Intelligent marketing verbal skill self-adaptive generation method based on multi-round interaction feedback

Abstract

The invention discloses an intelligent marketing verbal skill self-adaptive generation method based on multi-round interaction feedback, and relates to the technical field of intelligent dialogues. Comprising the following steps: S1, carrying out feature fusion on a CBSS operation log, voice emotion data and financial credit investigation data to obtain a client dynamic portrait matrix; s2, acquiring an initial dialogue state according to the client dynamic portrait matrix, and acquiring a dialogue state of a 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; and S4, according to the corrected dialogue state of the current dialogue, obtaining a compliance verbal skill candidate set, and through the compliance verbal skill candidate set and a reinforcement learning strategy, determining a personalized marketing verbal skill. According to the method, the problems of data splitting, intention drifting and rule stiffness in traditional marketing are solved, and the personalized recommendation effect and compliance are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent dialogue, and specifically to an intelligent marketing speech adaptive generation method based on multi-round interaction feedback. Background Art

[0002] Traditional sales scripts usually rely on manual design and are difficult to cover the personalized needs of different customer groups. For example, the same script may not be suitable for the communication scenarios of technical customers and non-technical customers. In addition, fixed scripts cannot be dynamically adjusted according to real-time customer feedback (such as tone, questions, objections), resulting in limited conversion rates and customer satisfaction.

[0003] At the same time, traditional methods rely on experience summary and manual optimization, lacking the ability to deeply analyze historical dialogue data. For example, it is impossible to mine which scripts are more effective or which strategies are likely to lead to customer loss through historical interaction data. And in complex sales scenarios, the needs of customers often need to be gradually clarified through multiple rounds of communication. However, traditional script design does not fully consider the continuity of context information in multi-round interactions, easily leading to subsequent recommendations or communications deviating from the real needs of users.

[0004] The Chinese invention patent with the publication number CN119807486A discloses a medical education guidance method and system based on artificial intelligence, including 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 several medical education resource data in the medical education resource library; generating a user profile according to the real-time basic information and real-time behavior data of the user; matching in several real-time resource classification results according to the real-time user profile; conducting medical education guidance according to the real-time user profile and the characteristics of several matched real-time resource data; sending the real-time medical education strategy and several target medical education resource data to the user terminal; optimizing the parameters of the artificial intelligence model according to the real-time feedback data sent by the user terminal. The present invention solves the problems of large cost investment, low efficiency, poor effect, lack of customization, and low intelligence level existing in the prior art.

[0005] In the current intelligent marketing scenario, due to the lack of the ability to jointly represent and learn multi-modal heterogeneous data (such as CBSS system operation logs, real-time voice emotion features, financial credit investigation tags), the process of mining customers' potential needs still remains at the stage of shallow semantic matching. At the same time, due to the problem of gradient decay in traditional sequence modeling methods when dealing with long conversation dependencies, it will lead to the phenomenon of intention recognition drift in the later stage of multi-round interaction, and this cognitive bias will accumulate with the increase of conversation rounds, ultimately resulting in a systematic deviation between the recommended strategy and the customer's true demands. As a result, it is difficult for existing conversation generation methods to achieve personalized adaptation, and the dynamic optimization effect in multi-round interaction is limited, that is, it is difficult to achieve the dynamic balance between business rule constraints (such as the compliance of package changes) and generation diversity. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent marketing conversation adaptive generation method based on multi-round interaction feedback to solve the problems proposed in the above background technology.

[0007] To achieve the above purpose, the present invention provides the following technical solutions: An intelligent marketing conversation adaptive generation method based on multi-round interaction feedback, including: S1: Feature fusion: Through the spatio-temporal attention mechanism, fuse the CBSS operation log, voice emotion data, and financial credit investigation data to obtain the customer dynamic portrait matrix; S2: Obtain the conversation state: According to the customer dynamic portrait matrix, obtain the initial conversation state, and through the constructed conversation state tracker and the initial conversation state, obtain the conversation state of the current conversation; S3: Conversation state correction: Determine the risk level of the current conversation through the Shannon entropy algorithm and the intention recognition sub-model, and correct the conversation state of the current conversation according to the risk level, including: S3.1: Obtain the entropy value and variance: According to the conversation state of the current conversation and the Shannon entropy algorithm, obtain the entropy value of the current conversation. At the same time, through multiple intention recognition sub-models, obtain multiple prediction confidence levels of the current conversation, and determine the variance of the current conversation according to the multiple prediction confidence levels; S3.2: Obtain the risk level: Compare the entropy value of the current conversation with the preset entropy threshold, compare the variance of the current conversation with the preset variance threshold, and determine the risk level of the current conversation according to the comparison results; S3.3: Risk Handling: According to the risk level of the current conversation, correct the conversation state tracker through the constructed contrastive learning framework, and correct the conversation state of the current conversation through the corrected conversation state tracker. At the same time, repeat steps S3.1 - S3.3 until the risk level of the current conversation is low risk; S4: Determine Marketing Script: According to the corrected conversation state of the current conversation, obtain a candidate set of compliant scripts through a conditional generative adversarial network model, and determine personalized marketing scripts through the candidate set of compliant scripts and a reinforcement learning strategy.

[0008] Furthermore, obtaining the customer dynamic portrait matrix includes: S1.1: Feature Extraction: Obtain the behavior pattern vector in the CBSS operation log through a Transformer encoder, obtain the emotional intensity in the speech emotion data through a 3D convolutional neural network, and obtain the comprehensive credit weight in the financial credit data through normalization processing; S1.2: Weight Allocation: Obtain the fusion weight according to the initial weight sizes corresponding to the pattern vector, emotional intensity, and comprehensive credit weight, specifically: Where: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th modality, is the total number of modalities, is the index of the modality, is the business scenario identifier, is the reliability score of the i-th modality; S1.3: Generate Customer Dynamic Portrait Matrix: According to the fusion weight, obtain the fusion value of each portrait dimension, specifically: Where: is the fusion value of the k-th portrait dimension, is the fusion weight of the i-th modality, is the eigenvalue of the i-th modality for the k-th portrait dimension, is the index of the portrait dimension, is the set of modalities participating in the fusion.

[0009] Furthermore, obtaining the conversation state of the current conversation includes: S2.1: Determine the intention intensity vector: Based on the real-time conversation text stream, obtain the text vectors of each round of conversation, construct a time series feature matrix, and combine the time series feature matrix with the customer dynamic portrait matrix to obtain a comprehensive feature vector. At the same time, use the comprehensive feature vector as the input of the short-term memory network model, output the obtained intention intensity, and perform normalization processing on the intention intensity to obtain the normalized intention intensity. Specifically: Where: is the normalized intensity of the m-th type of intention, is the natural exponential function, is the original score of the m-th type of intention, is the category index of the intention, is the total number of intention types; S2.2: Determine the combined intensity: Based on the normalized intention intensity and the conversation of the preset number of rounds, determine the similarity between different intention intensities, and obtain the combined intensity between different intention intensities according to the similarity. Specifically: Where: is the combined intensity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the similarity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the a-th normalized intention intensity, is the b-th normalized intention intensity.

[0010] Furthermore, based on the text vector of the current round of conversation, determine the combined intensity corresponding to the current round of conversation, including: S2.2.1: Determine the similarity: Based on the normalized intention intensity corresponding to each round of conversation, construct a semantic vector, and obtain the similarity between two different intention intensities according to the semantic vector. At the same time, compare the similarity with a preset similarity threshold, and perform intention combination according to the comparison result. Specifically: When the similarity is greater than the preset similarity threshold, the two different intention intensities corresponding to the similarity perform intention combination, and execute the next step S2.2.2. Otherwise, no intention combination is performed; S2.2.2: Determine the effective combination: Based on the two different intention intensities that perform intention combination, obtain the combined intensity between the two different intention intensities. At the same time, compare the combined intensity with a preset intensity threshold, and determine the effective combination according to the comparison result. Specifically: When the combined strength is greater than the preset strength threshold, the combination between the two intention strengths corresponding to the combined strength is a valid combination; otherwise, the combination between the two intention strengths corresponding to the combined strength is an invalid combination. S2.2.3: Update of combined strength: According to the text vector of the current round of conversation, repeat steps S2.2.1 - S2.2.3 to update the similarity and combined strength between different intention strengths, and determine the combined strength corresponding to the current round of conversation.

[0011] Furthermore, determine the risk level of the current conversation, specifically: When the entropy value of the current conversation is greater than the preset entropy threshold and the variance of the current conversation is greater than the preset variance threshold, the risk level of the current conversation is a 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 a medium risk. Otherwise, the risk level of the current conversation is a low risk.

[0012] Furthermore, correct the conversation state of the current conversation according to the risk level of the current conversation, specifically: When the risk level is a low risk, execute step S4 to determine the personalized marketing words. When the risk level is a medium risk, adjust the intention strength after normalization of the current conversation through the attention heat map, and repeat steps S2.2 - S3.3 until the risk level is a low risk. When the risk level is a high risk, match the sample features of the current conversation with the historical sample data features through the contrastive learning framework, determine the matching historical sample data features, and re - obtain the customer dynamic portrait matrix according to the matching historical sample data features, and repeat steps S1.1 - S3.3 until the risk level is a low risk.

[0013] Furthermore, adjust the intention strength after normalization of the current conversation through the attention heat map, specifically: W1: Determine the attenuation coefficient: According to the user value coefficient and the emotion correction coefficient, determine the final attenuation coefficient, specifically: Where: 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: According to 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 strength 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: According to 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 m-th intent, is the normalized strength of the m-th intent, is the final attenuation coefficient, is the final attention weight.

[0014] Going further, determine the personalized marketing language, 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 the candidate set of compliant speech: obtain the intention state of the current conversation from the text content of the current conversation corresponding to the low risk, and splice the intention state with the user portrait to obtain a feature vector, and generate an adversarial network model based on the feature vector and conditions to determine the compliant speech and construct a candidate set of compliant speech; 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 determine the maximum comprehensive reward value. The compliant word corresponding to the maximum comprehensive reward value is the final marketing word.

[0015] Furthermore, according to the feature vector and the conditional generative adversarial network model, the compliant conversation scripts are determined, including: Use the feature vector as the input of the generator in the conditional generative adversarial network model to output a set of candidate conversation scripts, and repeat step S4.1 to obtain the compliance scores of each candidate conversation script in the set of candidate conversation scripts; Use the compliance score of the current conversation and the compliance scores of each candidate conversation script as the input of the discriminator in the conditional generative adversarial network model to output the naturalness and compliance scores between each candidate conversation script and the current conversation. At the same time, compare the naturalness and compliance scores with the preset natural threshold and preset score threshold respectively, and determine the compliant conversation script according to the comparison results, specifically: When the naturalness is not less than the preset natural threshold and the compliance score is not less than the preset score threshold, the candidate conversation script corresponding to the naturalness and compliance scores is the compliant conversation script; otherwise, the candidate conversation script is not a compliant conversation script.

[0016] Compared with the prior art, the beneficial effects of the present invention are: First: Through the spatio-temporal attention mechanism, the present invention dynamically allocates multi-modal weights to the CBSS operation logs, voice emotion data, and financial credit data. At the same time, through the feature fusion matrix, the user dynamic portrait is quantified, thereby improving the accuracy of the user portrait and solving the problem of intention misjudgment caused by shallow semantic matching; Second: Through the Shannon entropy algorithm and the model confidence variance, the present invention quantifies the risk level, and according to different risk levels, adjusts the intention intensity through the attention heat map, and reconstructs the user image through the contrast learning framework, thereby not only improving the stability of intention recognition in multi-round conversations, but also reducing the intention drift rate in multi-round conversations; Third: Through the conditional generative adversarial network, the present invention can not only obtain a set of candidate conversation scripts, but also perform double verification on the naturalness and compliance scores of the generated set of candidate conversation scripts. At the same time, through reinforcement learning optimization, personalized marketing conversation scripts are determined from the verified set of candidate conversation scripts, thereby not only reducing the compliance violation rate, but also improving the conversion rate of personalized conversation scripts. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of obtaining personalized marketing conversation scripts in the present invention; Figure 2 It is a flow chart of determining effective combinations in the present invention; Figure 3 It is a line graph of the change of entropy value in the present invention; Figure 4 It is a distribution diagram of conversation script compliance in the present invention. Detailed implementation manners

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] In the current intelligent marketing scenario, due to the lack of the ability of joint representation learning for multi-modal heterogeneous data (such as CBSS system operation logs, real-time voice emotion features, financial credit investigation labels), the process of mining customers' potential needs still remains at the stage of shallow semantic matching. At the same time, due to the problem of gradient attenuation in the traditional sequence modeling method when dealing with long dialogue dependencies, the phenomenon of intention recognition drift will occur in the later stage of multi-round interaction, and this cognitive bias will accumulate with the increase of dialogue rounds, ultimately resulting in a systematic deviation between the recommendation strategy and the customer's true demands. As a result, it is difficult for the existing conversation generation methods to achieve personalized adaptation, and the dynamic optimization effect in multi-round interaction is limited, that is, it is difficult to achieve the dynamic balance between business rule constraints (such as package change compliance) and generation diversity. The technical solution of this application dynamically integrates CBSS operation logs, voice emotion data and financial credit investigation data through a spatio-temporal attention mechanism to generate a customer dynamic portrait matrix, evaluates the dialogue risk level in real time through Shannon entropy and variance analysis, corrects the intention deviation through an attention heat map or a contrast learning framework, and simultaneously generates candidate conversation sentences through a conditional generative adversarial network, and combines a reinforcement learning strategy to screen out the optimal conversation sentences, thus solving the problems of data fragmentation, intention drift and rule rigidity in traditional marketing, and improving the personalized recommendation effect and compliance.

[0020] Embodiment 1 Refer to Figures 1 - 4 , this embodiment provides an intelligent marketing conversation sentence adaptive generation method based on multi-round interaction feedback. The intelligent marketing conversation sentence adaptive generation method includes the following steps: Step S1: Feature fusion. That is, through a Transformer encoder, temporal features are extracted from CBSS operation logs, spatio-temporal emotion features are extracted from voice emotion data through a 3D convolutional neural network, and the extracted temporal features, spatio-temporal emotion features and financial credit investigation data are subjected to multi-modal feature fusion through a spatio-temporal attention mechanism to obtain a customer dynamic portrait matrix. Specifically as follows: Step S1.1: Feature extraction. That is, standardize the business records of the user within 12 months (including but not limited to the time series of package changes and fluctuations in traffic usage) to normalize the consumption amount, traffic usage rate, etc. to the interval [0, 1], and use the normalized data as the input of the Transformer encoder to output the corresponding behavior pattern vector in the CBSS operation log.

[0021] Furthermore, take the MFCC features in each frame of speech emotion data as the input of the 3D convolutional neural network to output and obtain its corresponding emotion intensity.

[0022] Furthermore, after normalizing the financial credit investigation data (including but not limited to central bank credit investigation and consumption data), according to the size of the normalized credit score and the size of consumption stability, obtain the corresponding comprehensive credit investigation weight through preset weight allocation.

[0023] Step S1.2: Weight allocation. That is, according to the initial weight sizes corresponding to the CBSS operation log, speech emotion data, and financial credit investigation data, dynamically adjust the set initial weight sizes through the spatio-temporal attention mechanism to respectively obtain the corresponding fusion weight sizes. It should be noted that after normalizing the dynamically adjusted weight sizes, the sum of the three normalized weight sizes is 1.

[0024] In this embodiment, the adjustment formula of the initial weight is specifically: Where: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th modality, is the total number of modalities, is the index of the modality, is the business scenario identifier, is the reliability score of the i-th modality.

[0025] Furthermore, adjust the original score function of each modality according to the scene benchmark score corresponding to each modality, specifically: Where: is the original score function of the i-th modality, is the scene benchmark score of the i-th modality, is the reliability coefficient, is the business scenario identifier, is the reliability score of the i-th modality.

[0026] In the process of specific implementation, the business scenario identifiers are set to three types, namely package recommendation, complaint handling, and credit assessment. Among them, the scenario baseline score corresponding to package recommendation is set to 0.6, the scenario baseline score corresponding to complaint handling is set to 0.7, and the scenario baseline score corresponding to credit assessment is set to 0.5. It should be noted that in the business scenario of package recommendation, CBSS operation logs are the dominant data; in the business scenario of complaint handling, voice emotion data are the dominant data; and in the business scenario of credit assessment, financial credit investigation data are the dominant data.

[0027] Furthermore, the reliability coefficient is set to 0.3. At the same time, the reliability score of CBSS operation logs is 0.9, the reliability score of voice emotion data is 0.4, and the reliability score of financial credit investigation data is 0.7. Then, in the business scenario of package recommendation, the original score function of CBSS operation logs is 0.87, the original score function of voice emotion data is 0.72, and the original score function of financial credit investigation data is 0.81. That is to say, in the business scenario of package recommendation, the fusion weight of CBSS operation logs is 0.36, the fusion weight of voice emotion data is 0.28, and the fusion weight of financial credit investigation data is 0.36.

[0028] Step S1.3: Generate the customer dynamic portrait matrix. That is, according to the fusion weights obtained in step S1.2, obtain the fusion value of each portrait dimension, specifically: Where: is the fusion value of the k-th portrait dimension, is the fusion weight of the i-th modality, is the eigenvalue of the i-th modality for the k-th portrait dimension, is the index of the portrait dimension, is the set of modalities participating in the fusion.

[0029] Furthermore, according to the reliability scores of CBSS operation logs, voice emotion data, and financial credit investigation data corresponding to each portrait dimension and the corresponding fusion values, construct a multi-dimensional feature fusion matrix.

[0030] In the process of specific implementation, the reliability score of the CBSS operation log is 0.8, the reliability score of the voice emotion data is 0.7, and the reliability score of the financial credit investigation data is 0.6. At the same time, in the business scenario of complaint handling, the fusion weight of the voice emotion data is 0.8, and the fusion weight of the CBSS operation log is 0.1. That is to say, when obtaining the fusion value of the portrait dimension, the fusion value corresponding to the portrait dimension of the emotional state is 0.56, the fusion value corresponding to the portrait dimension of price sensitivity is 0.14, and the fusion value corresponding to the portrait dimension of credit reliability is 0.06. Therefore, the corresponding customer dynamic portrait matrix is shown in Table 1 below: Table 1: Customer Dynamic Portrait Matrix Table Dimension CBSS Feature Voice Emotion Credit Investigation Feature Fusion Value Price Sensitivity 0.8 - 0.6 0.14 Emotional State - 0.7 - 0.56 Credit Reliability - - 0.6 0.06 Step S2: Obtain the dialogue state. That is, combine the customer dynamic portrait matrix obtained in step S1.3 with the real-time dialogue text stream, and use the combined customer dynamic portrait matrix and real-time dialogue text stream as the input of the constructed dialogue state tracker to output the dialogue state of the current dialogue. Specifically as follows: Step S2.1: Determine the intention intensity vector. That is, according to the real-time dialogue text stream, extract the text vector of each round of dialogue. That is to say, according to the dialogue round and the corresponding text vector, construct a time series feature matrix. At the same time, combine the customer dynamic portrait matrix obtained in step S1.3 with the time series feature matrix to obtain a comprehensive feature vector.

[0031] Furthermore, use the comprehensive feature vector as the input of the short-term memory network model and output the corresponding intention intensity. At the same time, normalize the obtained intention intensity to obtain the normalized intention intensity.

[0032] In this embodiment, the formula for obtaining the normalized intention intensity is specifically: Where: is the normalized intensity of the m-th type of intention, is the natural exponential function, is the original score of the m-th type of intention, is the category index of the intention, is the total number of intention types.

[0033] Furthermore, the formula for obtaining the original score of the m-th type of intention is specifically: Where: is the original score of the m-th type of intention, is the time step index of the current dialogue round, is the time window size, is the convolutional weight for the m-th type of intent in the n-th round of conversation, is the eigenvalue for the m-th type of intent in the n-th round of conversation, and n is the round index of the conversation.

[0034] In the process of specific implementation, the real-time conversation text stream of this round is "How much is the international roaming charge", the text vector of this round of conversation is [0.2, 0.1, 0.9], and at the same time, the conversation text streams of the previous two rounds are "Insufficient traffic", and the text vector of this round of conversation is [0.8, 0.1, 0.1]. It should be noted that the text vector data of each round of conversation respectively represents the traffic demand dimension, price sensitivity dimension, and international roaming dimension. Then the time series feature matrix in this embodiment is shown in Table 2 below, specifically: Table 2: Time Series Feature Matrix Table Round Text Vector The Previous Two Rounds [0.8,0.1,0.1] The Previous Round [0.3,0.6,0.1] The Current Round [0.2,0.1,0.9] Furthermore, the portrait data in the customer dynamic portrait matrix is set as: emotional state 0.56, price sensitivity 0.14. Then, after splicing the portrait data in the customer dynamic portrait matrix with the above time series feature matrix, a comprehensive feature vector is obtained, as shown in Table 3 below, specifically: Table 3: Comprehensive Feature Vector Matrix Table Round Comprehensive Feature Vector The Previous Two Rounds [0.8,0.1,0.1,0.56,0.14] The Previous Round [0.3,0.6,0.1,0.56,0.14] The Current Round [0.2,0.1,0.9,0.56,0.14] Furthermore, taking the comprehensive feature vector as the input of the short-term memory network model, the eigenvalues corresponding to the traffic demand scores in three rounds are: 0.8, 0.3, and 0.2, and the corresponding original score is 0.53. The eigenvalues corresponding to the price sensitivity scores in three rounds are: 0.1, 0.6, and 0.1, and the corresponding original score is 0.4. The eigenvalues corresponding to the international roaming scores in three rounds are: 0.1, 0.1, and 0.9, and the corresponding original score is 0.74. That is to say, the normalized intensity of the traffic demand intensity is 0.32, the normalized intensity of the price sensitivity intensity is 0.28, and the normalized intensity of the international roaming intensity is 0.4.

[0035] Step S2.2: Determine the combination strength. That is, according to the normalized intent strength obtained in step S2.1 and the conversations of the preset number of rounds, determine the similarity between different intent strengths. At the same time, according to the determined similarity, obtain the combination strength between different intent strengths, specifically: where: is the combination strength between the a-th normalized intent strength and the b-th normalized intent strength, is the similarity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the a-th normalized intention intensity, is the a-th normalized intention intensity.

[0036] Step S3: Dialogue state correction. That is, through the Shannon entropy algorithm and the intention recognition sub-model, the deviation risk of the current dialogue is quantified, and according to the quantification result, the risk level corresponding to the current dialogue is determined. At the same time, according to the determined risk level, the dialogue state of the current dialogue is corrected. Specifically as follows: Step S3.1: Obtain the entropy value and variance. That is, through the Shannon entropy algorithm and the normalized intention intensity corresponding to the current dialogue text, the entropy value corresponding to the current dialogue is determined, specifically: Where: is the entropy value of the current dialogue, is the normalized intensity of the m-th type of intention, is the category index of the intention, is the total number of intention types.

[0037] In the process of specific implementation, the normalized intention intensity of the traffic demand corresponding to the current dialogue is 0.4, and the normalized intention intensity of the low-price demand is 0.6, then the entropy value corresponding to the current dialogue is 0.97.

[0038] Furthermore, through multiple intention recognition sub-models, multiple prediction confidence levels of the current dialogue are obtained. In this embodiment, the current dialogue is used as the input of the BERT model, the RNN model, and the rule engine, and the confidence levels corresponding to the BERT model, the RNN model, and the rule engine are obtained respectively. At the same time, according to the confidence levels corresponding to the BERT model, the RNN model, and the rule engine, the variance of the current dialogue is obtained, specifically: Where: is the variance of the current dialogue, is the number of intention recognition sub-models, is the confidence level obtained by the j-th intention recognition sub-model for the current dialogue, is the average confidence level corresponding to the current dialogue, is the index of the intention recognition sub-model serial number.

[0039] In the process of specific implementation, the confidence levels of traffic demand in the current conversation obtained by the BERT model, RNN model, and rule engine are 0.6, 0.5, and 0.7 respectively. Then, the average confidence level corresponding to the current conversation is 0.6, and the variance of the current conversation is 0.007.

[0040] Step S3.2: Obtain the risk level. That is, compare the entropy value of the current conversation obtained in step S3.1 with a preset entropy threshold, compare the variance of the current conversation obtained in step S3.1 with a preset variance threshold, and determine the risk level of the current conversation according to the comparison results. Specifically: When the entropy value of the obtained current conversation is greater than the preset entropy threshold, and at the same time the variance of the obtained current conversation is greater than the preset variance threshold, the risk level of the current conversation is a high risk.

[0041] When the entropy value of the obtained current conversation is greater than the preset entropy threshold, or the variance of the obtained current conversation is greater than the preset variance threshold, the risk level of the current conversation is a medium risk.

[0042] When the entropy value of the obtained current conversation is not greater than the preset entropy threshold, and at the same time the variance of the obtained current conversation is not greater than the preset variance threshold, the risk level of the current conversation is a low risk.

[0043] In the process of specific implementation, 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 to say, the risk level of the current conversation is a medium risk.

[0044] Step S3.3: Risk handling. That is, perform corresponding processing for different risk levels according to the risk level of the current conversation determined in step S3.2. Specifically: When the risk level is a low risk, directly execute step S4.

[0045] When the risk level is a medium risk, adjust the intention intensity after normalization of the current conversation through an attention heat map, and repeat steps S2.2 - S3.3 until the risk level is a low risk.

[0046] When the risk level is a high risk, re-obtain the customer dynamic portrait matrix through the constructed contrastive learning framework, and at the same time repeat steps S1.1 - S3.3 until the risk level of the current conversation is a low risk.

[0047] In this embodiment, when the risk level is high risk, the dialogue sample features of the current round corresponding to the high risk are matched with the historical sample data features to obtain the matched historical sample data features. At the same time, according to the determined matched historical sample data features, the customer dynamic portrait matrix is re-obtained, that is, steps S1.1 - S3.3 are repeated until the risk level of the current dialogue is low risk.

[0048] It should be noted that the matching rules between the dialogue sample features of the current round and the historical sample data features include intention type, conversion rate, and user portrait similarity. Specifically, the matched historical sample data features are: having the same intention conflict type, a conversion rate of not less than 80%, and a user portrait similarity of not less than 70%.

[0049] Reference Figure 3 , Figure 3 is the line chart of the change of entropy value in this embodiment. It can be seen from Figure 3 that when the actually obtained entropy value is greater than the preset entropy threshold, after being corrected by the attention heat map or the contrast learning framework, the entropy value can be reduced to no greater than the preset entropy threshold. For example, the entropy value of the 3rd round of dialogue is 0.95, and the entropy value of the 4th round of dialogue is 0.7 (preset entropy threshold). At the same time, after the 7th round of dialogue, the entropy value stabilizes below 0.6, that is to say, the multi-round interaction feedback mechanism effectively suppresses intention drift.

[0050] Step S4: Determine the marketing words. That is, according to the text content of the current dialogue corresponding to the low risk determined in step S3.3, obtain its corresponding total compliance loss, and through the conditional generative adversarial network model according to the obtained total compliance loss and the knowledge graph, construct a candidate set of compliance words. At the same time, through the reinforcement learning strategy and the candidate set of compliance words, determine the personalized marketing words. Specifically as follows: Step S4.1: Obtain the compliance score. That is, according to the text content of the current dialogue corresponding to the low risk determined in step S3.3, determine the corresponding total compliance loss, specifically: Where: 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.

[0051] In the process of specific implementation, the relevant business rules are generated as follows: If the "5G package contract period ≥ 12 months", the corresponding formula is: max(0, 12 - contract period). If the "international roaming package requires a prepaid amount ≥ 300 yuan", the corresponding formula is: max(0, 300 - prepaid amount). Then the total compliance loss corresponding to these two business conversations is: 0.8 * max(0, 12 - contract period) and 0.5 * max(0, 300 - prepaid amount).

[0052] Furthermore, based on the obtained total compliance loss, the corresponding compliance score is determined, specifically: Where: is the total compliance loss, is the total compliance loss, is the weight of the q-th rule, is the threshold of the q-th rule.

[0053] In the process of specific implementation, if the current conversation content is "upgrade to 100GB package (24-month contract)", the corresponding total compliance loss is 50, and the total compliance loss is 0.69.

[0054] Step S4.2: Determine the candidate set of compliant conversation scripts. That is, according to the text content of the current conversation corresponding to the low risk determined in step S3.3, obtain the intent state of the current conversation, splice the intent state with the user profile to obtain a feature vector. And use the feature vector as the input of the generator in the conditional generative adversarial network model to output and obtain the corresponding candidate set of conversation scripts. At the same time, according to each candidate conversation script in the candidate set of conversation scripts, repeat step S4.1 to obtain the compliance score corresponding to each candidate conversation script.

[0055] Furthermore, take the determined compliance score of the current conversation and the compliance scores corresponding to each candidate conversation script as the input of the discriminator in the conditional generative adversarial network model, and output and obtain the naturalness and compliance scores between each candidate conversation script and the current conversation. Compare the obtained naturalness with the preset natural threshold, compare the obtained compliance score with the preset score threshold, and determine the compliant conversation script according to the comparison results. At the same time, construct a candidate set of compliant conversation scripts based on the determined multiple compliant conversation scripts. Specifically: When the obtained naturalness is not less than the preset natural threshold and the obtained compliance score is not less than the preset score threshold, the candidate conversation script corresponding to the naturalness and compliance score is the compliant conversation script. Otherwise, the corresponding candidate conversation script is not a compliant conversation script.

[0056] Refer to Figure 4 , Figure 4 is the distribution diagram of the compliance of the conversation script in this embodiment, from Figure 4It can be known that when the set compliance threshold is 0.8, only about 20%-30% of the words can pass the compliance detection, that is, most of the obtained candidate words need to be further optimized.

[0057] Step S4.3: Determine the marketing words. That is, according to the candidate set of compliance words determined in step S4.2, through the reinforcement learning strategy, obtain the reward function corresponding to each compliance word, specifically: Among them: is the comprehensive reward value, is the conversion probability weight, is the predicted conversion probability, is the compliance weight, is the total compliance loss, is the round penalty coefficient, is the current dialogue round.

[0058] Furthermore, according to the comprehensive reward value corresponding to each obtained compliance word, determine the maximum comprehensive reward value from them. The compliance word corresponding to the maximum comprehensive reward value is the final marketing word.

[0059] Embodiment 2 This embodiment provides an intelligent marketing word adaptive generation method based on multi-round interaction feedback. The specific implementation method is the same as that of Embodiment 1. The difference is that in step S2.2, according to the text vector of the current round of dialogue, update the normalized intention intensity corresponding to the current round of dialogue, that is, according to the text vector of the current round of dialogue, repeat steps S2.1 and S2.2, update the similarity and combination intensity between different intention intensities, and determine the combination intensity corresponding to the current round of dialogue. The present invention will be illustrated below in conjunction with the specific implementation manners of this embodiment.

[0060] In this embodiment, determine the combination intensity corresponding to the current round of dialogue, specifically as follows: Step S2.2.1: Determine the similarity. That is, according to the normalized intention intensity obtained in step S2.1, and at the same time combining the dialogue of the preset number of rounds, determine the normalized intention intensity corresponding to each round of dialogue, and construct the corresponding semantic vector. At the same time, according to the constructed semantic vector, obtain the similarity between two different intention intensities, specifically: Among them: is the similarity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the a-th normalized intention intensity, is the intention intensity after the b-th normalization process.

[0061] Furthermore, compare the obtained similarity with a preset similarity threshold, and perform intention combination according to the comparison result. Specifically: When the obtained similarity is greater than the preset similarity threshold, perform intention combination on the two different intention intensities corresponding to the similarity, that is, execute the next step S2.2.2; otherwise, do not perform intention combination.

[0062] In the process of 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]. Then the similarity between traffic demand and international roaming is 0.345. Furthermore, the preset similarity threshold in this embodiment is set to 0.3. Therefore, traffic demand and international roaming in this embodiment are combined in terms of intention.

[0063] Step S2.2.2: Determine the valid combination. That is, according to the two different intention intensities determined to perform intention combination in step S2.2.1, obtain the corresponding combined intensity through the intention intensity with the largest value among the two different intention intensities in the semantic vector. Specifically: Where: is the combined intensity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the similarity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the a-th normalized intention intensity, is the a-th normalized intention intensity.

[0064] Furthermore, compare the obtained combined intensity with a preset intensity threshold, and determine the valid combination according to the comparison result. Specifically: When the obtained combined intensity is greater than the preset intensity threshold, the combination between the two intention intensities corresponding to the combined intensity is a valid combination; otherwise, the combination between the two intention intensities corresponding to the combined intensity is an invalid combination.

[0065] In the process of specific implementation, the intention intensity with the largest value for traffic demand is 0.8, the intention intensity with the largest value for international roaming is 0.9, and the similarity between traffic demand and international roaming is 0.345. Then the corresponding combined intensity is 0.345 * 0.8 = 0.276. Furthermore, the preset intensity threshold in this embodiment is set to 0.2. Therefore, the combination between traffic demand and international roaming in this embodiment is a valid combination.

[0066] Step S2.2.3: Update of combined strength. That is, based on the text vector of the current round of conversation, repeat Step S2.1 to update the normalized intention strength corresponding to the current round of conversation. At the same time, based on the valid combinations determined in Step S2.2.2 and the updated normalized intention strength, repeat Steps S2.2.1 - S2.2.3 to update the similarity and combined strength between different intention strengths, so as to determine the combined strength corresponding to the current round of conversation.

[0067] Embodiment 3 This embodiment provides an intelligent marketing conversation adaptive generation method based on multi-round interaction feedback. The specific implementation method is the same as that of Embodiment 1, except that in Step S3.3, when adjusting the normalized intention strength of the current conversation through the attention heat map, the corresponding attenuation coefficient and attention weight are adjusted. The following is an example of the present invention in combination with the specific implementation manner of this embodiment.

[0068] In this embodiment, the normalized intention strength of the current conversation is adjusted through the attention heat map, specifically as follows: Step W1: Determine the attenuation coefficient. That is, based on the basic attenuation coefficients corresponding to different risk levels, and in combination with the user value coefficient and the emotion correction coefficient, determine the final attenuation coefficient, specifically: Where: is the final attenuation coefficient, is the basic attenuation coefficient, is the user value coefficient, is the emotion correction coefficient.

[0069] Furthermore, the basic attenuation coefficient is set according to different risk levels. Specifically, the basic attenuation coefficient corresponding to low risk is 0.1, the basic attenuation coefficient corresponding to medium risk is 0.3, and the basic attenuation coefficient corresponding to high risk is 0.5.

[0070] Furthermore, the user value coefficient is set according to different user types. Specifically, the user value coefficient corresponding to the high-value user group is 0.8, the user value coefficient corresponding to the potential churn user group is 1.2, and the user value coefficient corresponding to the ordinary user group is 1.

[0071] Furthermore, an emotion correction coefficient is set according to 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 in the range of [0.5, 0.7], the corresponding emotion correction coefficient is 1.1. When the user's emotion value is in the range of (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.

[0072] In the process of specific implementation, the current risk level is medium risk, the user type is a high-value user group, and the user's emotion value is 0.8. Then the corresponding final attenuation coefficient is 0.3 * 0.8 * 1.2 = 0.288.

[0073] Step W2: Determine the attention weight. That is, according to the intention intensity corresponding to the current round of conversation, the historical average intention intensity, and the difference between the current round of conversation and the conflict round of conversation, the adjusted attention weight is determined. Specifically: Where: is the final attention weight, is the intention intensity mutation value, is the historical average intensity, is the natural exponent, is the attenuation rate coefficient, is the time interval.

[0074] In the process of specific implementation, the intention intensity corresponding to the third round of conversation is 0.7, the intention intensity corresponding to the fourth round of conversation is 0.8, and the historical average intensity corresponding to the past five rounds is 0.6. Then the corresponding intention mutation degree is 0.17. Furthermore, the attenuation rate coefficient is set to 0.5 and the time interval is 1, so the corresponding time attenuation factor is 0.61. Therefore, the corresponding final attention weight is 0.78.

[0075] Step W3: Determine the final intention intensity. That is, according to the final attenuation coefficient obtained in Step W1 and the final attention weight obtained in Step W2, the adjusted and normalized intention intensity is obtained. Specifically: Where: is the adjusted normalized intensity of the mth type of intention, is the normalized intensity of the mth type of intention, is the final attenuation coefficient, is the final attention weight.

[0076] In the process of specific implementation, 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. Then the adjusted normalized intention intensity is: 0.7 * (1 - 0.78 * 0.288) = 0.54.

[0077] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended embodiments and their equivalents.

Claims

1. An intelligent marketing speech adaptive generation method based on multi-round interactive feedback, characterized in that, It includes: S1: Feature fusion: Through the spatio-temporal attention mechanism, fuse the CBSS operation logs, voice emotion data, and financial credit investigation data to obtain the customer dynamic portrait matrix; S2: Obtain the dialogue state: According to the customer dynamic portrait matrix, obtain the initial dialogue state, and through the constructed dialogue state tracker and the initial dialogue state, obtain the dialogue state of the current dialogue; S3: Dialogue state correction: Through the Shannon entropy algorithm and the intent recognition sub-model, determine the risk level of the current dialogue, and according to the risk level, correct the dialogue state of the current dialogue, including: S3.1: Obtain the entropy value and variance: According to the dialogue state of the current dialogue and the Shannon entropy algorithm, obtain the entropy value of the current dialogue. At the same time, through multiple intent recognition sub-models, obtain multiple prediction confidence levels of the current dialogue, and according to the multiple prediction confidence levels, determine the variance of the current dialogue; S3.2: Obtain the risk level: Compare the entropy value of the current dialogue with the preset entropy threshold, compare the variance of the current dialogue with the preset variance threshold, and according to the comparison results, determine the risk level of the current dialogue; S3.3: Risk handling: According to the risk level of the current dialogue, through the constructed contrastive learning framework, correct the dialogue state tracker, and through the corrected dialogue state tracker, correct the dialogue state of the current dialogue. At the same time, repeat steps S3.1 - S3.3 until the risk level of the current dialogue is low risk; S4: Determine the marketing script: According to the corrected dialogue state of the current dialogue, through the conditional generative adversarial network model, obtain the candidate set of compliant scripts, and through the candidate set of compliant scripts and the reinforcement learning strategy, determine the personalized marketing script.

2. The adaptive generation method of intelligent marketing conversation based on multi-round interaction feedback according to claim 1, characterized in that Obtaining the customer dynamic portrait matrix includes: S1.1: Feature extraction: Through the Transformer encoder, obtain the behavior pattern vector in the CBSS operation logs, through the 3D convolutional neural network, obtain the emotion intensity in the voice emotion data, and through normalization processing, obtain the comprehensive credit weight in the financial credit investigation data; S1.2: Weight assignment: According to the initial weight sizes corresponding to the pattern vector, emotion intensity, and comprehensive credit weight, obtain the fusion weight, specifically: Wherein: is the fusion weight of the i-th modality, is the natural exponential function, is the original score function of the i-th modality, is the total number of modalities, is the index of the modality, is the business scenario identifier, is the reliability score of the i-th modality; S1.3: Generate the customer dynamic portrait matrix: According to the fusion weight, obtain the fusion value of each portrait dimension, specifically: Wherein: 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 for the k-th image dimension, is the index of the image dimension, is the set of modalities participating in the fusion.

3. The adaptive generation method of intelligent marketing words based on multi-round interaction feedback according to claim 1, wherein Obtaining the dialogue state of the current dialogue includes: S2.1: Determine the intent intensity vector: According to the real-time dialogue text stream, obtain the text vector of each round of dialogue, construct the temporal feature matrix, and combine the temporal feature matrix and the customer dynamic portrait matrix to obtain the comprehensive feature vector. At the same time, use the comprehensive feature vector as the input of the short-term memory network model, output to obtain the intent intensity, and perform normalization processing on the intent intensity to obtain the normalized intent intensity, specifically: Wherein: is the normalized intensity of the m-th type of intention, is the natural exponential function, is the original score of the m-th type of intention, is the category index of the intention, is the total number of intention types; S2.2: Determine the combination strength: Determine the similarity between different intention strengths according to the normalized intention strength and the preset number of rounds of dialogue, and obtain the combination strength between different intention strengths according to the similarity, specifically: Wherein: is the combined strength between the a-th normalized intention intensity and the b-th normalized intention intensity, is the similarity between the a-th normalized intention intensity and the b-th normalized intention intensity, is the a-th normalized intention intensity, is the a-th normalized intention intensity.

4. The adaptive generation method of intelligent marketing words based on multi-round interactive feedback according to claim 3, characterized in that According to 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 dialogue, and obtain the similarity between two different intent strengths based on the semantic vector, and compare the similarity with a preset similarity threshold, and perform intent combination based on the comparison result, specifically: When the similarity is greater than a preset similarity threshold, the two different intention strengths corresponding to the similarity are combined and the next step S2.2.2 is executed. Otherwise, the intention is not combined. S2.2.2: Determine a valid combination: According to two different intention strengths for intention combination, obtain the combination strength between the two different intention strengths, and compare the combination strength with a preset strength threshold, and determine a valid combination according to 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: According to 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 intention strengths to determine the combination strength corresponding to the current round of dialogue.

5. The adaptive generation method of intelligent marketing conversation skills 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 adaptive generation method of intelligent marketing words based on multi-round interactive feedback according to claim 1 or 5, characterized in that According to the risk level of the current conversation, the conversation state of the current conversation is corrected, specifically: When the risk level is low risk, step S4 is executed to determine personalized marketing words; When the risk level is medium risk, the normalized intention 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; 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 ​​reacquired, and steps S1.1-S3.3 are repeated until the risk level is low risk.

7. The adaptive generation method of intelligent marketing words based on multi-round interactive feedback according to claim 6, 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: Wherein: 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 intensity of the current round of conversation, the historical average intention intensity, and the difference between the current round of conversation and the conflict round of conversation, determine the adjusted attention weight, specifically: Wherein: is the final attention weight, is the intention intensity mutation value, is the historical average intensity, is the natural exponent, is the attenuation rate coefficient, is the time interval; W3: Determine the final intention intensity: Based on the final attenuation coefficient and the final attention weight, determine the adjusted and normalized intention intensity, specifically: Wherein: is the normalized intensity after the adjustment of the m-th type of intention, is the normalized intensity of the m-th type of intention, is the final attenuation coefficient, is the final attention weight.

8. A method for adaptively generating intelligent marketing scripts based on multi-round interactive feedback according to claim 1, characterized in that, Determine the personalized marketing words, including: S4.1: Obtain the compliance score: Based on the text content of the current conversation corresponding to low risk, determine the total compliance loss, and based on the total compliance loss, obtain the compliance score, specifically: Wherein: is the total compliance loss, is the weight of the q-th rule, is the threshold of the q-th rule, is the actual value of the q-th rule, is the total number of rules, is the index of the rule, is the total compliance loss; S4.2: Determine the compliance words candidate set: For the text content of the current conversation corresponding to low risk, obtain the intention state of the current conversation, splice the intention state and the user portrait to obtain a feature vector, and at the same time, based on the feature vector and the conditional generative adversarial network model, determine the compliance words and construct a compliance words candidate set; S4.3: Determine the marketing words: Through the reinforcement learning strategy and the compliance words, obtain the comprehensive reward value corresponding to each compliance word in the compliance words candidate set, and at the same time determine the maximum comprehensive reward value. The compliance word corresponding to the maximum comprehensive reward value is the final marketing word.

9. The adaptive generation method of intelligent marketing words based on multi-round interactive feedback according to claim 8, characterized in that Based on the feature vector and the conditional generative adversarial network model, determine the compliance words, including: Use the feature vector as the input of the generator in the conditional generative adversarial network model, output to obtain a words candidate set, and repeat step S4.1 to obtain the compliance score of each candidate word in the words candidate set; Use the compliance score of the current conversation and the compliance score of each candidate word as the input of the discriminator in the conditional generative adversarial network model, output to obtain the naturalness and compliance score between each candidate word and the current conversation, and at the same time compare the naturalness and compliance score with the preset natural threshold and the preset score threshold respectively, and based on the comparison results, determine the compliance words, specifically: When the naturalness is not less than the preset natural threshold and the compliance score is not less than the preset score threshold, the candidate word corresponding to the naturalness and compliance score is the compliance word; otherwise, the candidate word is not the compliance word.

Images