Multi-modal real-time sentiment analysis method and system for 5G new call

Abstract

The invention discloses a multi-modal real-time sentiment analysis method and system for a 5G new call, and relates to the technical field. The method comprises the following steps: constructing a training set containing multi-modal features such as intonations, expressions and vocabularies based on 5G features, and dividing data subsets by adopting a stratified sampling technology; calculating a time period-emotion association value and optimizing data distribution through timestamp coding and emotion label mapping; during real-time analysis, cross-modal context features are dynamically extracted, and a prediction model is updated in combination with a semantic emotion association rule; establishing a double-sample library mechanism, storing the features which are verified to be consistent into a reinforcement learning library, marking the inconsistent features as misleading / noise features, and automatically removing the misleading / noise features; and performing triple label labeling on the misjudgment sample, and through misjudgment density screening, conflict degree calculation and clustering analysis, positioning a high-risk error mode and generating a correction rule. According to the embodiment, the sentiment analysis real-time performance and accuracy are improved by utilizing the 5G low-delay characteristic, and key technical support is provided for scenes such as intelligent customer service and emergency response.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field, and particularly relates to a multi-modal real-time emotion analysis method and system for 5G new calls. BACKGROUND

[0002] With the popularization of 5G communication technology, 5G new calls (such as VoNR+) support real-time transmission of multi-modal data such as voice, video and text, and provide a new carrier for emotional interaction scenarios, thanks to high bandwidth and low latency.

[0003] However, the traditional emotion analysis technology has the following limitations, which cannot meet the real-time, multi-modal fusion and dynamic adaptability requirements in the 5G new call scenario:

[0004] Insufficient real-time and scenario adaptability: the 5G new call scenario (such as business negotiation, emergency call, daily casual conversation) has very high real-time requirements, while the traditional model needs offline training and cannot dynamically adapt to scenario changes. For example, the negative emotion of a user in the "workday evening + business call" scenario may be caused by overtime pressure, but if the model does not distinguish between time periods and scenarios, it is easy to misjudge as neutral emotion.

[0005] Data division and error processing are rough: traditional methods usually divide the data set by random sampling, resulting in unbalanced subset distribution (such as too few negative emotion samples), and do not consider the joint distribution of time-scenario-emotion. In addition, misjudgment samples are simply excluded or manually reviewed, without in-depth analysis of misjudgment patterns (such as the contradictory features of "flat tone but text containing aggressive words"), which limits the iteration efficiency of the model.

[0006] Interference of noise features and conflicting modalities: multi-modal data often contain noise (such as background noise interfering with voice analysis) or conflicts between modalities (such as video showing a smile but text containing negative words). Traditional methods lack a mechanism to quantify the degree of conflict, resulting in unstable results from the model under contradictory features.

[0007] To solve the above problems, the application provides a multi-modal real-time emotion analysis method and system for 5G new calls. SUMMARY

[0008] In order to overcome the shortcomings and deficiencies of the prior art, the first object of the application is to provide a multi-modal real-time emotion analysis method for 5G new calls; and the second object of the application is to provide a multi-modal real-time emotion analysis system for 5G new calls.

[0009] The first object of the application adopts the following technical solutions:

[0010] The multi-modal real-time emotion analysis method for 5G new calls has the following process:

[0011] Based on the characteristics of the 5G network, multi-modal data such as voice, video and text in the call are collected, and a historical emotional expression training set is constructed.

[0012] The training set is divided into a test set, a validation set and a training subset by stratified sampling technology, and a timestamp and a corresponding user emotional state label are labeled for each sample in the training set. The emotional state label includes positive, negative and neutral.

[0013] The correlation value between the sample acquisition time and the user emotional state is analyzed to establish the mapping relationship between the time dimension and the emotional pattern.

[0014] Multi-modal features are extracted and matched with emotional types to generate an emotional analysis model.

[0015] The matching degree of the model output result and the actual emotional state of the user is verified, and effective features and misleading features are classified and stored.

[0016] According to the scene category, the multi-modal sample is marked, the frequency is counted and sorted.

[0017] Real-time call data is collected, modalities are divided, and context features and semantic emotional correlation rules are analyzed.

[0018] The emotional analysis result is optimized by secondary division of error feature data, and the random gradient descent method is used to iteratively train the emotional prediction model.

[0019] Preferably, the stratified sampling technology is as follows:

[0020] The test set, the validation set and the training subset are divided according to the proportion of the emotional state category, and the data division tool is used to generate independent subsets according to the preset proportion. The stratification dimensions include time, scene and emotional category, and the emotional category distribution in each subset is consistent with the original training set.

[0021] Preferably, the correlation value between the sample acquisition time and the user emotional state is generated by the following steps:

[0022] The timestamp is encoded into four categories: weekday daytime, weekday evening, weekend and special period;

[0023] The global emotional proportion is calculated And the time period emotional proportion The correlation strength between the time period and the emotion is quantified by the standardized difference

[0024] Preferably, the matching degree of the model output result and the actual emotional state of the user includes:

[0025] If the analysis result is consistent with the user feedback, the key multi-modal feature is extracted and marked as an effective feature and stored in the reinforcement learning sample library; ​

[0026] If not, the feature is marked as a misleading feature or a noise feature and stored in an error analysis sample library; the error feature library is automatically removed and disabled in subsequent analysis.

[0027] Preferably, the optimization of the sentiment analysis result by secondary division of the error feature data comprises:

[0028] The misjudgment sample is labeled with a ternary group label (T, S, E), the misjudgment frequency C (T, S, E) is counted, and the misjudgment density D (T, S, E) is calculated.

[0029] The high-density combination is screened as a secondary division unit, and the K-means clustering is used to find the subdivided error mode.

[0030] The representative rule R is extracted, and the misjudgment probability is calculated Wherein, R is a sample set satisfying the rule R, and the weight is dynamically adjusted in real-time analysis.

[0031] Preferably, the optimization of the sentiment analysis result by secondary division of the error feature data further comprises:

[0032] The conflict degree of the multi-modal feature is calculated According to the conflict degree threshold, the sample is divided into high conflict misjudgment and low conflict misjudgment, and the high correlation feature is extracted

[0033]

[0034] Preferably, the division mode comprises:

[0035] The voice, video and text data are divided into independent segments, and the cross-modal context feature is analyzed.

[0036] The transmission index is recorded and the sustainability verification value is calculated.

[0037] According to the verification value and the predicted feasibility, the transmission path is dynamically adjusted to ensure the real-time performance of the sentiment analysis.

[0038] Preferably, the training step of the sentiment analysis model comprises:

[0039] The random gradient descent method is used to optimize the model parameters, the correlation feature between the user emotion and the sample is collected, the data feature vector, the node adaptation degree, the path matching degree and the model output are stored by the blockchain, and the transparent and traceable full life cycle sentiment analysis model is constructed based on the stored data.

[0040] The second object of the application adopts the following technical scheme:

[0041] The multi-modal real-time sentiment analysis system for 5G new calls is used to realize the multi-modal real-time sentiment analysis method for 5G new calls, and the system comprises

[0042] Big data information acquisition module: for acquiring the historical emotional expression data of the preset user in the 5G new call, and generating a training set corresponding to the voice, video and text multi-modal data, and then generating an emotional analysis and recognition model;

[0043] Comprehensive data processing module: the acquired multi-modal samples are distinguished and marked, and the distinguished data information is input into the recognition model, and the recognition model is updated; the training set is divided into a test set and a validation set, and the acquisition time of each sample in the training set and the emotional state of the user are marked;

[0044] Fusion analysis module: for acquiring real-time multi-modal data of the user in the 5G new call, dividing the data and analyzing the semantic emotional features and user emotional data information piece by piece, which includes a data statistics submodule, an analysis submodule, a verification submodule and a correction submodule: the data statistics submodule: according to the 5G new call multi-modal data of the user in different time periods, the historical emotional state of the user in each time period is counted;

[0045] Analysis submodule: analyze the features between different emotional states of the user and multi-modal data samples, classify the emotional states, and match the collected multi-modal features according to the emotional types;

[0046] Verification submodule: generate emotional analysis results according to the emotional types matched by the multi-modal features, verify the results with the actual emotional state of the user, if correct, store and upload, otherwise regenerate until correct;

[0047] Correction submodule: analyze the data features between correct and incorrect emotional analysis results, and automatically exclude the data features contained in the incorrect results in future analysis;

[0048] Data discrimination module: in response to the multi-modal data divided piece by piece by the fusion analysis module, according to the user emotional data information and the corresponding data, the emotional analysis and recognition results of the data are judged and processed to verify the emotional analysis and recognition results of the data.

[0049] As described above, due to the adoption of the above technical solutions, the beneficial effects of the present application are:

[0050] 1、The present application constructs a full-dimensional emotional feature library by fusing multi-modal data such as voice (intonation, speech rate), video (facial expression, body movement), and text (vocabulary, sentence pattern), overcoming the one-sidedness of traditional single-modal analysis. For example, when the user shows contradictory features of "flat intonation but text containing aggressive vocabulary", the model can comprehensively judge it as a negative emotion, avoiding single-modal misjudgment. At the same time, the hierarchical sampling technology is adopted to divide the data set according to time-scene-emotion categories, ensuring the balanced distribution of each subset, so as to improve the recognition accuracy of the model in specific scenes such as "weekday evening + business call".

[0051] 2、By encoding timestamps into 4 categories of time periods (e.g. "special time period" covering holidays, emergencies) and combining with scene labels (daily / conversational / emergency call), the model can dynamically adapt to the complex scenarios of 5G new calls. In addition, noise features (such as background noise interfering with speech analysis) and high conflict misjudgment samples (such as speech-video prediction contradiction) are automatically removed in real-time analysis, reducing the delay of single-call sentiment analysis and meeting real-time interaction needs.

[0052] 3、Introducing triad labels (time period T, scene S, predicted emotion E) and misjudgment density calculation (e.g. "weekday evening + business call + negative emotion" combination misjudgment density) for misjudgment samples to accurately locate high-risk error patterns. By quantifying modal conflict weights and mutual information to filter high-correlation features, samples are divided into feature-driven misjudgment and noise-driven misjudgment. Further, K-means clustering is used to extract sub-rules, shortening the model iteration period. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0054] Figure 1 The flowchart of the multi-modal real-time emotion analysis method for 5G new calls of the present application is shown;

[0055] Figure 2 The module diagram of the multi-modal real-time emotion analysis system for 5G new calls of the present application is shown. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0058] Example 1:

[0059] See Figure 1 As shown in the figure, the multimodal real-time sentiment analysis method for new 5G calls in this embodiment has the following process:

[0060] Based on big data, historical emotional expression data of preset users in 5G new calls is obtained, and corresponding multimodal data of voice, video, and text are used to generate a training set. Multimodal information of each sample is extracted from the training set, including tone and speed of speech, facial expressions and body movements in video, and vocabulary and sentence structure in text, to generate a sentiment analysis and recognition model, and the sample text is input into the model.

[0061] In the recognition model, stratified sampling is employed, dividing the training set into a test set, a validation set, and a retained training subset based on the distribution ratio of emotional state categories. A data partitioning tool is used to segment the training set data according to a preset ratio, generating three independent subsets.

[0062] The acquisition time of each sample in the training set is labeled with the user's emotional state: a timestamp T is labeled for each sample. i =[T start ,T end ], and tagged with the emotion tag L i ∈{positive, negative, neutral}.

[0063] Stratified sampling: The dataset is divided into stratified subsets based on time, scene, and sentiment category to ensure a balanced distribution of each subset.

[0064] The correlation between sample acquisition time and user emotional state is analyzed, and correlation values ​​between sample acquisition time and user emotional state are generated. The timestamps are then converted into four types of time period codes.

[0065] T1: Weekday daytime (09:00-18:00)

[0066] T2: Weekday evenings (6:00 PM - 12:00 AM)

[0067] T3: Weekends (00:00-24:00)

[0068] T4: Special period (holidays, emergencies, etc.)

[0069] Emotion label coding: map the emotion state to 3-class numerical labels:

[0070] E = 1: Positive emotion

[0071] E = 0: Neutral emotion

[0072] E = -1: Negative emotion

[0073] For example: Obtain the 5G new call multi-modal data of users in different periods such as daytime, evening, weekend, etc. on weekdays, and count the historical emotion states of users in each period, including happy, sad, angry, calm, etc. emotion categories. Collect the features between different emotion states and multi-modal data samples, such as high pitch and loud volume features of voice under angry emotion, frown and squinting expression features of face, and intense vocabulary features in text; divide the emotion state into three categories: positive emotion, negative emotion and neutral emotion.

[0074] Calculate the global emotion proportion:

[0075]

[0076] Calculate the emotion proportion of period T i

[0077]

[0078] Calculate the standardized difference:

[0079]

[0080] Where N is the total number of samples; the correlation value SPD i The greater, the period T i The more significant the emotion correlation.

[0081] According to the emotion type matched by the multi-modal features, generate the emotion analysis result, and verify it with the actual emotion state of the user. If the analysis result is consistent with the user feedback, extract the key multi-modal features of the data unit, mark it as valid features, and store it in the reinforcement learning sample library. If the analysis result is inconsistent with the user feedback, extract the key multi-modal features of the data unit, mark it as misleading features or noise features, and store it in the error analysis sample library.

[0082] Analyze the data features between correct and incorrect emotion analysis results, and automatically exclude the data features contained in the verification error results in the analysis process when performing 5G new call emotion analysis in the future.

[0083] ​The obtained multi-modal samples are distinguished and labeled, including being distinguished into scene categories such as daily communication modal data, business call modal data and emergency call modal data; meanwhile, the use frequency of historical multi-modal data of the user in the distinguished sample data is counted, and the distinguished sample data is sorted according to the use frequency.

[0084] Real-time multi-modal data of the user in the 5G new call is obtained, the data is divided and analyzed piece by piece. The voice data, video data and text data are divided into [θ v1 ,θ v2 ,θ v3 ,...,θ vn ], [θ f1 ,θ f2 ,θ f3 ,...,θ fn ], [θ t1 ,θ t2 ,θ t3 ,...,θ tn ] respectively, the context features between different modal data and between different pieces of data of the same modal are analyzed, such as the association between the change of voice tone and the emotional tendency of text vocabulary, and the current divided data is recorded and stored.

[0085] According to the divided multi-modal data, the emotional data information of the user is collected, and the emotional data information obtained is processed according to the corresponding data. The semantic emotional features of each piece of multi-modal data are analyzed, and the semantic emotional features are divided into positive semantic features, negative semantic features and neutral semantic features, the association rules between the emotional features are obtained, and the association rules are input into the recognition model, wherein the semantic emotional features include the emotional semantic features conveyed by the video picture.

[0086] According to the emotional data information of the user obtained and the corresponding data, the judgment processing is carried out to verify the emotional analysis and recognition result of the data. When the emotional analysis and recognition result of the corresponding data is consistent with the emotional data information of the user obtained, the data features between the corresponding data and the emotional data information are further obtained, and the data features are summarized as correct recognition data features; otherwise, when they are inconsistent, the data features when they are inconsistent are obtained, and they are summarized as incorrect recognition data features.

[0087] In the incorrect recognition data features, the corresponding data is further divided twice, and new emotional semantic analysis is carried out according to the twice divided data, and is matched with the user emotional data information, the emotional analysis result of the user is reacquired, specifically,

[0088] The misjudgment sample is labeled with a triple tag: (T, S, E), wherein:

[0089] T E {T1, T2, T3, T4};

[0090] S E {daily communication, business call, emergency call};

[0091] E E {positive, neutral, negative}.

[0092] Count the misjudgment frequency C(T, S, E) of each triple combination, and calculate the misjudgment density: Screen combinations with misjudgment density greater than a set proportion, and mark them as high-density combinations as secondary division units.

[0093] Measure the degree of contradiction of different modal characteristics to emotional judgment by the formula Where m E {speech, video, text}; f m (x) is the emotional prediction result of modal m; ω m,m ' is the inter-modal conflict weight (e.g. speech-video conflict weight is higher); I(·) is the indicator function division standard: divide the samples into high-conflict misjudgment and low-conflict misjudgment according to the Conflict(x) threshold.

[0094] For misjudgment sample x, extract its feature vector v(x) = [v1, v2, v3,..., vn]. Calculate the mutual information of the feature and the true label E true Sort by MI(vi, E true ) in descending order, screen features with MI(vi, E true ) value greater than a set threshold, and mark them as high-correlation features. Divide the samples into feature-driven misjudgment and noise-driven misjudgment according to whether they contain high-correlation features.

[0095] For samples in high-density triple (T, S, E), use K-means clustering (k = 3-5) based on feature vector v(x) to further group. Goal: find the sub-error patterns under the same time-scene-emotion combination (e.g. in "workday evening + business call + negative emotion", samples with flat tone but text containing aggressive words).

[0096] For each clustering result, extract representative rules (e.g. "if T = T2 and S = business call and Conflict(x) >= 2, then the misjudgment probability increases by 40%"), and the misjudgment probability formula is: Where R is the sample set that satisfies the rule R.

[0097] In the training set, use the stochastic gradient descent method to train the samples therein, and collect the correlation features between user emotions and samples, to generate an emotional prediction model based on correlation features.

[0098] ​The beneficial effects of the embodiment are: the system utilizes the characteristics of low latency and large bandwidth of 5G network, collects and fuses multi-modal data of voice, video and text in real time, analyzes the emotional state (positive, negative and neutral) of the user through advanced deep learning model, and provides key decision support for improving call service quality, intelligent customer service, emotional care, emergency response and other applications.

[0099] Referring to Figure 2 The multi-modal real-time emotional analysis system for 5G new call of the embodiment comprises

[0100] Big data information acquisition module: used for acquiring historical emotional expression data of a preset user in 5G new call, generating a training set of corresponding multi-modal data of voice, video and text, and further generating an emotional analysis and recognition model.

[0101] Comprehensive data processing module: the acquired multi-modal samples are distinguished and marked, and the distinguished data information is input into the recognition model, and the recognition model is updated; the training set is divided into a test set and a validation set, and the acquisition time of each sample in the training set and the emotional state of the user are marked.

[0102] Fusion analysis module: used for acquiring real-time multi-modal data of the user in 5G new call, dividing the data and analyzing the semantic emotional features and user emotional data information of each piece of data, which comprises a data statistics submodule, an analysis submodule, a verification submodule and a correction submodule:

[0103] Data statistics submodule: according to the multi-modal data of 5G new call of the user in different time periods, the historical emotional state of the user in each time period is counted.

[0104] Analysis submodule: analyze the features between different emotional states of the user and multi-modal data samples, classify the emotional states, and match the collected multi-modal features according to the emotional types.

[0105] Verification submodule: generate emotional analysis results according to the emotional types matched by the multi-modal features, verify the results with the actual emotional state of the user, if correct, store and upload, otherwise regenerate until correct.

[0106] Correction submodule: analyze the data features between correct and incorrect emotional analysis results, and automatically exclude the data features contained in the incorrect results in future analysis.

[0107] Data discrimination module: in response to the multi-modal data divided piece by piece by the fusion analysis module, the acquired user emotional data information and the corresponding data are judged and processed to verify the emotional analysis and recognition results of the data.

[0108] The beneficial effects of the embodiment are: through multi-modal fusion and dynamic scene adaptation, the accuracy of emotion recognition is improved, real-time error correction ensures smooth interaction; combined with conflict feature quantization and closed-loop iterative learning, the model is automatically optimized and noise interference is removed, the accuracy of emotion analysis is improved, the response delay is reduced, and the real-time demand of complex 5G call scene is adapted.

[0109] The weight of the application is used to measure the influence degree of different factors or variables on a certain result or decision. The definition of weight refers to the value assigned to each factor when comparing and evaluating multiple factors to reflect its importance or priority. These weights can be determined according to specific circumstances and needs, usually by professionals or relevant stakeholders together to develop and confirm. By reasonably setting the weight, the program or system can make more accurate decisions or predictions.

[0110] The above is only the preferred embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

[0111] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and limit the application to the specific embodiments. Obviously, according to the content of the specification, many modifications and changes can be made. The specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited by the claims and their entire scope and equivalents.

Claims

1. A multimodal real-time sentiment analysis method for 5G new call, characterized in that, The method flow is as follows: Based on the characteristics of 5G network, the voice, video, text multi-modal data in the call are collected, and a historical emotional expression training set is constructed; Through stratified sampling technology, the training set is divided into test set, validation set and training subset, and each sample in the training set is labeled to obtain timestamp and corresponding user emotional state label, the emotional state label includes positive, negative and neutral; The correlation value of sample acquisition time and user emotional state is analyzed, and the mapping relationship between time dimension and emotional mode is established; Multi-modal features are extracted and matched with emotional types to generate an emotional analysis model; The matching degree of the model output result and the actual emotional state of the user is verified, and effective features and misleading features are classified and stored; According to the scene category, the multi-modal samples are marked, the usage frequency is counted and sorted; Real-time call data is collected, modalities are divided, and context features and semantic emotional correlation rules are analyzed; The emotional analysis result is optimized by secondary division of error feature data, and the random gradient descent method is used to iteratively train the emotional prediction model.

2. The multi-modal real-time sentiment analysis method for 5G new call according to claim 1, characterized in that, The stratified sampling technology is as follows: The test set, validation set and training subset are divided according to the proportion of emotional state category, and the data division tool is used to generate independent subsets according to the preset proportion; the stratification dimensions include time, scene and emotional category, and the emotional category distribution in each subset is consistent with the original training set. 3.The multi-modal real-time sentiment analysis method for 5G new call orientation of claim 1, wherein, The correlation value of sample acquisition time and user emotional state is generated by the following steps: The timestamp is encoded into four categories: weekday daytime, weekday evening, weekend and special period; Computing global sentiment proportions and time period sentiment proportions Quantifying time period and sentiment association strength by standardizing differences Quantifying time period and sentiment association strength by standardizing differences 4. The multi-modal real-time sentiment analysis method for 5G new call according to claim 1, characterized in that, The matching degree of the model output result and the actual emotional state of the user includes: If the analysis result is consistent with the user feedback, the key multi-modal features are extracted and marked as effective features and stored in the reinforcement learning sample library; If it is inconsistent, the features are extracted and marked as misleading features or noise features and stored in the error analysis sample library; the features in the error feature library are automatically removed, and are disabled in subsequent analysis.

5. The multi-modal real-time sentiment analysis method for 5G new call oriented of claim 1, wherein, The optimization of emotional analysis result by secondary division of error feature data includes: The misjudgment sample is labeled with a three-tuple label (T, S, E), the misjudgment frequency C(T, S, E) is counted, and the misjudgment density D(T, S, E) is calculated; High-density combinations are selected as secondary division units, and K-means clustering is used to find sub-error patterns; extracting representative rules R, computing false positive probabilities where R is a set of samples satisfying the rule R, and the weights are dynamically adjusted in real-time analysis.

6. The multi-modal real-time sentiment analysis method for 5G new call oriented of claim 1, wherein, The optimization of emotional analysis result by secondary division of error feature data further includes: Computing multimodal feature conflict Separating samples into high conflict misjudgments and low conflict misjudgments by conflict threshold, and extracting high correlation features 7. The multi-modal real-time sentiment analysis method for 5G new call oriented of claim 1, wherein, The modalities are divided as follows: The voice, video and text data are divided into independent segments, and the cross-modal context features are analyzed; The transmission indicators are recorded and the sustainability verification value is calculated; According to the verification value and the prediction feasibility, the transmission path is dynamically adjusted to ensure the real-time performance of emotional analysis.

8. The multi-modal real-time sentiment analysis method for 5G new call oriented of claim 1, wherein, The training steps of the emotional analysis model include: The random gradient descent method is used to optimize the model parameters, and the correlation features between user emotions and samples are collected; the data feature vectors, node adaptation degree, path matching degree and model output are stored through blockchain; and a transparent and traceable full life cycle emotional analysis model is constructed based on the stored data.

9. The multi-modal real-time sentiment analysis system for 5G new call oriented, for implementing the multi-modal real-time sentiment analysis method for 5G new call oriented as claimed in claim 1, characterized in that, The system includes Big data information acquisition module: for acquiring the historical emotional expression data of the preset user in the 5G new call, and generating a training set of corresponding voice, video and text multi-modal data, and then generating an emotional analysis and recognition model; Comprehensive data processing module: the acquired multi-modal samples are distinguished and marked, and the distinguished data information is input into the recognition model, and the recognition model is updated; the training set is divided into a test set and a validation set, and the acquisition time of each sample in the training set and the emotional state of the user are marked; Fusion analysis module: for acquiring real-time multi-modal data of the user in the 5G new call, dividing the data and analyzing the semantic emotional features and user emotional data information piece by piece, which includes a data statistics submodule, an analysis submodule, a verification submodule and a correction submodule: data statistics submodule: according to the 5G new call multi-modal data of the user in different time periods, the historical emotional state of the user in each time period is counted; Analysis submodule: analyze the characteristics between different emotional states of the user and multi-modal data samples, classify the emotional states, and match the collected multi-modal features according to the emotional types; Verification submodule: generate emotional analysis results according to the emotional types matched by the multi-modal features, verify the results with the actual emotional state of the user, if correct, store and upload, otherwise regenerate until correct; Correction submodule: analyze the data features between correct and incorrect emotional analysis results, and automatically exclude the data features contained in the incorrect results in future analysis; Data discrimination module: in response to the multi-modal data divided piece by piece by the fusion analysis module, judge and process according to the acquired user emotional data information and corresponding data to verify the emotional analysis and recognition results of the data.

Images