Personalized interaction system based on user emotion perception

Abstract

The invention relates to the technical field of emotion calculation, and discloses a personalized interaction system based on user emotion perception. According to the system, a dynamic emotion flow is formed by integrating physiological signals, interactive behaviors and environmental parameters of a user through a dynamic emotion flow construction module. The emotion flow is then projected to an emotion space network, and after node and weight dynamic processing, an emotion space coordinate and an energy distribution diagram are generated. The system analyzes the emotional evolution direction and intensity trend by calculating the divergence and rotation of the emotional vector field. The trend drives a plurality of parallel interaction strategy generation sub-networks to output a multi-modal preliminary strategy vector; and finally, generating a final executable interaction instruction sequence through strategy fusion and conflict resolution. According to the method, continuous dynamic modeling and evolution prediction of emotions are realized, and coordinated and coherent personalized interaction responses can be generated.

Description

Technical Field

[0001] This invention relates to the field of affective computing technology, specifically a personalized interaction system based on user emotion perception. Background Technology

[0002] Existing personalized interaction systems typically rely on static analysis of single or limited modal signals for emotion perception, classifying emotional outcomes into discrete, pre-defined basic emotion categories. These methods often extract isolated emotional feature points from the interaction process and identify them using pattern matching or classification models. When generating interaction strategies, these systems often employ rule-based mapping or single decision models, directly mapping identified emotion labels to pre-defined interaction commands, lacking in-depth modeling and utilization of the continuous dynamic evolution of emotions.

[0003] These existing technical solutions have shortcomings. Static, discrete emotion classification cannot accurately depict the continuous flow and complex mixture of user emotions in real interactions, resulting in coarse and delayed emotion perception. The link from emotion perception to strategy generation is fragmented; emotional state is merely a trigger condition or input parameter, and the strategy generation process fails to incorporate consideration of the dynamic evolution of emotions. Therefore, the interaction strategies generated by the system are often rigid and reactive, lacking foresight and contextual coherence, making it difficult to achieve deeply personalized and naturally fluid emotional interaction.

[0004] This invention aims to address two core issues: how to continuously, dynamically, and quantitatively model and represent users' emotions during interaction, and how to proactively generate personalized interaction strategies that integrate multiple modalities and are consistent, based on the inherent laws of the dynamic evolution of emotions rather than static labels. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized interaction system based on user emotion perception to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a personalized interaction system based on user emotion perception, the system comprising: The dynamic emotion stream construction module is used to construct a dynamic emotion stream based on the real-time interaction process of the target user. The dynamic emotion stream integrates multimodal physiological signal stream, interaction behavior time sequence stream and environmental parameter stream. The emotional space network processing module is used to project the dynamic emotional flow onto a preset emotional space network for processing. The emotional space network outputs the emotional space coordinates and emotional energy distribution map of the target user through node activation and connection weight adjustment. The emotion evolution trend calculation module is used to analyze the emotion spatial coordinates and emotion energy distribution map. This analysis process relies on the divergence and curl calculation of the emotion vector field to generate an emotion evolution trend that includes the direction and intensity of emotion evolution. An interaction strategy generation network module is used to input the emotion evolution trend into the interaction strategy generation network. The interaction strategy generation network contains multiple parallel strategy generation sub-networks. Each strategy generation sub-network corresponds to an interaction mode and outputs a preliminary set of interaction strategy vectors. The strategy fusion and conflict resolution module is used to perform strategy fusion and conflict resolution on the preliminary set of interaction strategy vectors to generate the final executable sequence of interaction instructions.

[0007] Preferably, the step of projecting the dynamic emotion stream onto a preset emotion space network for processing includes the following sub-steps: Feature extraction and time alignment are performed on the multimodal physiological signal stream to form a standardized physiological feature tensor; The interactive behavior time-series is segmented into actions and identified into intents to form a behavior intent sequence with timestamps; The environmental parameter stream is subjected to event detection and context-related annotation to form a structured environmental event sequence; The physiological feature tensor, the behavioral intention sequence, and the structured environmental event sequence are fused using a multi-stream fusion mechanism with an attention weight allocation mechanism to generate a fused emotional feature vector sequence. The emotional feature vector sequence is input into the emotional space network, which is constructed from an emotional basis vector matrix. The emotional space coordinates are determined by calculating the projection of the emotional feature vector sequence onto the emotional basis vector matrix, and the emotional energy distribution map is generated by mapping through an activation function.

[0008] Preferably, the multi-stream fusion employs an attention weight allocation mechanism, including: Configure physiological attention weights for the physiological feature tensor, configure behavioral attention weights for the behavioral intention sequence, and configure environmental attention weights for the structured environmental event sequence; The physiological attention weight is calculated based on the signal-to-noise ratio of the physiological signal and the energy proportion of a specific frequency band. The calculation of the behavioral attention weight is based on the clarity of the behavioral intention and the frequency of the behavior. The environmental attention weight is calculated based on the suddenness of environmental events and the strength of their historical correlation. The weighted physiological feature tensor, the behavioral intention sequence, and the structured environmental event sequence are concatenated along the time dimension, and the concatenated high-dimensional vector is subjected to dimensionality reduction and smoothing to form the fused emotion feature vector sequence.

[0009] Preferably, the step of analyzing the emotional spatial coordinates and emotional energy distribution map includes: In the emotional space network, starting from the emotional space coordinates, continuous emotional space coordinate segments are extracted along the time axis; For each of the aforementioned emotional space coordinate segments, the gradient of its corresponding emotional energy distribution map in the emotional space is calculated to obtain the local emotional energy change gradient field. The divergence of the local emotional energy change gradient field is calculated to characterize the aggregation or diffusion trend of emotional energy, and the emotional energy divergence value is obtained. The curl of the local emotional energy change gradient field is calculated to characterize the rotation or vortex trend of the emotional state, and the emotional energy curl value is obtained. By combining the emotional energy divergence value and the emotional energy curl value, and through vector synthesis and normalization, the emotional evolution direction and the emotional evolution intensity, which characterize the global trend of change, are obtained.

[0010] Preferably, inputting the emotion evolution trend into the interaction strategy generation network includes: The interaction strategy generation network includes a voice interaction strategy subnetwork, a visual content presentation strategy subnetwork, and an operation process adjustment strategy subnetwork. The voice interaction strategy subnetwork receives the emotion evolution trend and outputs voice strategy components including voice intonation parameters, word choice vectors, and speech rate rhythm vectors. The visual content presentation strategy subnetwork receives the emotion evolution trend and outputs visual strategy components including interface color parameters, information layout vectors, and dynamic effect vectors. The operation process adjustment strategy sub-network receives the emotion evolution trend and outputs operation strategy components including task step sequence vector, interaction feedback delay parameter and option complexity parameter. The voice strategy component, the visual strategy component, and the operation strategy component together constitute the preliminary interaction strategy vector set.

[0011] Preferably, the step of performing policy fusion and conflict resolution on the preliminary set of interaction policy vectors includes: A policy conflict detection rule base is established, which defines the incompatibility conditions between policy vectors of different interaction modalities; The speech policy component, the visual policy component, and the operation policy component in the preliminary interaction policy vector set are compared pairwise, and conflicting policy pairs are identified based on the policy conflict detection rule base. For each conflicting policy pair identified, the policy priority arbitrator is invoked. The policy priority arbitrator assigns an execution priority to the conflicting policy pair based on the main objective of the current interaction task, user historical preferences, and system resource status. Based on the execution priority, the parameters of the low-priority policy components are adjusted or replaced until all conflicting policy pairs meet the compatibility conditions. All adjusted strategy components are sorted and encapsulated according to the interaction timeline to form the final executable interactive instruction sequence.

[0012] Preferably, the invocation strategy priority arbitrator includes: The strategy priority arbitrator has an embedded task target matching degree calculation unit, which is used to calculate the matching degree of each conflict strategy pair with the main target of the current interaction task and output the task target matching degree score. The strategy priority arbitrator has an embedded user preference compliance calculation unit, which is used to query the user's historical interaction logs, calculate the degree of compliance between each conflicting strategy pair and the user's historical preferences, and output the user preference compliance score. The strategy priority arbitrator has an embedded system load assessment unit, which is used to assess the computational resources and time costs required to execute each conflicting strategy and output a system load assessment score. The task objective matching score, the user preference compliance score, and the system load assessment score are weighted and summed to obtain the total arbitration score for each conflict strategy pair; Based on the total arbitration score, execution priorities are assigned to conflicting strategy pairs, with the strategy component with the higher total arbitration score receiving a higher execution priority.

[0013] Preferably, the system further includes an update step for the emotion space network: After executing the executable interactive instruction sequence, subsequent interactive feedback data of the user is collected, including task completion efficiency indicators and user satisfaction indicators. The emotion perception effectiveness evaluation value of the current interaction cycle is calculated based on the subsequent interaction feedback data. When the emotion perception effectiveness evaluation value is lower than the preset update threshold, the parameter update process of the emotion space network is triggered. The parameter update process includes: extracting the dynamic emotion flow, the emotion evolution trend, and the subsequent interaction feedback data within the current interaction cycle to form network training samples; and using the network training samples to perform backpropagation adjustment on the emotion basis vector matrix in the emotion space network to update the connection weights of the emotion basis vector matrix.

[0014] Preferably, the step of calculating the emotion perception effectiveness evaluation value for the current interaction cycle based on the subsequent interaction feedback data includes: The task completion efficiency index is normalized to obtain a standardized task efficiency value. The user satisfaction index is quantified and encoded to obtain a standardized satisfaction value. A performance weighting coefficient is assigned to the standardized task efficiency value and the standardized satisfaction value, and the performance weighting coefficient is dynamically configured according to the type of interactive task; The weighted standardized task efficiency value is added to the standardized satisfaction value to obtain the emotion perception efficacy evaluation value.

[0015] Preferably, the task completion efficiency index includes task completion time and number of erroneous operations; the user satisfaction index is derived from direct user ratings or implicit satisfaction inference.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Dynamic emotion streams are projected onto a pre-defined emotion space network. Through dynamic activation of network nodes and real-time adjustment of connection weights, multidimensional signals are fused and transformed into a continuous emotion space coordinate and energy distribution map. This method breaks through the limitations of traditional discrete emotion classification, achieving a high-dimensional, continuous quantitative representation of users' emotional states. It can capture the intensity, mixing degree, and precise location of emotions in the abstract space more delicately, providing a much more refined data foundation than classification labels for subsequent processing.

[0017] By introducing vector field analysis tools, the divergence and curl of emotional spatial coordinates and energy distribution maps are calculated, thereby revealing the convergence and diffusion trends of emotional energy and the rotational characteristics of its evolutionary direction. This technology transforms the dynamic changes of emotions into a computable physical field model, quantifying the direction and intensity of emotional evolution. It enables predictive judgment of short-term emotional trends, allowing the system to move beyond passive reactions based on current emotional states and instead make forward-looking decisions based on the inherent trends of emotional evolution.

[0018] The interaction strategy generation network employs a structure of multiple parallel subnetworks, each specializing in policy generation for a specific interaction modality and receiving a unified emotional evolution trend as its core input. This architecture ensures that policy generation for different interaction modalities is based on the same understanding of the dynamic evolution of emotions. Subsequent fusion and conflict resolution modules collaborate on this foundation, ultimately generating a multimodal interaction instruction sequence that is consistent in timing and intent, achieving deep synchronization and natural integration between system response and user emotional changes. Attached Figure Description

[0019] Figure 1 This is a sequence diagram of the personalized interaction system based on user emotion perception described in this invention. Figure 2 A flowchart for Emotion Space Network Processing; Figure 3 Generate a flowchart of network operation for the interaction strategy; Figure 4 A composite chart of strategy priority arbitration scores; Figure 5 A line graph showing the results of the standardization process for the indicators. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1This invention provides a personalized interaction system based on user emotion perception. The system includes: a dynamic emotion flow construction module that captures multimodal physiological signal streams, interactive behavior time-series streams, and environmental parameter streams during real-time interaction with a target user, integrating these data streams into a unified dynamic emotion flow; an emotion space network processing module that receives this dynamic emotion flow and projects it onto a preset emotion space network, which is composed of an emotion basis vector matrix; and a module that processes input data through node activation and connection weight adjustment to output the target user's specific coordinates and emotional energy distribution map within the emotion space; and an emotion evolution trend calculation module that calculates the emotion space... The coordinate and emotional energy distribution map is analyzed. This analysis process relies on calculating the divergence and curl of the emotional vector field to generate an emotional evolution trend that includes a clear direction of emotional evolution and a quantified intensity of emotional evolution. The interaction strategy generation network module receives this emotional evolution trend and contains multiple parallel strategy generation sub-networks. Each sub-network corresponds to an interaction mode, and each sub-network independently outputs a preliminary set of interaction strategy vectors. The strategy fusion and conflict resolution module processes this preliminary set of interaction strategy vectors, executes the logic of strategy fusion and conflict resolution, and finally generates a coordinated, conflict-free, executable sequence of interaction instructions and delivers it to the executor.

[0022] Example 1: See Figure 2 The dynamic emotional stream is projected onto a pre-defined emotional space network for processing. This process performs feature extraction and time alignment operations on the multimodal physiological signal stream to form a standardized physiological feature tensor. It also performs action segmentation and intent recognition operations on the temporal stream of interactive behaviors to form a behavioral intent sequence with precise timestamps. Finally, it performs event detection and context-related annotation operations on the environmental parameter stream to form a structured environmental event sequence. The physiological feature tensor, behavioral intent sequence, and structured environmental event sequence are then fed into a multi-stream fusion unit. This unit employs an attention weight allocation mechanism, assigning physiological attention weights to the physiological feature tensor based on the signal-to-noise ratio and energy proportion of specific frequency bands of the physiological signal; assigning behavioral attention weights to the behavioral intent sequence based on the clarity of the behavioral intent and the frequency of the behavior; and assigning environmental attention weights to the structured environmental event sequence based on the suddenness of the environmental event and the strength of its historical association. The weighted physiological feature tensor, behavioral intent sequence, and structured environmental event sequence are concatenated along the time dimension. The resulting high-dimensional vector is then subjected to dimensionality reduction and smoothing processing to form a fused emotional feature vector sequence. The sequence of emotion feature vectors is input into the emotion space network, which is constructed from a pre-trained emotion basis vector matrix. The emotion space coordinates are determined by calculating the projection of the emotion feature vector sequence onto the emotion basis vector matrix, and an emotion energy distribution map is generated by mapping through an activation function.

[0023] In practical implementation, a user is using the in-vehicle infotainment system for navigation. The dynamic emotion flow construction module collects multimodal physiological signal streams, interactive behavior time-series streams, and environmental parameter streams in real time. The multimodal physiological signal streams originate from the hand grip pressure sensor on the steering wheel and the heart rate sensor on the driver's seat. The interactive behavior time-series streams originate from the user's interaction logs with the central control touchscreen. The environmental parameter streams originate from the in-vehicle camera and positioning system. Feature extraction and time alignment are performed on the multimodal physiological signal streams to form standardized physiological feature tensors. For example, heart rate variability features are extracted from the heart rate sensor signals, and pressure change frequency features are extracted from the hand grip pressure sensor signals. These two types of features are aligned and combined under the same millisecond-level timestamp to form a physiological feature tensor. Action segmentation and intent recognition are performed on the interactive behavior time-series streams to form a sequence of behavioral intents with timestamps. For example, continuous tapping and swiping operations by the user on the touchscreen are segmented into independent input events, and "rapidly tapping a point on the map multiple times" is identified as a "zooming the map" behavior with a clear intent. Event detection and context-related annotation are performed on the environmental parameter stream to form a structured environmental event sequence. For example, if a camera detects a sudden road construction obstacle on the road ahead, the positioning system can associate it with the current road segment as a congested section, thus annotating it as a "sudden road obstacle" environmental event.

[0024] In practical implementation, physiological feature tensors, behavioral intention sequences, and structured environmental event sequences are fused using a multi-stream fusion mechanism. This multi-stream fusion employs an attention weight allocation mechanism. Physiological attention weights are assigned to the physiological feature tensor, behavioral attention weights to the behavioral intention sequence, and environmental attention weights to the structured environmental event sequence. The physiological attention weight is calculated based on the signal-to-noise ratio (SNR) of the physiological signal and the proportion of energy in a specific frequency band. For example, a heart rate signal with a high SNR and prominent low-frequency energy is assigned a higher physiological attention weight. The behavioral attention weight is calculated based on the clarity of the behavioral intention and the frequency of the behavior. For example, if a user frequently performs the explicit action of "cancel route" within a short period, a higher behavioral attention weight is assigned. The environmental attention weight is calculated based on the suddenness of the environmental event and the strength of its historical association. For example, a detected "sudden road obstacle" environmental event is often associated with negative user emotions in historical data, thus assigning a higher environmental attention weight. When calculating the fusion weights of the three types of attention, the following relationship is introduced: in: This represents the fusion weighting factor used for subsequent splicing. This represents the calculated physiological attention weights. This represents the calculated behavioral attention weights. This represents the calculated environmental attention weights. , , These are dynamic adjustment coefficients related to the current interaction stage. The weighted physiological feature tensor, behavioral intention sequence, and structured environmental event sequence are concatenated along the time dimension, and the concatenated high-dimensional vector is subjected to dimensionality reduction and smoothing to form a fused emotion feature vector sequence.

[0025] It is understandable that the fused emotional feature vector sequence is input into the emotional space network, which is constructed from the emotional basis vector matrix. For example, by calculating the projection of the emotional feature vector sequence onto the emotional basis vector matrix, the user's current emotional space coordinates are determined as (V, A), where V represents the valence dimension coordinates and A represents the arousal dimension coordinates. Through activation function mapping, an emotional energy distribution map centered on this coordinate is generated, which presents the probability density distribution in the emotional space. In some embodiments, the results of the attention weight allocation mechanism can be illustrated through data comparison. Assuming that at the first moment, the physiological signal is stable, the behavioral intention is ambiguous, and there are no environmental events, the calculated physiological attention weight is 0.6, the behavioral attention weight is 0.1, and the environmental attention weight is 0.3. At the second moment, the physiological signal fluctuates significantly, and at the same time, a clear "frequent error correction" operation and a "traffic congestion" environmental event are detected, resulting in a calculated physiological attention weight of 0.4, a behavioral attention weight of 0.4, and an environmental attention weight of 0.2. This dynamic change in weights reflects the multi-stream fusion mechanism's differentiated consideration of the credibility and importance of information sources in different contexts, and directly affects the composition of the fused emotion feature vector sequence.

[0026] Example 2: See Figure 3The analysis process of emotional space coordinates and emotional energy distribution maps unfolds within an emotional space network. Starting with emotional space coordinates, continuous emotional space coordinate segments are extracted along the time axis. For each emotional space coordinate segment, the gradient of its corresponding emotional energy distribution map in the emotional space is calculated, yielding a local emotional energy change gradient field. The divergence of this local emotional energy change gradient field is calculated to characterize the aggregation or diffusion trend of emotional energy, resulting in an emotional energy divergence value. Simultaneously, the curl of this gradient field is calculated to characterize the rotation or eddy current trend of the emotional state, yielding an emotional energy curl value. The emotional energy divergence and curl values ​​are then vectorized and normalized to obtain the emotional evolution direction and intensity, representing the global trend of change. The emotional evolution trend is input into an interaction strategy generation network, which includes a voice interaction strategy subnetwork, a visual content presentation strategy subnetwork, and an operation flow adjustment strategy subnetwork. The voice interaction strategy subnetwork receives the emotion evolution trend and outputs voice strategy components including voice intonation parameters, word choice vectors, and speech rate rhythm vectors. The visual content presentation strategy subnetwork receives the emotion evolution trend and outputs visual strategy components including interface color parameters, information layout vectors, and dynamic effect vectors. The operation flow adjustment strategy subnetwork receives the emotion evolution trend and outputs operation strategy components including task step sequence vectors, interaction feedback delay parameters, and option complexity parameters. The voice strategy components, visual strategy components, and operation strategy components together constitute the initial set of interaction strategy vectors.

[0027] In practice, a user is engaging in continuous human-computer dialogue practice through an online learning system. The emotion space network processing module has already output the user's emotion space coordinates and emotion energy distribution map in the current dialogue round. The process of parsing the emotion space coordinates and emotion energy distribution map unfolds within the emotion space network. Starting from the current emotion space coordinates, continuous emotion space coordinate segments containing the last five dialogue rounds are extracted along the time axis. Each coordinate segment is associated with the emotion energy distribution map at the time of its generation. In practice, for each emotion space coordinate segment, the gradient of its corresponding emotion energy distribution map in the emotion space is calculated to obtain the local emotion energy change gradient field. The divergence of the local emotion energy change gradient field is calculated to characterize the aggregation or diffusion trend of emotion energy, yielding the emotion energy divergence value. Simultaneously, the curl of the local emotion energy change gradient field is calculated to characterize the rotation or eddy trend of the emotional state, yielding the emotion energy curl value. The emotion evolution trend calculation module combines the emotion energy divergence value and the emotion energy curl value, and through vector synthesis and normalization operations, obtains the emotion evolution direction and emotion evolution intensity, representing the global change trend. Within one calculation cycle, the emotional energy divergence value sequence is [+0.8, +0.5, +0.2], and the emotional energy curl value sequence is [-0.1, +0.3, +0.6]. The vector composition and normalization calculations follow the following relationship: in: This represents the calculated intensity scalar of emotion evolution. This represents the emotional energy divergence value for the i-th time slice. This represents the curl value of emotional energy in the i-th time slice. This represents the number of time slices. The calculation results indicate that emotional energy exhibits a complex trend of weakening diffusion in space, accompanied by enhanced state rotation.

[0028] The emotion evolution trend is input into the interaction strategy generation network, which includes a voice interaction strategy subnetwork, a visual content presentation strategy subnetwork, and an operation flow adjustment strategy subnetwork. The voice interaction strategy subnetwork receives the emotion evolution trend and outputs voice strategy components including voice intonation parameters, word choice vectors, and speech rate / rhythm vectors. The visual content presentation strategy subnetwork receives the emotion evolution trend and outputs visual strategy components including interface color parameters, information layout vectors, and dynamic effect vectors. The operation flow adjustment strategy subnetwork receives the emotion evolution trend and outputs operation strategy components including task step sequence vectors, interaction feedback delay parameters, and option complexity parameters. The voice strategy components, visual strategy components, and operation strategy components together constitute a preliminary set of interaction strategy vectors.

[0029] In some embodiments, data comparison can illustrate the dynamic adjustments of the interaction strategy generation network. When the calculated intensity of emotion evolution... When the emotional intensity is high and the direction of emotional evolution points towards the "increasing frustration" region, the word choice vector output by the voice interaction strategy subnetwork will switch from "neutral statement" to "highly supportive and guiding," the interface color parameter output by the visual content presentation strategy subnetwork will be adjusted from "high contrast" to "soft, low-saturation color scheme," and the option complexity parameter output by the operation flow adjustment strategy subnetwork will indicate "providing step-by-step prompts." It can be understood that when the intensity of emotional evolution... When the emotional evolution is low and points towards the "increased focus" region, the strategy components generated by each subnetwork will be adjusted accordingly. For example, the wording selection vector regression will be "neutral and professional," the interface color parameters will remain "clear and efficient," and the option complexity parameters will allow for "displaying the complete challenge." Optionally, the voice interaction strategy subnetwork contains a prosodic model based on the emotional evolution trend, which maps the direction and intensity of emotional evolution to specific phonetic parameters. The visual content presentation strategy subnetwork contains a graphics rendering strategy engine that calculates the visual attributes of interface elements in real time based on the emotional evolution trend. The outputs of each subnetwork in the interaction strategy generation network are parallel; the voice strategy components, visual strategy components, and operational strategy components are temporally aligned, collectively forming a multimodal response plan for the current emotional evolution stage.

[0030] Example 3: Policy fusion and conflict resolution are performed on the initial set of interaction policy vectors. First, a policy conflict detection rule base is established, defining incompatibility conditions between policy vectors of different interaction modalities. The voice policy components, visual policy components, and operation policy components in the initial set of interaction policy vectors are compared pairwise, and conflicting policy pairs are identified based on the policy conflict detection rule base. For each identified conflicting policy pair, a policy priority arbitrator is invoked. This arbitrator assigns execution priorities to the conflicting policy pairs based on the main objective of the current interaction task, user historical preferences, and system resource status. Based on the assigned execution priorities, the parameters of lower-priority policy components are adjusted or replaced until all conflicting policy pairs meet the compatibility conditions. All adjusted policy components are sorted and encapsulated according to the interaction timeline to form the final executable interaction instruction sequence.

[0031] In practice, a user is practicing math problems through an online learning system. The interaction strategy generation network module has output a preliminary set of interaction strategy vectors. This set includes voice strategy components generated by the voice interaction strategy sub-network, visual strategy components generated by the visual content presentation strategy sub-network, and operation strategy components generated by the operation flow adjustment strategy sub-network. The strategy fusion and conflict resolution module establishes a strategy conflict detection rule base. This rule base defines incompatible conditions between strategy vectors of different interaction modalities. For example, the rule base may include "high-awakening positive voice feedback" and "cold-toned interface" as incompatible conditions, or "simplified operation steps" and "add detailed explanation pop-up" as incompatible conditions.

[0032] In practice, the initial set of interaction strategy vectors, including voice strategy components, visual strategy components, and operational strategy components, is compared pairwise. A strategy conflict detection rule base is used to identify conflicting strategy pairs. For example, a voice strategy component might include "providing immediate encouragement using a rapid, high-pitched tone," a visual strategy component might include "highlighting incorrect answer areas with flashing red," and an operational strategy component might include "automatically jumping to an easier question after the current question is answered." After comparing with the rule base, it was found that "providing immediate encouragement using a rapid, high-pitched tone" and "highlighting incorrect answer areas with flashing red" conflict in terms of emotional guidance. The former aims to enhance positive emotions, while the latter may exacerbate the user's frustration. Therefore, a conflicting strategy pair is identified. The conflict detection process can be formally represented by a matching function: in: Represents policy components With strategy components The conflict detection results It is a binary indicator function, when the policy component With strategy components The parameter combination takes a value of 1 if it satisfies any incompatible condition in the policy conflict detection rule base, and a value of 0 otherwise. This indicates the conflict level identifier corresponding to the incompatible condition that was triggered.

[0033] For each identified conflicting strategy pair, the strategy priority arbitrator is invoked. The arbitrator assigns execution priorities to the conflicting strategy pairs based on the primary objective of the current interaction task, user historical preferences, and system resource status. According to the assigned execution priorities, the parameters of lower-priority strategy components are adjusted or replaced until all conflicting strategy pairs meet compatibility requirements. For example, the strategy priority arbitrator determines that the primary objective of the current interaction task is to maintain the user's learning continuity rather than immediately correcting errors; user historical preferences show low ratings for abrupt visual alerts; and the system currently has sufficient computing resources. Therefore, the "automatically jump to a simpler question" operation strategy component is assigned the highest priority, and the "blinking red highlight" visual strategy component is assigned the lowest priority. The parameters of the lower-priority "blinking red highlight" visual strategy component are adjusted, replacing it with "providing a gentle yellow border as a prompt." In practice, after all conflicting strategy pairs have been processed according to this logic, all adjusted strategy components are sorted and encapsulated according to the interaction timeline to form the final executable interaction instruction sequence.

[0034] In some embodiments, data comparison can illustrate the effectiveness of strategy fusion and conflict resolution. Assuming that in the first conflict detection, three conflicting strategy pairs are identified, after strategy priority arbitration and adjustment, the final executable interactive instruction sequence ensures that the voice, visual, and operational feedback remain consistent in emotional orientation. In another interaction, since the main goal of the current task becomes "accurately mastering knowledge points," and the user's historical preferences display detailed preference analysis, the strategy priority arbitrator may assign high priority to the "add detailed analysis pop-up" operation strategy component, while assigning low priority to the "automatic question skipping" strategy and replacing it. It can be understood that the strategy fusion and conflict resolution process is dynamic, and the strategy priority arbitrator's decisions change based on real-time input. Optionally, the strategy conflict detection rule base can be incrementally updated based on online learning data to incorporate newly discovered incompatible patterns. In some embodiments, parameter adjustment or replacement operations may include directly replacing a functionally equivalent but parameter-different strategy component from a preset compatible strategy template library. Optionally, when a conflict cannot be eliminated through adjustment, the entire interaction modality containing the lowest-priority strategy component can be temporarily disabled. Understandably, the final generated sequence of executable interactive instructions is sent to the various executors of the system to drive the speech synthesis engine, graphical interface, and application programming interface to produce a consistent user experience.

[0035] See Figure 4 This is a composite chart of strategy priority arbitration scores, primarily used to demonstrate the performance of different strategy components across the arbitration dimension. System load assessment scored the highest (close to 95), and task objective matching was also at a high level (approximately 85); the total arbitration score was approximately 82, indicating a relatively good overall performance. All dimensions scored the lowest, with a total arbitration score of approximately 67, making it the worst performing of the three strategy categories. Task objective matching scored the highest (approximately 90), and user preference compliance was also at a high level (approximately 80); the total arbitration score was approximately 82, close to that of the voice strategy. This type of chart is commonly found in personalized interaction systems and strategy decision-making scenarios, used to quantify the priority of different strategies, aiding in subsequent strategy fusion and conflict resolution. Strategies with high scores are assigned higher priority, while low-scoring strategies may be adjusted or replaced.

[0036] Example 4: The invocation process of the strategy priority arbitrator involves its internal functional units. The strategy priority arbitrator embeds a task target matching degree calculation unit, used to calculate the degree of matching between each conflicting strategy pair and the main target of the current interaction task, outputting a task target matching degree score. It also embeds a user preference compliance calculation unit, used to query user historical interaction logs, calculate the degree of compliance between each conflicting strategy pair and the user's historical preferences, outputting a user preference compliance score. Finally, it embeds a system load assessment unit, used to assess the computational resources and time cost required to execute each conflicting strategy pair, outputting a system load assessment score. The task target matching degree score, user preference compliance score, and system load assessment score are weighted and summed to obtain the total arbitration score for each conflicting strategy pair. Based on the total arbitration score, execution priorities are assigned to the conflicting strategy pairs, with the strategy component with the higher total arbitration score receiving a higher execution priority.

[0037] In practice, a user is practicing math through an online learning system. The strategy fusion and conflict resolution module identifies a conflicting strategy pair. This conflict involves the voice strategy component "using fast, high-frequency encouraging voice feedback" generated by the voice interaction strategy sub-network and the visual strategy component "using dynamic red cross animation to mark errors" generated by the visual content presentation strategy sub-network. The strategy priority arbitrator is invoked. This arbitrator has an embedded task goal matching calculation unit to calculate the degree of matching between each conflicting strategy pair and the main goal of the current interaction task, outputting a task goal matching score. For example, if the main goal of the current interaction task is "to deepen understanding while maintaining learning motivation," the task goal matching calculation unit analyzes the positive contribution of "fast, high-frequency encouraging voice feedback" to maintaining motivation and the potential negative impact of "dynamic red cross animation to mark errors" on motivation, providing quantitative scores for each.

[0038] The strategy priority arbitrator embeds a user preference compliance calculation unit. This unit queries historical user interaction logs, calculates the degree of compliance between each conflicting strategy and the user's historical preferences, and outputs a user preference compliance score. For example, log analysis might reveal that in the past ten similar scenarios, the user gave a higher satisfaction score to the feedback combination of "gentle prompt tone with a static yellow exclamation mark" and a lower score to the feedback combination of "sharp prompt tone with a red animation." The user preference compliance calculation unit uses this historical data to calculate compliance scores for the two strategy components of the current conflict. The strategy priority arbitrator also embeds a system load assessment unit. This unit assesses the computational resources and time cost required to execute each conflicting strategy and outputs a system load assessment score. For example, generating "fast, high-frequency encouraging voice feedback" requires real-time speech synthesis and rendering, resulting in a high computational load, while presenting "dynamic red cross animation" requires calling a graphics engine, resulting in a lower time cost. The system load assessment unit evaluates the impact of these two operations on the system's real-time performance and converts it into a score.

[0039] The arbitration score for each conflicting strategy pair is obtained by weighted summing of the task objective matching score, user preference compliance score, and system load assessment score, following the relationship below: in: This represents the total arbitration score calculated. The task target matching score is the output of the task target matching score calculation unit. This represents the user preference conformity score output by the user preference conformity calculation unit. The system load assessment score represents the output of the system load assessment unit. , , These are dynamic weight coefficients related to the context of the current interaction phase. Based on the total arbitration score, execution priorities are assigned to conflicting strategy pairs, with strategies having higher total arbitration scores receiving higher execution priorities. Referring to Table 1, in a specific conflict resolution cycle, assume the system calculates the scores and results shown in Table 1.

[0040] Table 1: Calculation Table for Strategy Conflict Arbitration Conflict strategies for description Task target matching score User preference compliance score System load assessment score Arbitration total score Execution priority Voice strategy: rapid, high-frequency encouragement 0.85 0.60 0.70 0.72 high Visual strategy: Dynamic red cross 0.40 0.20 0.90 0.50 Low Operation Strategy A: Instant pop-up analysis 0.90 0.80 0.50 0.73 high Operation Strategy B: Delayed Prompt Button 0.75 0.70 0.95 0.80 high In practice, Table 1 shows the arbitration calculation for two independent policy conflict pairs. In the first conflict pair, the total arbitration score of the speech policy component is calculated based on the given dynamic weighting coefficients. The total arbitration score for the visual strategy component is 0.72. The score is 0.50, therefore a high execution priority is assigned to the speech policy component and a low execution priority to the visual policy component, and the parameters of the visual policy component are then adjusted. In the second conflict pair, the total arbitration score of operation policy B is calculated. It is higher, therefore it receives a higher execution priority.

[0041] In some embodiments, when the system detects that the current interactive task is at a critical teaching node... The value will be increased, making the task target matching degree have a greater impact on the arbitration result. Optionally, in low-power mode where system resources are scarce, The value will be increased, making system load a higher priority consideration. The matching degree evaluation of the task target matching degree calculation unit can be calculated based on the cosine similarity between the policy component and the preset task target feature vector. The conformity evaluation of the user preference conformity calculation unit can be obtained by querying the user profile database and weighting the historical acceptance scores of various interaction strategies. It can be understood that the evaluation of the system load evaluation unit can be modeled based on the expected execution time and memory usage of the computational tasks triggered by the policy components in the processor. The policy priority arbitrator independently calculates the arbitration total score for the policy components in each conflict pair. Priorities are assigned by comparing total scores, a process that is parallel and real-time. In some embodiments, when two conflicting strategies are arbitrated based on their total scores... When the difference is less than a preset threshold, the policy priority arbitrator will initiate an additional random selection or fallback mechanism to the default policy. Optional, the total arbitration score... The calculation results and scores of each item will be recorded in the log for subsequent offline analysis of the adjustment of the weight coefficients of the strategy priority arbitrator.

[0042] Example 5: The system includes an update step for the emotion space network. After executing an executable interactive instruction sequence, subsequent user interaction feedback data is collected, including task completion efficiency indicators and user satisfaction indicators. Based on the subsequent interaction feedback data, an emotion perception effectiveness evaluation value for the current interaction cycle is calculated. The task completion efficiency indicator is normalized to obtain a standardized task efficiency value, and the user satisfaction indicator is quantified and encoded to obtain a standardized satisfaction value. Efficacy weight coefficients are assigned to the standardized task efficiency value and the standardized satisfaction value. The efficiency weight coefficients are dynamically configured according to the interaction task type. The weighted standardized task efficiency value and the standardized satisfaction value are added together to obtain the emotion perception effectiveness evaluation value. When the emotion perception effectiveness evaluation value is lower than a preset update threshold, the parameter update process of the emotion space network is triggered. The parameter update process includes extracting the dynamic emotion flow, emotion evolution trend, and subsequent interaction feedback data within the current interaction cycle to form a network training sample. This network training sample is used to perform backpropagation adjustment on the emotion basis vector matrix in the emotion space network, updating the connection weights of the emotion basis vector matrix. The task completion efficiency indicator includes task completion time and number of erroneous operations, and the user satisfaction indicator is derived from direct user ratings or implicit satisfaction inference.

[0043] In practice, when a user completes an interactive task cycle called "Application of Geometric Theorems" in an online learning system, after executing a sequence of executable interactive instructions generated based on emotion perception, the system initiates an update step on the emotion space network. Subsequent interactive feedback data is collected, including task completion efficiency metrics and user satisfaction metrics. Task completion efficiency metrics include task completion time and the number of errors. User satisfaction metrics are derived from direct user ratings or implicit satisfaction inferences. For example, the system records that a user completed a set of five practice questions in 420 seconds, made 3 errors, and collected a direct rating of 7 out of 10 from a pop-up window.

[0044] Based on subsequent interaction feedback data, the emotion perception effectiveness evaluation value of the current interaction cycle is calculated. The task completion efficiency index is normalized to obtain a standardized task efficiency value; for example, historical data is used to map task completion time and the number of erroneous operations to values ​​between 0 and 1. The user satisfaction index is quantified and encoded to obtain a standardized satisfaction value; for example, a direct user rating of 7 is linearly converted to 0.7, or a value of 0.65 is obtained through model inference of implicit user behaviors such as dwell time and repeated learning counts. Efficiency weight coefficients are assigned to the standardized task efficiency value and the standardized satisfaction value. These coefficients are dynamically configured according to the interaction task type. For example, for "knowledge learning" tasks, the efficiency weight coefficient for the standardized task efficiency value is configured as 0.4, and the efficiency weight coefficient for the standardized satisfaction value is configured as 0.6; for "skill challenge" tasks, the efficiency weight coefficient for the standardized task efficiency value is configured as 0.7, and the efficiency weight coefficient for the standardized satisfaction value is configured as 0.3. The calculation follows the following relationship: in: This represents the calculated emotional perception efficacy rating. This represents a standardized task efficiency value. Represents a standardized satisfaction score. The efficiency weighting coefficient represents the performance value assigned to the standardized task efficiency value. This represents the efficacy weighting coefficient assigned to the standardized satisfaction value. The weighted standardized task efficiency value is added to the standardized satisfaction value to obtain the emotion perception efficacy evaluation value.

[0045] In practical implementation, assuming that the current "geometric theorem application" task is classified as "knowledge learning", the standardized task efficiency value is calculated. The standardized satisfaction score is 0.72. The efficiency weighting coefficient is 0.70. It is 0.4. The value is 0.6. Substituting this value into the formula, we obtain the emotional perception efficacy evaluation score. The value is 0.708. The system's preset update threshold is 0.75, when the emotion perception efficacy evaluation value... When the value falls below the preset update threshold of 0.75, the parameter update process of the emotion space network is triggered. The parameter update process includes extracting the dynamic emotion flow, emotion evolution trend, and subsequent interaction feedback data within the current interaction cycle to form network training samples. The network training samples are then used to perform backpropagation adjustment on the emotion basis vector matrix in the emotion space network to update the connection weights of the emotion basis vector matrix.

[0046] In some embodiments, data comparison can illustrate the impact of dynamic configuration of efficacy weight coefficients. For example, for the same user, the emotion perception efficacy evaluation value is calculated when completing a "knowledge learning" type task. The calculated emotional perception efficacy score is 0.71. However, when completing "skill challenge" tasks, due to different efficacy weighting coefficient configurations, even if the original efficiency and satisfaction data are similar, the calculated score differs. The value might be 0.68, reflecting the adaptability of the assessment criteria to the task objectives. This is understandable; only when the emotion perception efficacy rating is... Only when the value falls below a threshold is the current mapping relationship of the emotion space network considered to have failed to effectively support the interaction goal, thus triggering an update. Optionally, the update threshold can also be dynamically fine-tuned based on task type or long-term user performance. The goal of backpropagation adjustment is to ensure that the emotion space network's prediction of emotion evolution trends for similar dynamic emotion stream inputs leads to higher future emotion perception efficacy evaluation values. .

[0047] See Figure 5 This is a line chart showing the results of standardized indicator processing, corresponding to the "indicator standardization stage" in the emotion space network update process. Its core purpose is to display the changing trends of standardized indicators across different task cycles. Overall, the values ​​are in the relatively high range of 0.7-0.85, with small fluctuations, indicating strong stability in user feedback. Larger fluctuations (0.5-0.85), such as a significant increase in efficiency values ​​in cycle 2 and a noticeable decrease in cycle 7, reflect the cyclical changes in task execution efficiency. In most cycles, efficiency values ​​and satisfaction values ​​show a weak inverse correlation, with one increasing as the other decreases. This type of chart is a preliminary step in emotion perception effectiveness assessment. By mapping different dimensions of "task efficiency" and "user satisfaction" to the 0-1 range, it provides standardized basic data for subsequent calculation of the "emotion perception effectiveness evaluation value," ensuring the rationality of the weight allocation and weighted calculation of different indicators.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A personalized interaction system based on user emotion perception, characterized in that, The system includes: The dynamic emotion stream construction module is used to construct a dynamic emotion stream based on the real-time interaction process of the target user. The dynamic emotion stream integrates multimodal physiological signal stream, interaction behavior time sequence stream and environmental parameter stream. The emotional space network processing module is used to project the dynamic emotional flow onto a preset emotional space network for processing. The emotional space network outputs the emotional space coordinates and emotional energy distribution map of the target user through node activation and connection weight adjustment. The emotion evolution trend calculation module is used to analyze the emotion spatial coordinates and emotion energy distribution map. This analysis process relies on the divergence and curl calculation of the emotion vector field to generate an emotion evolution trend that includes the direction and intensity of emotion evolution. An interaction strategy generation network module is used to input the emotion evolution trend into the interaction strategy generation network. The interaction strategy generation network contains multiple parallel strategy generation sub-networks. Each strategy generation sub-network corresponds to an interaction mode and outputs a preliminary set of interaction strategy vectors. The strategy fusion and conflict resolution module is used to perform strategy fusion and conflict resolution on the preliminary set of interaction strategy vectors to generate the final executable sequence of interaction instructions.

2. The personalized interaction system based on user emotion perception according to claim 1, characterized in that, The process of projecting the dynamic emotion stream onto a preset emotion space network for processing includes the following sub-steps: Feature extraction and time alignment are performed on the multimodal physiological signal stream to form a standardized physiological feature tensor; The interactive behavior time-series is segmented into actions and identified into intents to form a behavior intent sequence with timestamps; The environmental parameter stream is subjected to event detection and context-related annotation to form a structured environmental event sequence; The physiological feature tensor, the behavioral intention sequence, and the structured environmental event sequence are fused using a multi-stream fusion mechanism with an attention weight allocation mechanism to generate a fused emotional feature vector sequence. The emotional feature vector sequence is input into the emotional space network, which is constructed from an emotional basis vector matrix. The emotional space coordinates are determined by calculating the projection of the emotional feature vector sequence onto the emotional basis vector matrix, and the emotional energy distribution map is generated by mapping through an activation function.

3. The personalized interaction system based on user emotion perception according to claim 2, characterized in that, The multi-stream fusion employs an attention weight allocation mechanism, including: Configure physiological attention weights for the physiological feature tensor, configure behavioral attention weights for the behavioral intention sequence, and configure environmental attention weights for the structured environmental event sequence; The physiological attention weight is calculated based on the signal-to-noise ratio of the physiological signal and the energy proportion of a specific frequency band. The calculation of the behavioral attention weight is based on the clarity of the behavioral intention and the frequency of the behavior. The environmental attention weight is calculated based on the suddenness of environmental events and the strength of their historical correlation. The weighted physiological feature tensor, the behavioral intention sequence, and the structured environmental event sequence are concatenated along the time dimension, and the concatenated high-dimensional vector is subjected to dimensionality reduction and smoothing to form the fused emotion feature vector sequence.

4. The personalized interaction system based on user emotion perception according to claim 2, characterized in that, The analysis of the emotional spatial coordinates and emotional energy distribution map includes: In the emotional space network, starting from the emotional space coordinates, continuous emotional space coordinate segments are extracted along the time axis; For each of the aforementioned emotional space coordinate segments, the gradient of its corresponding emotional energy distribution map in the emotional space is calculated to obtain the local emotional energy change gradient field. The divergence of the local emotional energy change gradient field is calculated to characterize the aggregation or diffusion trend of emotional energy, and the emotional energy divergence value is obtained. The curl of the local emotional energy change gradient field is calculated to characterize the rotation or vortex trend of the emotional state, and the emotional energy curl value is obtained. By combining the emotional energy divergence value and the emotional energy curl value, and through vector synthesis and normalization, the emotional evolution direction and the emotional evolution intensity, which characterize the global trend of change, are obtained.

5. The personalized interaction system based on user emotion perception according to claim 1, characterized in that, The step of inputting the emotion evolution trend into the interaction strategy generation network includes: The interaction strategy generation network includes a voice interaction strategy subnetwork, a visual content presentation strategy subnetwork, and an operation process adjustment strategy subnetwork. The voice interaction strategy subnetwork receives the emotion evolution trend and outputs voice strategy components including voice intonation parameters, word choice vectors, and speech rate rhythm vectors. The visual content presentation strategy subnetwork receives the emotion evolution trend and outputs visual strategy components including interface color parameters, information layout vectors, and dynamic effect vectors. The operation process adjustment strategy sub-network receives the emotion evolution trend and outputs operation strategy components including task step sequence vector, interaction feedback delay parameter and option complexity parameter. The voice strategy component, the visual strategy component, and the operation strategy component together constitute the preliminary interaction strategy vector set.

6. The personalized interaction system based on user emotion perception according to claim 5, characterized in that, The step of performing policy fusion and conflict resolution on the preliminary set of interaction policy vectors includes: A policy conflict detection rule base is established, which defines the incompatibility conditions between policy vectors of different interaction modalities; The speech policy component, the visual policy component, and the operation policy component in the preliminary interaction policy vector set are compared pairwise, and conflicting policy pairs are identified based on the policy conflict detection rule base. For each conflicting policy pair identified, the policy priority arbitrator is invoked. The policy priority arbitrator assigns an execution priority to the conflicting policy pair based on the main objective of the current interaction task, user historical preferences, and system resource status. Based on the execution priority, the parameters of the low-priority policy components are adjusted or replaced until all conflicting policy pairs meet the compatibility conditions. All adjusted strategy components are sorted and encapsulated according to the interaction timeline to form the final executable interactive instruction sequence.

7. The personalized interaction system based on user emotion perception according to claim 6, characterized in that, The invocation strategy priority arbitrator includes: The strategy priority arbitrator has an embedded task target matching degree calculation unit, which is used to calculate the matching degree of each conflict strategy pair with the main target of the current interaction task and output the task target matching degree score. The strategy priority arbitrator has an embedded user preference compliance calculation unit, which is used to query the user's historical interaction logs, calculate the degree of compliance between each conflicting strategy pair and the user's historical preferences, and output the user preference compliance score. The strategy priority arbitrator has an embedded system load assessment unit, which is used to assess the computational resources and time costs required to execute each conflicting strategy and output a system load assessment score. The task objective matching score, the user preference compliance score, and the system load assessment score are weighted and summed to obtain the total arbitration score for each conflict strategy pair; Based on the total arbitration score, execution priorities are assigned to conflicting strategy pairs, with the strategy component with the higher total arbitration score receiving a higher execution priority.

8. The personalized interaction system based on user emotion perception according to claim 1, characterized in that, It also includes an update step for the emotion space network: After executing the executable interactive instruction sequence, subsequent interactive feedback data of the user is collected, including task completion efficiency indicators and user satisfaction indicators. The emotion perception effectiveness evaluation value of the current interaction cycle is calculated based on the subsequent interaction feedback data. When the emotion perception effectiveness evaluation value is lower than the preset update threshold, the parameter update process of the emotion space network is triggered. The parameter update process includes: extracting the dynamic emotion stream, the emotion evolution trend, and the subsequent interaction feedback data within the current interaction cycle to form network training samples; The emotion basis vector matrix in the emotion space network is adjusted by backpropagation using the network training samples, and the connection weights of the emotion basis vector matrix are updated.

9. The personalized interaction system based on user emotion perception according to claim 8, characterized in that, The calculation of the emotion perception effectiveness evaluation value for the current interaction cycle based on the subsequent interaction feedback data includes: The task completion efficiency index is normalized to obtain a standardized task efficiency value. The user satisfaction index is quantified and encoded to obtain a standardized satisfaction value. A performance weighting coefficient is assigned to the standardized task efficiency value and the standardized satisfaction value, and the performance weighting coefficient is dynamically configured according to the type of interactive task; The weighted standardized task efficiency value is added to the standardized satisfaction value to obtain the emotion perception efficacy evaluation value.

10. The personalized interaction system based on user emotion perception according to claim 9, characterized in that, The task completion efficiency index includes task completion time and the number of erroneous operations; the user satisfaction index is derived from direct user ratings or implicit satisfaction inference.

Images