A CRM sales speech intelligent analysis method based on natural language processing

Abstract

The application discloses a CRM sales dialogue intelligent analysis method based on natural language processing. The method constructs a non-Euclidean emotional space, uses hyperspherical embedding to map the emotional state of the customer into a point in a high-dimensional space, further extracts the potential features of the customer's emotion through a graph neural network, and generates personalized sales dialogue by combining a multi-layer attention mechanism and a conditional decoding structure. The method can dynamically identify the emotional changes of the customer and optimize the dialogue generation in real time according to the emotional state of the customer, improving the adaptability and accuracy of the sales dialogue. Through deep reinforcement learning optimization pruning strategy, the application can effectively improve the computing efficiency, reduce redundant data processing and improve the system response speed.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and customer relationship management, and more particularly to a method for intelligent analysis of CRM sales scripts based on natural language processing. Background Technology

[0002] With the rapid development of information technology, Customer Relationship Management (CRM) systems have been widely applied across various industries, becoming an important tool for optimizing sales processes and customer management. Especially in sales and customer service, CRM systems help companies improve customer relationship maintenance efficiency and sales success rates. As market competition intensifies, 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 key factors in improving customer interaction quality and sales performance.

[0003] Traditional sales script generation methods primarily rely on manual design and rule matching, developing scripts through simple analysis of customer needs and problems. However, these methods have several problems. First, traditional script generation methods struggle to accurately capture and adapt to changes in customer emotions, resulting in scripts that are often fixed and simplistic, failing to match customer emotional needs. Second, existing methods struggle to adjust to real-time emotional fluctuations when faced with complex customer emotional changes, thus failing to efficiently generate sales scripts that meet customer needs in dynamic environments. Furthermore, traditional methods ignore the multi-dimensional characteristics of customer emotional states, failing to effectively integrate local and global features, leading to a lack of personalization and precision in the generated scripts.

[0004] In the field of Natural Language Processing (NLP), many intelligent sales script generation methods based on machine learning and deep learning have emerged in recent years. These methods attempt to provide personalized and customized sales scripts by analyzing customers' historical dialogues and sentiment data. However, existing NLP-based sales script generation methods still face the following problems. First, although these methods can analyze customers' sentiment data to a certain extent, they often fail to fully consider the complexity of emotional states and lack the ability to respond to emotional fluctuations in real time, resulting in generated scripts that cannot effectively adapt to changes in customers' emotions. Second, existing technologies often neglect the multi-layered relationships of emotions during sentiment analysis, failing to fully understand customers' emotional needs, thus affecting the accuracy and effectiveness of the scripts.

[0005] Existing sales script generation technologies have significant limitations when dealing with complex emotional changes and large-scale customer data, failing to provide accurate and personalized sales scripts in dynamic environments. To effectively improve the quality and adaptability of sales scripts, a new sales script generation method is urgently needed, capable of combining the customer's emotional state and the context of the conversation to generate more accurate, flexible, and customer-specific scripts.

[0006] Based on the aforementioned problems, this invention proposes an intelligent analysis method for sales scripts based on customer emotional states, combining graph neural networks and deep reinforcement learning techniques. This method deeply analyzes customers' emotional states, captures emotional changes in real time, and combines this with customers' personalized needs to generate more accurate and efficient sales scripts, thereby improving customer interaction quality and sales conversion rates. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent analysis method for CRM sales scripts based on natural language processing. This method constructs a non-Euclidean sentiment space with adjustable curvature, introduces knowledge-driven Gaussian decay contrastive learning to generate sentence embeddings, establishes a sentiment state graph based on sentiment features and semantic similarity, and obtains node embedding matrices and graph-level convergence vectors through a graph neural network. It then employs a multi-layer attention mechanism and a conditional decoding structure to fuse local and global features, generating sales scripts with sentiment control conditions. This technical approach is geared towards multi-turn interaction scenarios, providing a unified solution to the problems of traditional rule templates and single-sequence models being unable to represent sentiment evolution, integrate domain knowledge, or achieve dynamic control.

[0008] A CRM sales script intelligent analysis method based on natural language processing according to an embodiment of the present invention includes the following steps:

[0009] Step 1: Collect and standardize customer interaction data in the CRM system to obtain standardized interaction data. Based on the standardized interaction data, calculate the emotional feature vectors at different time points and construct a non-Euclidean emotional space.

[0010] Step 2: Based on the non-Euclidean emotional space and emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space through hyperspherical embedding to obtain the customer's emotional state;

[0011] Step 3: Based on a knowledge-driven unsupervised learning method, combined with Gaussian decay contrastive learning, sentences in customer dialogues are embedded and generated; the distance constraint of the embedding space is optimized through contrastive learning, and knowledge information is introduced into the embedding process by combining an external knowledge base to generate sentence embedding vectors.

[0012] Step 4: Construct a graph neural network model. Based on the distance relationship and similarity measure in the emotional space, generate an emotional state graph. Perform multi-layer graph convolution and readout operations on the emotional state graph to obtain a node embedding matrix and a graph-level convergence vector. The graph-level convergence vector serves as a potential representation of the customer's emotional state evolution.

[0013] Step 5: Based on the node embedding matrix and graph-level convergence vector, through a multi-layer attention mechanism and conditional decoding structure, the local and global features of the customer's emotional state are fused to generate a sales script model and output sales scripts with emotional control conditions.

[0014] Step Six: Using a dataset that includes customer history dialogues, sentiment state annotations, and sales result feedback, supervised training of the sales script model is conducted through supervised learning and multi-objective optimization methods. The model parameters are iteratively optimized to obtain the final sales script model.

[0015] Step 1: Collect and standardize customer interaction data in the CRM system to obtain standardized interaction data. Based on the standardized interaction data, calculate the emotional feature vectors at different time points and construct a non-Euclidean emotional space.

[0016] Step 2: Based on the non-Euclidean emotional space and emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space through hyperspherical embedding to obtain the customer's emotional state;

[0017] Step 3: Based on a knowledge-driven unsupervised learning method, combined with Gaussian decay contrastive learning, sentences in customer dialogues are embedded and generated; the distance constraint of the embedding space is optimized through contrastive learning, and knowledge information is introduced into the embedding process by combining an external knowledge base to generate sentence embedding vectors.

[0018] Step 4: Construct a graph neural network model. Based on the distance relationship and similarity measure in the emotional space, generate an emotional state graph. Perform multi-layer graph convolution and readout operations on the emotional state graph to obtain a node embedding matrix and a graph-level convergence vector. The graph-level convergence vector serves as a potential representation of the customer's emotional state evolution.

[0019] Step 5: Based on the node embedding matrix and graph-level convergence vector, through a multi-layer attention mechanism and conditional decoding structure, the local and global features of the customer's emotional state are fused to generate a sales script model and output sales scripts with emotional control conditions.

[0020] Step Six: Using a dataset that includes customer history dialogues, sentiment state annotations, and sales result feedback, supervised training of the sales script model is conducted through supervised learning and multi-objective optimization methods. The model parameters are iteratively optimized to obtain the final sales script model.

[0021] Optionally, step one specifically includes:

[0022] 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;

[0023] 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.

[0024] 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.

[0025] Optionally, step two specifically includes:

[0026] 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;

[0027] 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.

[0028] 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.

[0029] The mapping allows us to obtain the customer's emotional state in a non-Euclidean emotional space.

[0030] Optionally, the optimization constraints for the hyperspherical embedding include:

[0031] In a high-dimensional hyperspherical space, the angular distance between sentiment feature vectors is defined as a similarity measure;

[0032] Construct a weighted angular distance loss function

[0033] ;

[0034] 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.

[0035] 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.

[0036] Optionally, step three specifically includes:

[0037] 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.

[0038] 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;

[0039] 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.

[0040] Knowledge-driven and Gaussian decay contrastive learning are combined through a joint loss function.

[0041] ;

[0042] in, For scalar weights, Gaussian attenuation contrast loss, For knowledge constraint loss, For emotional alignment loss, This is the time-series smoothing loss;

[0043] Sales domain knowledge information is incorporated into the sentence embedding process to generate sentence embedding vectors.

[0044] Optionally, the optimization objective of the Gaussian decay contrastive learning model includes:

[0045] ;

[0046] 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 sentence 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

[0047] ;

[0048] in, Sentence and sentences Time difference in the dialogue sequence The standard deviation of time decay;

[0049] 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.

[0050] Optionally, step four specifically includes:

[0051] 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.

[0052] 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.

[0053] 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.

[0054] 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.

[0055] Optionally, step five specifically includes:

[0056] 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.

[0057] The generated sales script model is optimized based on the customer's current emotional state, and outputs sales scripts with emotional control conditions.

[0058] Optionally, the attention mechanism and conditional decoding are specifically as follows:

[0059] 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.

[0060] ;

[0061] 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.

[0062] The conditional decoding structure utilizes the aforementioned fusion features Using customer emotional state information as input, the final sales script is generated.

[0063] ;

[0064] 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, The sales script to be delivered.

[0065] Optionally, step six specifically includes:

[0066] We used a dataset that included customer history dialogues, sentiment status annotations, and sales result feedback to conduct supervised training on the sales script model.

[0067] During the training process, a joint evaluation function is constructed by comprehensively considering emotion recognition accuracy, sales conversion rate, and script generation quality through a multi-objective optimization method.

[0068] The model parameters are iteratively adjusted based on the joint evaluation function to continuously optimize the emotional matching degree and semantic coherence in the sales script generation process, and the final sales script model is obtained.

[0069] The beneficial effects of this invention are:

[0070] This invention achieves dynamic modeling of customer emotional states and intelligent sales script generation by introducing a multi-layer feature fusion mechanism that combines non-Euclidean emotional space modeling, knowledge-driven Gaussian decay contrastive learning, and graph neural networks, resulting in significant beneficial effects.

[0071] This invention constructs a non-Euclidean sentiment space and introduces a hyperspherical embedding method to maintain geometric consistency of customer sentiment features in a high-dimensional space, thereby more accurately depicting the trend of customer sentiment changes and significantly improving the accuracy of sentiment recognition. Utilizing a knowledge-driven Gaussian decay contrastive learning model, a domain knowledge and temporal weight decay mechanism are introduced during the embedding stage to achieve adaptive enhancement of semantic and sentiment information, effectively improving the discriminability and robustness of the embedded representation.

[0072] By using graph neural networks to structurally model multi-turn customer dialogue data, the temporal dependencies and semantic relationships between nodes are captured, and a global potential representation of emotional states is obtained, providing more context-aware feature support for sales script generation.

[0073] This invention achieves multi-layered fusion of local and global features of customer emotions through a multi-layered attention mechanism and conditional decoding structure. This allows the generated sales scripts to be dynamically adjusted based on the customer's real-time emotions, enhancing the emotional relevance and naturalness of the language. A multi-objective optimization training framework is employed to jointly optimize emotion recognition accuracy, semantic coherence, and sales conversion rate, resulting in more personalized and commercially valuable outputs in real-world business applications.

[0074] This invention breaks through the limitations of traditional template-based script generation methods and enables intelligent script generation that addresses real customer emotions, thus having broad application prospects in the fields of intelligent customer service and intelligent marketing. Attached Figure Description

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0076] Figure 1 This is a flowchart of a CRM sales script intelligent analysis method based on natural language processing proposed in this invention;

[0077] Figure 2 This is a schematic diagram of Gaussian decay contrastive learning for a CRM sales script intelligent analysis method based on natural language processing proposed in this invention. Detailed Implementation

[0078] Combination Figures 1-2The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides a CRM sales script intelligent analysis method based on natural language processing, comprising the following steps:

[0079] Step 1: Collect and standardize customer interaction data in the CRM system to obtain standardized interaction data. Based on the standardized interaction data, calculate the emotional feature vectors at different time points and construct a non-Euclidean emotional space.

[0080] Step 2: Based on the non-Euclidean emotional space and emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space through hyperspherical embedding to obtain the customer's emotional state;

[0081] Step 3: Based on a knowledge-driven unsupervised learning method, combined with Gaussian decay contrastive learning, sentences in customer dialogues are embedded and generated; the distance constraint of the embedding space is optimized through contrastive learning, and knowledge information is introduced into the embedding process by combining an external knowledge base to generate sentence embedding vectors.

[0082] Step 4: Construct a graph neural network model. Based on the distance relationship and similarity measure in the emotional space, generate an emotional state graph. Perform multi-layer graph convolution and readout operations on the emotional state graph to obtain a node embedding matrix and a graph-level convergence vector. The graph-level convergence vector serves as a potential representation of the customer's emotional state evolution.

[0083] Step 5: Based on the node embedding matrix and graph-level convergence vector, through a multi-layer attention mechanism and conditional decoding structure, the local and global features of the customer's emotional state are fused to generate a sales script model and output sales scripts with emotional control conditions.

[0084] Step Six: Using a dataset that includes customer history dialogues, sentiment state annotations, and sales result feedback, supervised training of the sales script model is conducted through supervised learning and multi-objective optimization methods. The model parameters are iteratively optimized to obtain the final sales script model.

[0085] In this embodiment, step one specifically includes:

[0086] 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;

[0087] 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.

[0088] 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.

[0089] In this embodiment, step two specifically includes:

[0090] 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;

[0091] 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.

[0092] 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.

[0093] The mapping allows us to obtain the customer's emotional state in a non-Euclidean emotional space.

[0094] In this embodiment, the optimization constraints of the hyperspherical embedding include:

[0095] In a high-dimensional hyperspherical space, the angular distance between sentiment feature vectors is defined as a similarity measure;

[0096] Construct a weighted angular distance loss function

[0097] ;

[0098] 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.

[0099] 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.

[0100] In this embodiment, step three specifically includes:

[0101] 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.

[0102] 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;

[0103] 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.

[0104] Knowledge-driven and Gaussian decay contrastive learning are combined through a joint loss function.

[0105] ;

[0106] in, For scalar weights, Gaussian attenuation contrast loss, For knowledge constraint loss, For emotional alignment loss, This is the time-series smoothing loss;

[0107] Sales domain knowledge information is incorporated into the sentence embedding process to generate sentence embedding vectors.

[0108] In this embodiment, the optimization objective of the Gaussian decay contrastive learning model includes:

[0109] ;

[0110] 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 sentence 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

[0111] ;

[0112] in, Sentence and sentences Time difference in the dialogue sequence The standard deviation of time decay;

[0113] 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.

[0114] In this embodiment, step four specifically includes:

[0115] 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.

[0116] 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.

[0117] 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.

[0118] 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.

[0119] In this embodiment, step five specifically includes:

[0120] 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.

[0121] The generated sales script model is optimized based on the customer's current emotional state, and outputs sales scripts with emotional control conditions.

[0122] In this embodiment, the attention mechanism and conditional decoding are specifically as follows:

[0123] 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.

[0124] ;

[0125] 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.

[0126] The conditional decoding structure utilizes the aforementioned fusion features Using customer emotional state information as input, the final sales script is generated.

[0127] ;

[0128] 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, The sales script to be delivered.

[0129] In this embodiment, step six specifically includes:

[0130] We used a dataset that included customer history dialogues, sentiment status annotations, and sales result feedback to conduct supervised training on the sales script model.

[0131] During the training process, a joint evaluation function is constructed by comprehensively considering emotion recognition accuracy, sales conversion rate, and script generation quality through a multi-objective optimization method.

[0132] The model parameters are iteratively adjusted based on the joint evaluation function to continuously optimize the emotional matching degree and semantic coherence in the sales script generation process, and the final sales script model is obtained.

[0133] Example 1:

[0134] The intelligent analysis method for CRM sales scripts based on natural language processing, as proposed in this invention, has been applied to the customer relationship management system of a nationwide auto finance service company. This company has customer service centers in over twenty provinces, handling more than 800,000 interactions daily, encompassing various scenarios such as loan inquiries, insurance renewals, after-sales service, and complaint handling. Traditional sales script systems primarily rely on keyword matching and template replies, failing to provide granular perception of customer emotions. When customer emotions fluctuate significantly or needs are ambiguous, the generated scripts are often stiff and lack relevance, leading to decreased customer satisfaction and low sales conversion rates. To address these issues, the company adopted the intelligent analysis method proposed in this invention to achieve dynamic modeling of customer emotional states and emotion-driven script generation.

[0135] In practical applications, the system first automatically collects historical interaction records between customers and sales consultants from the CRM database. The raw data, including text, emotion tags, and feedback ratings, is standardized, removing redundant characters, punctuation, and irrelevant fields. By analyzing customer interaction records at different time points, the system extracts emotional feature vectors, including emotion type, intensity, and trends, and constructs a non-Euclidean emotion space based on the distribution patterns of these features. This space is used to represent the emotional evolution trajectory of customers in long-term interactions, enabling the effective differentiation of distance relationships between different emotional states.

[0136] Subsequently, the system uses a hyperspherical embedding algorithm to map customer sentiment features at different time points into state points in a high-dimensional space, forming a computable sentiment state representation. Based on this, the system generates unsupervised semantic embeddings for each customer sentence. To enhance the temporal consistency and domain adaptability of sentiment semantics, the embedding model employs a Gaussian decay-based contrastive learning strategy, optimizing the embedding space through time intervals and semantic similarity. This results in sentences with similar semantics and similar time periods being more concentrated, while sentences with significant semantic differences and large time spans are kept at a greater distance. By combining sales scenario terminology and a sentiment dictionary from an external enterprise knowledge base, the model further improves the domain relevance of sentence representations.

[0137] After obtaining the customer's emotional state points and sentence embedding vectors, the system constructs an emotional state graph. This graph uses the customer's emotional state at each time point as nodes and the emotional evolution relationships between adjacent time points as edges, with the edge weights determined by both time difference and semantic similarity. A graph neural network performs multi-layer graph convolution operations on this graph structure to update node representations, thereby obtaining a node embedding matrix and a graph-level convergence vector. The node embedding matrix characterizes the changing trends of customer emotions within a short time window, while the graph-level convergence vector reflects the overall characteristics of the customer's long-term emotional evolution. Through this two-layer structure, the system can simultaneously capture local emotional fluctuations and global psychological tendencies.

[0138] During the script generation phase, the system employs a multi-layered attention mechanism and a conditional decoding structure. This integrates local emotional features embedded in the node matrix with global information from the graph-level convergence vector, forming a unified emotional feature expression. The attention mechanism automatically assigns weights among different node features, enabling the model to focus on historical emotional fragments most relevant to the customer's current state. The conditional decoding structure uses the current emotional state as a constraint, controlling the tone, sentence structure, and vocabulary intensity of the generated content to achieve script output with emotional control. For example, when a customer is anxious or hesitant, the system automatically generates a persuasive script with a gentle tone and thorough explanation; while when a customer exhibits positive and trusting emotions, the system generates a more transaction-oriented expression, enhancing the naturalness of sales conversion.

[0139] To verify the performance of this method, the company selected approximately 15 million customer interaction data points accumulated in the first half of 2025 as training and validation samples. Of these, 70% were used for training and 30% for testing. During the deployment phase, the model underwent supervised learning using a multi-objective optimization method, jointly optimizing three indicators: emotion recognition accuracy, sales conversion rate, and script generation quality. After two months of validation experiments, the results showed that the method of this invention significantly outperformed traditional template methods in terms of emotion recognition accuracy, script matching degree, and customer satisfaction. Specifically, the emotion recognition accuracy increased to 93%, the sales conversion rate improved by approximately 6 percentage points compared to traditional systems, customer satisfaction increased from 3.8 points to 4.6 points, the average response time was reduced to 0.7 seconds, and the system maintained stable operation even under high-concurrency scenarios.

[0140]

[0141] In real-world customer scenarios, the method of this invention demonstrates strong adaptability. For example, when customers frequently change loan plans or express negative emotions, the system can generate reassuring yet fully explanatory language through real-time updates to the emotional state map and dynamic adjustments to the attention mechanism, avoiding mechanical responses from sales consultants. Simultaneously, the model excels in identifying emotional turning points, anticipating potential customer dissatisfaction and hesitation, thereby generating preventative, emotionally guided communication that significantly reduces potential complaints and cancellations.

[0142] This embodiment fully demonstrates that the present invention can achieve high-precision emotion recognition and speech generation in complex multi-turn dialogue environments. By organically combining non-Euclidean emotion space, hyperspherical embedding, Gaussian decay learning, graph neural networks, and attention decoding mechanisms, the system can not only accurately capture the trend of customer emotion changes, but also generate optimal sales speech in real time based on the emotion state. Compared with traditional methods, the present invention significantly improves emotion matching accuracy, language naturalness, and business completion rate, and has high commercial application value and promotion potential.

Claims

1. A CRM sales script intelligent analysis method based on natural language processing, characterized in that, Includes the following steps: Step 1: Collect and standardize customer interaction data in the CRM system to obtain standardized interaction data. Based on the standardized interaction data, calculate the emotional feature vectors at different time points and construct a non-Euclidean emotional space. Step 2: Based on the non-Euclidean emotional space and emotional feature vector, the customer's emotional state is mapped to points in a high-dimensional space through hyperspherical embedding to obtain the customer's emotional state; 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 is used to obtain the customer's emotional state in the non-Euclidean emotional space; Step 3: Based on a knowledge-driven unsupervised learning method, combined with Gaussian decay contrastive learning, sentences in customer dialogues are embedded and generated; the distance constraint of the embedding space is optimized through contrastive learning, and knowledge information is introduced into the embedding process by combining an external knowledge base to generate sentence embedding vectors. Step 4: Construct a graph neural network model. Based on the distance relationship and similarity measure in the emotional space, generate an emotional state graph. Perform multi-layer graph convolution and readout operations on the emotional state graph to obtain a node embedding matrix and a graph-level convergence vector. The graph-level convergence vector serves as a potential representation of the customer's emotional state evolution. Step 5: Based on the node embedding matrix and graph-level convergence vector, through a multi-layer attention mechanism and conditional decoding structure, the local and global features of the customer's emotional state are fused to generate a sales script model and output sales scripts with emotional control conditions. Step Six: Using a dataset that includes customer history dialogues, sentiment state annotations, and sales result feedback, supervised training of the sales script model is conducted through supervised learning and multi-objective optimization methods. The model parameters are iteratively optimized to obtain the final sales script model.

2. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, 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.

3. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, 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 using Gaussian decay contrastive learning, the distance constraints in the sentence embedding space are optimized, which constrains semantically similar sentences to be closer in the vector space, and sentences with greater semantic differences to be farther apart. Sales domain knowledge information is incorporated into the sentence embedding process to generate sentence embedding vectors.

4. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, 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.

5. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, 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.

6. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, Step six specifically includes: We used a dataset that included customer history dialogues, sentiment status annotations, and sales result feedback to conduct supervised training on the sales script model. During the training process, a joint evaluation function is constructed by comprehensively considering emotion recognition accuracy, sales conversion rate, and script generation quality through a multi-objective optimization method. The model parameters are iteratively adjusted based on the joint evaluation function to continuously optimize the emotional matching degree and semantic coherence in the sales script generation process, and the final sales script model is obtained.

7. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 1, characterized in that, 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.

8. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 3, characterized in that, The optimization objectives of the Gaussian decay contrastive learning model include: ； 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 positive samples that are semantically similar or sentimentally consistent Indicates and sentence 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 semantics closer in the embedding space, while sentences with a large time span or large semantic differences are farther apart in the embedding space.

9. The intelligent analysis method for CRM sales scripts based on natural language processing according to claim 5, characterized in that, 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. : ; 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, The sales script to be delivered.

Images