Adaptive generation method of intelligent marketing speech based on multi-round interactive feedback

Abstract

The present invention discloses a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback, which relates to the field of intelligent dialogue technology. It includes: S1: fusing CBSS operation logs, voice emotion data and financial credit data to obtain a customer dynamic portrait matrix; S2: obtaining the initial dialogue state according to the customer dynamic portrait matrix, and obtaining the dialogue state of the current dialogue through a constructed dialogue state tracker; S3: determining the risk level of the current dialogue and correcting the dialogue state of the current dialogue; S4: obtaining a set of compliant speech candidates based on the corrected dialogue state of the current dialogue, and determining personalized marketing speech through the compliant speech candidate set and reinforcement learning strategy. The present invention solves the problems of data fragmentation, intention drift and rule rigidity in traditional marketing, and improves the effect and compliance of personalized recommendations.

Description

Technical Field

[0001] The present invention relates to the field of intelligent dialogue technology, and specifically to a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback. Background Art

[0002] Traditional sales scripts often rely on manual design, making them difficult to address the personalized needs of different customer groups. For example, the same script may not be suitable for communicating with both technically savvy and non-technical customers. Furthermore, fixed scripts cannot be dynamically adjusted based on real-time customer feedback (such as tone, questions, and objections), which limits closing rates and customer satisfaction.

[0003] Furthermore, traditional approaches rely on empirical experience and manual optimization, lacking the ability to deeply analyze historical conversation data. For example, historical interaction data cannot be used to identify which sales pitches are most effective, or which strategies are most likely to lead to customer churn. Furthermore, in complex sales scenarios, customer needs often require gradual clarification through multiple rounds of communication. However, traditional sales pitch design fails to fully consider the continuity of contextual information across multiple rounds of interaction, which can lead to subsequent recommendations or communications deviating from the user's true needs.

[0004] The Chinese invention patent, publication number CN119807486A, discloses an artificial intelligence-based medical education guidance method and system, comprising the following steps: setting up a medical education resource library in a cloud data center and constructing an artificial intelligence model; extracting and classifying features of a number of medical education resource data in the medical education resource library; generating a user profile based on the user's real-time basic information and real-time behavior data; matching a number of real-time resource classification results based on the real-time user profile; providing medical education guidance based on the real-time user profile and the features of a number of matched real-time resource data; sending a real-time medical education strategy and a number of target medical education resource data to a user terminal; and optimizing the parameters of the artificial intelligence model based on the real-time feedback data sent by the user terminal. This invention solves the problems of the existing technology, such as high cost investment, low efficiency, poor effect, lack of customization, and low degree of intelligence.

[0005] In current intelligent marketing scenarios, due to the lack of joint representation learning capabilities for multimodal, heterogeneous data (such as CBSS system operation logs, real-time voice emotion features, and financial credit labels), the process of identifying potential customer needs remains at a shallow semantic matching stage. Furthermore, traditional sequence modeling methods suffer from gradient decay when handling long conversational dependencies, leading to drift in intent recognition in the later stages of multi-round interactions. This cognitive bias accumulates with increasing conversational rounds, ultimately causing a systematic deviation between recommendation strategies and actual customer needs. Consequently, existing speech generation methods struggle to achieve personalized adaptation and are limited in their dynamic optimization effectiveness during multi-round interactions. This makes it difficult to achieve a dynamic balance between business rule constraints (such as compliance with package changes) and generation diversity. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback, comprising:

[0008] S1: Feature Fusion: Through the spatiotemporal attention mechanism, CBSS operation logs, voice emotion data, and financial credit data are integrated to obtain a dynamic customer portrait matrix.

[0009] S2: Obtaining the conversation state: Obtaining the initial conversation state based on the customer dynamic portrait matrix, and obtaining the conversation state of the current conversation through the constructed conversation state tracker and the initial conversation state;

[0010] S3: Conversation Status Correction: Determine the risk level of the current conversation using the Shannon entropy algorithm and the intent recognition sub-model, and correct the conversation status of the current conversation based on the risk level, including:

[0011] S3.1: Obtaining Entropy and Variance: Obtaining the entropy of the current conversation based on the conversation state and the Shannon entropy algorithm. Simultaneously, obtaining multiple prediction confidences for the current conversation using multiple intent recognition sub-models, and determining the variance of the current conversation based on the multiple prediction confidences.

[0012] S3.2: Obtaining a risk level: comparing the entropy value of the current conversation with a preset entropy threshold, comparing the variance of the current conversation with a preset variance threshold, and determining the risk level of the current conversation based on the comparison results;

[0013] S3.3: Risk handling: Based on the risk level of the current conversation, the conversation state tracker is modified using the constructed contrastive learning framework. The conversation state of the current conversation is corrected using the modified conversation state tracker. Steps S3.1 to S3.3 are repeated until the risk level of the current conversation is low risk.

[0014] S4: Determine marketing words: Based on the corrected conversation state of the current conversation, obtain a set of compliant words candidate through conditional generative adversarial network model, and determine personalized marketing words through the compliant words candidate set and reinforcement learning strategy.

[0015] Furthermore, we can obtain a customer dynamic profile matrix, including:

[0016] S1.1: Feature Extraction: Use the Transformer encoder to obtain behavioral pattern vectors from CBSS operation logs, use a 3D convolutional neural network to obtain the emotional intensity from voice emotion data, and use normalization processing to obtain the comprehensive credit weight from financial credit data.

[0017] S1.2: Weight allocation: Obtain the fusion weight based on the initial weight corresponding to the pattern vector, emotion intensity, and credit comprehensive weight, specifically:

[0018]

[0019] in: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th mode, is the total number of modes, is the index of the modal, It is the business scenario identifier. Score the reliability of the i-th mode;

[0020] S1.3: Generate a customer dynamic portrait matrix: Based on the fusion weights, obtain the fusion value of each portrait dimension, specifically:

[0021]

[0022] in: is the fusion value of the k-th image dimension, is the fusion weight of the i-th modality, is the eigenvalue of the i-th modality to the k-th image dimension, is the index of the portrait dimension, is the set of modalities involved in fusion.

[0023] Furthermore, the current conversation status is obtained, including:

[0024] S2.1: Determine the intent strength vector: Based on the real-time conversation text stream, obtain the text vector of each round of conversation, construct a time series feature matrix, and combine the time series feature matrix with the customer dynamic profile matrix to obtain a comprehensive feature vector. Simultaneously, use the comprehensive feature vector as the input of the short-term memory network model, output the obtained intent strength, and normalize the intent strength to obtain the normalized intent strength, specifically:

[0025]

[0026] in: is the normalized intensity of the m-th type of intent, is the natural exponential function, is the original score of the mth type of intent, is the category index of the intent, is the total number of intent types;

[0027] S2.2: Determine the combination strength: Based on the normalized intent strength and a preset number of rounds of conversation, determine the similarity between different intent strengths, and based on the similarity, obtain the combination strength between the different intent strengths, specifically:

[0028]

[0029] in: is the combined strength of the ath normalized intention strength and the bth normalized intention strength, is the similarity between the ath normalized intention strength and the bth normalized intention strength, is the ath normalized intention intensity, is the ath normalized intention intensity.

[0030] Furthermore, based on the text vector of the current round of dialogue, the combination strength corresponding to the current round of dialogue is determined, including:

[0031] S2.2.1: Determine Similarity: Construct a semantic vector based on the normalized intent strength corresponding to each round of conversation. Based on the semantic vector, obtain the similarity between two intents of different strengths. Compare the similarity with a preset similarity threshold and, based on the comparison result, perform intent combination. Specifically,

[0032] When the similarity is greater than a preset similarity threshold, the two intentions with different strengths corresponding to the similarity are combined and the next step S2.2.2 is executed. Otherwise, no intention combination is performed.

[0033] S2.2.2: Determine a valid combination: Based on the two different intention strengths for the intention combination, obtain the combined strength between the two different intention strengths, compare the combined strength with a preset strength threshold, and determine a valid combination based on the comparison result, specifically:

[0034] When the combination strength is greater than a preset strength threshold, the combination between the two intention strengths corresponding to the combination strength is a valid combination; otherwise, the combination between the two intention strengths corresponding to the combination strength is an invalid combination;

[0035] S2.2.3: Combination strength update: Based on the text vector of the current round of dialogue, repeat steps S2.2.1-S2.2.3 to update the similarity and combination strength between different intent strengths to determine the combination strength corresponding to the current round of dialogue.

[0036] Furthermore, the risk level of the current conversation is determined, specifically:

[0037] When the entropy value of the current conversation is greater than a preset entropy threshold, and the variance of the current conversation is greater than a preset variance threshold, the risk level of the current conversation is high risk. When the entropy value of the current conversation is greater than the preset entropy threshold, or the variance of the current conversation is greater than the preset variance threshold, the risk level of the current conversation is medium risk. Otherwise, the risk level of the current conversation is low risk.

[0038] Furthermore, according to the risk level of the current conversation, the conversation state of the current conversation is corrected, specifically:

[0039] When the risk level is low, step S4 is executed to determine personalized marketing tactics;

[0040] When the risk level is medium, the normalized intent strength of the current conversation is adjusted using the attention heat map, and steps S2.2 to S3.3 are repeated until the risk level is low.

[0041] When the risk level is high risk, the sample features of the current conversation are matched with the historical sample data features through a comparative learning framework to determine the matching historical sample data features, and based on the matching historical sample data features, the customer dynamic portrait matrix is ​​re-obtained, and steps S1.1-S3.3 are repeated until the risk level is low risk.

[0042] Furthermore, the normalized intent strength of the current conversation is adjusted through the attention heat map, specifically:

[0043] W1: Determine the attenuation coefficient: Based on the user value coefficient and the emotion correction coefficient, determine the final attenuation coefficient, specifically:

[0044]

[0045] in: is the final attenuation coefficient, is the basic attenuation coefficient, is the user value coefficient, is the emotion correction coefficient;

[0046] W2: Determine the attention weight: Based on the intention strength of the current round of dialogue and the historical average intention strength, and the difference between the current round of dialogue and the conflicting round of dialogue, determine the adjusted attention weight, specifically:

[0047]

[0048] in: is the final attention weight, is the intention intensity mutation value, is the historical average intensity, is the natural index, is the decay rate coefficient, is the time interval;

[0049] W3: Determine the final intention strength: Based on the final attenuation coefficient and the final attention weight, determine the adjusted normalized intention strength, specifically:

[0050]

[0051] in: is the normalized intensity after adjustment for the mth class of intent, is the normalized intensity of the m-th type of intent, is the final attenuation coefficient, is the final attention weight.

[0052] Going further, determine personalized marketing strategies, including:

[0053] S4.1: Obtaining a compliance score: Determine the total compliance loss based on the text content of the current conversation corresponding to the low risk, and obtain a compliance score based on the total compliance loss, specifically:

[0054]

[0055] in: is the total compliance loss, is the weight of the qth rule, is the threshold of the qth rule, is the actual value of the qth rule, is the total number of rules, is the index of the rule, is the total compliance loss;

[0056] S4.2: Determine a set of candidate compliant speech phrases: The text content of the current conversation corresponding to low risk is used to obtain the intent state of the current conversation. This intent state is then combined with the user profile to obtain a feature vector. Based on this feature vector and conditions, an adversarial network model is generated to determine compliant speech phrases and construct a candidate set of compliant speech phrases.

[0057] S4.3: Determine the marketing words: Through the reinforcement learning strategy and the compliant words, obtain the comprehensive reward value corresponding to each compliant word in the compliant word candidate set, and at the same time determine the maximum comprehensive reward value. The compliant word corresponding to the maximum comprehensive reward value is the final marketing word.

[0058] Furthermore, an adversarial network model is generated based on the feature vector and conditions to determine compliant speech, including:

[0059] The feature vector is used as the input of the generator in the conditional generative adversarial network model, and the output is a candidate set of speech phrases. Step S4.1 is repeated to obtain the compliance score of each candidate speech phrase in the candidate set of speech phrases.

[0060] The compliance score of the current conversation and the compliance score of each candidate speech are used as inputs to the discriminator in the conditional generative adversarial network model. The naturalness and compliance scores between each candidate speech and the current conversation are obtained as outputs. At the same time, the naturalness and compliance scores are compared with the preset naturalness threshold and the preset score threshold respectively. Based on the comparison results, the compliant speech is determined, specifically as follows:

[0061] When the naturalness is not less than the preset naturalness threshold and the compliance score is not less than the preset score threshold, the candidate speech corresponding to the naturalness and compliance score is a compliant speech; otherwise, the candidate speech is not a compliant speech.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] First, this invention uses a spatiotemporal attention mechanism to dynamically assign multimodal weights to CBSS operation logs, voice emotion data, and financial credit data. Simultaneously, it quantifies user dynamic profiles through a feature fusion matrix, thereby improving user profile accuracy and resolving the problem of misjudgment of intent caused by shallow semantic matching.

[0064] Second, this invention quantifies risk levels using the Shannon entropy algorithm and model confidence variance. Based on the risk level, it adjusts intent intensity using attention heatmaps and reconstructs user profiles using a comparative learning framework. This not only improves the stability of intent recognition in multi-round conversations, but also reduces the drift rate of intent in multi-round conversations.

[0065] Third: The present invention uses a conditional generative adversarial network to not only obtain a candidate set of speech words, but also perform dual verification of the naturalness and compliance scores of the generated candidate set of speech words. At the same time, through reinforcement learning optimization, personalized marketing speech words can be determined from the verified candidate set of speech words, which can not only reduce the compliance violation rate, but also improve the conversion rate of personalized speech words. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of the process of obtaining personalized marketing techniques in the present invention;

[0067] Figure 2 A schematic diagram of a process for determining an effective combination in the present invention;

[0068] Figure 3 It is a line graph showing changes in entropy value in the present invention;

[0069] Figure 4 This is a distribution diagram of the compliance of the speech in the present invention. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0071] In current intelligent marketing scenarios, due to the lack of joint representation learning capabilities for multimodal, heterogeneous data (such as CBSS system operation logs, real-time voice emotion features, and financial credit labels), the process of identifying potential customer needs remains at a shallow semantic matching stage. Furthermore, traditional sequence modeling methods suffer from gradient decay when handling long conversational dependencies, leading to drift in intent recognition in the later stages of multi-round interactions. This cognitive bias accumulates with increasing conversational rounds, ultimately causing a systematic deviation between recommendation strategies and actual customer needs. Consequently, existing speech generation methods struggle to achieve personalized adaptation and are limited in their dynamic optimization effectiveness during multi-round interactions. This makes it difficult to achieve a dynamic balance between business rule constraints (such as compliance with package changes) and generation diversity. The technical solution of this application dynamically integrates CBSS operation logs, voice emotion data and financial credit data through the spatiotemporal attention mechanism to generate a customer dynamic portrait matrix. Through Shannon entropy and variance analysis, it evaluates the conversation risk level in real time, and corrects the intention deviation through attention heat map or comparative learning framework. At the same time, it generates candidate speech through conditional generative adversarial network and combines reinforcement learning strategy to screen out the optimal speech, thereby solving the problems of data fragmentation, intention drift and rigid rules in traditional marketing, and improving the effect and compliance of personalized recommendations.

[0072] Example 1

[0073] refer to Figures 1-4 This embodiment provides a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback. The method includes the following steps:

[0074] Step S1: Feature Fusion. This involves extracting temporal features from the CBSS operation log using a Transformer encoder, extracting spatiotemporal emotional features from the voice emotion data using a 3D convolutional neural network, and then performing a multimodal feature fusion of the extracted temporal features, spatiotemporal emotional features, and financial credit data using a spatiotemporal attention mechanism to obtain a customer dynamic profile matrix. The details are as follows:

[0075] Step S1.1: Feature Extraction. This involves normalizing the user's 12-month transaction records (including but not limited to the time series of package changes and data usage fluctuations) to normalize the consumption amount and data usage rate to the range [0, 1]. This normalized data is then used as input for the Transformer encoder, which outputs the corresponding behavior pattern vector in the CBSS operation log.

[0076] Furthermore, the MFCC features in each frame of speech emotion data are used as the input of the 3D convolutional neural network, and the output obtains its corresponding emotion intensity.

[0077] Furthermore, after the financial credit data (including but not limited to central bank credit and consumption data) is normalized, the corresponding comprehensive credit weight is obtained through preset weight allocation based on the normalized credit score and consumption stability.

[0078] Step S1.2: Weight Assignment. This involves dynamically adjusting the initial weights of the CBSS operation log, voice emotion data, and financial credit data using the spatiotemporal attention mechanism to obtain the corresponding fused weights. It is worth noting that after normalizing the dynamically adjusted weights, the sum of the three normalized weights is 1.

[0079] In this embodiment, the initial weight adjustment formula is specifically:

[0080]

[0081] in: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th mode, is the total number of modes, is the index of the modal, It is the business scenario identifier. Score the reliability of the i-th mode.

[0082] Furthermore, the original score function of each modality is adjusted according to the scene benchmark score corresponding to each modality, specifically:

[0083]

[0084] in: is the original score function of the i-th mode, is the scene benchmark score of the i-th modality, is the reliability coefficient, It is the business scenario identifier. Score the reliability of the i-th mode.

[0085] During implementation, three business scenarios were identified: package recommendation, complaint handling, and credit assessment. The benchmark scores for package recommendation, complaint handling, and credit assessment were set at 0.6, 0.7, and 0.5, respectively. It's worth noting that in the package recommendation scenario, CBSS operation logs are the primary data, in the complaint handling scenario, voice and emotion data are the primary data, and in the credit assessment scenario, financial credit data are the primary data.

[0086] Specifically, let's set the reliability coefficient to 0.3, the reliability score for CBSS operation logs to 0.9, the reliability score for voice emotion data to 0.4, and the reliability score for financial credit data to 0.7. In the package recommendation scenario, the raw score function for CBSS operation logs is 0.87, the raw score function for voice emotion data is 0.72, and the raw score function for financial credit data is 0.81. In other words, in the package recommendation scenario, the fusion weight for CBSS operation logs is 0.36, the fusion weight for voice emotion data is 0.28, and the fusion weight for financial credit data is 0.36.

[0087] Step S1.3: Generate a customer dynamic portrait matrix. That is, based on the fusion weights obtained in step S1.2, obtain the fusion value of each portrait dimension, specifically:

[0088]

[0089] in: is the fusion value of the k-th image dimension, is the fusion weight of the i-th modality, is the eigenvalue of the i-th modality to the k-th image dimension, is the index of the portrait dimension, is the set of modalities involved in fusion.

[0090] Furthermore, a multi-dimensional feature fusion matrix is ​​constructed based on the reliability score of the CBSS operation log, the reliability score of the voice emotion data, the reliability score of the financial credit data and the corresponding fusion value corresponding to each portrait dimension.

[0091] During the specific implementation process, the reliability score of the CBSS operation log was 0.8, the reliability score of the voice emotion data was 0.7, and the reliability score of the financial credit data was 0.6. At the same time, in the business scenario of complaint handling, the fusion weight of the voice emotion data was 0.8, and the fusion weight of the CBSS operation log was 0.1. In other words, when obtaining the fusion value of the portrait dimension, the fusion value corresponding to the portrait dimension of the emotional state was 0.56, the fusion value corresponding to the portrait dimension of price sensitivity was 0.14, and the fusion value corresponding to the portrait dimension of credit reliability was 0.06. Therefore, the corresponding customer dynamic portrait matrix is ​​shown in Table 1 below:

[0092] Table 1: Customer dynamic portrait matrix

[0093] Dimensions CBSS characteristics Voice emotion Credit characteristics Fusion Value Price sensitivity 0.8 - 0.6 0.14 emotional state - 0.7 - 0.56 Creditworthiness - - 0.6 0.06

[0094] Step S2: Obtain the conversation state. This involves combining the customer dynamic profile matrix obtained in step S1.3 with the real-time conversation text stream. This combined matrix and real-time conversation text stream serve as the input for the constructed conversation state tracker, which outputs the conversation state of the current conversation. This is as follows:

[0095] Step S2.1: Determine the intent strength vector. This involves extracting the text vector for each conversation turn based on the real-time conversation text stream. In other words, a time series feature matrix is ​​constructed based on the conversation turns and the corresponding text vectors. Simultaneously, the customer dynamic profile matrix obtained in step S1.3 is combined with the time series feature matrix to obtain a comprehensive feature vector.

[0096] Furthermore, the comprehensive feature vector is used as the input of the short-term memory network model, and the corresponding intention strength is obtained as the output. At the same time, the obtained intention strength is normalized to obtain the normalized intention strength.

[0097] In this embodiment, the formula for obtaining the normalized intention strength is specifically:

[0098]

[0099] in: is the normalized intensity of the m-th type of intent, is the natural exponential function, is the original score of the mth type of intent, is the category index of the intent, is the total number of intent types.

[0100] Furthermore, the formula for obtaining the original score of the mth type of intent is as follows:

[0101]

[0102] in: is the original score of the mth type of intent, is the time step index of the current dialogue turn, is the time window size, For the The convolution weight of the round dialogue for the m-th type of intent, For the The feature value of the m-th intention in the round dialogue, The turn index of the conversation.

[0103] During the specific implementation, the text stream of this round of real-time conversation is "How is international roaming charged?", and the text vector of this round of conversation is [0.2, 0.1, 0.9]. At the same time, the text stream of the previous two rounds of conversation is "Insufficient data flow", and the text vector of this round of conversation is [0.8, 0.1, 0.1]. It is worth noting that the text vector data of each round of conversation represents the data flow demand dimension, price sensitivity dimension, and international roaming dimension respectively. The time series feature matrix in this embodiment is shown in Table 2 below, specifically:

[0104] Table 2: Timing feature matrix

[0105] Round Text Vector First two rounds [0.8,0.1,0.1] Previous round [0.3,0.6,0.1] Current round [0.2,0.1,0.9]

[0106] Furthermore, the profile data in the customer dynamic profile matrix is ​​set as follows: emotional state 0.56, price sensitivity 0.14, and after splicing the profile data in the customer dynamic profile matrix with the above-mentioned time series feature matrix, a comprehensive feature vector is obtained, as shown in Table 3 below:

[0107] Table 3: Comprehensive eigenvector matrix

[0108] Round Comprehensive feature vector First two rounds [0.8,0.1,0.1,0.56,0.14] Previous round [0.3,0.6,0.1,0.56,0.14] Current round [0.2,0.1,0.9,0.56,0.14]

[0109] Furthermore, using the comprehensive eigenvector as input to the short-term memory network model, the eigenvalues ​​of the traffic demand score in the three rounds are 0.8, 0.3, and 0.2, respectively, resulting in a raw score of 0.53. The eigenvalues ​​of the price sensitivity score in the three rounds are 0.1, 0.6, and 0.1, respectively, resulting in a raw score of 0.4. The eigenvalues ​​of the international roaming score in the three rounds are 0.1, 0.1, and 0.9, respectively, resulting in a raw score of 0.74. In other words, the normalized strength of traffic demand intensity is 0.32, the normalized strength of price sensitivity intensity is 0.28, and the normalized strength of international roaming intensity is 0.4.

[0110] Step S2.2: Determine the combination strength. That is, based on the normalized intent strengths obtained in step S2.1 and the preset number of rounds of dialogue, determine the similarity between different intent strengths. Based on the determined similarity, obtain the combination strength between different intent strengths, specifically:

[0111]

[0112] in: is the combined strength of the ath normalized intention strength and the bth normalized intention strength, is the similarity between the ath normalized intention strength and the bth normalized intention strength, is the ath normalized intention intensity, is the ath normalized intention intensity.

[0113] Step S3: Correcting the conversation state. This involves quantifying the deviation risk of the current conversation using the Shannon entropy algorithm and the intent recognition sub-model. Based on the quantification results, the corresponding risk level of the current conversation is determined. Furthermore, based on the determined risk level, the conversation state of the current conversation is corrected. The details are as follows:

[0114] Step S3.1: Obtain entropy and variance. This is done by using the Shannon entropy algorithm and the normalized intent strength corresponding to the current conversation text to determine the entropy corresponding to the current conversation. Specifically,

[0115]

[0116] in: is the entropy value of the current conversation, is the normalized intensity of the m-th type of intent, is the category index of the intent, is the total number of intent types.

[0117] During the specific implementation process, the normalized intention strength of the traffic demand corresponding to the current conversation is 0.4, and the normalized intention strength of the low-price demand is 0.6, so the entropy value corresponding to the current conversation is 0.97.

[0118] Furthermore, multiple intent recognition sub-models are used to obtain multiple prediction confidences for the current conversation. In this embodiment, the current conversation is used as input to the BERT model, RNN model, and rule engine, and the corresponding confidences of the BERT model, RNN model, and rule engine are obtained. Simultaneously, based on the corresponding confidences of the BERT model, RNN model, and rule engine, the variance of the current conversation is obtained, specifically:

[0119]

[0120] in: is the variance of the current conversation, is the number of intent recognition sub-models, To obtain the confidence level corresponding to the current conversation through the j-th intent recognition sub-model, is the mean confidence value corresponding to the current dialogue, An index that identifies the submodel number for the intent.

[0121] During the implementation process, the BERT model, RNN model, and rule engine obtain the traffic demand confidence levels in the current conversation as 0.6, 0.5, and 0.7, respectively. The mean confidence level corresponding to the current conversation is 0.6, and the variance of the current conversation is 0.007.

[0122] Step S3.2: Obtaining the risk level. Specifically, the entropy value of the current conversation obtained in step S3.1 is compared with a preset entropy threshold, and the variance of the current conversation obtained in step S3.1 is compared with a preset variance threshold. Based on the comparison results, the risk level of the current conversation is determined. Specifically:

[0123] When the obtained entropy value of the current conversation is greater than a preset entropy threshold, and the obtained variance of the current conversation is greater than a preset variance threshold, the risk level of the current conversation is high risk.

[0124] When the obtained entropy value of the current conversation is greater than a preset entropy threshold, or the obtained variance of the current conversation is greater than a preset variance threshold, the risk level of the current conversation is medium risk.

[0125] When the obtained entropy value of the current conversation is not greater than the preset entropy threshold, and the obtained variance of the current conversation is not greater than the preset variance threshold, the risk level of the current conversation is low risk.

[0126] In the specific implementation process, the preset entropy threshold in this embodiment is set to 0.7, and the preset variance threshold is set to 0.05. At the same time, the entropy value of the current conversation is 0.97, and the variance of the current conversation is 0.007, that is, the risk level of the current conversation is medium risk.

[0127] Step S3.3: Risk processing. That is, according to the risk level of the current conversation determined in step S3.2, corresponding processing of different risk levels is performed. Specifically:

[0128] When the risk level is low risk, step S4 is directly executed.

[0129] When the risk level is medium risk, the normalized intent strength of the current conversation is adjusted through the attention heat map, and steps S2.2 to S3.3 are repeated until the risk level is low risk.

[0130] When the risk level is high, the customer dynamic portrait matrix is ​​re-obtained through the constructed comparative learning framework, and steps S1.1 to S3.3 are repeated until the risk level of the current conversation is low.

[0131] In this embodiment, when the risk level is high, the corresponding features of the current round of conversation samples are matched with the features of the historical sample data to obtain matching features of the historical sample data. Based on the determined matching features of the historical sample data, the customer dynamic profile matrix is ​​re-obtained, i.e., steps S1.1 through S3.3 are repeated until the risk level of the current conversation is low.

[0132] It's worth noting that the matching rules between the current round's conversation sample features and historical sample data features include intent type, conversion rate, and user profile similarity. Specifically, matching historical sample data features include: the same intent conflict type, a conversion rate of no less than 80%, and a user profile similarity of no less than 70%.

[0133] refer to Figure 3 , Figure 3 is a line graph showing the change of entropy value in this embodiment, Figure 3 It can be seen that when the actual entropy value obtained is greater than the preset entropy threshold, correction through the attention heatmap or contrastive learning framework can reduce the entropy value to no more than the preset entropy threshold. For example, the entropy value of the third round of dialogue is 0.95, and the entropy value of the fourth round of dialogue is 0.7 (the preset entropy threshold). Furthermore, after the seventh round of dialogue, the entropy value stabilizes below 0.6, indicating that the multi-round interactive feedback mechanism effectively suppresses intent drift.

[0134] Step S4: Determine the marketing pitch. This involves obtaining the corresponding total compliance loss based on the text content of the current conversation corresponding to the low-risk criteria identified in step S3.3. Based on this total compliance loss and the knowledge graph, a conditional generative adversarial network model is used to construct a candidate set of compliant pitches. Furthermore, personalized marketing pitches are determined using a reinforcement learning strategy and the candidate set of compliant pitches. The details are as follows:

[0135] Step S4.1: Obtaining the compliance score. That is, based on the text content of the current conversation corresponding to the low risk determined in step S3.3, determine the corresponding total compliance loss, specifically:

[0136]

[0137] in: is the total compliance loss, is the weight of the qth rule, is the threshold of the qth rule, is the actual value of the qth rule, is the total number of rules, The index of the rule.

[0138] During the specific implementation process, the relevant business rules are generated as follows:

[0139] For a 5G package with a contract period of ≥12 months, the corresponding formula is: max(0.12 - contract period). For an international roaming package with a pre-deposit of ≥300 yuan, the corresponding formula is: max(0.300 - pre-deposit amount). The total compliance loss for these two business conversations is: 0.8*max(0.12 - contract period) and 0.5*max(0.300 - pre-deposit amount).

[0140] Furthermore, based on the total compliance loss obtained, the corresponding compliance score is determined, specifically:

[0141]

[0142] in: is the total compliance loss, is the total compliance loss, is the weight of the qth rule, is the threshold of the qth rule.

[0143] During the specific implementation process, the current conversation content is "Upgrade to 100GB package (24-month contract)", then the corresponding total compliance loss is 50, and the total compliance loss is 0.69.

[0144] Step S4.2: Determine the set of compliant speech candidates. This involves obtaining the intent state of the current conversation based on the text content corresponding to the low-risk criteria determined in step S3.3. This intent state is then combined with the user profile to obtain a feature vector. This feature vector is then used as the input to the generator in the conditional generative adversarial network model, which outputs the corresponding speech candidate set. Repeat step S4.1 for each candidate speech in the speech candidate set to obtain the corresponding compliance score for each candidate speech.

[0145] Furthermore, the determined compliance score of the current conversation and the compliance score corresponding to each candidate speech are used as inputs to the discriminator in the conditional generative adversarial network model, and the naturalness and compliance score between each candidate speech and the current conversation are obtained as output. The obtained naturalness is compared with the preset natural threshold, and the obtained compliance score is compared with the preset score threshold. Based on the comparison results, the compliant speech is determined. At the same time, based on the multiple compliant speech determined, a candidate set of compliant speech is constructed. Specifically:

[0146] If the obtained naturalness is not less than the preset naturalness threshold, and the obtained compliance score is not less than the preset score threshold, then the candidate speech corresponding to the naturalness and compliance score is a compliant speech. Otherwise, the corresponding candidate speech is not a compliant speech.

[0147] refer to Figure 4 , Figure 4 This is the compliance distribution diagram of the speech in this embodiment, Figure 4 It can be seen that when the compliance threshold is set to 0.8, only about 20%-30% of the speech can pass the compliance test, that is, most of the candidate speech obtained need to be further optimized.

[0148] Step S4.3: Determine the marketing tactics. That is, based on the set of compliant tactics candidate determined in step S4.2, the reward function corresponding to each compliant tactic is obtained through reinforcement learning strategy. Specifically, it is:

[0149]

[0150] in: is the comprehensive reward value, is the conversion probability weight, To predict the conversion probability, is the compliance weight, is the total compliance loss, is the round penalty coefficient, The current conversation turn.

[0151] Furthermore, based on the comprehensive reward value corresponding to each compliant speech obtained, the maximum comprehensive reward value is determined, and the compliant speech corresponding to the maximum comprehensive reward value is the final marketing speech.

[0152] Example 2

[0153] This embodiment provides a method for adaptively generating intelligent marketing scripts based on multi-round interactive feedback. Its specific implementation method is the same as that of Example 1, except that, in step S2.2, the normalized intent strength corresponding to the current round of conversation is updated based on the text vector of the current round of conversation. That is, based on the text vector of the current round of conversation, steps S2.1 and S2.2 are repeated to update the similarities and combined strengths between different intent strengths, thereby determining the combined strength corresponding to the current round of conversation. The present invention is illustrated below with reference to the specific implementation methods of this embodiment.

[0154] In this embodiment, the combination strength corresponding to the current round of dialogue is determined as follows:

[0155] Step S2.2.1: Determine similarity. This involves determining the normalized intent strength obtained in step S2.1 and combining it with a preset number of conversation rounds to determine the normalized intent strength corresponding to each round of conversation and construct a corresponding semantic vector. Based on the constructed semantic vector, the similarity between two different intent strengths is determined, specifically:

[0156]

[0157] in: is the similarity between the ath normalized intention strength and the bth normalized intention strength, is the ath normalized intention intensity, is the bth normalized intention strength.

[0158] Furthermore, the obtained similarity is compared with the preset similarity threshold, and the intention combination is performed based on the comparison result. Specifically:

[0159] When the obtained similarity is greater than the preset similarity threshold, the two different intention intensities corresponding to the similarity are intentionally combined, that is, the next step S2.2.2 is executed; otherwise, the intention is not combined.

[0160] In the specific implementation, the semantic vector of traffic demand is [0.8, 0.1, 0.1], and the semantic vector of international roaming is [0.2, 0.1, 0.9]. The similarity between traffic demand and international roaming is 0.345. Furthermore, the preset similarity threshold in this embodiment is set to 0.3, so the traffic demand and international roaming are intentionally combined in this embodiment.

[0161] Step S2.2.2: Determine a valid combination. That is, based on the two different intention strengths determined for the intention combination in step S2.2.1, obtain the corresponding combination strength based on the maximum intention strength of the two different intention strengths in the semantic vector, specifically:

[0162]

[0163] in: is the combined strength of the ath normalized intention strength and the bth normalized intention strength, is the similarity between the ath normalized intention strength and the bth normalized intention strength, is the ath normalized intention intensity, is the ath normalized intention intensity.

[0164] Furthermore, the obtained combination strength is compared with a preset strength threshold, and a valid combination is determined based on the comparison result. Specifically:

[0165] When the acquired combination strength is greater than the preset strength threshold, the combination between the two intention strengths corresponding to the combination strength is a valid combination; otherwise, the combination between the two intention strengths corresponding to the combination strength is an invalid combination.

[0166] In a specific implementation, the maximum traffic demand intention strength is 0.8, the maximum international roaming intention strength is 0.9, and the similarity between traffic demand and international roaming is 0.345. The corresponding combination strength is 0.345 * 0.8 = 0.276. Furthermore, the preset strength threshold in this embodiment is set to 0.2, so the combination of traffic demand and international roaming in this embodiment is a valid combination.

[0167] Step S2.2.3: Update combination strength. That is, based on the text vector of the current conversation round, repeat step S2.1 to update the normalized intent strength corresponding to the current conversation round. Simultaneously, based on the valid combinations determined in step S2.2.2 and the updated normalized intent strength, repeat steps S2.2.1-S2.2.3 to update the similarities and combination strengths between different intent strengths to determine the combination strength corresponding to the current conversation round.

[0168] Example 3

[0169] This embodiment provides a method for adaptively generating intelligent marketing speech based on multi-round interactive feedback. Its specific implementation method is the same as that of Example 1, except that, in step S3.3, when adjusting the normalized intent intensity of the current conversation using the attention heat map, the corresponding attenuation coefficient and attention weight are adjusted. The present invention is illustrated below with reference to the specific implementation methods of this embodiment.

[0170] In this embodiment, the normalized intent strength of the current conversation is adjusted using the attention heat map, as follows:

[0171] Step W1: Determine the attenuation coefficient. That is, based on the basic attenuation coefficients corresponding to different risk levels, combined with the user value coefficient and the sentiment correction coefficient, determine the final attenuation coefficient, specifically:

[0172]

[0173] in: is the final attenuation coefficient, is the basic attenuation coefficient, is the user value coefficient, is the emotion correction coefficient.

[0174] Furthermore, a basic attenuation coefficient is set according to the risk level. Specifically, the basic attenuation coefficient for low risk is 0.1, the basic attenuation coefficient for medium risk is 0.3, and the basic attenuation coefficient for high risk is 0.5.

[0175] Furthermore, user value coefficients are set based on different user types. Specifically, the user value coefficient for the high-value user group is 0.8, the user value coefficient for the potential churn user group is 1.2, and the user value coefficient for the average user group is 1.

[0176] Furthermore, an emotion correction coefficient is set based on the user's emotion value. Specifically, when the user's emotion value is less than 0.5, the corresponding emotion correction coefficient is 1. When the user's emotion value is between [0.5, 0.7], the corresponding emotion correction coefficient is 1.1. When the user's emotion value is between (0.7, 0.9], the corresponding emotion correction coefficient is 1.2. When the user's emotion value is greater than 0.9, the corresponding emotion correction coefficient is 1.5.

[0177] During the implementation, the current risk level is medium risk, the user type is a high-value user group, and the user sentiment value is 0.8. The corresponding final attenuation coefficient is 0.3*0.8*1.2=0.288.

[0178] Step W2: Determine the attention weight. That is, based on the intention strength corresponding to the current round of dialogue and the historical average intention strength, and the difference between the current round of dialogue and the conflicting round of dialogue, determine the adjusted attention weight, specifically:

[0179]

[0180] in: is the final attention weight, is the intention intensity mutation value, is the historical average intensity, is the natural index, is the decay rate coefficient, is the time interval.

[0181] In the specific implementation, the intent strength corresponding to the third round of conversation was 0.7, and the intent strength corresponding to the fourth round of conversation was 0.8. Meanwhile, the historical average strength for the past five rounds was 0.6. This resulted in a corresponding intent mutation of 0.17. Furthermore, with a decay rate coefficient of 0.5 and a time interval of 1, the corresponding time decay factor was 0.61. Therefore, the corresponding final attention weight was 0.78.

[0182] Step W3: Determine the final intention strength. That is, based on the final attenuation coefficient obtained in step W1 and the final attention weight obtained in step W2, obtain the adjusted normalized intention strength, specifically:

[0183]

[0184] in: is the normalized intensity after adjustment for the mth class of intent, is the normalized intensity of the m-th type of intent, is the final attenuation coefficient, is the final attention weight.

[0185] During the specific implementation process, the normalized intensity of the initial intention is 0.7, the final attention weight is 0.78, and the final attenuation coefficient is 0.288. The corresponding adjusted normalized intensity of the intention is: 0.7*(1-0.78*0.288)=0.54.

[0186] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is limited by the accompanying embodiments and their equivalents.

Claims

1. A method for adaptively generating intelligent marketing speech based on multi-round interactive feedback, characterized in that: Includes: S1: Feature Fusion: Through the spatiotemporal attention mechanism, CBSS operation logs, voice emotion data, and financial credit data are integrated to obtain a dynamic customer portrait matrix. S2: Obtaining the conversation state: The customer dynamic profile matrix and the real-time conversation text stream are combined to obtain a comprehensive feature vector, and the comprehensive feature vector is used as the input of the constructed conversation state tracker. The output obtains the conversation state of the current conversation, and the intention strength vector and the combination strength are determined based on the conversation state of the current conversation, including step S2.1 determining the intention strength vector and step S2.2 determining the combination strength; S3: Conversation Status Correction: Determine the risk level of the current conversation using the Shannon entropy algorithm and the intent recognition sub-model, and correct the conversation status of the current conversation based on the risk level, including: S3.1: Obtaining Entropy and Variance: Obtaining the entropy of the current conversation based on the conversation state and the Shannon entropy algorithm. Simultaneously, obtaining multiple prediction confidences for the current conversation using multiple intent recognition sub-models, and determining the variance of the current conversation based on the multiple prediction confidences. S3.2: Obtaining a risk level: comparing the entropy value of the current conversation with a preset entropy threshold, comparing the variance of the current conversation with a preset variance threshold, and determining the risk level of the current conversation based on the comparison results; S3.3: Risk Management: Based on the risk level of the current conversation, the conversation state tracker is modified using the constructed contrastive learning framework, and the conversation state of the current conversation is corrected using the modified conversation state tracker, specifically: When the risk level is low risk, step S4 is executed; When the risk level is medium, a final attenuation coefficient is determined based on the user value coefficient and the emotion correction coefficient. An adjusted attention weight is determined based on the intent strength of the current conversation round, the historical average intent strength, and the difference between the current conversation round and the conflicting conversation round. Furthermore, an adjusted normalized intent strength is determined based on the final attenuation coefficient and the final attention weight. At the same time, the normalized intent strength of the current conversation is adjusted according to the adjusted normalized intent strength, and steps S2.2 to S3.3 are repeated until the risk level is low risk; When the risk level is high, the sample features of the current conversation are matched with the features of the historical sample data through a comparative learning framework to determine the matching features of the historical sample data. Based on the matching features of the historical sample data, the customer dynamic profile matrix is ​​re-obtained, and steps S1 to S3.3 are repeated until the risk level is low. S4: Determine marketing words: Based on the corrected conversation state of the current conversation, obtain a set of compliant words candidate through conditional generative adversarial network model, and determine personalized marketing words through the compliant words candidate set and reinforcement learning strategy.

2. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 1, characterized in that: Get the customer dynamic portrait matrix, including: S1.1: Feature Extraction: Use the Transformer encoder to obtain behavioral pattern vectors from CBSS operation logs, use a 3D convolutional neural network to obtain the emotional intensity from voice emotion data, and use normalization processing to obtain the comprehensive credit weight from financial credit data. S1.2: Weight allocation: Obtain the fusion weight based on the initial weight corresponding to the pattern vector, emotion intensity, and credit comprehensive weight, specifically: in: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th mode, is the total number of modes, is the index of the modal, It is the business scenario identifier. Score the reliability of the i-th mode; S1.3: Generate a customer dynamic portrait matrix: Based on the fusion weights, obtain the fusion value of each portrait dimension, specifically: in: is the fusion value of the k-th image dimension, is the fusion weight of the i-th modality, is the eigenvalue of the i-th modality to the k-th image dimension, is the index of the portrait dimension, is the set of modalities involved in fusion.

3. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 1, characterized in that: Get the conversation status of the current conversation, including: S2.1: Determine the intent strength vector: Based on the real-time conversation text stream, obtain the text vector of each round of conversation, construct a time series feature matrix, and combine the time series feature matrix with the customer dynamic profile matrix to obtain a comprehensive feature vector. Simultaneously, use the comprehensive feature vector as the input of the short-term memory network model, output the obtained intent strength, and normalize the intent strength to obtain the normalized intent strength, specifically: in: is the normalized intensity of the m-th type of intent, is the natural exponential function, is the original score of the mth type of intent, is the category index of the intent, is the total number of intent types; S2.2: Determine the combination strength: Based on the normalized intent strength and a preset number of rounds of conversation, determine the similarity between different intent strengths, and based on the similarity, obtain the combination strength between the different intent strengths, specifically: in: is the combined strength of the ath normalized intention strength and the bth normalized intention strength, is the similarity between the ath normalized intention strength and the bth normalized intention strength, is the ath normalized intention intensity, is the ath normalized intention intensity.

4. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 3, characterized in that: Based on the text vector of the current round of dialogue, the combination strength corresponding to the current round of dialogue is determined, including: S2.2.1: Determine Similarity: Construct a semantic vector based on the normalized intent strength corresponding to each round of conversation. Based on the semantic vector, obtain the similarity between two intents of different strengths. Compare the similarity with a preset similarity threshold and, based on the comparison result, perform intent combination. Specifically, When the similarity is greater than a preset similarity threshold, the two intentions with different strengths corresponding to the similarity are combined and the next step S2.2.2 is executed. Otherwise, no intention combination is performed. S2.2.2: Determine a valid combination: Based on the two different intention strengths for the intention combination, obtain the combined strength between the two different intention strengths, compare the combined strength with a preset strength threshold, and determine a valid combination based on the comparison result, specifically: When the combination strength is greater than a preset strength threshold, the combination between the two intention strengths corresponding to the combination strength is a valid combination; otherwise, the combination between the two intention strengths corresponding to the combination strength is an invalid combination; S2.2.3: Combination strength update: Based on the text vector of the current round of dialogue, repeat steps S2.2.1-S2.2.3 to update the similarity and combination strength between different intent strengths to determine the combination strength corresponding to the current round of dialogue.

5. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 1, characterized in that: Determine the risk level of the current conversation, specifically: When the entropy value of the current conversation is greater than a preset entropy threshold, and the variance of the current conversation is greater than a preset variance threshold, the risk level of the current conversation is high risk. When the entropy value of the current conversation is greater than the preset entropy threshold, or the variance of the current conversation is greater than the preset variance threshold, the risk level of the current conversation is medium risk. Otherwise, the risk level of the current conversation is low risk.

6. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 1, characterized in that: Through the attention heat map, the normalized intent strength of the current conversation is adjusted as follows: W1: Determine the attenuation coefficient: Based on the user value coefficient and the emotion correction coefficient, determine the final attenuation coefficient, specifically: in: is the final attenuation coefficient, is the basic attenuation coefficient, is the user value coefficient, is the emotion correction coefficient; W2: Determine the attention weight: Based on the intention strength of the current round of dialogue and the historical average intention strength, and the difference between the current round of dialogue and the conflicting round of dialogue, determine the adjusted attention weight, specifically: in: is the final attention weight, is the intention intensity mutation value, is the historical average intensity, is the natural index, is the decay rate coefficient, is the time interval; W3: Determine the final intention strength: Based on the final attenuation coefficient and the final attention weight, determine the adjusted normalized intention strength, specifically: in: is the normalized intensity after adjustment for the mth class of intent, is the normalized intensity of the m-th type of intent, is the final attenuation coefficient, is the final attention weight.

7. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 1, characterized in that: Determine personalized marketing tactics, including: S4.1: Obtaining a compliance score: Determine the total compliance loss based on the text content of the current conversation corresponding to the low risk, and obtain a compliance score based on the total compliance loss, specifically: in: is the total compliance loss, is the weight of the qth rule, is the threshold of the qth rule, is the actual value of the qth rule, is the total number of rules, is the index of the rule, is the total compliance loss; S4.2: Determine a set of candidate compliant speech phrases: The text content of the current conversation corresponding to low risk is used to obtain the intent state of the current conversation. This intent state is then combined with the user profile to obtain a feature vector. Based on this feature vector and conditions, an adversarial network model is generated to determine compliant speech phrases and construct a candidate set of compliant speech phrases. S4.3: Determine the marketing words: Through the reinforcement learning strategy and the compliant words, obtain the comprehensive reward value corresponding to each compliant word in the compliant word candidate set, and at the same time determine the maximum comprehensive reward value. The compliant word corresponding to the maximum comprehensive reward value is the final marketing word.

8. The method for adaptively generating intelligent marketing speech based on multi-round interactive feedback according to claim 7, characterized in that: Based on the feature vector and conditions, an adversarial network model is generated to determine compliant speech, including: The feature vector is used as the input of the generator in the conditional generative adversarial network model, and the output is a candidate set of speech phrases. Step S4.1 is repeated to obtain the compliance score of each candidate speech phrase in the candidate set of speech phrases. The compliance score of the current conversation and the compliance score of each candidate speech are used as inputs to the discriminator in the conditional generative adversarial network model. The naturalness and compliance scores between each candidate speech and the current conversation are obtained as outputs. At the same time, the naturalness and compliance scores are compared with the preset naturalness threshold and the preset score threshold respectively. Based on the comparison results, the compliant speech is determined, specifically as follows: When the naturalness is not less than the preset naturalness threshold and the compliance score is not less than the preset score threshold, the candidate speech corresponding to the naturalness and compliance score is a compliant speech; otherwise, the candidate speech is not a compliant speech.

Images