Large-model complaint intention recognition method based on sentiment analysis

Abstract

The invention discloses a large-model complaint intention recognition method based on sentiment analysis, and the method comprises the steps: obtaining call voice data of a customer, and converting the call voice data into text data; preprocessing the text data, and removing noise and marking components; performing emotion feature analysis on the text data, extracting an emotion feature value, and generating an emotion vector; the emotion vectors and the text data are input into a large model together, and the large model combines the emotion feature values and context semantics to generate intention feature vectors; constructing an emotion-intention state vector, inputting the emotion-intention state vector into an asynchronous dominant actor commentator algorithm model, and generating a corresponding complaint intention probability value; judging a complaint intention probability, and generating risk early warning; and preferentially distributing high-risk customers and responding to customer demands. Through combination of voice data preprocessing, text semantic feature extraction, emotion intensity quantitative analysis and a multi-modal fusion algorithm and efficient risk assessment based on an asynchronous dominant actor reviewer model, the early warning and response capabilities of customer complaint risks are significantly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and relates to a method for identifying complaint intentions of a large model based on sentiment analysis. Background Art

[0002] With the rapid development of artificial intelligence technology, customer service systems are gradually transforming towards intelligence, and sentiment analysis and semantic understanding technologies have become the key to improving customer service quality. In traditional customer service scenarios, customer service staff mainly rely on personal experience to subjectively judge the emotions and intentions of customers. This method may be effective when dealing with individual customers, but its limitations become increasingly prominent when facing a large number of customers. On the one hand, it is difficult for manual judgment to maintain consistency. Especially when dealing with customers with complex emotions, it is easily interfered by subjective factors, resulting in inaccurate judgment results. On the other hand, the efficiency of manual service is low, and it is difficult to meet the needs of large-scale customer complaint prediction and handling. Especially in the early stage when customer emotions intensify, it is often impossible to achieve timely and accurate identification of complaint intentions and effective customer diversion.

[0003] To overcome these limitations, some intelligent customer service systems have emerged in the current market. These systems attempt to introduce sentiment analysis technology to classify the emotions of customers. However, most of them only stay at the level of emotion detection, mainly identifying customer emotions through keyword matching or simple sentiment classification methods, lacking in-depth semantic understanding and intention judgment. Therefore, these systems fail to deeply explore the true intentions behind customer emotions in emotion recognition, making it difficult to effectively predict customer complaint intentions. For example, when a customer shows dissatisfaction due to product quality problems, existing systems often have difficulty accurately distinguishing whether this will translate into a complaint intention.

[0004] In response to the above problems, there are currently the following several technical solutions, but each has its own deficiencies:

[0005] Sentiment analysis and intention recognition method based on a static corpus: This method establishes the association between sentiment and intention through learning a large amount of labeled data. Since it is based on a static corpus, it cannot adapt to the changes in customer emotions and language expression methods in real time, resulting in a decline in the performance of the model when dealing with new data. In addition, this method has low adaptability to rare scenarios and cannot capture context associations, and can only perform isolated analysis on single inputs.

[0006] Complaint intention prediction method based on emotion feature statistics: This method predicts complaint intentions by extracting emotion features in customer inputs and combining statistical rules. However, relying only on emotion feature statistics ignores the semantic structure and deep emotional meaning of language, resulting in insufficient processing ability for complex sentence patterns or metaphorical expressions. At the same time, the generalization ability of this method is limited, and the capture of customer emotion changes is limited to a single moment, and it cannot identify the process of emotion progression or intensification.

[0007] Rule-driven intelligent shunting system: This method realizes intelligent shunting through predefined rules, simplifying the shunting logic. However, the design and maintenance of rules rely on manual experience, which not only involves a large amount of work but also is difficult to cover complex scenarios. Moreover, this method lacks flexibility and scalability. When new business scenarios or complaint types emerge, rules need to be redefined, resulting in insufficient system adaptability. At the same time, the accuracy of shunting decisions is also relatively low, unable to combine multi-level semantics and emotional dynamics, leading to a low matching degree of shunting strategies to the actual needs of customers.

[0008] In summary, the existing technologies have significant deficiencies in the depth of emotion recognition, the accuracy of intent prediction, and the connection of intelligent shunting. There is an urgent need for a new method to achieve more accurate sentiment analysis and intent recognition, provide efficient complaint risk management and intelligent shunting strategies, so as to better meet the needs of complaint intent recognition in customer service scenarios. Summary of the Invention

[0009] The purpose of the present invention is to solve the problems existing in the prior art in the depth of emotion recognition, the accuracy of intent prediction, and the connection of intelligent shunting, and to provide a method for identifying complaint intent based on a large model of sentiment analysis.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A method for identifying complaint intent based on a large model of sentiment analysis, comprising the following steps:

[0012] S1. Obtain the voice data of the customer call, and convert the voice data into text data through a voice-to-text module;

[0013] S2. Perform text preprocessing on the text data, remove noise and mark semantic components, and generate text data with clear semantics;

[0014] S3. Based on the sentiment analysis model, perform sentiment feature analysis on the preprocessed text data, extract the sentiment feature values in the text data, where the sentiment feature values include the intensity values of negative emotions such as anger, dissatisfaction, and anxiety, and generate an emotion vector containing sentiment features, and the emotion vector is used to characterize the emotional state of the customer during the call;

[0015] S4. Use the emotion vector as the input data for initially judging the complaint risk, and input it together with the text data into the large model. The large model combines the sentiment feature values in the emotion vector and the context semantic information in the text data to perform in-depth semantic understanding on the potential complaint intent of the customer and generate an intent feature vector;

[0016] S5. Construct an emotion-intention state vector based on the emotion vector and the intention feature vector, and input the emotion-intention state vector into the asynchronous advantage actor-critic algorithm model. The asynchronous advantage actor-critic algorithm model asynchronously processes the emotion-intention state vector in a multi-threaded environment to determine whether the customer has a complaint intention and generates a corresponding complaint intention probability value;

[0017] S6. In the asynchronous advantage actor-critic algorithm model, judge the complaint intention probability value based on a preset complaint intention threshold. When the complaint intention probability value is greater than the complaint intention threshold, mark the customer as a high complaint risk and generate a risk warning message for subsequent diversion and response;

[0018] S7. Input the risk warning message into the intelligent diversion module. After receiving the risk warning message, the intelligent diversion module preferentially assigns the customers with high complaint risks to full-time customer service for processing;

[0019] S8. After the customer is assigned, the intelligent diversion module generates real-time service guidelines and response strategies according to the risk warning message.

[0020] The specific steps in S1 include the following:

[0021] S1.1, Receive the real-time voice data stream X of the customer call t , where t represents the time step. The voice data stream is collected through the customer's voice input device and is input into the voice preprocessing module of the system in real time through the transmission network;

[0022] S1.2, Perform noise reduction processing on the real-time voice data stream X in the voice preprocessing module t , use a filter to eliminate background noise and high-frequency components, and generate a noise-reduced voice signal X t ';

[0023] S1.3, Send the noise-reduced voice signal X t ' to the voice framing processing module, and segment the voice signal X t ' according to the time window ΔT to generate a voice frame sequence X t '(i), where i represents the i-th segmented voice segment;

[0024] S1.4, Use the Mel-frequency cepstral coefficient analysis method to extract features from each voice frame sequence X t '(i), extract the acoustic feature M i , and obtain a feature vector M = {M i}, where each M i represents the voice feature in the i-th frame;

[0025] S1.5, Input the feature vector M into the speech-to-text module, and process the feature vector through a pre-trained speech recognition model to generate corresponding text data T, where T = {T i} represents the text segment corresponding to each speech frame;

[0026] S1.6, Merge all text segments T i to form the complete text data T total .

[0027] The specific steps in S3 are as follows:

[0028] S2.1, Receive the text data T generated by the speech-to-text module total , and perform denoising processing on the text data T total , remove redundant non-verbal characters and low-confidence words through a noise filtering algorithm, and generate denoised text data T total ';

[0029] S2.2, Perform semantic tagging on the denoised text data T total ', use predefined semantic component tags to tag the semantic components in the text data, and generate tagged text data T total ″;

[0030] S2.3, Use a word segmentation algorithm to perform word segmentation on the tagged text data T total ″ to generate a word segmentation result sequence W;

[0031] S2.4, Perform stop word filtering based on the word segmentation result sequence W, remove common meaningless words, and obtain a semantically clear word segmentation sequence W';

[0032] S2.5, Perform part-of-speech tagging on the semantically clear word segmentation sequence W', and tag each word segmentation unit w i ' according to its grammatical role and semantic attributes to obtain text data T clean with clear semantics and part-of-speech information.

[0033] The common meaningless words are functional words that appear frequently in a sentence but have little impact on the sentence semantics, including auxiliary words, prepositions, structural words, conjunctions, and function words, pronouns, adverbs, modal particles, demonstrative words, and interrogative words.

[0034] The specific steps in S3 are as follows:

[0035] S3.1, Receive the preprocessed text data T clean , perform sentiment classification on the text data T clean based on a sentiment analysis model, identify the emotion categories in the text, and the emotion categories include negative emotions such as anger, dissatisfaction, and anxiety;

[0036] S3.2. Perform intensity analysis on the emotion categories and calculate the intensity value E of each emotion i , where i represents the emotion category, including anger, dissatisfaction, and anxiety, and the intensity value of each emotion is generated based on the occurrence frequency of emotional words and the context weight;

[0037] S3.3. Generate an emotion feature vector:

[0038] E = {E ang , E dis , E anx}

[0039] where E ang , E dis , E anx represent the intensity values of anger, dissatisfaction, and anxiety emotions respectively, and are used to characterize the customer's emotional state during the call.

[0040] The emotion intensity value is obtained by weighted summation of the occurrence frequency of emotional words and the context weight, and the specific expression is:

[0041]

[0042] where E i represents the intensity value of emotion category i; N i is the total number of emotional words belonging to emotion i in the text; f ij represents the occurrence frequency or degree of the jth emotional word in the text; w ij is the context weight corresponding to this emotional word, and this weight is calculated based on the position, dependency relationship, and context information of the sentence where the emotional word is located, so as to reflect its influence in the current context; b i is the bias term, used to adjust the overall effect; σ is the activation function, used to perform non - linear mapping on the output.

[0043] The specific steps in S4 include the following:

[0044] S4.1. Receive the emotion feature vector E generated by emotion analysis and the pre - processed text data T clean , and use the emotion feature vector E as the input data for preliminary judgment of the complaint risk;

[0045] S4.2. Input the emotion feature vector E and the text data T clean into the large model together, and through the multi - modal fusion mechanism, conduct in - depth semantic understanding of the customer's potential complaint intention to generate an intention feature vector I:

[0046] I = σ(W e E + W t φ(T clean) + b)

[0047] Among them, σ is the activation function, and W e is the emotional feature weight matrix, which maps the emotional feature vector E to the hidden space to capture the influence of emotional information on the complaint intention. W t is the text semantic feature weight matrix, which maps the semantic feature vector Φ(T clean ) to the hidden space to capture the influence of text content on the complaint intention. Φ(T clean ) is the text semantic feature extraction function, which uses a pre-trained word vector model to convert the text data. b is the bias vector;

[0048] S4.3, calculate the complaint intention probability value P comp , and use the softmax function to normalize the intention feature vector I:

[0049]

[0050] Among them, I comp is the eigenvalue of the complaint intention, indicating the degree of the customer's complaint intention. I noncomp is the eigenvalue of the non-complaint intention, indicating the degree of the customer having no complaint intention. P comp is the probability value that the customer has a complaint intention, and the range is between [0, 1];

[0051] S4.4, compare the complaint intention probability value P comp with the preset complaint intention threshold θ. When P comp > θ, it is determined that the customer has a high complaint risk, and the corresponding intention feature vector I is generated; otherwise, it is determined as a low complaint risk, and the intention feature vector I remains unchanged.

[0052] The specific steps in S5 include the following:

[0053] S5.1, construct the emotion-intention state vector S based on the emotional feature vector E and the intention feature vector I:

[0054]

[0055] Among them, Concat(x, y) represents the operation of splicing and combining two vectors into a whole. W i is the weight matrix of the intention feature, which maps the intention feature vector I to the hidden space. b e and b i are the bias vectors of emotion and intention. W ei is the weight matrix of the interaction term, indicating the interaction influence between the emotional feature and the intention feature. Denotes the element-wise product of the emotion feature vector E and the intention feature vector I, reflecting the specific interaction relationship between emotion and intention, and ε is a random noise term;

[0056] S5.2, Input the emotion-intention state vector S into the asynchronous advantage actor-critic algorithm model, and calculate the state value function V(S) and the advantage function A(S,a) of the customer's complaint intention:

[0057] V(S) = f(W v ·tanh(W vs ·S + b vs ) + b v )

[0058] A(S,a) = g(W a ·ReLU(W as ·S + b as ) - V(S))

[0059] Among them, f and g are activation functions respectively, used for the non-linear transformation of the state value and the advantage function, W v and W vs are the weight matrix and the mapping matrix of the state value respectively, b v and b vs are the corresponding bias terms, W a and W as are the weight matrix and the mapping matrix of the advantage function respectively, b as is the advantage bias term, and tanh and ReLU are used for non-linear transformation respectively;

[0060] S5.3, Based on the advantage function A(S,a) and the state value function V(S), calculate the customer's complaint intention probability value P adv :

[0061]

[0062] Among them, λ is a balance coefficient, used to adjust the weights of the state value and the sum of advantages, represents the advantage values of all potential response strategies, is an indicator function, which takes 1 when the policy a is equal to the specific policy a k and 0 otherwise, and P adv is the final probability value that the customer has a complaint intention.

[0063] The specific steps in S6 include the following:

[0064] S6.1, Receive the customer complaint intention probability value P adv output by the asynchronous advantage actor-critic algorithm model, and compare it with the preset complaint intention threshold θ, where θ represents the system-predefined complaint risk judgment standard;

[0065] S6.2. When the probability value P of the complaint intention adv satisfies P adv > θ, the system marks this customer as a high-complaint-risk customer and generates a corresponding risk warning message R warn for subsequent shunting and response;

[0066] S6.3. If the probability value P of the complaint intention adv ≤ θ, the customer is marked as a low-complaint-risk, and no risk warning message R warn is generated, and resources are preferentially allocated to high-risk customers.

[0067] The specific steps in S7 include the following:

[0068] S7.1. Input the risk warning message R warn into the intelligent shunting module and parse the risk warning message to obtain the parsed customer complaint risk level L risk :

[0069] L risk = η·Log(1 + α r R warn + β c C hist ) + ε 0

[0070] where η is the risk level amplification coefficient, α r and β c are the weighting coefficients, corresponding to the risk warning value R warn and the customer's historical complaint frequency C hist , ε 0 is a constant offset used to balance the baseline values of different risk levels;

[0071] S7.2. Based on the parsed risk level L risk calculate the customer service priority P priority through the priority allocation mechanism:

[0072]

[0073] where γ is the risk square weighting coefficient, δ is the warning weighting coefficient, λ 1 is the historical weight coefficient, θ 1 is the priority offset;

[0074] S7.3. Compare the calculated priority P priority with the system-predefined priority threshold τ, and judge whether the customer is a high priority through the following conditions:

[0075] I priority= Heaviside(P priority -τ)

[0076] wherein, I priority is the priority determination indicator. When P priority ≥ τ, I priority = 1, indicating high priority; otherwise I priority = 0;

[0077] S7.4. When I priority = 1, generate a diversion instruction D assign , the content includes the calculation of the matching degree between the customer and the dedicated customer service and the response time window, and send the diversion instruction D assign to the customer service system for the priority processing of customers with high complaint risks.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] Improve the accuracy of emotion and intention recognition: By constructing emotion feature vectors and intention feature vectors, and adopting a multi-modal fusion mechanism to integrate the customer's emotion and intention features in the form of an emotion-intention state vector into the large model, the present invention can deeply excavate the customer's emotional information from the emotional level, and can also understand the customer's potential intention at the semantic level, effectively solving the problem of the disconnection between emotion analysis and intention recognition in the prior art. Through this improvement, the model has higher understanding ability when dealing with complex customer emotions, and can significantly improve the accuracy of complaint intention recognition.

[0080] Improve the processing efficiency and response speed: By introducing the asynchronous advantage actor-critic algorithm, using the multi-threaded parallel processing characteristics of the asynchronous advantage actor-critic algorithm to asynchronously process the emotion-intention state vector, the asynchronous advantage actor-critic algorithm updates the policy and value function in parallel through independent threads, enabling the system to maintain efficient and stable processing capabilities when dealing with high-concurrency and large-scale customer data. Compared with traditional single-threaded or synchronous algorithms, the multi-threaded algorithm of the asynchronous advantage actor-critic algorithm can dynamically optimize the complaint intention recognition process, achieve faster emotion and intention judgment, and effectively improve the efficiency and response speed of the system in dealing with real-time customer service scenarios.

[0081] Optimize resource allocation and risk management: After identifying that the customer has a high complaint risk, the intelligent diversion module assigns priorities to the customer's complaint intentions, calculates the service priority of the customer using factors such as priority weights and risk levels to ensure the priority processing of customers with high complaint risks. It also combines the customer's historical complaint data and current emotional fluctuations to generate risk warning information, and assigns the customer to a dedicated customer service, which can effectively reduce the risk of customer churn and complaint escalation, ensure the reasonable allocation of customer service resources, and further improve the intelligence level of complaint management. Brief Description of the Drawings

[0082] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0083] Figure 1 It is a flowchart of a method for identifying complaint intentions of a large model based on sentiment analysis of the present invention. Detailed Embodiments

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0085] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present invention to be protected, but only represents the selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0086] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0087] The following will further describe the present invention in detail with reference to the drawings:

[0088] See Figure 1 , which is a method for identifying complaint intentions of a large model based on sentiment analysis of the present invention, characterized by including the following steps:

[0089] S1. Obtain the voice data of the customer call and convert the voice data into text data through a voice-to-text module.

[0090] S1.1, Receive the real-time voice data stream X of the customer call t , where t represents the time step. The voice data stream is collected through the customer's voice input device and is input into the voice preprocessing module of the system in real time through the transmission network;

[0091] S1.2, In the voice preprocessing module, perform operations on the real-time voice data stream Xt Perform noise reduction processing, use a filter to eliminate background noise and high-frequency components, and generate a noise-reduced speech signal X t ′;

[0092] S1.3, Send the noise-reduced speech signal X t ′ into the speech framing processing module, and segment the speech signal X t ′ according to the time window ΔT to generate a speech frame sequence X t ′(i), where i represents the i-th speech segment after framing;

[0093] S1.4, Use the Mel Frequency Cepstral Coefficient (MFCC) analysis method to extract features from each speech frame sequence X t ′(i), and extract the acoustic features M i , to obtain a feature vector M = {M i} containing speech spectrum information, where each M i represents the speech feature in the i-th frame;

[0094] S1.5, Input the feature vector M into the speech-to-text module, and process the feature vector through a pre-trained speech recognition model to generate the corresponding text data T, where T = {T i} represents the text segment corresponding to each speech frame;

[0095] S1.6, Merge all text segments T i to form the complete text data T total .

[0096] S2. Perform text preprocessing on the text data to remove noise and mark semantic components, and generate text data with clear semantics.

[0097] S2.1, Receive the text data T generated by the speech-to-text module total , perform denoising processing on the text data T total , and remove redundant non-verbal characters and low-confidence words through a noise filtering algorithm to generate denoised text data T total ′;

[0098] S2.2, Perform semantic marking on the denoised text data T total ′, and mark the semantic components in the text data using predefined semantic component tags to generate marked text data T total ″;

[0099] S2.3, Use a word segmentation algorithm to perform word segmentation on the marked text data T total ″ to generate a word segmentation result sequence W;

[0100] S2.4. Perform stop word filtering based on the word segmentation result sequence W, remove common meaningless words, and obtain a word segmentation sequence W' with clear semantics.

[0101] S2.5. Perform part-of-speech tagging on the word segmentation sequence W' with clear semantics, and tag each word segmentation unit w i ' according to its grammatical role and semantic attributes, and obtain text data T with clear semantics and part-of-speech information. clean .

[0102] In Chinese text processing, common meaningless words (i.e., stop words) include those functional words that appear frequently in a sentence but have little impact on the main semantics of the sentence. For example:

[0103] Auxiliary words, prepositions and structural words: de, di, de, le, zhe, guo

[0104] Conjunctions and function words: he, huo, dan, suoyi, ruguo, yinwei, suiran

[0105] Pronouns: wo, ni, ta, ta, ta, women, tamen

[0106] Adverbs and modal particles: jiu, ye, hai, dou, you, ne, a, ba

[0107] Demonstrative words and interrogative words: zhege, nage, shenme, nali

[0108] In actual business, the stop word list is adjusted and extended according to specific corpora and requirements.

[0109] S3. Perform sentiment feature analysis on the preprocessed text data based on the sentiment analysis model, extract the sentiment feature values in the text data, the sentiment feature values include the intensity values of negative emotions such as anger, dissatisfaction and anxiety, and generate an emotion vector containing sentiment features, and the emotion vector is used to represent the emotion state of the customer during the call.

[0110] S3.1. Receive the preprocessed text data T clean , perform sentiment classification on the text data T based on the sentiment analysis model clean , identify the emotion categories in the text, and the emotion categories include negative emotions such as anger, dissatisfaction and anxiety;

[0111] S3.2. Perform intensity analysis on the emotion categories, calculate the intensity value E of each emotion i , where i represents the emotion category, including anger, dissatisfaction and anxiety, and the intensity value of each emotion is generated according to the frequency of occurrence of sentiment words and context weights; the specific expression is:

[0112]

[0113] Among them, E iRepresents the intensity value of emotion category i; N i is the total number of sentiment words belonging to emotion i in the text; f ij Represents the occurrence frequency or degree of occurrence of the j-th sentiment word in the text; w ij is the context weight corresponding to this sentiment word, which is calculated based on the position, dependency relationship, and context information of the sentence where the sentiment word is located, and is used to reflect its influence in the current context; b i is the bias term used to adjust the overall effect; σ is the activation function used for non-linear mapping of the output.

[0114] S3.3, Generate emotion feature vectors:

[0115] E = {E ang , E dis , E anx}

[0116] where E ang , E dis , E anx respectively represent the intensity values of anger, dissatisfaction, and anxiety emotions, and are used to characterize the customer's emotional state during the call.

[0117] S4. Use the emotion vector as the input data for preliminary judgment of the complaint risk, and input it together with the text data into the large model. The large model combines the emotional eigenvalue in the emotion vector and the context semantic information in the text data to perform in-depth semantic understanding of the customer's potential complaint intention and generate an intention feature vector.

[0118] S4.1, Receive the emotion feature vector E generated by sentiment analysis and the preprocessed text data T clean , and use the emotion feature vector E as the input data for preliminary judgment of the complaint risk;

[0119] S4.2, Input the emotion feature vector E and the text data T clean together into the large model, and perform in-depth semantic understanding of the customer's potential complaint intention through a multi-modal fusion mechanism to generate an intention feature vector I:

[0120] I = σ(W e E + W t Φ(T clean ) + b)

[0121] where σ is the activation function, W e is the emotion feature weight matrix, mapping the emotion feature vector E to the hidden space to capture the influence of emotion information on the complaint intention, W t is the text semantic feature weight matrix, mapping the semantic feature vector Φ(T clean) into the latent space to capture the impact of text content on the complaint intention, Φ(T clean ) is a text semantic feature extraction function that uses a pre-trained word vector model to convert text data b into a bias vector;

[0122] S4.3. Calculate the complaint intention probability value P comp , and use the softmax function to normalize the intention feature vector I:

[0123]

[0124] where I comp is the eigenvalue of the complaint intention, representing the degree to which the customer has a complaint intention, and I noncomp is the eigenvalue of the non-complaint intention, representing the degree to which the customer has no complaint intention. P comp is the probability value that the customer has a complaint intention, ranging from [0, 1];

[0125] S4.4. Compare the complaint intention probability value P comp with the preset complaint intention threshold θ. When P comp > θ, it is determined that the customer has a high complaint risk, and the corresponding intention feature vector I is generated; otherwise, it is determined as a low complaint risk, and the intention feature vector I remains unchanged.

[0126] S5. Based on the emotion vector and the intention feature vector, construct an emotion-intention state vector, and input the emotion-intention state vector into the asynchronous advantage actor-critic algorithm model. The asynchronous advantage actor-critic algorithm model performs asynchronous processing on the emotion-intention state vector in a multi-threaded environment to determine whether the customer has a complaint intention and generate the corresponding complaint intention probability value.

[0127] S5.1. Based on the emotion feature vector E and the intention feature vector I, construct an emotion-intention state vector S:

[0128]

[0129] where Concat(x, y) represents the operation of concatenating two vectors into a whole. W i is the weight matrix of the intention feature, mapping the intention feature vector I to the latent space, b e and b i are the bias vectors of emotion and intention, and W ei is the weight matrix of the interaction term, representing the interaction effect between the emotion feature and the intention feature, represents the element-wise product of the emotion feature vector E and the intention feature vector I, reflecting the specific interaction relationship between emotion and intention, and ε is a random noise term;

[0130] S5.2, input the emotion-intention state vector S into the asynchronous advantage actor-critic algorithm model, and calculate the state value function V(S) and the advantage function A(S,a) of the customer's complaint intention:

[0131] V(S) = f(W v ·tanh(W vs ·S + b vs ) + b v )

[0132] A(S,a) = g(W a ·ReLU(W as ·S + b as ) - V(S))

[0133] where f and g are activation functions respectively used for the non-linear transformation of the state value and the advantage function, W v and W vs are the weight matrix and the mapping matrix of the state value respectively, b v and b vs are the corresponding bias terms, W a and W as are the weight matrix and the mapping matrix of the advantage function respectively, b as is the advantage bias term, and tanh and ReLU are used for non-linear transformation respectively;

[0134] S5.3, based on the advantage function A(S,a) and the state value function V(S), calculate the complaint intention probability value P adv :

[0135]

[0136] where λ is a balance coefficient used to adjust the weights of the state value and the total advantage, represents the advantage values of all potential response strategies, is an indicator function that takes 1 when the policy a is equal to a specific policy a k and 0 otherwise, and P adv is the final probability value that the customer has a complaint intention.

[0137] S6. In the asynchronous advantage actor-critic algorithm model, judge the complaint intention probability value based on a preset complaint intention threshold. When the complaint intention probability value is greater than the complaint intention threshold, mark this customer as a high complaint risk and generate a risk warning message for subsequent diversion and response.

[0138] S6.1, receive the customer complaint intention probability value P adv output by the asynchronous advantage actor-critic algorithm model, and compare it with the preset complaint intention threshold θ, where θ represents the predefined complaint risk judgment criterion of the system;

[0139] S6.2, When the probability value P of the complaint intention adv satisfies P adv > θ, the system marks the customer as a high-complaint-risk customer and generates a corresponding risk warning message R warn for subsequent diversion and response;

[0140] S6.3, If the probability value P of the complaint intention adv ≤ θ, the customer is marked as a low-complaint-risk, and no risk warning message R warn is generated, and the resources are preferentially allocated to high-risk customers.

[0141] S7. Input the risk warning message into the intelligent diversion module. After receiving the risk warning message, the intelligent diversion module preferentially allocates customers with high complaint risks to dedicated customer service for processing.

[0142] S7.1, Input the risk warning message R warn into the intelligent diversion module and parse the risk warning message to obtain the parsed customer complaint risk level L risk :

[0143] L risk = η·Log(1 + α r R warn + β c C hist ) + ε 0

[0144] where η is the risk level amplification coefficient, α r and β c are weighting coefficients, corresponding to the risk warning value R warn and the customer's historical complaint frequency C hist , ε 0 is a constant offset used to balance the reference values of different risk levels;

[0145] S7.2, Based on the parsed risk level L risk calculate the customer service priority P priority through the priority allocation mechanism:

[0146]

[0147] where γ is the risk square weighting coefficient, δ is the warning weighting coefficient, λ 1 is the historical weight coefficient, and θ 1 is the priority offset;

[0148] S7.3, Input the calculated priority P priorityCompare with the system - preset priority threshold τ, and judge whether the customer is of high priority through the following conditions:

[0149] I priority = Heaviside(P priority -τ)

[0150] where I priority is the priority determination indicator. When P priority ≥τ, I priority = 1, indicating high priority; otherwise I priority = 0;

[0151] S7.4, when I priority = 1, generate a shunt instruction D assign , the content includes the calculation of the matching degree between the customer and the dedicated customer service and the response time window, and send the shunt instruction D assign to the customer service system for the priority processing of customers with high complaint risks.

[0152] S8. After the customer is assigned, the intelligent shunt module generates real - time service guidelines and coping strategies according to the risk warning information.

[0153] The present invention realizes key improvements in many aspects in the customer complaint intention recognition and high - risk customer processing flow, effectively improving the deep - fusion analysis ability of emotional and semantic features, the accuracy and real - time performance of complaint intention recognition, and the rationality and efficiency of high - complaint - risk customer resource allocation. By combining the pre - processing of voice data, text semantic feature extraction, emotional intensity quantification analysis, multi - modal fusion algorithm, and efficient risk assessment based on the asynchronous advantage actor - critic model, the present invention significantly improves the early warning and response ability of customer complaint risks. The technical solution is applicable to the high - complaint - risk prediction and shunt processing in complex scenarios of the customer service system, such as online customer service systems, call center management systems, and other intelligent service scenarios that need to judge the customer's emotions and intentions in real time, and can effectively meet the complaint - risk control requirements in large - scale customer services.

[0154] Through the organic combination of modular design, multi - modal fusion, and reinforcement learning framework, the present invention takes into account the efficiency, accuracy, and scalability of complaint intention recognition, and significantly improves the applicability and effect in practical applications.

[0155] Another embodiment of the present invention adopts a complaint intention prediction algorithm based on a graph neural network (GNN): GNN realizes the comprehensive judgment of complaint intention by constructing a relationship graph of customer emotional features and semantic features.

[0156] Another embodiment of the present invention adopts an end-to-end sequence-to-sequence model: This solution directly inputs speech into the Transformer model, captures emotion and intention information through the self-attention mechanism, and omits the intermediate module.

[0157] Another embodiment of the present invention adopts a dynamic weight allocation model based on reinforcement learning: The reinforcement learning framework can optimize the weight allocation of emotion and semantic features through historical data.

[0158] Another embodiment of the present invention adopts a distributed emotion-intention evaluation architecture: This solution reduces the load on the central node by dispersing the emotion feature evaluation task on edge devices.

[0159] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A complaint intention recognition method based on a large model of sentiment analysis, characterized in that: The following steps are involved: S1. Obtain the voice data of the customer's call and convert the voice data into text data through the voice-to-text module; S2. Perform text preprocessing on the text data to remove noise and mark semantic components to generate semantically clear text data; S3. Perform sentiment feature analysis on the preprocessed text data based on the sentiment analysis model, extract sentiment feature values ​​from the text data, the sentiment feature values ​​include intensity values ​​of negative emotions such as anger, dissatisfaction and anxiety, and generate sentiment vectors containing sentiment features, the sentiment vectors are used to characterize the emotional state of the customer during the call; S4. The sentiment vector is used as input data for preliminary judgment of complaint risk, and is input into the big model together with the text data. The big model combines the emotional feature value in the sentiment vector and the contextual semantic information in the text data to conduct a deep semantic understanding of the customer's potential complaint intention and generate an intention feature vector. S5. construct an emotion-intention state vector based on the emotion vector and the intention feature vector, and input the emotion-intention state vector into the asynchronous advantage actor critic algorithm model. The asynchronous advantage actor critic algorithm model asynchronously processes the emotion-intention state vector in a multi-threaded environment to determine whether the customer has a complaint intention and generate a corresponding complaint intention probability value; S6. In the asynchronous advantage actor-critic algorithm model, the complaint intention probability value is judged based on the preset complaint intention threshold. When the complaint intention probability value is greater than the complaint intention threshold, the customer is marked as having a high complaint risk, and risk warning information is generated for subsequent diversion and response; S7. Input the risk warning information into the intelligent diversion module. After receiving the risk warning information, the intelligent diversion module will prioritize the customers with high complaint risks to the dedicated customer service for processing; S8. After customer allocation, the intelligent diversion module generates real-time service guidance and response strategies based on risk warning information.

2. A method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S1 specifically includes the following steps: S1.1, receiving real-time voice data stream X of customer calls t , where t represents the time step, the voice data stream is collected through the customer's voice input device and input into the system's voice preprocessing module in real time through the transmission network; S1.2, in the voice preprocessing module, the real-time voice data stream X t Perform noise reduction processing, use filters to eliminate background noise and high-frequency components, and generate a noise-reduced speech signal X t ′; S1.3, the noise-reduced speech signal X t ′ is sent to the speech frame processing module, and the speech signal X is processed according to the time window ΔT. t ' is segmented to generate a speech frame sequence X t ′(i), where i represents the speech segment after the i-th frame; S1.4, using the Mel frequency cepstral coefficient analysis method to analyze each speech frame sequence X t ′(i) Perform feature extraction to extract acoustic features M i , and obtain the feature vector M containing speech spectrum information = {M i }, each M i Represents the speech features in the i-th frame; S1.5, input the feature vector M into the speech-to-text module, and process the feature vector through the pre-trained speech recognition model to generate corresponding text data T, where T = {T i } represents the text segment corresponding to each speech frame; S1.6, merge all text segments T i Form a complete text data T total .

3. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S3 specifically includes the following steps: S2.1, receiving text data T generated by the speech-to-text module total , for text data T total Perform denoising processing, remove redundant non-language characters and low-confidence words through noise filtering algorithm, and generate denoised text data T total ′; S2.2, in the denoised text data T total ′, and use predefined semantic component labels to label the semantic components in the text data, generating the labeled text data T total ″; S2.3, use the word segmentation algorithm to analyze the marked text data T total ″Perform word segmentation processing to generate a word segmentation result sequence W; S2.4, based on the word segmentation result sequence W, stop word filtering is performed to remove common meaningless words to obtain a semantically clear word segmentation sequence W'; S2.5, perform part-of-speech tagging on the semantically clear word segmentation sequence W', and mark each word segmentation unit w i 'Annotate according to its grammatical role and semantic attributes to obtain text data with clear semantics and part of speech information clean .

4. A method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 3, characterized in that: The common meaningless words are functional words that appear frequently in sentences but have little impact on the semantics of the sentences, including auxiliary words, prepositions and structural words, conjunctions and function words, pronouns, adverbs and modal particles, demonstrative words and question words.

5. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S3 specifically includes the following steps: S3.1, receiving preprocessed text data T clean , based on sentiment analysis model for text data T clean Perform sentiment classification to identify the emotional categories in the text, including anger, dissatisfaction, and anxiety; S3.2, perform intensity analysis on emotion categories and calculate the intensity value E of each emotion i , where i represents the emotion category, including anger, dissatisfaction and anxiety, and the intensity value of each emotion is generated according to the frequency of occurrence of emotional words and context weight; S3.3, generate emotion feature vector: And={And ang ,AND dis ,AND anx } Among them, E ang 、E dis 、E anx They represent the intensity values ​​of anger, dissatisfaction, and anxiety, respectively, and are used to characterize the customer's emotional state during the call.

6. A method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 5, characterized in that: The emotion intensity value is obtained by weighted summing the occurrence frequency of the emotion words and the context weight, which can be specifically expressed as: Among them, E i Indicates the intensity value of emotion category i; N i is the total number of sentiment words belonging to sentiment i in the text; f ij represents the frequency or degree of occurrence of the jth sentiment word in the text; w ij is the context weight corresponding to the sentiment word, which is calculated based on the position, dependency, and context information of the sentence where the sentiment word is located, and is used to reflect its influence in the current context; b i is a bias term used to adjust the overall effect; σ is an activation function used to perform nonlinear mapping on the output.

7. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S4 specifically includes the following steps: S4.1, receiving the emotion feature vector E generated by sentiment analysis and the preprocessed text data T clean , the emotional feature vector E is used as the input data for preliminary judgment of complaint risk; S4.2, the emotion feature vector E and the text data T clean They are input into the big model together, and a deep semantic understanding of the customer's potential complaint intention is performed through the multimodal fusion mechanism to generate the intention feature vector I: I=σ(W e E+W t Φ(T clean )+b) Among them, σ is the activation function, W e is the emotional feature weight matrix, mapping the emotional feature vector E to the implicit space, capturing the impact of emotional information on complaint intention, W t is the text semantic feature weight matrix, mapping the semantic feature vector Φ(T clean ) to the implicit space, capturing the impact of text content on complaint intention, Φ(T clean ) is a text semantic feature extraction function, which uses a pre-trained word vector model to convert text data b into a bias vector; S4.3, calculate the customer's complaint intention probability value P comp , the intention feature vector I is normalized using the soft maximization function: Among them, I comp is the characteristic value of complaint intention, indicating the degree to which customers have complaint intention. noncomp is the characteristic value of non-complaint intention, indicating the degree to which customers have no intention to complain, P comp is the probability value of the customer's intention to complain, ranging from [0,1]; S4.4, the complaint intention probability value P comp Compared with the preset complaint intention threshold θ, when P comp >θ, the customer is judged to have a high complaint risk and the corresponding intention feature vector I is generated; otherwise, it is judged to have a low complaint risk and the intention feature vector I remains unchanged.

8. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S5 specifically includes the following steps: S5.1, construct the emotion-intention state vector S based on the emotion feature vector E and the intention feature vector I: Concat(x,y) represents the operation of concatenating two vectors into a whole. i is the weight matrix of the intent feature, mapping the intent feature vector I to the implicit space, b e and b i is the bias vector of emotion and intention, W ei is the weight matrix of the interaction term, which represents the interaction between emotion features and intention features. EoI represents the element-by-element product of the emotion feature vector E and the intention feature vector I, which reflects the specific interaction relationship between emotion and intention. ε is the random noise term. S5.2, input the emotion-intention state vector S into the asynchronous advantage actor-critic algorithm model to calculate the state value function V(S) and advantage function A(S,a) of the customer's complaint intention: V(S)=f(W v ·tanh(W vs ·S+b vs )+b v ) A(S,a)g(W a ·ReLU(W as ·S+b as )-V(S)) Among them, f and g are activation functions, which are used for nonlinear transformation of state value and advantage function, respectively. v and W vs are the weight matrix and mapping matrix of the state value, respectively, v and b vs is the corresponding bias term, W a and W as is the weight matrix and mapping matrix of the advantage function, b as is the dominant bias term, tanh and ReLU are used for nonlinear transformation respectively; S5.3, based on the advantage function A(S,a) and the state value function V(S), calculate the customer's complaint intention probability value P adv : Among them, λ is the balance coefficient, which is used to adjust the weight of the state value and the sum of advantages. represents the advantage value of all potential response strategies, is the indicator function, when strategy a is equal to specific strategy a k 1 when P is 0 otherwise. adv is the final probability value of the customer having the intention to complain.

9. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S6 specifically includes the following steps: S6.1, receiving the customer complaint intention probability value P output by the asynchronous advantage actor-critic algorithm model adv , and compare it with the preset complaint intention threshold θ, where θ represents the complaint risk judgment standard predefined by the system; S6.2, when the complaint intention probability value P adv Satisfy P adv >θ, the system marks the customer as a high complaint risk customer and generates corresponding risk warning information R warn For subsequent diversion and response; S6.3, if the probability value of complaint intention is P adv ≤θ, the customer is marked as having low complaint risk and no risk warning information R is generated warn , prioritizing resources to high-risk customers.

10. The method for identifying complaint intentions based on a large model of sentiment analysis as claimed in claim 1, characterized in that: The S7 specifically includes the following steps: S7.1, Risk warning information R warn Input to the intelligent diversion module, and analyze the risk warning information to obtain the analyzed customer complaint risk level L risk : L risk =η·Log(1+α r R warn +b c C hist )+ε0 Among them, η is the risk level magnification factor, α r and β c are weighted coefficients, corresponding to the risk warning values ​​R warn and customer historical complaint frequency C hist , ε0 is a constant offset used to balance the benchmark values ​​of different risk levels; S7.2, based on the analyzed risk level L risk Calculate customer service priority P through priority allocation mechanism priority : Among them, γ is the risk square weighting coefficient, δ is the warning weighting coefficient, λ1 is the historical weighting coefficient, and θ1 is the priority offset; S7.3, the calculated priority P priority Compare with the system preset priority threshold τ and determine whether the customer is high priority based on the following conditions: I priority =Heaviside(P priority -t) Among them, I priority is the priority indicator, when P priority ≥τ, I priority =1, indicating high priority; otherwise I priority =0; S7.4, when I priority =1, generate diversion instruction D assign , including the calculation of the matching degree between the customer and the full-time customer service and the response time window, and the diversion instruction D assign Sent to the customer service system, giving priority to customers with high complaint risks.

Images