CRM sales verbal skill intelligent analysis method based on natural language processing

Abstract

The invention discloses a CRM sales talk skill intelligent analysis method based on natural language processing. According to the method, a non-Euclidean emotional space is constructed, emotional states of customers are mapped into points in a high-dimensional space by adopting hyperspherical embedding, potential features of customer emotions are further extracted through a graph neural network, and a personalized sales verbal skill is generated in combination with a multi-layer attention mechanism and a condition decoding structure. The method can dynamically recognize the emotional change of the customer, optimizes the verbal skill generation in real time according to the emotional state of the customer, and improves the adaptability and accuracy of the sales verbal skill. The pruning strategy is optimized through deep reinforcement learning, the calculation efficiency can be effectively improved, redundant data processing is reduced, and the system response speed is increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing and customer relationship management, and particularly relates to a CRM sales script intelligent analysis method based on natural language processing. BACKGROUND

[0002] With the rapid development of information technology, customer relationship management systems have been widely used in various industries, becoming an important tool for optimizing sales processes and customer management. Especially in the field of sales and customer service, CRM systems help enterprises improve the efficiency of customer relationship maintenance and the success rate of sales. With the increasing competition in the market, sales teams urgently need more intelligent and efficient tools to improve sales conversion rates and customer satisfaction. Therefore, intelligent sales script generation and optimization have become a key factor in improving customer interaction quality and sales performance.

[0003] Traditional sales script generation methods mainly rely on manual design and rule matching, and develop scripts through simple analysis of customer needs and problems. However, these methods have many problems. First, the traditional script generation method is difficult to accurately capture and adapt to the emotional changes of customers, and the generated script is often fixed and single, which cannot match the emotional needs of customers. Second, existing methods are difficult to adjust according to real-time emotional fluctuations when facing complex customer emotional changes, so they cannot efficiently generate sales scripts that meet customer needs in a dynamic environment. In addition, traditional methods ignore the multi-dimensional characteristics of customer emotional states, and cannot effectively integrate local features and global features, resulting in the lack of personalization and precision of the generated scripts.

[0004] In the field of natural language processing, in recent years, many intelligent script generation methods based on machine learning and deep learning have emerged. These methods attempt to provide personalized and customized sales scripts for customers by analyzing customer historical conversations and emotional data. However, existing NLP-based sales script generation methods still face the following problems. First, although these methods can analyze customer emotional data to some extent, they often fail to fully consider the complexity of emotional states and lack real-time response capabilities to emotional fluctuations, resulting in generated scripts that cannot effectively adapt to customer emotional changes. Second, existing technologies often ignore the multi-level relationship of emotions when analyzing emotions, and cannot fully understand customer emotional needs, thereby affecting the accuracy and effectiveness of the script.

[0005] Existing sales script generation technologies have great limitations in dealing with complex emotional changes and large-scale customer data, and cannot provide accurate and personalized sales scripts in a dynamic environment. In order to effectively improve the quality and adaptability of sales scripts, a new type of sales script generation method is needed, which can combine customer emotional states and conversation context to generate more accurate, flexible and customer demand-oriented scripts.

[0006] The present application is based on the above problems, combined with graph neural network and deep reinforcement learning technology, and proposes a customer sentiment state-based sales script intelligent analysis method. The method deeply analyzes the customer's emotional state, captures the emotional changes in real time, and generates more accurate and efficient sales scripts by combining the customer's personalized needs, thereby improving the customer interaction quality and sales conversion rate. SUMMARY

[0007] One object of the present application is to propose a CRM sales script intelligent analysis method based on natural language processing. The method constructs a non-Euclidean emotional space with adjustable curvature, introduces knowledge-driven Gaussian decay contrast learning to generate sentence embedding, establishes an emotional state graph based on emotional features and semantic similarity, and obtains node embedding matrix and graph-level aggregation vector through graph neural network. A multi-layer attention mechanism and a conditional decoding structure are used to fuse local features and global features, and a sales script with emotional control conditions is generated. This technical route is aimed at multi-round interaction scenarios, and gives a unified solution to the problems that traditional rule templates and single sequence models are difficult to represent emotional evolution, difficult to integrate domain knowledge, and difficult to realize dynamic regulation.

[0008] According to the CRM sales script intelligent analysis method based on natural language processing, the method comprises the following steps: Step one: collect the interaction data of the customer in the CRM system and standardize it to obtain the standardized interaction data, calculate the emotional feature vector at different time nodes according to the standardized interaction data, and construct a non-Euclidean emotional space; Step two: according to the non-Euclidean emotional space and the emotional feature vector, the emotional state of the customer is mapped to a point in a high-dimensional space through hyperspherical embedding, and the emotional state of the customer is obtained; Step three: based on the knowledge-driven unsupervised learning method, combined with Gaussian decay contrast learning, the sentences in the customer conversation are embedded to generate; through contrast learning to optimize the distance constraint of the embedding space, combined with the external knowledge base, knowledge information is introduced in the embedding process to generate sentence embedding vectors; Step four: construct a graph neural network model, generate an emotional state graph according to the distance relationship and similarity measurement in the emotional space, perform multi-layer graph convolution and readout operation on the emotional state graph, obtain node embedding matrix and graph-level aggregation vector, and the graph-level aggregation vector is used as the potential representation of the evolution of the customer's emotional state; Step five: based on the node embedding matrix and the graph-level aggregation vector, the local features and global features of the customer's emotional state are fused through the multi-layer attention mechanism and the conditional decoding structure to generate a sales script model, and the sales script with emotional control conditions is output; Step six: using a data set containing customer historical dialogue, emotion state labeling, and sales result feedback, the sales dialogue model is supervised trained and the model parameters are iteratively optimized through supervised learning and multi-objective optimization method, and the final sales dialogue model is obtained.

[0009] Step one: collect and standardize the interaction data of customers in the CRM system to obtain standardized interaction data, calculate the emotion feature vector at different time nodes according to the standardized interaction data, and construct a non-Euclidean emotion space; Step two: according to the non-Euclidean emotion space and the emotion feature vector, the emotion state of the customer is mapped to a point in the high-dimensional space through hyperspherical embedding, and the emotion state of the customer is obtained; Step three: based on the knowledge-driven unsupervised learning method, the sentence in the customer dialogue is embedded and generated by combining Gaussian decay contrast learning; the distance constraint of the embedding space is optimized by contrast learning, and knowledge information is introduced in the embedding process by combining an external knowledge base to generate a sentence embedding vector; Step four: construct a graph neural network model, generate an emotion state graph according to the distance relationship and similarity measurement in the emotion space, perform multi-layer graph convolution and readout operation on the emotion state graph, obtain a node embedding matrix and a graph-level aggregation vector, and the graph-level aggregation vector is used as a latent representation of the evolution of the customer's emotion state; Step five: based on the node embedding matrix and the graph-level aggregation vector, the local features and global features of the customer's emotion state are fused through a multi-layer attention mechanism and a conditional decoding structure to generate a sales dialogue model and output a sales dialogue with emotion control conditions; Step six: using a data set containing customer historical dialogue, emotion state labeling, and sales result feedback, the sales dialogue model is supervised trained and the model parameters are iteratively optimized through supervised learning and multi-objective optimization method, and the final sales dialogue model is obtained.

[0010] Optionally, the step one specifically includes: Collecting the interaction data of customers in the CRM system, including the historical emotion data, dialogue content, customer feedback information and emotion label of the customers, and performing standardization operation on the interaction data to obtain standardized interaction data; According to the standardized interaction data, the emotion feature vector of the customer at different time nodes is extracted, and the emotion feature vector includes emotion type, emotion intensity and emotion change trend; According to the distribution relationship of the emotion feature vector, the dimension number and curvature parameter of the emotion space are determined, and a non-Euclidean emotion space is constructed.

[0011] Optionally, the step two specifically includes: Based on the non-Euclidean emotional space and the emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space using the hyperspherical embedding algorithm; During the mapping process, the distribution of sentiment features in the non-Euclidean sentiment space is optimized by using hyperspherical geometric constraints, so that the distance between points with similar sentiment features is closer and the distance between points with different sentiment features is farther. The mapping result is the coordinates of the customer's emotional characteristics at different time points in the non-Euclidean emotional space. The coordinates represent the customer's emotional type, emotional intensity, emotional change trend, and emotional change pattern at different time points. The mapping allows us to obtain the customer's emotional state in a non-Euclidean emotional space.

[0012] Optionally, the optimization constraints for the hyperspherical embedding include: In a high-dimensional hyperspherical space, the angular distance between sentiment feature vectors is defined as a similarity measure; Construct a weighted angular distance loss function ; in, To optimize the overall objective function, we define the degree of aggregation and separation of sentiment features in the hyperspherical space. For any two sentiment feature vectors and The angle between them The weighting coefficients, determined by both temporal difference and semantic similarity, are used to control the contribution of different features to the loss. By minimizing the weighted angular distance loss function This allows points with similar emotional characteristics to aggregate on a hypersphere, while significantly different emotional characteristics remain separated in space, resulting in a structurally stable and highly discriminative distribution of customer emotional states.

[0013] Optionally, step three specifically includes: Based on a knowledge-driven unsupervised learning method, combined with a Gaussian decay contrastive learning model, each sentence in the customer dialogue is embedded and generated. Preprocess sentences in customer conversations to extract semantic features and sentiment tags, and enhance sentiment semantics by combining them with a sentiment dictionary from an external knowledge base; By employing Gaussian decay contrastive learning, the distance constraints in the sentence embedding space are optimized, ensuring that semantically similar sentences are placed closer together in the vector space, while sentences with significant semantic differences are placed further apart. Knowledge-driven and Gaussian decay contrastive learning are combined through a joint loss function. ; in, For scalar weights, Gaussian attenuation contrast loss, For knowledge constraint loss, For emotional alignment loss, For time-series smoothing loss; Sales domain knowledge information is incorporated into the sentence embedding process to generate sentence embedding vectors.

[0014] Optionally, the optimization objective of the Gaussian decay contrastive learning model includes: ; in, To optimize the overall objective loss function, They represent the first The and the first The embedding vector of each sentence. Representation and Sample A set of semantically similar positive samples Indicates and sentences Any other sentence embedding vector to be compared , To balance the temperature parameter representing the distance between similar and dissimilar samples in the embedding space, For use in sentences With all other sentences Gaussian decay weights in the relationship between them Sentence and sentences Cosine similarity between them Sentence and sentences Cosine similarity between them For use in sentences With all other sentences The relationship between Gaussian decay weights ; in, Sentence and sentences Time difference in the dialogue sequence The standard deviation of time decay; By minimizing the overall optimization objective loss function This makes sentences that are close in time and semantically similar closer in the embedding space, while sentences with a larger time span and greater semantic differences are farther apart in the embedding space.

[0015] Optionally, step four specifically includes: A graph neural network is constructed to generate an emotional state graph based on the distance relationships and similarity measures in the non-Euclidean emotional space. The emotional states of customers at different time points are used as nodes in the graph, and the correlation between emotional states between adjacent time points is used as edges in the graph. The connection between nodes is determined based on the emotional change trend between adjacent time points and the semantic similarity between sentence embedding vectors, and each edge is assigned a weight representing the intensity of emotional change and time relevance. The constructed sentiment state graph is input into a graph neural network, multi-layer graph convolution operations are performed on the nodes, the node representation is updated based on the node adjacency relationship, and a node embedding matrix is ​​generated. By using the readout operation of the graph neural network, the features of the node layer are globally aggregated to generate a graph-level aggregated vector.

[0016] Optionally, step five specifically includes: Based on node embedding matrices and graph-level convergence vectors, a sales script model is generated by fusing local and global features of customer emotional state through a multi-layer attention mechanism and conditional decoding structure. The generated sales script model is optimized based on the customer's current emotional state, and outputs sales scripts with emotional control conditions.

[0017] Optionally, the attention mechanism and conditional decoding are specifically as follows: Attention mechanisms calculate the weighted relationship between local and global features, and then use multi-layer attention mechanisms to perform weighted fusion of local and global features to obtain fused features.

[0018] ; in, For the first Local features of each node For the first A graph-level convergence vector, and These are the weights calculated by the attention mechanism. For the number of nodes, This represents the number of graph-level convergence vectors. The conditional decoding structure utilizes the aforementioned fusion features Using customer emotional state information as input, the final sales script is generated. ; in, and These are the weight matrices for local features and emotion control conditions, respectively. This indicates the conditions for emotional control, including information about the customer's emotional state. For bias terms, For activation function, For the output of the sales pitch.

[0019] Optionally, the step six specifically comprises: The sales pitch model is supervised training by using the data set containing customer history dialogue, emotion state labeling and sales result feedback; In the training process, the emotion recognition accuracy, the sales conversion rate and the pitch generation quality are comprehensively considered by a multi-objective optimization method to construct a joint evaluation function; According to the joint evaluation function, the model parameters are iteratively adjusted to continuously optimize the emotion matching degree and the semantic coherence in the sales pitch generation process, and the final sales pitch model is obtained.

[0020] The beneficial effects of the present application are: The present application realizes dynamic modeling of customer emotional state and intelligent sales pitch generation by introducing a multi-layer feature fusion mechanism combining non-Euclidean emotional space modeling, knowledge-driven Gaussian decay contrast learning and graph neural network, which has significant beneficial effects.

[0021] The present application constructs a non-Euclidean emotional space and introduces a hyperspherical embedding method, so that the customer emotional features maintain geometric consistency in high-dimensional space, thereby more accurately depicting the customer emotional change trend and significantly improving the accuracy of emotion recognition. The knowledge-driven Gaussian decay contrast learning model is used to introduce the field knowledge and time weight decay mechanism in the embedding stage, realize the adaptive enhancement of semantic and emotional information, and effectively improve the discriminability and robustness of the embedding representation.

[0022] The graph neural network is used to structure the customer multi-round dialogue data, capture the temporal dependence and semantic association between nodes, and obtain the global latent representation of the emotional state, which provides more context-aware feature support for sales pitch generation.

[0023] The present application realizes the multi-layer fusion of customer emotional local features and global features through a multi-layer attention mechanism and a conditional decoding structure, so that the generated pitch content can be dynamically regulated according to the real-time emotion of the customer, and the emotional fit and language naturalness of the sales pitch are enhanced. A multi-objective optimization training framework is used to realize joint optimization in terms of emotion recognition accuracy, semantic coherence and sales conversion rate, so that the generated result has higher individuality and commercial value in actual business application.

[0024] The present application as a whole breaks through the limitations of traditional template-based pitch generation methods, can realize intelligent pitch generation for real customer emotions, and has wide application prospects in the fields of intelligent customer service and intelligent marketing. BRIEF DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of the specification, illustrate embodiments of the application and are used to explain the application, but are not intended to limit the application. In the drawings:

[0026] Fig. 1 A CRM sales talk intelligent analysis method based on natural language processing is proposed for the application, and a flow chart is shown in the figure. Fig. 2 A Gaussian decay contrast learning diagram of a CRM sales talk intelligent analysis method based on natural language processing is shown in the figure. DETAILED DESCRIPTION

[0027] In combination Figs. 1-2 The application is further described in detail. These drawings are simplified schematic diagrams and only schematically show the basic structure of the application, showing the main components related to the application. According to the drawings, Fig. 1 and Fig. 2 A CRM sales talk intelligent analysis method based on natural language processing includes the following steps:

[0028] Step one: Collect the interaction data of customers in the CRM system and standardize it to obtain standardized interaction data. According to the standardized interaction data, calculate the emotional feature vector at different time nodes, and construct a non-Euclidean emotional space; Step two: According to the non-Euclidean emotional space and the emotional feature vector, the emotional state of the customer is mapped to a point in the high-dimensional space through hyperspherical embedding, and the emotional state of the customer is obtained; Step three: Based on the knowledge-driven unsupervised learning method, combined with Gaussian decay contrast learning, the sentences in the customer conversation are embedded to generate; through contrast learning to optimize the distance constraint of the embedding space, combined with the external knowledge base, introduce knowledge information in the embedding process, generate sentence embedding vector; Step four: Construct a graph neural network model, generate an emotional state graph according to the distance relationship and similarity measurement in the emotional space, perform multi-layer graph convolution and readout operation on the emotional state graph, obtain node embedding matrix and graph-level aggregation vector, and the graph-level aggregation vector is used as the potential representation of the evolution of the customer's emotional state; Step five: Based on the node embedding matrix and the graph-level aggregation vector, through the multi-layer attention mechanism and the conditional decoding structure, the local features and global features of the customer's emotional state are fused to generate a sales talk model, and the sales talk with emotional control conditions is output; Step six: Use a data set containing customer historical conversation, emotional state labeling and sales result feedback to supervise the training of the sales talk model through supervised learning and multi-objective optimization method, and iteratively optimize the model parameters to obtain the final sales talk model.

[0029] In this embodiment, step one specifically includes: Collect customer interaction data in the CRM system, including customer historical sentiment data, dialogue content, customer feedback information and sentiment tags, and perform standardization operations on the interaction data to obtain standardized interaction data; Based on the standardized interaction data, emotional feature vectors of customers at different time points are extracted. The emotional feature vectors include emotional type, emotional intensity, and emotional change trend. The dimension and curvature parameters of the emotional space are determined based on the distribution relationship of the emotional feature vectors, and a non-Euclidean emotional space is constructed.

[0030] In this embodiment, step two specifically includes: Based on the non-Euclidean emotional space and the emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space using the hyperspherical embedding algorithm; During the mapping process, the distribution of sentiment features in the non-Euclidean sentiment space is optimized by using hyperspherical geometric constraints, so that the distance between points with similar sentiment features is closer and the distance between points with different sentiment features is farther. The mapping result is the coordinates of the customer's emotional characteristics at different time points in the non-Euclidean emotional space. The coordinates represent the customer's emotional type, emotional intensity, emotional change trend, and emotional change pattern at different time points. The mapping allows us to obtain the customer's emotional state in a non-Euclidean emotional space.

[0031] In this embodiment, the optimization constraints of the hyperspherical embedding include: In a high-dimensional hyperspherical space, the angular distance between sentiment feature vectors is defined as a similarity measure; Construct a weighted angular distance loss function ; in, To optimize the overall objective function, we define the degree of aggregation and separation of sentiment features in the hyperspherical space. For any two sentiment feature vectors and The angle between them The weighting coefficients, determined by both temporal difference and semantic similarity, are used to control the contribution of different features to the loss. By minimizing the weighted angular distance loss function This allows points with similar emotional characteristics to aggregate on a hypersphere, while significantly different emotional characteristics remain separated in space, resulting in a structurally stable and highly discriminative distribution of customer emotional states.

[0032] In this embodiment, step three specifically includes: Based on the knowledge-driven unsupervised learning method, the Gaussian decay contrast learning model is combined to generate embedding of each sentence in the customer conversation; The sentences in the customer conversation are preprocessed, the semantic features and emotional labels of the sentences are extracted, and the emotional semantic enhancement is combined with the emotional dictionary in the external knowledge base; Through Gaussian decay contrast learning, the distance constraint in the sentence embedding space is optimized, and the sentences with similar semantics are closer in the vector space, and the sentences with large semantic difference are farther apart, The knowledge-driven and Gaussian decay contrast learning are combined through a joint loss function, ; Wherein, is a scalar weight, is a Gaussian decay contrast loss, is a knowledge constraint loss, is an emotional alignment loss, is a timing smoothing loss; In the sentence embedding process, the sales field knowledge information is introduced to generate sentence embedding vectors.

[0033] In this embodiment, the optimization target of the Gaussian decay contrast learning model includes: ; Wherein, is the overall optimization target loss function, respectively represent the embedding vectors of the first and the second sentence, represents a set of positive samples similar in semantics to the sample , represents any other sentence embedding vector compared with the sentence , is a temperature parameter for balancing the distance of similar samples and non-similar samples in the embedding space, is a Gaussian decay weight for the relationship between sentence and all other sentences , represents the cosine similarity between sentence and sentence , represents the cosine similarity between sentence and sentence , is a Gaussian decay weight for the relationship between sentence and all other sentences ​ ; in, Sentence and sentences Time difference in the dialogue sequence The standard deviation of time decay; By minimizing the overall optimization objective loss function This makes sentences that are close in time and semantically similar closer in the embedding space, while sentences with a larger time span and greater semantic differences are farther apart in the embedding space.

[0034] In this embodiment, step four specifically includes: A graph neural network is constructed to generate an emotional state graph based on the distance relationships and similarity measures in the non-Euclidean emotional space. The emotional states of customers at different time points are used as nodes in the graph, and the correlation between emotional states between adjacent time points is used as edges in the graph. The connection between nodes is determined based on the emotional change trend between adjacent time points and the semantic similarity between sentence embedding vectors, and each edge is assigned a weight representing the intensity of emotional change and time relevance. The constructed sentiment state graph is input into a graph neural network, multi-layer graph convolution operations are performed on the nodes, the node representation is updated based on the node adjacency relationship, and a node embedding matrix is ​​generated. By using the readout operation of the graph neural network, the features of the node layer are globally aggregated to generate a graph-level aggregated vector.

[0035] In this embodiment, step five specifically includes: Based on node embedding matrices and graph-level convergence vectors, a sales script model is generated by fusing local and global features of customer emotional state through a multi-layer attention mechanism and conditional decoding structure. The generated sales script model is optimized based on the customer's current emotional state, and outputs sales scripts with emotional control conditions.

[0036] In this embodiment, the attention mechanism and conditional decoding are specifically as follows: Attention mechanisms calculate the weighted relationship between local and global features, and then use multi-layer attention mechanisms to perform weighted fusion of local and global features to obtain fused features.

[0037] ; in, For the first Local features of each node For the first A graph-level convergence vector, and are weights calculated by attention mechanism respectively, is the number of nodes, is the number of graph-level aggregation vectors; The conditional decoding structure uses the fusion features and customer sentiment state information as input to generate the final sales script ; wherein, and are weight matrices of local features and emotional control conditions respectively, represents the emotional control condition, which contains the customer's emotional state information, is a bias term, is an activation function, is the output sales script.

[0038] In the embodiment, step six specifically includes: A dataset containing customer historical conversations, sentiment state labels, and sales result feedback is used to supervise the training of the sales script model; During the training process, the accuracy of sentiment recognition, the sales conversion rate, and the quality of script generation are considered comprehensively by a multi-objective optimization method to construct a joint evaluation function; According to the joint evaluation function, the model parameters are iteratively adjusted to continuously optimize the emotional matching degree and semantic coherence in the sales script generation process, and the final sales script model is obtained.

[0039] Example 1: The CRM sales script intelligent analysis method based on natural language processing of the application is applied to the customer relationship management system of a nationwide automobile financial service company. The company has customer service centers in more than twenty provinces, with an average daily interaction volume of more than eight hundred thousand, including loan consultation, insurance renewal, after-sales service, and complaint handling, etc. The traditional sales script system mainly relies on keyword matching and template reply, which cannot perceive the customer's emotion in detail. When the customer's emotion changes greatly and the demand is ambiguous, the generated script is often harsh and lacks pertinence, resulting in a decrease in customer satisfaction and a low sales conversion rate. In order to solve these problems, the enterprise adopts the intelligent analysis method proposed by the application to realize dynamic modeling of the customer's emotional state and emotion-driven script generation.

[0040] In practical applications, the system first automatically collects historical interaction records between customers and sales consultants from the CRM database, standardizes the raw data containing text, sentiment labels, and feedback scores, and removes redundant characters, punctuation, and irrelevant fields. By analyzing the interaction records of customers at different time nodes, the system extracts sentiment feature vectors, including sentiment types, intensity, and trend, and constructs a non-Euclidean sentiment space based on the distribution of these features. This space is used to represent the emotional evolution trajectory of customers in long-term interactions, allowing the distance relationships between different emotional states to be effectively distinguished.

[0041] Subsequently, the system maps the sentiment features of customers at different time nodes into state points in a high-dimensional space using a hyperspherical embedding algorithm, forming a computable sentiment state representation. On this basis, the system generates an unsupervised semantic embedding for each customer sentence. To enhance the temporal consistency and domain adaptability of the sentiment semantics, the embedding model uses a Gaussian decay contrast learning strategy, which optimizes the embedding space based on time intervals and semantic similarity, allowing sentences with similar semantics and close time to be more concentrated, while sentences with significant semantic differences and large time spans remain distant. Combined with the sales scenario terms and sentiment dictionaries in the enterprise external knowledge base, the model further improves the domain relevance of sentence representation.

[0042] After obtaining the sentiment state points and sentence embedding vectors of customers, the system constructs a sentiment state graph. This graph takes the sentiment states of customers at each time node as nodes and the emotional evolution relationship between adjacent time nodes as edges, with the edge weights determined by the time difference and semantic similarity. Graph neural networks perform multi-layer graph convolution operations on this graph structure to update node representations, obtaining a node embedding matrix and a graph-level aggregation vector. The node embedding matrix depicts the changing trend of customer sentiment within a short time window, while the graph-level aggregation vector reflects the overall characteristics of long-term emotional evolution of customers. Through this double-layer structure, the system can capture both local emotional fluctuations and global psychological tendencies.

[0043] In the script generation phase, the system uses a multi-layer attention mechanism and a conditional decoding structure to integrate local sentiment features in the node embedding matrix with global information in the graph-level aggregation vector, forming a unified sentiment feature representation. The attention mechanism automatically assigns weights between different node features, allowing the model to focus on the historical emotional segments most relevant to the current state of the customer. The conditional decoding structure uses the current sentiment state as a constraint to control the tone, sentence structure, and vocabulary intensity of the generated content, achieving script output with emotional control conditions. For example, when the customer is in an anxious and hesitant emotional state, the system automatically generates a gentle tone, fully explained, and persuasive script; when the customer shows positive and trusting emotions, the system generates more interaction-promoting expressions, enhancing the naturalness of sales conversion.

[0044] To verify the performance of the method, the enterprise selected about 15 million customer interaction data in the first half of 2025 as training and verification samples. Among them, 70% is used for training, and 30% is used for testing. In the deployment stage, the model is supervised learning through multi-objective optimization method, considering the accuracy of sentiment recognition, sales conversion rate and the quality of language generation three indicators for joint optimization. After two months of verification experiment, the results show that the method of the invention is significantly better than the traditional template method in sentiment recognition accuracy, language matching degree and customer satisfaction. Specifically, the sentiment recognition accuracy is improved to 93%, the sales conversion rate is improved by about 6 percentage points compared with the traditional system, the customer satisfaction is improved from 3.8 to 4.6, the average response time is shortened to 0.7 seconds, and the system still maintains stable operation in high concurrency scene.

[0045]

[0046] In real customer scenarios, the method of the invention shows strong adaptive ability. For example, in the case of frequent changes in loan schemes and negative emotions of customers, the system can generate language that is both soothing and fully explained through real-time updating of the emotion state graph and dynamic adjustment of the attention mechanism, avoiding mechanical responses of sales consultants. At the same time, the model performs outstandingly in emotion turning point identification, can judge in advance the dissatisfaction and hesitation of customers, and generate preventive language with emotional guidance, greatly reducing potential complaints and cancellations.

[0047] This embodiment fully demonstrates that the invention can realize high-precision sentiment recognition and language generation in a complex multi-round dialogue environment. By combining non-Euclidean emotion space, hyperspherical embedding, Gaussian decay learning, graph neural network and attention decoding mechanism, the system can not only accurately capture the trend of customer emotional changes, but also generate optimal sales language in real time according to the emotional state. Compared with traditional methods, the invention significantly improves the emotion matching degree, language naturalness and business achievement rate, and has high commercial application value and promotion potential.

Claims

1. A method for intelligent analysis of CRM sales scripts based on natural language processing, characterized in that, The method comprises the following steps: Step 1: collecting customer interaction data in a CRM system and standardizing the data to obtain standardized interaction data, and calculating emotional feature vectors at different time nodes based on the standardized interaction data to construct a non-Euclidean emotional space; Step 2: mapping the emotional state of the customer to a point in a high-dimensional space by hyperspherical embedding according to the non-Euclidean emotional space and the emotional feature vectors to obtain the emotional state of the customer; Step 3: embedding and generating sentences in the customer conversation based on a knowledge-driven unsupervised learning method combined with Gaussian decay contrast learning; optimizing the distance constraint of the embedding space through contrast learning, introducing knowledge information in the embedding process combined with an external knowledge base to generate sentence embedding vectors; Step 4: constructing a graph neural network model, generating an emotional state graph according to the distance relationship and similarity measurement in the emotional space, performing multi-layer graph convolution and readout operation on the emotional state graph to obtain a node embedding matrix and a graph-level aggregation vector, and the graph-level aggregation vector as a latent representation of the evolution of the emotional state of the customer; Step 5: based on the node embedding matrix and the graph-level aggregation vector, fusing the local features and global features of the emotional state of the customer through a multi-layer attention mechanism and a conditional decoding structure to generate a sales script model and output a sales script with emotional control conditions; Step 6: using a data set containing customer historical conversations, emotional state labels and sales result feedback, supervising and training the sales script model through a supervised learning and multi-objective optimization method, and iteratively optimizing the model parameters to obtain the final sales script model. 2.The CRM sales pitch intelligent analysis method based on natural language processing of claim 1, wherein, The step 1 specifically comprises: Collecting customer interaction data in a CRM system, including the customer's historical emotional data, conversation content, customer feedback information and emotional labels, and performing standardization operations on the interaction data to obtain standardized interaction data; According to the standardized interaction data, extracting emotional feature vectors of the customer at different time nodes, the emotional feature vectors including emotional type, emotional intensity and emotional change trend; According to the distribution relationship of the emotional feature vectors, determine the dimension number and curvature parameters of the emotional space, and construct a non-Euclidean emotional space. 3.The CRM sales pitch intelligent analysis method based on natural language processing of claim 1, wherein, The step 2 specifically comprises: According to the non-Euclidean emotional space and the emotional feature vectors, mapping the emotional state of the customer to a point in a high-dimensional space by a hyperspherical embedding algorithm; In the mapping process, the distribution of emotional features in the non-Euclidean emotional space is optimized by using the hyperspherical geometric constraint, so that the distance between similar emotional features is closer and the distance between different emotional features is farther; The result of the mapping is the coordinates of the emotional features of the customer at different time nodes in the non-Euclidean emotional space, which represent the emotional type, emotional intensity, emotional change trend and emotional change pattern of the customer at different time nodes; Through the mapping, the emotional state of the customer in the non-Euclidean emotional space is obtained.

4. The CRM sales pitch intelligent analysis method based on natural language processing according to claim 3, characterized in that, The optimization constraint of the hyperspherical embedding comprises: In the high-dimensional hyperspherical space, define the angular distance between emotional feature vectors as a similarity measurement; Construct a weighted angular distance loss function ； wherein, is the overall optimization objective function, representing the degree of aggregation and separation of emotional features in the hypersphere space, is the angle between any two emotional feature vectors and is the weighted coefficient determined by the time difference and semantic similarity, used to control the contribution of different features to the loss;​ by minimizing the weighted angular distance loss function The points of similar sentiment features are aggregated on the hypersphere, while the points of significantly different sentiment features are kept separated in the space, obtaining a customer sentiment state distribution with stable structure and high discrimination. 5.The CRM sales pitch intelligent analysis method based on natural language processing of claim 1, wherein, The third step specifically comprises: Based on the knowledge-driven unsupervised learning method, the Gaussian decay contrast learning model is combined to generate embedding for each sentence in the customer dialogue; The sentences in the customer dialogue are preprocessed, the semantic features and emotional labels of the sentences are extracted, and the emotional semantic enhancement is performed in combination with the emotional dictionary in the external knowledge base; Through Gaussian decay contrast learning, the distance constraint in the sentence embedding space is optimized, and the distance between sentences with similar semantics in the vector space is closer, and the distance between sentences with large semantic difference is farther, The knowledge-driven and Gaussian decay contrast learning are combined through a joint loss function, ； wherein, is a scalar weight, is a Gaussian decay contrast loss, is a knowledge constraint loss, is a sentiment alignment loss, is a temporal smoothing loss; In the sentence embedding process, the sales field knowledge information is introduced to generate the sentence embedding vector.

6. The CRM sales pitch intelligent analysis method based on natural language processing according to claim 5, characterized in that, The optimization target of the Gaussian decay contrast learning model comprises: ； in, To optimize the overall objective loss function, They represent the first The and the first The embedding vector of each sentence. Representation and Sample A set of semantically similar positive samples Indicates and sentences Any other sentence embedding vector to be compared , To balance the temperature parameter representing the distance between similar and dissimilar samples in the embedding space, For use in sentences With all other sentences Gaussian decay weights in the relationship between them Sentence and sentences Cosine similarity between them Sentence and sentences Cosine similarity between them For use in sentences With all other sentences The relationship between Gaussian decay weights ； wherein, represents a sentence and a sentence a time difference in a dialog sequence, is the standard deviation of the time decay.

7. The CRM sales pitch intelligent analysis method based on natural language processing according to claim 1, characterized in that, The fourth step specifically comprises: A graph neural network is constructed to generate an emotional state graph according to the distance relationship and similarity measurement in the non-Euclidean emotional space; The emotional state of the customer at different time nodes is taken as a node in the graph, the association relationship between the emotional states of adjacent time nodes is taken as an edge in the graph, the connection relationship between the nodes is determined according to the emotional change trend between adjacent time nodes and the semantic similarity between sentence embedding vectors, and each edge is assigned a weight representing the emotional change intensity and time correlation; The constructed emotional state graph is input into the graph neural network, multi-layer graph convolution operation is performed on the nodes, the node representation is updated based on the node adjacency relationship, and a node embedding matrix is generated; Through the readout operation of the graph neural network, the node layer features are globally aggregated to generate a graph-level aggregation vector. 8.The CRM sales pitch intelligent analysis method based on natural language processing of claim 1, wherein, The fifth step specifically comprises: Based on the node embedding matrix and the graph-level aggregation vector, a multi-layer attention mechanism and a conditional decoding structure are used to fuse the local features and global features of the customer emotional state, and a sales script model is generated; The generated sales script model is optimized according to the current emotional state of the customer, and outputs a sales script with emotional control conditions. 9.The CRM sales pitch intelligent analysis method based on natural language processing of claim 8, wherein, The attention mechanism and the conditional decoding specifically are: The attention mechanism calculates a weighted relationship between the local feature and the global feature, and the local feature and the global feature are weighted and fused through a plurality of attention mechanisms to obtain a fused feature : ; wherein, is a local feature of the th node, is a graph-level aggregation vector of the th node, and are weights calculated by an attention mechanism, respectively, is the number of nodes, is the number of graph-level aggregation vectors; The conditional decoding structure utilizes the fused features described above and customer sentiment state information as inputs to generate final sales scripts ; wherein, and are weight matrices for local features and emotional control conditions, respectively, represents an emotional control condition, including customer emotional state information, is a bias term, is an activation function, is an output sales pitch. 10.The CRM sales pitch intelligent analysis method based on natural language processing of claim 1, wherein, The sixth step specifically comprises: A data set containing customer historical dialogue, emotional state labeling and sales result feedback is used to supervise the training of the sales script model; During the training process, the multi-objective optimization method is used to comprehensively consider the emotional recognition accuracy, the sales conversion rate and the script generation quality, and a joint evaluation function is constructed; According to the joint evaluation function, the model parameters are iteratively adjusted to continuously optimize the emotional matching degree and semantic coherence in the sales script generation process, and the final sales script model is obtained.

Images