Emotion labeling method and terminal based on multi-source verification

Abstract

The application provides a multi-source verification emotion labeling method and a terminal. The method comprises the following steps: obtaining emotion inducing content and expert pre-evaluation results corresponding to the emotion inducing content, and generating a benchmark emotion coordinate according to the expert pre-evaluation results; collecting subjective self-evaluation results and physiological response signals of a target user in the process of experiencing the emotion inducing content; performing consistency verification on the benchmark emotion coordinate, the subjective self-evaluation results and the physiological response signals; labeling multi-dimensional emotion labels for the subjective self-evaluation results and the physiological response signals that pass the consistency verification, and adding the subjective self-evaluation results and the physiological response signals to a training sample set. Through cross verification of the expert benchmark, the subjective self-evaluation and the physiological response signals, the application effectively eliminates abnormal samples, solves the problems of unstable single subjective self-evaluation and high cost of pure expert labeling, and provides a high-quality training data basis for an emotion recognition model.

Description

Technical Field

[0001] This invention relates to the field of emotion recognition technology, and in particular to a multi-source verification emotion labeling method and terminal. Background Technology

[0002] In the construction of emotion recognition models based on electroencephalogram (EEG) signals, the quality of the training sample labels directly affects the model's recognition accuracy, generalization ability, and stability. Existing methods for constructing emotion labels mainly include three approaches: user subjective self-assessment, expert manual annotation, and automatic annotation of single signals. However, these methods still have the following shortcomings: 1. The labels are highly subjective and noisy: The subjective self-evaluation methods of users are greatly affected by individual differences in understanding, subjective preferences and immediate state, and the labels are unstable and difficult to serve as a reliable source of emotional labels; 2. Manual annotation is costly and lacks personalization: Although expert manual annotation has a certain degree of professionalism, it is costly, inefficient, and fails to reflect the differences in emotional experience among different individuals, thus failing to meet the needs of constructing large-scale training samples. 3. Lack of interpretability in single physiological signal labeling: The single signal automatic labeling method relies solely on physiological signals such as EEG to judge emotions, which lacks a psychological explanation basis, is prone to bias, and is difficult to verify the accuracy of the label. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a multi-source verification emotion labeling method and terminal that can generate highly reliable emotion labels based on multi-source consistency verification.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-source validation sentiment annotation method, characterized by comprising: Obtain the emotional triggering content and the corresponding expert pre-assessment results, and generate a baseline emotional coordinate based on the expert pre-assessment results; Collect the subjective self-evaluation results and physiological response signals of target users during the process of experiencing the emotionally induced content; The consistency of the baseline emotional coordinates, the subjective self-evaluation results, and the physiological response signals is verified. The subjective self-evaluation results and physiological response signals that pass the consistency check are labeled with multidimensional emotion tags and added to the training sample set.

[0005] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A multi-source verification sentiment annotation terminal includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the various steps of the aforementioned multi-source verification sentiment annotation method.

[0006] The beneficial effects of this invention are as follows: By acquiring emotion-inducing content and corresponding expert pre-assessment results and generating benchmark emotion coordinates, a stable and unified objective reference technology for emotion labeling is provided, solving the label drift problem of single subjective self-assessment; by collecting subjective self-assessment and physiological response signals of target users, individual differences are quantified under a unified benchmark, making up for the deficiency of pure expert labeling that ignores the user's real experience; by verifying the consistency of benchmark emotion coordinates, subjective self-assessment results, and physiological response signals, a multi-source cross-validation screening mechanism is constructed to eliminate abnormal samples and improve data purity; after passing the verification, multi-dimensional emotion labels are labeled and added to the sample set, ensuring that each sample is jointly verified by three types of information sources, providing a highly reliable and low-noise training data foundation for emotion recognition and intervention models, thereby improving the model's convergence speed, recognition accuracy, and cross-scenario generalization ability. This invention, through the mutual constraint and verification of multi-source information, elevates the construction of emotion labels from single-point subjective judgment to multi-source collaborative confirmation, realizing a substantial leap in label quality from uncontrollable to verifiable, screenable, and guaranteed, providing a standardized high-quality data production paradigm for emotion computing systems. Attached Figure Description

[0007] Figure 1 This is a flowchart of a multi-source validation sentiment annotation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-source verification emotion labeling terminal according to an embodiment of the present invention. Detailed Implementation

[0008] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0009] In existing technologies, emotion recognition model training relies on high-quality emotion labels. Current label construction methods include user subjective self-assessment, expert manual annotation, and automatic annotation of single signals. Subjective self-assessment is influenced by individual differences, resulting in unstable labels and high noise levels; expert annotation is costly, inefficient, and lacks personalization; single-signal annotation lacks a psychological explanation basis and is prone to bias. None of these methods integrate multi-source information for cross-validation, leading to deficiencies in the objectivity, consistency, and individual suitability of emotion labels, thus limiting the accuracy and generalization ability of emotion recognition models.

[0010] To address at least some of the aforementioned issues, this invention generates benchmark emotion coordinates by acquiring expert pre-assessment results of emotion-inducing content. Simultaneously, it collects users' subjective self-assessment results and physiological response signals, performs consistency verification on all three, and only labels multi-dimensional emotion tags on data that pass the verification and adds them to the training sample set. In this way, multi-source cross-validation using expert benchmarks, user self-assessments, and physiological response signals can eliminate abnormal samples such as those exhibiting subjective arbitrariness or inattentiveness. This solves the problems of unstable single subjective self-assessment labels, high cost and lack of personalization with pure expert labeling, and lack of interpretability with single physiological signal labeling. It provides a highly reliable, low-noise training data foundation for emotion recognition and intervention models, effectively improving the model's convergence speed, recognition accuracy, and cross-scenario generalization ability.

[0011] The following details a multi-source validation sentiment annotation method from this invention. Please refer to [link / reference]. Figure 1 The method 100 includes steps 110 to 140.

[0012] Step 110: Obtain the emotional triggering content and the corresponding expert pre-assessment results, and generate a baseline emotional coordinate based on the expert pre-assessment results.

[0013] For example, standardized music, videos, or digital content can be selected as emotional triggers. Experts can pre-evaluate each content segment, and the expert pre-evaluation results can be used as the baseline emotional coordinates for that content segment.

[0014] Step 120: Collect the subjective self-evaluation results and physiological response signals of the target users during the process of experiencing emotionally induced content.

[0015] For example, subjective ratings of target users during the experience of the aforementioned emotionally induced content are collected to obtain subjective self-evaluation results; at the same time, the target users' electroencephalogram (EEG) signals are collected as physiological response signals.

[0016] Step 130: Verify the consistency of the baseline emotional coordinates, subjective self-evaluation results, and physiological response signals.

[0017] For example, for each emotional trigger, the corresponding baseline emotional coordinates, subjective self-evaluation results, and physiological response signals are obtained, and the consistency of the three is verified.

[0018] Step 140: Label the subjective self-evaluation results and physiological response signals that have passed the consistency check with multidimensional emotion labels and add them to the training sample set.

[0019] For example, label multidimensional emotion tags on data that pass the consistency check, and add the labeled multidimensional emotion tag samples to the training sample set.

[0020] As described above, firstly, an objective and unified baseline emotion coordinate system is constructed based on expert pre-assessment results, providing a stable reference scale for emotion labels. Secondly, users' subjective self-assessment results and physiological response signals are collected simultaneously, incorporating individual differences in emotional experience into the labeling system. Then, multi-source consistency verification is performed on all three, constructing a cross-validation screening mechanism to effectively eliminate abnormal samples such as those exhibiting subjective arbitrariness or inattention. Finally, only data that passes verification are labeled with multi-dimensional emotion labels and added to the training sample set, ensuring that each sample undergoes collaborative verification from three information sources. Compared to existing schemes that rely solely on subjective self-assessment or pure expert labeling, this invention, through the mutual constraints and verification of multi-source information, elevates the construction of emotion labels from single-point subjective judgment to multi-source collaborative confirmation. This achieves a substantial leap in label quality from uncontrollable to verifiable, screenable, and guaranteed, providing a highly reliable, low-noise training data foundation for emotion recognition and intervention models, effectively improving the model's convergence speed, recognition accuracy, and cross-scenario generalization ability.

[0021] In one alternative implementation, step 110 includes steps 111 to 113.

[0022] Step 111: Obtain at least one of standardized music, video, and digital content as emotional triggering content.

[0023] For example, standardized music clips, short video clips, or other digital media content can be selected from a pre-set material library to ensure that the selected content has a uniform duration, format, and emotional induction effect, and is used as emotional induction content.

[0024] Step 112: Receive scores from experts in the field of emotion research on each emotion-inducing content on a preset emotion dimension, and obtain the expert pre-scoring results.

[0025] For example, experts in psychology or emotion research can score each emotion-inducing content on a preset emotion dimension, and the expert's score data on the preset emotion dimension can be received as the expert's pre-scoring result.

[0026] Step 113: Use the expert pre-score results as the baseline emotional coordinates for the emotionally induced content.

[0027] For example, the expert pre-score results corresponding to each emotion-inducing content are stored as the baseline emotion coordinates for that emotion-inducing content segment, which are then used for consistency comparison with user self-assessment results and physiological response signals.

[0028] As described above, by selecting standardized emotion-inducing content and receiving expert ratings on preset emotion dimensions, a stable and unified benchmark emotion coordinate system was constructed. This benchmark coordinate system is unaffected by individual user differences, providing an objective and reliable reference standard for subsequent multi-source consistency verification, fundamentally solving the label drift problem caused by a single subjective self-assessment method.

[0029] In one alternative implementation, step 112 includes step 1121.

[0030] Step 1121: Receive scores from experts in the field of emotion research on each emotion-evoking content in terms of emotional valence, emotional arousal, and sense of control, and obtain the expert pre-scoring results.

[0031] For example, experts in psychology or emotion research rate each emotion-inducing content on the following dimensions: emotional valence (pleasure or intensity of positive or negative emotion), emotional arousal (excitement or calmness), and perceived control (the individual's degree of control over the emotion). The system receives the experts' ratings on these three dimensions and generates pre-rated expert scores.

[0032] As described above, a three-dimensional baseline emotional coordinate system was constructed by receiving expert ratings on three dimensions: emotional valence, emotional arousal, and perceived control. This three-dimensional coordinate system can more precisely and comprehensively describe the characteristics of emotional states, encompassing both the positive and negative attributes of emotions, as well as the intensity of emotions and the individual's degree of control. It provides a multi-dimensional basis for comparison in subsequent consistency verification, making the construction of emotional labels more scientific and accurate.

[0033] In one optional implementation, step 120 collects the target user's subjective self-evaluation results during the experience of emotionally induced content, including steps 121 and 122.

[0034] Step 121: After each emotionally evoked content session, receive the target user's subjective ratings on the emotional valence dimension, emotional arousal dimension, and sense of control dimension to obtain the subjective self-evaluation results.

[0035] For example, after each piece of music or video content finishes playing, a rating interface pops up, prompting the target user to rate the content segment based on their own feelings in three dimensions: emotional valence, emotional arousal, and sense of control. The system receives the rating data input by the user and uses it as the subjective self-evaluation result for that content segment.

[0036] Step 122: Record the offset between the subjective self-evaluation result and the baseline emotional coordinate as individualized emotional offset information.

[0037] For example, the difference between the user's subjective rating and the expert's baseline rating in each dimension is calculated. This difference is stored as individualized emotional offset information to reflect the user's personalized emotional response characteristics when experiencing specific content.

[0038] Step 130 verifies the consistency between the subjective self-assessment results and the baseline emotional coordinates, including steps 1301 and 1302.

[0039] Step 1301: If the individualized emotion offset information is within the preset threshold range, it indicates that the subjective self-evaluation result is consistent with the baseline emotion coordinate.

[0040] For example, the preset threshold is that the offset of each dimension does not exceed 2 points. If the user's offset in the three dimensions of valence, arousal and sense of control is within the preset threshold range, it is determined that the subjective self-evaluation result is consistent with the baseline emotional coordinate, indicating that the user's emotional reaction is basically consistent with the expert's pre-evaluation result.

[0041] Step 1302: If the individualized emotion offset information exceeds the preset threshold range, it indicates that the subjective self-evaluation result is inconsistent with the baseline emotion coordinate.

[0042] For example, if a user's offset in any dimension exceeds a preset threshold, it is determined that the subjective self-assessment result is inconsistent with the baseline emotional coordinates, indicating that the user's emotional response deviates significantly from the expert's pre-assessment result. This sample may be affected by factors such as the user's subjective arbitrariness and lack of concentration.

[0043] As described above, by receiving users' subjective ratings across three dimensions and calculating their offset from the baseline coordinates, individual differences are quantified. By setting thresholds to judge the offsets, consistency verification between subjective self-assessment results and expert benchmarks is achieved. This mechanism can effectively identify the credibility of users' subjective self-assessments, providing a basis for subsequent removal of outlier samples, while preserving individualized emotional offset information, thus achieving personalized emotional expression within a unified benchmark framework.

[0044] In one optional implementation, step 120 collects physiological response signals of the target user during the process of experiencing emotionally induced content, including steps 123 to 126.

[0045] Step 123: Simultaneously collect physiological response signals from the left and right prefrontal cortex regions of the target user.

[0046] For example, after the target user wears a dual prefrontal EEG acquisition device, EEG signals from two channels, the left and right prefrontal regions, are acquired simultaneously. The sampling rate is set to 256 Hz or higher to ensure that neurophysiological activities related to emotional changes can be captured, and the acquired raw EEG signals are used as physiological response signals.

[0047] Step 124: Perform time-series preprocessing on the collected physiological response signals.

[0048] For example, the acquired raw EEG signals are sequentially subjected to time-series preprocessing operations such as baseline drift removal, filtering and noise reduction, removal of electrooculogram artifacts, segmentation and normalization to remove noise interference in the signals and obtain clean and usable EEG time-series data.

[0049] Step 125: Extract the temporal or implicit feature representations of the physiological response signals that reflect emotional changes.

[0050] For example, time-domain features such as mean, variance, and peak value can be extracted from preprocessed EEG signals, or frequency-domain features such as power spectral density of each frequency band can be extracted, or implicit feature representations can be learned through models such as deep autoencoders, as physiological response features reflecting emotional changes.

[0051] Step 126: Map the temporal feature representation or implicit feature representation to the preset emotion category to establish a typical physiological response pattern for each preset emotion category.

[0052] For example, the preset emotion categories include common emotion types such as joy, sadness, anger, and calmness. The extracted temporal feature representations or implicit feature representations are associated with the corresponding emotion categories. A typical EEG response pattern library corresponding to each emotion category is established through machine learning methods, which is used to compare the physiological response signals with the emotion dimension in subsequent consistency verification.

[0053] As described above, by acquiring EEG signals from both prefrontal cortex regions, neurophysiological activity data closely related to emotion processing were obtained. Through temporal preprocessing and feature extraction, the raw EEG signals were transformed into quantifiable emotional response features. By mapping these features to preset emotion categories, typical physiological response patterns for different emotion categories were established. This mechanism provides an objective third-party verification basis for subsequent multi-source consistency verification, enabling physiological response signals to independently verify the authenticity of users' subjective self-assessments and effectively solving the problem of lack of interpretability in automatic labeling of single signals.

[0054] In one alternative implementation, step 130 includes steps 131 to 136.

[0055] Step 131: For each emotion-inducing content, obtain the corresponding baseline emotion coordinates, subjective self-evaluation results, and physiological response signals.

[0056] For example, for the currently processed emotionally induced content segment, the system reads from the storage module the expert pre-assessment baseline emotional coordinates corresponding to the segment, the subjective self-assessment results submitted by the user after the segment ends, and the physiological response signals collected synchronously by the user during the experience of the segment.

[0057] Step 132: Compare the baseline emotional coordinates with the subjective self-evaluation results at the first level to obtain the first result.

[0058] For example, the baseline emotional coordinates scores on the three dimensions of valence, arousal, and sense of control are compared with the user's subjective self-evaluation scores on the corresponding dimensions. The deviation values ​​of each dimension are calculated, and the deviations of which dimensions exceed the preset threshold are recorded as the first result.

[0059] Step 133: The subjective self-evaluation results are compared with the physiological response signals to obtain the second result.

[0060] For example, the emotional state reflected by the user's subjective self-assessment is compared with the emotional category corresponding to the typical physiological response pattern identified through physiological response signals. The results determine whether the two match and record any deviations, serving as a second result.

[0061] Step 134: Based on the first and second results, determine whether there is a significant inconsistency between the baseline emotional coordinates, subjective self-evaluation results, and physiological response signals.

[0062] For example, if the first result shows that the user's subjective self-evaluation results deviate significantly from the baseline emotional coordinates in any dimension, and the second result shows that the subjective self-evaluation results deviate from the emotional dimension reflected by the physiological response signals, then it is determined that there is a significant inconsistency among the three.

[0063] Step 135: If there is a significant inconsistency among the three, the consistency check is deemed to have failed.

[0064] For example, if a significant inconsistency is found among the three, the sample is marked as a low-confidence sample, the consistency check result is failed, and the sample will not be included in the final emotion label training sample set.

[0065] Step 136, otherwise, the consistency check is deemed successful.

[0066] For example, if neither the first nor the second result triggers a condition of significant inconsistency, then it is determined that there is good consistency among the three, the verification result is passed, and the sample can proceed to the subsequent multidimensional emotion labeling stage.

[0067] As described above, a two-level comparison mechanism is used to verify three types of information sources layer by layer: the first-level comparison verifies the consistency between subjective self-assessment and expert benchmarks, and the second-level comparison verifies the consistency between subjective self-assessment and physiological response signals. Only when neither level of comparison shows significant anomalies is the sample deemed to have passed the verification. This mechanism achieves cross-validation of multi-source information, effectively identifying and eliminating anomalous samples such as those with arbitrary subjective opinions, lack of focus, or mismatches between self-assessment and physiological responses, thereby improving the data purity and label reliability of the training sample set.

[0068] In one alternative implementation, step 134 determines whether there is a significant inconsistency between the baseline emotional coordinates, subjective self-evaluation results, and physiological response signals, including steps 1341 and 1342.

[0069] Step 1341: When the first result is that the deviation between the subjective self-evaluation result and the benchmark emotional coordinate in any dimension exceeds a preset threshold, and the second result is that the subjective self-evaluation result deviates from the emotional dimension reflected by the physiological response signal, a significant inconsistency is determined.

[0070] For example, if a user's subjective rating on the emotional valence dimension differs from the expert benchmark rating by more than a preset threshold (e.g., 2 points), and the emotion category identified by the physiological response signal (e.g., sadness) is inconsistent with the user's subjective self-rated emotional state (e.g., pleasure), then it is determined that there is a significant inconsistency among the three, indicating that the user's emotional experience does not match the objective benchmark and physiological response, and the sample may be affected by factors such as subjective randomness or lack of concentration.

[0071] Step 1342, otherwise, determine that there is no significant inconsistency.

[0072] For example, if the first result shows that the deviations of all dimensions are within the preset threshold range, or the second result shows that the subjective self-evaluation results are basically matched with the emotional dimensions reflected by the physiological response signals, then it is determined that there is no significant inconsistency among the three, indicating that the sample has passed the multi-source consistency check and has high credibility.

[0073] As described above, significant inconsistency is determined by a combination of two conditions: first, a significant deviation between subjective self-assessment and expert benchmarks; and second, a dimensional deviation between subjective self-assessment and physiological response signals. Both conditions are indispensable; only when both subjective and objective information are abnormal is the sample deemed unreliable. This judgment strategy avoids overly stringent screening that could lead to the loss of valid samples, while ensuring accurate identification of truly problematic anomalous samples, achieving a reasonable balance between sample quality and quantity.

[0074] In one alternative implementation, step 140 labels the subjective self-evaluation results and physiological response signals that have passed the consistency check with multidimensional emotion tags, including steps 141 to 143.

[0075] Step 141: Based on the consistency results between the baseline emotional coordinates and physiological response signals, determine the objective emotion type label by referring to the baseline emotional coordinates.

[0076] For example, based on the expert pre-assessment results in the baseline emotional coordinates, and combined with the emotional categories corresponding to the typical physiological response patterns identified by the physiological response signals, the emotional type that matches both is taken as the objective emotional type label, such as discrete emotional categories like pleasure, sadness, anger, or calmness.

[0077] Step 142: For the subjective self-evaluation results that pass the consistency check, determine the subjective continuous dimension labels on the emotional valence dimension, emotional arousal dimension, and sense of control dimension, respectively.

[0078] For example, the subjective self-evaluation results submitted by users on the three dimensions of valence, arousal, and sense of control can be directly used as subjective continuous dimension labels, preserving the individualized emotional expression information of users under a unified benchmark framework without additional correction.

[0079] Step 143: Combine the objective emotion type labels with the subjective continuous dimension labels to generate multidimensional emotion labels.

[0080] For example, discrete objective emotion type labels can be associated and stored with continuous three-dimensional subjective ratings to form a complete sample label record that includes both interpretable emotion categories and detailed dimensional ratings, thus constituting a multi-dimensional emotion label.

[0081] As described above, the multidimensional emotion tags generated in this implementation include both objective emotion type tags and subjective continuous dimension tags: objective emotion type tags are determined based on the consistency between expert benchmarks and physiological responses, possessing stability and interpretability; subjective continuous dimension tags directly adopt users' self-assessment results, preserving individual differences. These two types of tags complement each other and are organically combined, allowing for use in training both emotion classification and emotion regression models, significantly improving the information density and applicability of the training samples. In an alternative implementation, after step 140 adds the training sample set, step 150 is also included.

[0082] Step 150: Use the physiological response signals in the training sample set as model input and the corresponding multidimensional emotion labels in the training sample set as target reference to train the emotion recognition model or emotion intervention strategy model.

[0083] For example, the bifrontal EEG physiological response signals collected from the training sample set can be used as input features, and the objective emotion type labels or subjective continuous dimension labels in the corresponding multidimensional emotion labels can be used as supervision signals. The emotion recognition model can be trained using supervised learning. Alternatively, multidimensional emotion labels can be used as intervention targets to train a personalized emotion intervention strategy model, so that the model learns the mapping relationship from physiological signals to multidimensional emotion labels.

[0084] As described above, applying the high-quality training sample set generated through multi-source consistency verification to model training—using physiological response signals as input and multi-dimensional emotion labels as target references—achieves end-to-end model training. Because the training samples have undergone multi-source cross-validation, they possess high reliability and low noise characteristics, significantly improving the model's convergence speed, recognition accuracy, and cross-scenario generalization ability, providing a solid model foundation for emotion recognition and emotion intervention applications.

[0085] The multi-source verification emotion labeling method and terminal described above are applicable to various scenarios that require the construction of high-quality emotion training samples. They are particularly advantageous in training emotion recognition models based on EEG signals, developing personalized emotion intervention systems, and conducting cross-user emotion calculation research. The following is a detailed description of the implementation methods.

[0086] The following details the application examples of this application. This application can apply the above solution to the data construction scenario of a brain-computer interface emotion recognition system, especially in the fields of intelligent psychological counseling and affective computing, where large-scale, high-quality emotion-labeled samples are needed to improve model performance. Taking an online emotional health monitoring platform developed by an artificial intelligence company as an example, it needs to build a high-quality EEG sample library covering multiple emotional states to train the user's emotion recognition model. The specific implementation may include the following steps: S1. In the emotional content preparation stage, the system selects music clips, short video clips, and digital image content from a standardized multimedia material library, covering various emotional categories such as joy, sadness, anger, and calmness. Three psychology experts independently score each content clip on the dimensions of emotional valence, emotional arousal, and sense of control. The average score of the three experts is taken as the baseline emotional coordinate for that content clip. This corresponds to steps 110 and 111 to 113 above.

[0087] S2. During the data acquisition phase, the system recruited 100 target users to participate in the experiment. Each user wore a bifrontal EEG acquisition device and sequentially experienced the prepared emotion-inducing content. After each content segment, the system displayed a scoring interface, receiving the user's subjective ratings on three dimensions: emotional valence, arousal, and sense of control, thus obtaining the subjective self-assessment result. Simultaneously, during the user experience, the system synchronously acquired EEG signals from the user's left and right prefrontal cortex regions as physiological response signals. The system recorded the offset of the user's subjective self-assessment result from the baseline emotional coordinates in each dimension, as individualized emotional offset information. This is equivalent to steps 120 and 121 to 126 above.

[0088] S3. In the multi-source consistency verification stage, the system acquires the corresponding baseline emotional coordinates, subjective self-evaluation results, and physiological response signals for each emotion-induced content segment. The baseline emotional coordinates are compared with the subjective self-evaluation results at the first level to determine if the deviation in each dimension exceeds a preset threshold (e.g., 2 points). The subjective self-evaluation results are compared with the physiological response signals at the second level to determine if the emotional dimensions reflected by the two match. If the first-level comparison shows that any dimension deviation exceeds the threshold, and the second-level comparison shows that the subjective self-evaluation and physiological response signal dimensions deviate, then a significant inconsistency is determined between the three, and the sample is marked as low confidence and removed; otherwise, the verification is deemed successful, and the sample is retained. This is equivalent to steps 130 and 131 to 136 above.

[0089] S4. In the multidimensional emotion labeling stage, the system determines objective emotion type labels (such as joy, sadness, anger, calm, etc.) for samples that have passed the consistency check based on the consistency results of the baseline emotion coordinates and physiological response signals. At the same time, the user's subjective self-rating results are used as subjective continuous dimension labels (continuous ratings of valence, arousal, and sense of control). The two are combined to generate multidimensional emotion labels, which are then added to the training sample set. This is equivalent to steps 140 and steps 141 to 143 above.

[0090] S5. During the model training phase, the system uses the physiological response signals from the training sample set as model input, the objective emotion type label from the corresponding multi-dimensional emotion labels as the classification target, and the subjective continuous dimension label as the regression target to perform multi-task training on the emotion recognition model, enabling the model to simultaneously output emotion category and emotion dimension scores. This is equivalent to step 150 above.

[0091] Through the above application examples, this invention improves the efficiency of emotion tag construction, reduces sample noise, enhances the accuracy of emotion recognition models, and strengthens cross-user generalization ability. It effectively solves typical problems such as the instability of single subjective self-assessment tags, the high cost and lack of personalization of pure expert annotation, and the lack of interpretability of single physiological signal annotation. This invention has significant engineering value in improving the performance of emotion computing systems, optimizing data governance processes, and reducing the cost of manual annotation.

[0092] Please refer to Figure 2 The present invention also provides a multi-source verification sentiment annotation terminal 200, including a memory 201, a processor 202, and a computer program stored in the memory 201 and executable on the processor 202. When the processor 202 executes the computer program, it implements the various steps of the multi-source verification sentiment annotation method described above.

[0093] In summary, the multi-source validated emotion labeling method and terminal provided by this invention generates baseline emotion coordinates by acquiring standardized emotion-inducing content and expert pre-assessment results, providing a stable and unified objective reference benchmark for emotion labels and solving the label drift problem of a single subjective self-assessment method. By simultaneously collecting the subjective self-assessment results of target users and physiological response signals from both prefrontal cortex regions, individual differences are quantified under a unified benchmark, compensating for the deficiency of pure expert labeling in ignoring the user's real experience. By performing multi-source consistency validation on the baseline emotion coordinates, subjective self-assessment results, and physiological response signals, a cross-validation screening mechanism is constructed, effectively eliminating abnormal samples such as subjective arbitrariness, lack of concentration, and mismatch between self-assessment and physiological response, thereby improving the data purity of the training sample set. By labeling the validated data with multi-dimensional emotion labels including objective emotion type labels and subjective continuous dimension labels, the training needs of both emotion classification models and emotion regression models are simultaneously supported. By elevating the construction of emotion labels from single-point subjective judgment to multi-source collaborative confirmation, a substantial leap has been achieved in label quality from uncontrollable to verifiable, screenable, and guaranteed. This provides a highly reliable and low-noise training data foundation for emotion recognition and intervention models, effectively improving the model's convergence speed, recognition accuracy, and cross-scenario generalization ability.

[0094] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-source validation sentiment annotation method, characterized in that, include: Obtain the emotional triggering content and the corresponding expert pre-assessment results, and generate a baseline emotional coordinate based on the expert pre-assessment results; Collect the subjective self-evaluation results and physiological response signals of target users during the process of experiencing the emotionally induced content; The consistency of the baseline emotional coordinates, the subjective self-evaluation results, and the physiological response signals is verified. The subjective self-evaluation results and physiological response signals that pass the consistency check are labeled with multidimensional emotion tags and added to the training sample set.

2. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, The process of acquiring the emotion-inducing content and the corresponding expert pre-assessment results, and generating a baseline emotion coordinate based on the expert pre-assessment results, includes: At least one of standardized music, video, and digital content is used as the said emotion-inducing content; The system receives scores from experts in the field of emotion research on each of the aforementioned emotion-inducing content on a preset emotion dimension, and obtains the expert pre-scoring results. The expert pre-scoring results are used as the baseline emotional coordinates for the emotionally induced content.

3. The sentiment annotation method based on multi-source verification according to claim 2, characterized in that, The experts in the field of emotion research score each of the emotion-inducing contents on a preset emotion dimension, resulting in expert pre-scoring results, including: Experts in the field of emotion research score each emotion-inducing content on the dimensions of emotional valence, emotional arousal, and sense of control, and the expert pre-scoring results are obtained.

4. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, The subjective self-evaluation results of the target users during their experience with the emotionally induced content include: After each emotionally induced content session, the target user's subjective ratings on the emotional valence dimension, emotional arousal dimension, and sense of control dimension are received to obtain subjective self-evaluation results. Record the offset between the subjective self-evaluation result and the baseline emotional coordinate as individualized emotional offset information; The consistency verification between the subjective self-assessment results and the baseline emotion coordinates includes: If the individualized emotion offset information is within a preset threshold range, it indicates that the subjective self-evaluation result is consistent with the baseline emotion coordinates; If the individualized emotion offset information exceeds the preset threshold range, it indicates that the subjective self-evaluation result is inconsistent with the baseline emotion coordinates.

5. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, The collection of physiological response signals from target users during their experience of the emotionally induced content includes: Simultaneously collect physiological response signals from the left and right prefrontal regions of the target user; The collected physiological response signals are subjected to time-series preprocessing; Extract temporal or implicit feature representations reflecting emotional changes from the physiological response signals; The temporal feature representation or the implicit feature representation is mapped to a preset emotion category to establish a typical physiological response pattern corresponding to each preset emotion category.

6. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, The process of verifying the consistency of the baseline emotional coordinates, the subjective self-evaluation results, and the physiological response signals includes: For each of the aforementioned emotion-inducing contents, the corresponding baseline emotion coordinates, subjective self-evaluation results, and physiological response signals are obtained respectively; The first result is obtained by comparing the baseline emotion coordinates with the subjective self-evaluation results at the first level. The subjective self-assessment results are compared with the physiological response signals at a second level to obtain a second result; Based on the first and second results, determine whether there is a significant inconsistency between the baseline emotional coordinates, the subjective self-evaluation results, and the physiological response signals; If there is a significant inconsistency among the three, the consistency check is deemed to have failed. Otherwise, the consistency check is deemed to have passed.

7. The sentiment annotation method based on multi-source verification according to claim 6, characterized in that, The determination of whether there is a significant inconsistency between the baseline emotional coordinates, the subjective self-evaluation results, and the physiological response signals includes: When the first result is that the deviation between the subjective self-evaluation result and the benchmark emotional coordinate in any dimension exceeds a preset threshold, and the second result is that the subjective self-evaluation result deviates from the emotional dimension reflected by the physiological response signal, it is determined that there is a significant inconsistency; Otherwise, it is determined that there is no significant inconsistency.

8. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, The process of labeling the subjective self-assessment results and physiological response signals that have passed the consistency check with multidimensional emotion tags includes: Based on the consistency between the baseline emotional coordinates and the physiological response signal, the objective emotional type label is determined with reference to the baseline emotional coordinates; For the subjective self-evaluation results that pass the consistency check, subjective continuous dimension labels are determined on the emotional valence dimension, emotional arousal dimension, and sense of control dimension, respectively. The objective emotion type label is combined with the subjective continuous dimension label to generate the multidimensional emotion label.

9. The sentiment annotation method based on multi-source verification according to claim 1, characterized in that, After adding the training sample set, the following is also included: The physiological response signals in the training sample set are used as model inputs, and the multidimensional emotion labels corresponding to the training sample set are used as target references to train the emotion recognition model or emotion intervention strategy model.

10. A multi-source verification sentiment annotation terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the multi-source verification sentiment labeling method according to any one of claims 1 to 9.

Images