loss function L, which simultaneously optimizes the accuracy of decision rewards and the alignment with values. .in, It is the timing difference error of traditional DQN. The target value score is provided by the value ontology model, and λ (e.g., set to 0.4) is the value bias penalty coefficient, used to control the importance of value alignment during training. By minimizing this joint loss, the network is driven to learn an optimal policy that simultaneously achieves high task rewards and high value alignment.

[0094] In an optional implementation, in step S40 above, the emotional state information is input into a deep reinforcement learning network, and the deep reinforcement learning network calculates the reward value for each candidate emotional interaction decision based at least on the value matching degree and the emotional matching degree, including: Based on the emotion type and adjusted emotion intensity in the emotion state information, the weights of the value matching factor and emotion matching factor used to calculate the reward value in the deep reinforcement learning network are dynamically adjusted. The reward value for each candidate emotional interaction decision is calculated based on the adjusted weights and by combining value matching and emotional matching.

[0095] Specifically, in the process of dynamically adjusting the weights of the DQN reward function based on emotional state information, the emotional type determines the direction of weight adjustment. Different emotional types correspond to different decision priority tendencies. For example, the emotional type of "worry" focuses on safety and adaptability, which corresponds to the strategy of "increasing the weight β of emotional matching". The emotional type of "urgency" focuses on response efficiency, which corresponds to the strategy of "increasing the weight γ of task efficiency". The emotional type of "mild" corresponds to the strategy of "maintaining the default weight".

[0096] The intensity of emotion determines the magnitude of weight adjustment. The higher the value of the emotion intensity (normalized to the range of [0,1]), the greater the magnitude of the corresponding weight adjustment. For example, when the intensity of "worry" is high, the β weight can be increased to 0.4.

[0097] In an alternative implementation, the AI-based emotion interaction method further includes: After executing the optimal emotional interaction decision, obtain the target object's feedback data on the optimal emotional interaction decision; The feedback data is correlated with the decision-making process, and the value ontology model is updated through a reinforcement learning mechanism. The update includes at least adjusting the priority parameters and / or the weights of the neural activation patterns of the value elements associated with the optimal emotional interaction decision.

[0098] Specifically, artificial intelligence systems, such as companion robots, acquire feedback data from the target object through multimodal sensors after making optimal emotional interaction decisions. These multimodal sensors can include, but are not limited to, devices such as cameras, microphones, and physiological sensors. For example, microphone arrays and natural language processing technology can be used to identify users' verbal feedback. For instance, if an elderly person says "thank you" after being helped up by a companion robot, this can be recognized as positive feedback. Furthermore, cameras and physiological sensors can monitor the user's subsequent state, such as whether their facial expressions are relaxed and whether their heart rate and blood pressure have returned to normal. When a painful expression or abnormal physiological indicators are detected in the elderly person, it can be considered negative feedback.

[0099] The working memory module associates the feedback data with the optimal emotional interaction decision. The feedback data is then quantified into a reward signal in a deep reinforcement learning network. For example, positive feedback is recorded as a +1 reward, and negative feedback as a -1 reward.

[0100] In some implementations, model updates can be driven in any of the following ways: Method 1: Adjust the priority parameters of value elements. Specifically, the priority of value elements highly related to the current optimal emotional interaction decision, such as the value of "safety principle" and the specific behavioral pattern of "slow and steady approach," will be adjusted. Positive feedback will increase its weight, making the element more strongly activated and having a greater impact on decision-making in similar future scenarios.

[0101] Method 2: Optimize the weights of neural activation patterns. Specifically, through backpropagation and transfer learning techniques, the model representation of the cortical activation pattern corresponding to the decision is strengthened. For example, when the "slowly stabilizing" behavior continues to receive positive feedback, the activation weight of the corresponding prefrontal cortex region will be enhanced.

[0102] Through the above approach, artificial intelligence systems can learn from interactive experiences and achieve personalized ethical alignment.

[0103] In an alternative implementation, the AI-based emotion interaction method further includes: In the process of executing optimal emotional interaction decisions, the computational resource allocation strategy for perceiving target objects and / or executing optimal emotional interaction decisions is adjusted based on the priority information of activated target benchmark value elements.

[0104] Specifically, during the decision-making and execution process, the activated target benchmark value elements and their activation intensity are monitored in real time in the current scenario. For example, in the "elderly person falls" scenario, the activation intensity of the "safety rule" can reach 0.95, which is significantly higher than other rules. This rule is also associated with constraints such as "contact force not exceeding 5N" and "movement speed not exceeding 1m / s".

[0105] Then, the central execution loop (which can be viewed as a resource scheduler) dynamically adjusts the allocation of computing resources based on the priority of the element association constraints in the value ontology model. When the "safety rules" are activated at a high intensity, the system will allocate most of the computing power (e.g., 80%) of the visual processing module (such as an RGB-D camera) to high-priority tasks such as "elderly body posture detection" and "contact force prediction", while reducing the proportion of computing power allocated to low-priority tasks such as motion speed detection.

[0106] The aforementioned scheme, by mimicking the human brain's attention allocation, allows the thalamus-prefrontal cortex pathway to automatically filter sensory information when facing high-risk scenarios, prioritizing attention on security threats and ignoring secondary details. This method, through real-time monitoring of value-related elements and their activation intensity, dynamically allocates limited computing resources, achieving self-awareness of task criticality and self-optimization of resource allocation, thereby improving the system's response speed and ethical safety in high-risk scenarios.

[0107] In an alternative implementation, the AI-based emotion interaction method further includes: During the process of making the optimal emotional interaction decision, corresponding execution behavior data is collected in real time; Calculate the behavioral matching degree between the execution behavior data and the rules associated with the target benchmark value elements; When the behavior matching degree is lower than the preset abnormal threshold, the abnormal handling mechanism is triggered.

[0108] Specifically, during the process of making optimal emotional interaction decisions, the system collects key data on behavior execution in real time through multiple types of sensors. For example, force sensors monitor the actual contact force between the robot and the target object, cameras or motion sensors monitor the robot's movement speed, and microphone arrays monitor the interaction volume.

[0109] Then, based on the activated target benchmark value element association constraint rules, a differential algorithm is used to calculate the matching degree according to the rule type. If there are multiple behavioral elements, a weighted average is performed to obtain the final matching degree. If the constraint rule is text-based, such as "actions should be gentle" or "interactions should be conducted in a soft voice", the text description of the real-time behavior is converted into a semantic vector. The cosine similarity algorithm is used to calculate the matching degree between the vector and the semantic vector of the constraint rule, with the value range being [0,1].

[0110] If the constraint rule is numerical, such as "contact force ≤ 5N" or "volume ≤ 50dB", first calculate the difference between the real-time data and the constraint threshold, and then normalize the difference to the [0,1] interval to obtain the single-element matching degree.

[0111] When the real-time calculated behavior matching degree is lower than the preset abnormal threshold, it is determined that the current execution behavior is abnormal and the abnormal handling mechanism is triggered.

[0112] The anomaly handling mechanism can be as follows: The central execution loop interrupts the current potentially harmful sequence of behaviors and calls alternative behaviors from a pre-set safety behavior library. For example, switching from "quickly help up" to "pause and quietly confirm again." Simultaneously, all information about this anomaly event is recorded, including the original decision content, real-time behavior data, the matching degree calculation process, and the triggered alternative behaviors. This data is then used as feedback information and fed into the value ontology model's update mechanism for subsequent optimization of decision-making strategies and rule constraint weights, preventing similar risks from recurring.

[0113] In the above scheme, the closed-loop intervention process can ensure that the behavior of the artificial intelligence system is always constrained within the scope allowed by the value system in a dynamic and uncertain environment.

[0114] Taking "medical companion robot handling a sudden fall in the elderly" as an example, the detailed process of this application is explained in detail. Please see [link to application]. Figure 3 , Figure 3 This is an architecture diagram of the AI-based emotion interaction method proposed in this application. The application is based on an architecture of "value ontology module --- emotion generation module --- DQN optimization module --- collaborative closed-loop module." Each module achieves data flow interoperability through standardized interfaces. The specific process is as follows: Step 1: Initialize the value ontology module.

[0115] Specifically, universal ethical rules are introduced: construct a triple of <subject, relationship, object>, including <robot, no harm, elderly bone safety> (target matching threshold Vtarget=0.8) and <robot, priority, emergency rescue>.

[0116] Import cultural adaptation rules: Import the <Robot, Respect, Elderly Body Autonomy> rule, which is based on the Eastern culture of respecting the elderly and corresponds to the activation mode of the A10 area of ​​the frontal lobe; Import individual preference rules: Based on historical interaction data, add the <Robot, Avoid, Noise> rule, which is set according to the individual characteristic that "the elderly auditory cortex has a 70% inhibition rate of sound above 80dB".

[0117] The above three layers of rules are integrated into a value ontology graph, stored in a Redis database, and interact with the emotion generation module and DQN optimization module through the gRPC interface.

[0118] Step 2: This step is completed by the emotion generation module, the core of which is to collect multimodal data and generate emotion state information.

[0119] The system collects physiological data of the elderly, such as blood pressure and heart rate, using millimeter-wave radar. Scene images are acquired using an RGB-D camera, and scene semantic information, such as "the elderly person is unsteady while holding onto the wall and has an angle of 30° with the ground," is output through target detection and 3D point cloud reconstruction.

[0120] Based on a pre-defined mapping model, the current scenario activates a "security rule," and a multi-dimensional feature vector X=[X 生理 X 环境 X 行为 X 价值观 ].

[0121] The input vector X is calculated using the Transformer-XL+GCN model, and the output emotional state information M=(worry, 9.2, 0.95) is generated, which corresponds to the emotional type, emotional intensity, and correlation with the "safety rule" respectively. Then, the emotional intensity is superimposed on the mirror system score. The behavior of "nurse gesturing for gentle support" is identified by motion capture, and the behavior value score S=0.92 is calculated. This score is superimposed on the emotional intensity, and the updated intensity is 9.5.

[0122] Step 3: This step is completed by the DQN optimization module, and its core is to select the optimal decision based on the three-dimensional reward function.

[0123] The emotional state information M is input into the DQN to make decisions on candidate interaction behaviors. Assume there are behaviors A and B.

[0124] Behavior A: Quickly rush forward to help them up and loudly ask questions (volume 75dB); Behavior B: Approach slowly (speed 0.5m / s) + ask softly (volume 50dB) + hold onto the arm firmly.

[0125] Then, the three-dimensional factor scoring items are calculated. Assuming that behavior A is due to excessive volume and too fast movement, the value matching degree V=0.26; behavior B conforms to the "gentle and low noise" rule, the value matching degree V=0.93; and assuming that behavior A does not match the "low interference and high safety" requirement of the "worry" emotion, the emotional matching degree E=0.3; behavior B highly matches the requirement, the emotional matching degree E=0.92; and assuming that behavior A is short in time but high in energy consumption, the task efficiency score T=0.29; behavior B is slightly longer in time but low in energy consumption, the task efficiency score T=0.14.

[0126] Total Reward Calculation and Behavior Selection: The total reward for behavior A is 0.278, and the total reward for behavior B is RB = 0.769. The DQN module outputs behavior B, which has the higher reward, as the optimal decision.

[0127] Step 4: This step is completed by the collaborative closed-loop module, the core of which is to realize decision execution monitoring, user feedback collection, and dynamic updating of the value ontology.

[0128] Real-time monitoring of the execution process: The central execution loop monitors the contact force between the robot and the elderly through an RGB-D camera. The contact force is ≤5N, which meets the "no harm" rule. The real-time value matching degree is stable at 0.92.

[0129] Multimodal feedback acquisition: The microphone array recognizes the elderly person's "thank you" voice feedback and determines it as positive feedback.

[0130] Value ontology model update: Increase the "individual preference weight" of behavior B by 10%, and write the association rule of "fall scenario → slowly steady" into the value ontology model ontology. Update the activation weight of frontal lobe A9 area to improve the decision response speed in similar scenarios.

[0131] The artificial intelligence-based emotion interaction method provided in this application has the following beneficial effects: (1) By integrating the Berkeley PrAGMATiC partition, the cortical voxels of values ​​are accurately labeled, solving the core industry problem of "difficulty in quantifying" values. At the same time, relying on transfer learning technology, the value rules are optimized online.

[0132] (2) Breaking away from the limitations of traditional DQN which only focuses on "single optimization of task efficiency", it takes value conformity as the core weight and introduces cortical activation similarity to calculate value matching degree V, so as to achieve three-dimensional collaborative optimization of "ethics-emotion-efficiency".

[0133] (3) Reproduce the value decision-making loop of human "orbitofrontal lobe-mirror neuron-thalamus", build a complete closed loop of "value screening attention → emotion quantification value → DQN optimization behavior", and greatly improve the naturalness and ethical safety of humanoid interaction.

[0134] (4) Provides a modular general AI ethics alignment solution that can directly generate robot decision-making systems, support the implementation of scenarios such as “human value-driven task execution”, support rapid migration to multiple scenarios such as medical care, home, and smart cockpit, and a dynamic update mechanism for value ontology.

[0135] (5) By replicating the neural mechanism of human value decision-making, the value system is used as the core anchor of the brain-like model’s human-like ability. The DQN reward function is optimized to realize the decision-making logic of “ethics first”, filling the technical gap of “value-emotion-decision” collaboration in the existing brain-like system. It can not only improve the interactive credibility of service AI, but also become a key technical support for the implementation of general artificial intelligence ethics.

[0136] Accordingly, this application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the artificial intelligence-based emotion interaction method in the above embodiments. The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. Program code contained on a computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (Radio Frequency), etc., or any suitable combination thereof. The aforementioned computer-readable storage medium may be contained within an electronic device; or it may exist independently, not assembled into an electronic device.

[0137] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0138] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Modules described in the embodiments of this application can be implemented in software or hardware. The names of modules do not, in some cases, constitute a limitation on the unit itself.

[0139] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described artificial intelligence-based emotional interaction method. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those provided in the above embodiments, and will not be repeated here.

[0140] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An artificial intelligence-based emotional interaction method, characterized in that, include: Construct a value ontology model, wherein the value ontology model includes at least two different levels of elements, the elements including at least a baseline value element based on general principles and an individual preference element based on the historical behavior data of the target object, and the elements at different levels are associated through semantic features; Acquire multimodal data of the target object, wherein the multimodal data includes at least the scene semantic information of the target object; Based on the semantic information of the scene, the target benchmark value elements activated in the current scene are selected from the value ontology model and their activation intensity is determined. The multimodal data and the activation intensity of the target benchmark value element are input into the emotion generation model to output emotion state information, wherein the emotion state information includes at least the emotion type and the correlation between the emotion type and the target benchmark value element. The emotional state information is input into a deep reinforcement learning network, which calculates the reward value of each candidate emotional interaction decision based at least on the value matching degree and the emotional matching degree, and outputs the candidate emotional interaction decision with the highest reward value as the optimal emotional interaction decision; wherein, the value matching degree is determined at least based on the matching degree between the candidate interaction decision and the target benchmark value element and / or the individual preference element.

2. The artificial intelligence-based emotion interaction method as described in claim 1, characterized in that, The at least two distinct levels include: At the macro level, it consists of elements based on pre-set ethical rules; The meso-level consists of cultural feature elements constructed based on cultural patterns and mapped to pre-defined brain map regions; The micro-layer consists of individual preference elements constructed from the historical behavioral data of the target object and mapped to the corresponding neural activation patterns.

3. The artificial intelligence-based emotion interaction method as described in claim 2, characterized in that, The step of filtering out the target baseline value elements activated in the current scene and determining their activation intensity from the value ontology model based on the scene semantic information includes: Based on the scene recognition semantic results for the target object in the scene semantic information, semantic matching is performed on the elements in the value ontology model to obtain matching elements and their matching degree that meet the preset matching conditions. If the matching element is not the ethical rule element, then the associated ethical rule element is determined based on the relationship between the elements, and the associated ethical rule element is used as the target benchmark value element. If the matching element is the ethical rule element, then the matched ethical rule element will be used as the target benchmark value element. The activation intensity of the target benchmark value element is output based on the matching degree of the matching element.

4. The artificial intelligence-based emotion interaction method as described in claim 1, characterized in that, The emotional state information also includes emotional intensity, and the method further includes: Based on the behavior recognition results of the auxiliary object in the scene semantic information, the behavior feature vector of the auxiliary object is extracted; The behavioral feature vector is semantically matched with a preset positive behavioral feature library to obtain a behavioral value score. The emotional intensity is adjusted based on the behavioral value score.

5. The artificial intelligence-based emotion interaction method as described in claim 4, characterized in that, The step of inputting the emotional state information into a deep reinforcement learning network, and having the deep reinforcement learning network calculate the reward value for each candidate emotional interaction decision based at least on value matching degree and emotional matching degree, includes: Based on the emotion type and adjusted emotion intensity in the emotional state information, the weights of value matching degree and emotion matching degree used to calculate reward value in the deep reinforcement learning network are dynamically adjusted. The reward value for each candidate emotional interaction decision is calculated based on the adjusted weights and the value matching degree and the emotional matching degree.

6. The artificial intelligence-based emotion interaction method as described in any one of claims 1 to 5, characterized in that, The method further includes: Based on a preset brain semantic map algorithm, the cortical activation mode corresponding to the candidate emotional interaction decision is determined; Calculate the first matching degree between the cortical activation pattern corresponding to the candidate emotional interaction decision and the cortical activation pattern corresponding to the individual preference element, and the second matching degree between the cortical activation pattern corresponding to the target benchmark value element. The value matching degree is determined based on the first matching degree and the second matching degree.

7. The artificial intelligence-based emotion interaction method as described in any one of claims 1 to 5, characterized in that, The method further includes: After executing the optimal emotional interaction decision, obtain the target object's feedback data on the optimal emotional interaction decision; The feedback data is associated with the optimal emotional interaction decision, and the value ontology model is updated through a reinforcement learning mechanism. The update includes at least adjusting the priority parameters and / or the weights of the neural activation patterns of the value elements associated with the optimal emotional interaction decision.

8. The artificial intelligence-based emotion interaction method as described in any one of claims 1 to 5, characterized in that, The method further includes: During the execution of the optimal emotional interaction decision, the computational resource allocation strategy for perceiving the target object and / or executing the optimal emotional interaction decision is adjusted based on the priority information of the activated target benchmark value elements.

9. The artificial intelligence-based emotion interaction method as described in any one of claims 1 to 5, characterized in that, The method further includes: During the execution of the optimal emotional interaction decision, corresponding execution behavior data is collected in real time; Calculate the behavioral matching degree between the execution behavior data and the rules associated with the target benchmark value elements; When the behavior matching degree is lower than a preset abnormal threshold, the abnormal handling mechanism is triggered.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the artificial intelligence-based emotion interaction method as described in any one of claims 1 to 9.

Images