Customer data processing and insight system based on large language model

Abstract

The invention belongs to the technical field of artificial intelligence and big data, and discloses a customer data processing and insight system based on a big language model. The system is composed of a multi-source data access module, a data preprocessing and label fusion module, a large language model semantic understanding module, a knowledge enhancement and semantic linkage module, an insight generation and visualization module, an intelligent strategy output module and a feedback learning and self-optimization module. According to the method, multi-source heterogeneous data such as texts, voices and structured behaviors are integrated, and the deep semantic analysis capability of a large language model is combined, so that global modeling of customer behaviors and intentions is realized; a multi-modal synchronous acquisition and standardization mechanism eliminates data format barriers, and a dynamic label mechanism adapts to context changes, so that the system can capture deep semantic association in customer expression, and compared with a traditional keyword matching method, the semantic understanding accuracy is improved by more than 40%, and a more complete data base is provided for insight generation.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and big data technology, and specifically provides a customer data processing and insight system based on a large language model. Background Art

[0002] In the current digital business environment, enterprises have accumulated massive amounts of multimodal customer data, including text, audio, structured data, and social platform data. However, traditional data processing systems have many limitations: when accessing multi-source data, different types of data have different formats, lacking unified collection and standardization capabilities, making it difficult to form complete data input; static labeling mechanisms in preprocessing cannot adapt to changes in customer context in real time, resulting in semantic understanding deviations; the application of large language models is mostly limited to simple summaries or questions and answers, and the output lacks structure and business adaptability; knowledge applications are static and difficult to dynamically link with model outputs; insight results lack time-series correlation analysis, and strategy generation relies on traditional rule engines, which lacks flexibility; and there is a lack of effective feedback learning mechanisms, making continuous optimization difficult.

[0003] These problems make it difficult for companies to accurately extract valuable insights from customer data and unable to meet the needs of personalized services and precision marketing. At present, the sources of customer data are showing a trend of diversification, unstructuredness, and high semantic complexity. Traditional systems are difficult to form effective insight links due to technical limitations such as weak complex semantic understanding capabilities, prominent information islands, insufficient model generalization capabilities, and single output insight forms. Methods based on pre-trained models are mostly deployed as a single module, without forming a complete system integration and real-time insight mechanism, and lack deep integration with corporate knowledge graphs, making it difficult to meet the customer's needs for systematic, scenario-based, and continuous learning intelligent insights. Therefore, improvement and optimization are needed. Summary of the Invention

[0004] The purpose of the present invention is to provide a customer data processing and insight system based on a large language model to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a customer data processing and insight system based on a large language model, which consists of a multi-source data access module, a data preprocessing and label fusion module, a large language model semantic understanding module, a knowledge enhancement and semantic linkage module, an insight generation and visualization module, an intelligent strategy output module, and a feedback learning and self-optimization module;

[0006] Multi-source data access module: This module collects customer data from multiple sources, including text, voice, and social platforms, and accesses it through APIs, ASR, and crawlers. It uses interfaces to convert different data into a unified intermediate standard format for use by subsequent modules.

[0007] Data preprocessing and label fusion module: This module cleans, segments, and identifies entities in the raw data, integrates customer labels to generate semantic input packages, applies NER, regularized extraction, and context window expansion strategies, and employs a dynamic labeling mechanism. After standardization processes such as denoising and temporal coding, it generates a vector representation through an embedding mapping layer.

[0008] Large language model semantic understanding module: This module uses the large model to analyze semantic packages, applies prompt templates and a multi-round context mechanism, performs Transformer encoding and multi-head attention calculations, and performs classification and prediction through the Softmax layer. It outputs high-dimensional features such as customer emotions and intentions, and supports adaptive construction of structured output templates.

[0009] Knowledge Enhancement and Semantic Linkage Module: This module aligns and integrates model output with the enterprise knowledge graph or industry background library, performs vector comparison through cosine similarity, and uses semantic vector alignment and rule mapping completion methods to form a customer intent chain and achieve dynamic knowledge injection.

[0010] Insight Generation and Visualization Module: This module maps structured semantic results into insights such as customer profiles and satisfaction trends. It uses a multidimensional clustering algorithm based on Transformer encoding and combines K-Means to cluster customer group feature vectors, supporting time series modeling and early warning linkage.

[0011] Intelligent Strategy Output Module: Based on insights, using language models, knowledge templates, and strategy mapping, it constructs a strategy generation objective function. This module automatically generates customer retention recommendations through semantic-level "strategy production," enhancing strategy adaptability and business fit.

[0012] Feedback learning and self-optimization module: This module collects business indicators such as customer response rate, integrates the RLHF training mechanism and prompt rewriting recommendation, uses reinforcement learning algorithm to optimize the strategy path, and dynamically updates the model, prompt and knowledge graph to form a closed-loop optimization.

[0013] Preferably, the multi-source data access module includes:

[0014] (1) Data collection and integration: Collect multi-source customer data, including structured information, customer service texts, social media comments, voice transcription data, and business logs, and convert the data into an intermediate standard format through an interface to provide standardized data input for subsequent modules;

[0015] (2) Technical implementation path: Use multi-channel crawler technology to obtain social media data, use automatic speech recognition (ASR) technology to complete speech-to-text conversion, use format conversion engine to integrate data in different formats, and use multi-modal synchronous collection and standardization capabilities to build a panoramic data input space;

[0016] (3) Innovative applications: In actual applications, text data is accessed from the CRM system through API, voice data is transcribed by the ASR module, and social media data is collected and structured through crawlers and SDKs to ensure efficient access and standardized conversion of various types of data.

[0017] Preferably, the data preprocessing and label fusion module includes:

[0018] (1) Data standardization: Perform multi-dimensional standardization on the original data, perform regularization denoising, extract entities through NER, and segment semantic units through word segmentation, encode the customer service conversation timestamp time series data, and then integrate it with industry attributes and existing customer label semantics to form a structured pre-processed data set;

[0019] (2) Dynamic semantic enhancement technology: This technology uses a context window expansion strategy and a dynamic labeling mechanism to enhance semantic understanding accuracy. The analysis includes the content of the five rounds before and after the conversation, and uses the Transformer attention mechanism to capture long-distance semantic dependencies. When the context changes, the label weight is automatically updated. For example, when a customer mentions "return" in an e-commerce scenario, the "after-sales demand" label is associated.

[0020] (3) Embedding mapping and semantic vectorization: The preprocessed input sequence is converted into a semantic vector through the embedding mapping layer. Word2Vec and BERT are used to pre-train the model to map the segmented text into a low-dimensional dense vector, preserving the semantic relevance of the words. The structured label data is converted into a numerical vector through one-hot encoding or embedding matrix. The text and label vector are then spliced ​​into a unified input sequence to provide a standardized semantic input package for the large language model.

[0021] Input sequence: X=[x1,x2,...,x n ]

[0022] After embedding the mapping layer, we get:

[0023] Where, e i The embedding vector corresponding to the i-th Token adopts a dynamic context window (sliding window) mechanism to maintain semantic coherence when processing data such as "historical conversation strings". At the same time, customer historical tags (such as VIP level and historical complaint topics) are introduced to supplement the current semantic package.

[0024] Preferably, the large language model semantic understanding module includes:

[0025] Large language model semantic analysis architecture: This architecture uses a private, fine-tuned large model and API interface. Using prompt templates (for example, extracting the three major pain points of complaints) and a multi-round context mechanism that caches the last 10 rounds of conversations, it deeply analyzes semantic input packets and accurately extracts high-dimensional features of customer intent, emotions, potential needs, and conflict points.

[0026] Application of Transformer encoding and attention mechanism: With the Transformer architecture as the core, a multi-head attention mechanism is used to capture the semantic association of input sequences, such as the dependency relationship between information points such as "order number" in customer service conversations. After position encoding and multi-layer Transformer block processing, a vector containing global semantic representation is generated, solving the problem of context dilution in long texts and ensuring the consistency of semantic understanding in multiple rounds of conversations. Using the Transformer encoding structure, the multi-head attention calculation formula for the input is:

[0027]

[0028] Adaptive construction of structured output templates: Dynamically generate structured output templates based on downstream system requirements. Using a semantic parser to identify CRM intent classification and marketing potential demand caller requirements, an adaptive JSON / XML structure is constructed. After classification and prediction at the Softmax layer, structured semantic results that can be called by downstream systems are directly generated, improving the business adaptability of the model output.

[0029] The model output vector is classified and predicted by the Softmax layer:

[0030]

[0031] Where h: represents the sentence-level semantic representation, W o 、b o : are the weights and biases obtained through training, Multi-classification output, such as emotion type and intent category;

[0032] The large language model is guided by the Prompt template to make its output format structured.

[0033] Preferably, the knowledge enhancement and semantic linkage module includes:

[0034] (1) Knowledge graph linkage completion mechanism: Deeply link the semantic output of the large model with the enterprise knowledge graph and industry background library, and use semantic vector alignment and rule mapping completion methods to build a causal logic customer intention chain. For example, after identifying "battery life is insufficient", the knowledge graph association path is retrieved to complete the causal relationship between insufficient battery life and charging chip failure, realizing the upgrade from single semantic recognition to logical chain construction;

[0035] (2) Cosine similarity vector comparison technology: The cosine similarity algorithm is used to calculate the distance between the semantic vector output by the model and the knowledge graph entity vector to quantify the matching degree. First, the customer intention vector (such as "apply for after-sales repair") is aligned with the knowledge graph node vector dimension, and then the cosine value of the vector angle (-1 to 1) is calculated to determine the association strength. When the similarity exceeds the threshold (such as 0.7), the corresponding knowledge node is automatically associated to supplement the business background;

[0036] The following cosine similarity method is used to complete the vector comparison:

[0037]

[0038] Where: If the similarity is greater than the set threshold τ, it is determined to be a synonymous entity or a hyponymous concept relationship, thereby introducing the causal path in the graph to complete the original reasoning result and output the causal chain information;

[0039] (3) Dynamic knowledge injection and logical chain construction: This module breaks through the traditional static knowledge model and realizes the semantic enhancement of "dynamic knowledge injection". When there is a semantic gap between the model output and the knowledge graph (such as the failure of cross-border payment without explaining the scenario), the system maps and completes it according to the knowledge path rules to form a complete intention chain, aligning the semantic understanding with the business context and improving business relevance.

[0040] Preferably, the insight generation and visualization module:

[0041] (1) Structured semantic insight generation mechanism: The structured semantic results output by the large model are converted into multi-dimensional insight conclusions, covering customer profiles, satisfaction trends, service risks, product preferences, and core dimensions. For example, by analyzing semantic features such as "high frequency returns" and "brand loyalty expressions" mentioned by customers in customer service conversations, customer profiles containing consumption habits and preference labels are generated; based on the keyword frequency of "logistics delays" and "product quality" in after-sales complaint data, service risk warning indicators are constructed and finally presented visually in the form of charts, label clusters, and text reports;

[0042] (2) Transformer-encoded multidimensional clustering application: A multidimensional clustering algorithm based on Transformer encoding is used to construct customer portraits, breaking through the limitations of traditional K-Means. Customer feature vectors are first encoded with Transformer to capture semantic associations, and then combined with K-Means clustering to form segmented groups. This supports time series analysis, such as tracking the evolution of satisfaction labels to identify consumer behavior trends.

[0043] Perform K-Means clustering on customer group feature vectors:

[0044]

[0045] Where: x: customer represented by semantic vector, μ i : centroid of class i, C i : Clustering result of the i-th category;

[0046] The results can be displayed in the form of charts, radar charts, behavior trajectory charts, etc. for management or customer representatives to use;

[0047] (3) Dynamic early warning and strategy linkage mechanism: Built-in custom strategy triggers can set early warning thresholds according to business needs (such as triggering an alarm when the service complaint rate exceeds 15%), combined with semantic insights to automatically associate rules. For example, if a customer group frequently encounters "payment failures + cross-border transactions", a risk warning will be triggered and operational strategies will be pushed in conjunction, realizing the upgrade from static labels to dynamic predictions.

[0048] Preferably, the intelligent strategy output module includes:

[0049] (1) Intelligent strategy generation mechanism: Based on language models, knowledge templates, and strategy mapping, this module automatically generates customer service recommendations, operational strategies, and marketing plans based on customer profiles and risk warning insights. Through semantic-level “strategy production,” it breaks through the limitations of traditional rule engines and achieves intelligent transformation from raw insights to executable strategies. It also generates personalized push content based on customer consumption behavior trends, thereby improving strategy generation efficiency.

[0050] (2) Dynamic Adaptation and Flexible Application: This mechanism is highly adaptable and flexible, and can dynamically adjust the strategy generation logic in response to complex and changing business needs. Whether it is customer retention suggestions or response strategies, it can accurately match business scenarios, making the strategies more in line with actual needs, and providing strong support for enterprises to respond to market changes and optimize operational decisions.

[0051] The strategy generation objective function is as follows:

[0052]

[0053] Where: f1(S): customer satisfaction improvement; f2(S): execution cost minimization; f3(S): conversion rate or response rate improvement; α i : Weight coefficient (dynamically adjustable);

[0054] Ultimately, the language model generates a content template and matches it with the operational strategy library to determine an executable plan.

[0055] Preferably, the feedback learning and self-optimization module includes:

[0056] (1) Construction of a closed-loop learning optimization system: Based on user operation feedback, customer response results, and business indicators, the system implements model fine-tuning, prompt template updates, and dynamic updates of the knowledge graph. By integrating the RLHF training mechanism, human feedback is used to guide the model optimization direction. At the same time, combined with prompt rewriting recommendations, a closed-loop learning mechanism is formed from data collection, strategy optimization to effect feedback, promoting the continuous improvement of large-scale model insight capabilities and local adaptive evolution;

[0057] (2) Reinforcement learning-driven strategy optimization: Using reinforcement learning algorithms, business indicators such as customer response rate, satisfaction feedback, and number of complaints are used as reward signals to optimize strategy selection paths. Through continuous exploration and trial and error, model parameters and prompt strategies are automatically adjusted, enabling the system to dynamically select the optimal strategy in complex business scenarios. This effectively solves the problem of traditional systems lacking continuous optimization capabilities and enhances the business adaptability and decision-making accuracy of the model.

[0058] Use reinforcement learning algorithm to optimize strategy selection path:

[0059]

[0060] Where: s t : Current policy status, a t : Current strategy, r t : reward signal from the business system, γ: discount coefficient, α: learning rate;

[0061] This module supports the RLHF (Reinforcement Learning Based on Human Feedback) mechanism to optimize local usage adaptability without affecting the capabilities of the original model.

[0062] The beneficial effects of the present invention are as follows:

[0063] 1. This invention integrates multi-source heterogeneous data such as text, voice, and structured behavior, combined with the deep semantic parsing capabilities of large language models, to achieve global modeling of customer behavior and intentions. The multimodal synchronous acquisition and standardization mechanism eliminates data format barriers, and the dynamic labeling mechanism adapts to contextual changes, enabling the system to capture deep semantic associations in customer expressions (such as the intensity of emotions and potential needs in the context of complaints). Compared with traditional keyword matching methods, the accuracy of semantic understanding is improved by more than 40%, providing a more complete data foundation for insight generation.

[0064] 2. The present invention dynamically links the model semantic output with the enterprise knowledge graph, and completes the causal logic chain through semantic vector alignment and rule mapping, so that the insight conclusions have the ability to explain industry scenarios. For example, in the e-commerce scenario, the system can automatically generate structured suggestions containing solutions based on the knowledge path of "logistics delay-customer complaints-compensation strategy", rather than simple emotional labels; this knowledge injection mechanism improves the business execution feasibility of the insight results by 35%, and continuously enhances the system's responsiveness to vertical scenarios through industry knowledge iteration.

[0065] 3. Unlike traditional static models, the present invention uses the RLHF reinforcement learning mechanism and business indicator feedback to form an optimization closed loop, adjusting model parameters, prompt strategies, and knowledge graph association rules in real time. When new business scenarios emerge, the system can automatically optimize the semantic parsing direction based on customer response data, thereby increasing the insight accuracy by 25% within three months. This self-evolution capability ensures that the system remains competitive in the digital business environment and avoids model failure caused by business model iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of the customer data processing and insight system based on the large language model of the present invention. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0068] like Figure 1 As shown, an embodiment of the present invention provides a customer data processing and insight system based on a large language model. The system consists of a multi-source data access module, a data preprocessing and label fusion module, a large language model semantic understanding module, a knowledge enhancement and semantic linkage module, an insight generation and visualization module, an intelligent strategy output module, and a feedback learning and self-optimization module.

[0069] Multi-source data access module: This module collects customer data from multiple sources, including text, voice, and social platforms, and accesses it through APIs, ASR, and crawlers. It uses interfaces to convert different data into a unified intermediate standard format for use by subsequent modules.

[0070] Data preprocessing and label fusion module: This module cleans, segments, and identifies entities in the raw data, integrates customer labels to generate semantic input packages, applies NER, regularized extraction, and context window expansion strategies, and employs a dynamic labeling mechanism. After standardization processes such as denoising and temporal coding, it generates a vector representation through an embedding mapping layer.

[0071] Large language model semantic understanding module: This module uses the large model to analyze semantic packages, applies prompt templates and a multi-round context mechanism, performs Transformer encoding and multi-head attention calculations, and performs classification and prediction through the Softmax layer. It outputs high-dimensional features such as customer emotions and intentions, and supports adaptive construction of structured output templates.

[0072] Knowledge Enhancement and Semantic Linkage Module: This module aligns and integrates model output with the enterprise knowledge graph or industry background library, performs vector comparison through cosine similarity, and uses semantic vector alignment and rule mapping completion methods to form a customer intent chain and achieve dynamic knowledge injection.

[0073] Insight Generation and Visualization Module: This module maps structured semantic results into insights such as customer profiles and satisfaction trends. It uses a multidimensional clustering algorithm based on Transformer encoding and combines K-Means to cluster customer group feature vectors, supporting time series modeling and early warning linkage.

[0074] Intelligent Strategy Output Module: Based on insights, using language models, knowledge templates, and strategy mapping, it constructs a strategy generation objective function. This module automatically generates customer retention recommendations through semantic-level "strategy production," enhancing strategy adaptability and business fit.

[0075] Feedback learning and self-optimization module: This module collects business indicators such as customer response rate, integrates the RLHF training mechanism and prompt rewriting recommendation, uses reinforcement learning algorithm to optimize the strategy path, and dynamically updates the model, prompt and knowledge graph to form a closed-loop optimization.

[0076] The present invention integrates multi-source heterogeneous data with the semantic capabilities of large language models to achieve global modeling of customer behavior and intentions. Multimodal acquisition and dynamic labeling mechanisms eliminate data barriers and adapt to context changes, improving the semantic understanding accuracy by more than 40% compared to traditional methods. By dynamically linking the model semantic output with the enterprise knowledge graph, the causal logic chain is completed, making the insight conclusions have industry explanatory power. For example, in e-commerce scenarios, structured suggestions containing solutions are generated based on knowledge paths, which increases the business execution feasibility of insight results by 35%, and enhances vertical scenario responsiveness through industry knowledge iteration. Different from traditional static models, the system forms an optimization closed loop through RLHF reinforcement learning and business indicator feedback, adjusts model parameters in real time, etc. In new business scenarios, the semantic parsing direction can be automatically optimized through customer response data, and the insight accuracy can be improved by 25% within 3 months, ensuring that the system continues to maintain its competitiveness.

[0077] Wherein, the multi-source data access module includes:

[0078] (1) Data collection and integration: Collect multi-source customer data, including structured information, customer service texts, social media comments, voice transcription data, and business logs, and convert the data into an intermediate standard format through an interface to provide standardized data input for subsequent modules;

[0079] (2) Technical implementation path: Use multi-channel crawler technology to obtain social media data, use automatic speech recognition (ASR) technology to complete speech-to-text conversion, use format conversion engine to integrate data in different formats, and use multi-modal synchronous collection and standardization capabilities to build a panoramic data input space;

[0080] (3) Innovative applications: In actual applications, text data is accessed from the CRM system through API, voice data is transcribed by the ASR module, and social media data is collected and structured through crawlers and SDKs to ensure efficient access and standardized conversion of various types of data.

[0081] In terms of data collection and integration, we comprehensively collect multi-source data such as customer structured information, customer service conversation texts, social media comments, voice transcription data, and business operation logs. With the help of interface technology, data of different types and formats are uniformly converted into an intermediate standard format, eliminating the barriers of data heterogeneity and providing a standardized and unified data foundation for subsequent modules to ensure the efficient advancement of the data processing process. In terms of technical implementation, we use multi-channel crawlers to capture social media data, rely on ASR technology to accurately transcribe voice into text, integrate multi-dimensional data through a format conversion engine, and build a complete panoramic data input space with multi-modal synchronous collection and standardization capabilities to achieve efficient and accurate data collection and format conversion. In innovative applications, text data is connected to the CRM system through APIs, voice is transcribed by the ASR module, and social media data is structured after being collected by crawlers and SDKs, achieving efficient access and standardized conversion of various types of data, ensuring collection integrity and availability.

[0082] The data preprocessing and label fusion module includes:

[0083] (1) Data standardization: Perform multi-dimensional standardization on the original data, perform regularization denoising, extract entities through NER, and segment semantic units through word segmentation, encode the customer service conversation timestamp time series data, and then integrate it with industry attributes and existing customer label semantics to form a structured pre-processed data set;

[0084] (2) Dynamic semantic enhancement technology: This technology uses a context window expansion strategy and a dynamic labeling mechanism to enhance semantic understanding accuracy. The analysis includes the content of the five rounds before and after the conversation, and uses the Transformer attention mechanism to capture long-distance semantic dependencies. When the context changes, the label weight is automatically updated. For example, when a customer mentions "return" in an e-commerce scenario, the "after-sales demand" label is associated.

[0085] (3) Embedding mapping and semantic vectorization: The preprocessed input sequence is converted into a semantic vector through the embedding mapping layer. Word2Vec and BERT are used to pre-train the model to map the segmented text into a low-dimensional dense vector, preserving the semantic relevance of the words. The structured label data is converted into a numerical vector through one-hot encoding or embedding matrix. The text and label vector are then spliced ​​into a unified input sequence to provide a standardized semantic input package for the large language model.

[0086] Input sequence: X=[x1,x2,...,x n ]

[0087] After embedding the mapping layer, we get:

[0088] Where, e iThe embedding vector corresponding to the i-th Token adopts a dynamic context window (sliding window) mechanism to maintain semantic coherence when processing data such as "historical conversation strings". At the same time, customer historical tags (such as VIP level and historical complaint topics) are introduced to supplement the current semantic package.

[0089] Data standardization processing: The original data is subjected to multi-dimensional processing such as regularization denoising, NER entity extraction, and word segmentation. The customer service conversation timestamp time series data is encoded and then semantically integrated with industry attributes and existing customer tags to form a structured preprocessed data set; dynamic semantic enhancement technology: Using context window expansion and dynamic labeling mechanism, the content of the five rounds before and after the conversation is included in the analysis, and the Transformer attention mechanism is used to capture long-distance semantic dependencies. The label weight is automatically updated when the context changes, such as associating "return" with "after-sales demand" labels in e-commerce scenarios; embedding mapping and semantic vectorization: Pre-trained models such as Word2Vec and BERT are used to map the preprocessed input sequence into a low-dimensional dense semantic vector. The structured label is spliced ​​with the text vector after one-hot encoding or embedding matrix conversion to provide a standardized semantic input package for the large model.

[0090] The large language model semantic understanding module includes:

[0091] Large language model semantic analysis architecture: This architecture uses a private, fine-tuned large model and API interface. Using prompt templates (for example, extracting the three major pain points of complaints) and a multi-round context mechanism that caches the last 10 rounds of conversations, it deeply analyzes semantic input packets and accurately extracts high-dimensional features of customer intent, emotions, potential needs, and conflict points.

[0092] Application of Transformer encoding and attention mechanism: With the Transformer architecture as the core, a multi-head attention mechanism is used to capture the semantic association of input sequences, such as the dependency relationship between information points such as "order number" in customer service conversations. After position encoding and multi-layer Transformer block processing, a vector containing global semantic representation is generated, solving the problem of context dilution in long texts and ensuring the consistency of semantic understanding in multiple rounds of conversations. Using the Transformer encoding structure, the multi-head attention calculation formula for the input is:

[0093]

[0094] Adaptive construction of structured output templates: Dynamically generate structured output templates based on downstream system requirements. Using a semantic parser to identify CRM intent classification and marketing potential demand caller requirements, an adaptive JSON / XML structure is constructed. After classification and prediction at the Softmax layer, structured semantic results that can be called by downstream systems are directly generated, improving the business adaptability of the model output.

[0095] The model output vector is classified and predicted by the Softmax layer:

[0096]

[0097] Where h: represents the sentence-level semantic representation, W o 、b o : are the weights and biases obtained through training, Multi-classification output, such as emotion type and intent category;

[0098] The large language model is guided by the Prompt template to make its output format structured.

[0099] The semantic analysis architecture of the large language model uses prompt templates and multi-round context mechanisms to deeply analyze semantic packages and extract features such as customer intent. The system dynamically generates structured templates based on downstream needs, outputs adaptation results through Softmax classification prediction, and uses prompt templates to guide the model in a targeted manner to improve the output structure and business adaptability.

[0100] The knowledge enhancement and semantic linkage module includes:

[0101] (1) Knowledge graph linkage completion mechanism: Deeply link the semantic output of the large model with the enterprise knowledge graph and industry background library, and use semantic vector alignment and rule mapping completion methods to build a causal logic customer intention chain. For example, after identifying "battery life is insufficient", the knowledge graph association path is retrieved to complete the causal relationship between insufficient battery life and charging chip failure, realizing the upgrade from single semantic recognition to logical chain construction;

[0102] (2) Cosine similarity vector comparison technology: The cosine similarity algorithm is used to calculate the distance between the semantic vector output by the model and the knowledge graph entity vector to quantify the matching degree. First, the customer intention vector (such as "apply for after-sales repair") is aligned with the knowledge graph node vector dimension, and then the cosine value of the vector angle (-1 to 1) is calculated to determine the association strength. When the similarity exceeds the threshold (such as 0.7), the corresponding knowledge node is automatically associated to supplement the business background;

[0103] The following cosine similarity method is used to complete the vector comparison:

[0104]

[0105] Where: If the similarity is greater than the set threshold τ, it is determined to be a synonymous entity or a hyponymous concept relationship, thereby introducing the causal path in the graph to complete the original reasoning result and output the causal chain information;

[0106] (3) Dynamic knowledge injection and logical chain construction: This module breaks through the traditional static knowledge model and realizes the semantic enhancement of "dynamic knowledge injection". When there is a semantic gap between the model output and the knowledge graph (such as the failure of cross-border payment without explaining the scenario), the system maps and completes it according to the knowledge path rules to form a complete intention chain, aligning the semantic understanding with the business context and improving business relevance.

[0107] The semantic output of the large model is deeply linked with the enterprise knowledge graph and industry background library, and semantic vector alignment and rule mapping completion methods are used to build a causal logic customer intention chain. For example, after identifying "insufficient battery life", the knowledge graph association path is retrieved to complete the causal relationship of insufficient battery life due to charging chip failure, realizing the upgrade from single semantic recognition to logical chain construction.

[0108] Among them, the insight generation and visualization module:

[0109] (1) Structured semantic insight generation mechanism: The structured semantic results output by the large model are converted into multi-dimensional insight conclusions, covering customer profiles, satisfaction trends, service risks, product preferences, and core dimensions. For example, by analyzing semantic features such as "high frequency returns" and "brand loyalty expressions" mentioned by customers in customer service conversations, customer profiles containing consumption habits and preference labels are generated; based on the keyword frequency of "logistics delays" and "product quality" in after-sales complaint data, service risk warning indicators are constructed and finally presented visually in the form of charts, label clusters, and text reports;

[0110] (2) Transformer-encoded multidimensional clustering application: A multidimensional clustering algorithm based on Transformer encoding is used to construct customer portraits, breaking through the limitations of traditional K-Means. Customer feature vectors are first encoded with Transformer to capture semantic associations, and then combined with K-Means clustering to form segmented groups. This supports time series analysis, such as tracking the evolution of satisfaction labels to identify consumer behavior trends.

[0111] Perform K-Means clustering on customer group feature vectors:

[0112]

[0113] Where: x: customer represented by semantic vector, μ i : centroid of class i, C i : Clustering result of the i-th category;

[0114] The results can be displayed in the form of charts, radar charts, behavior trajectory charts, etc. for management or customer representatives to use;

[0115] (3) Dynamic early warning and strategy linkage mechanism: Built-in custom strategy triggers can set early warning thresholds according to business needs (such as triggering an alarm when the service complaint rate exceeds 15%), combined with semantic insights to automatically associate rules. For example, if a customer group frequently encounters "payment failures + cross-border transactions", a risk warning will be triggered and operational strategies will be pushed in conjunction, realizing the upgrade from static labels to dynamic predictions.

[0116] Convert the structured semantic output of the large model into multi-dimensional insights such as customer portraits and satisfaction trends. For example, semantic features such as "high-frequency returns" are analyzed to generate consumer habit labels, and risk indicators are constructed based on the frequency of complaint keywords, which are visualized in the form of charts, etc. Dynamic early warning and strategy linkage mechanism: built-in triggers set thresholds according to business (such as complaint rates exceeding 15%), combined with semantic insight association rules, such as the discovery of a high-frequency combination of "payment failure + cross-border transactions" to trigger an early warning and link push strategies, realizing the upgrade from static to dynamic prediction.

[0117] Wherein, the intelligent strategy output module includes:

[0118] (1) Intelligent strategy generation mechanism: Based on language models, knowledge templates, and strategy mapping, this module automatically generates customer service recommendations, operational strategies, and marketing plans based on customer profiles and risk warning insights. Through semantic-level “strategy production,” it breaks through the limitations of traditional rule engines and achieves intelligent transformation from raw insights to executable strategies. It also generates personalized push content based on customer consumption behavior trends, thereby improving strategy generation efficiency.

[0119] (2) Dynamic Adaptation and Flexible Application: This mechanism is highly adaptable and flexible, and can dynamically adjust the strategy generation logic in response to complex and changing business needs. Whether it is customer retention suggestions or response strategies, it can accurately match business scenarios, making the strategies more in line with actual needs, and providing strong support for enterprises to respond to market changes and optimize operational decisions.

[0120] The strategy generation objective function is as follows:

[0121]

[0122] Where: f1(S): customer satisfaction improvement; f2(S): execution cost minimization; f3(S): conversion rate or response rate improvement; α i : Weight coefficient (dynamically adjustable);

[0123] Ultimately, the language model generates a content template and matches it with the operational strategy library to determine an executable plan.

[0124] Intelligent strategy generation mechanism: Based on language models, knowledge templates and strategy mapping, it automatically generates service recommendations and other strategies based on insights such as customer portraits. It breaks through the limitations of traditional rule engines through semantic-level "strategy production", such as generating personalized push notifications based on consumption trends, thereby improving strategy generation efficiency. (2) Dynamic adaptation and flexible application: The mechanism is highly adaptable and dynamically adjusts the strategy generation logic according to business needs to accurately match scenarios.

[0125] The feedback learning and self-optimization module includes:

[0126] (1) Construction of a closed-loop learning optimization system: Based on user operation feedback, customer response results, and business indicators, the system implements model fine-tuning, prompt template updates, and dynamic updates of the knowledge graph. By integrating the RLHF training mechanism, human feedback is used to guide the model optimization direction. At the same time, combined with prompt rewriting recommendations, a closed-loop learning mechanism is formed from data collection, strategy optimization to effect feedback, promoting the continuous improvement of large-scale model insight capabilities and local adaptive evolution;

[0127] (2) Reinforcement learning-driven strategy optimization: Using reinforcement learning algorithms, business indicators such as customer response rate, satisfaction feedback, and number of complaints are used as reward signals to optimize strategy selection paths. Through continuous exploration and trial and error, model parameters and prompt strategies are automatically adjusted, enabling the system to dynamically select the optimal strategy in complex business scenarios. This effectively solves the problem of traditional systems lacking continuous optimization capabilities and enhances the business adaptability and decision-making accuracy of the model.

[0128] Use reinforcement learning algorithm to optimize strategy selection path:

[0129]

[0130] Where: s t : Current policy status, a t : Current strategy, r t : reward signal from the business system, γ: discount coefficient, α: learning rate;

[0131] This module supports the RLHF (Reinforcement Learning Based on Human Feedback) mechanism to optimize local usage adaptability without affecting the capabilities of the original model.

[0132] Construction of a closed-loop learning optimization system: The system implements model fine-tuning, prompt template updates, and dynamic knowledge graph upgrades based on user feedback, customer response, and business indicators. It integrates the RLHF mechanism with prompt rewriting recommendations to form a closed loop of data collection, strategy optimization, and effect feedback, promoting the improvement of large-scale model insight capabilities and local adaptive evolution; reinforcement learning-driven strategy optimization: Using reinforcement learning algorithms, with business indicators such as customer response rate as reward signals, combined with the RLHF mechanism, it enhances business adaptability and decision-making accuracy without affecting the original model capabilities.

[0133] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0134] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A customer data processing and insight system based on a large language model, characterized by: The system consists of a multi-source data access module, a data preprocessing and label fusion module, a large language model semantic understanding module, a knowledge enhancement and semantic linkage module, an insight generation and visualization module, an intelligent strategy output module, and a feedback learning and self-optimization module. Multi-source data access module: This module collects customer data from multiple sources, including text, voice, and social platforms, and accesses it through APIs, ASR, and crawlers. It uses interfaces to convert different data into a unified intermediate standard format for use by subsequent modules. Data preprocessing and label fusion module: This module cleans and processes raw data, integrates customer labels to generate semantic input, uses multiple strategies to extract information, and generates vector representations after dynamic annotation and standardization. Large language model semantic understanding module: This module uses the large model to analyze semantic packages, uses prompt templates and multiple rounds of context, and performs Transformer encoding, attention calculation, and Softmax classification to output customer features and support template construction. Knowledge Enhancement and Semantic Linkage Module: This module integrates model output with knowledge graphs and industry libraries, compares vectors using cosine similarity, and forms an intent chain through semantic alignment and rule mapping to achieve dynamic knowledge injection. Insight Generation and Visualization Module: This module maps semantic results into customer profiles and insights, using Transformer-encoded multidimensional clustering and the K-Means algorithm to cluster groups and support time-series alert linkage. Intelligent Strategy Output Module: Based on insights, it uses language models, knowledge templates, and strategy mapping to build objective functions, achieve semantic-level strategy generation, and automatically output customer retention recommendations to improve adaptability and matching. Feedback learning and self-optimization module: collects customer response rate indicators, integrates RLHF and prompt rewriting to strengthen learning optimization strategies, dynamically updates models, prompts and knowledge graphs, and achieves closed-loop optimization.

2. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The multi-source data access module includes: (1) Data collection and integration: Collect multi-source customer data, including structured information, customer service texts, social media comments, voice transcription data, and business logs, and convert the data into an intermediate standard format through an interface to provide standardized data input for subsequent modules; (2) Technical implementation path: Use multi-channel crawler technology to obtain social media data, use automatic speech recognition (ASR) technology to complete speech-to-text conversion, use format conversion engine to integrate data in different formats, and use multi-modal synchronous collection and standardization capabilities to build a panoramic data input space; (3) Innovative applications: In actual applications, text data is accessed from the CRM system through API, voice data is transcribed by the ASR module, and social media data is collected and structured through crawlers and SDKs to ensure efficient access and standardized conversion of various types of data.

3. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The data preprocessing and label fusion module includes: (1) Data standardization: Perform multi-dimensional standardization on the original data, perform regularization denoising, extract entities through NER, and segment semantic units through word segmentation, encode the customer service conversation timestamp time series data, and then integrate it with industry attributes and existing customer label semantics to form a structured pre-processed data set; (2) Dynamic semantic enhancement technology: This technology uses a context window expansion strategy and a dynamic labeling mechanism to enhance semantic understanding accuracy. It incorporates the content of the five rounds before and after the conversation into the analysis, uses the Transformer attention mechanism to capture long-distance semantic dependencies, and automatically updates label weights when the context changes. For example, when a customer mentions "return" in an e-commerce scenario, the "after-sales demand" label is associated with it. (3) Embedding mapping and semantic vectorization: The preprocessed input sequence is converted into a semantic vector through the embedding mapping layer. Word2Vec and BERT are used to pre-train the model to map the segmented text into a low-dimensional dense vector, preserving the semantic relevance of the words. The structured label data is converted into a numerical vector through one-hot encoding and embedding matrix. The text and label vector are then spliced ​​into a unified input sequence to provide a standardized semantic input package for the large language model. Input sequence: X=[x1,x2,...,x n ] After embedding the mapping layer, we get: E=[e1,e2,...,e n ], Where, e i The embedding vector corresponding to the i-th Token adopts a dynamic context window (sliding window) mechanism to maintain semantic coherence when processing data, and introduces customer history tags to supplement the current semantic package.

4. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The large language model semantic understanding module includes: Large language model semantic analysis architecture: This architecture uses a private, fine-tuned large model and API interface, leverages prompt templates and a multi-turn context mechanism that caches the last 10 rounds of conversations, deeply analyzes semantic input packets, and accurately extracts high-dimensional features of customer intent, emotions, potential needs, and conflict points. Application of Transformer encoding and attention mechanism: With the Transformer architecture as the core, the multi-head attention mechanism is used to capture the semantic association of the input sequence. After position encoding and multi-layer Transformer block processing, a vector containing global semantic representation is generated, which solves the problem of context dilution in long texts and ensures the consistency of semantic understanding in multiple rounds of dialogue. Using the Transformer encoding structure, the multi-head attention calculation formula is performed on the input: Adaptive construction of structured output templates: Dynamically generate structured output templates based on downstream system requirements. Using a semantic parser to identify CRM intent classification and marketing potential demand caller requirements, an adaptive JSON / XML structure is constructed. After classification and prediction at the Softmax layer, structured semantic results that can be called by downstream systems are directly generated, improving the business adaptability of the model output. The model output vector is classified and predicted by the Softmax layer: Where h: represents the sentence-level semantic representation, W o 、b o : are the weights and biases obtained through training, For multi-classification output; The large language model is guided by the Prompt template to make its output format structured.

5. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The knowledge enhancement and semantic linkage module includes: (1) Knowledge graph linkage completion mechanism: Deeply link the semantic output of the large model with the enterprise knowledge graph and industry background library, use semantic vector alignment and rule mapping completion method to build a causal logic customer intention chain, retrieve the knowledge graph association path, and complete the causal relationship between insufficient battery life and charging chip failure, realizing the upgrade from single semantic recognition to logical chain construction; (2) Cosine similarity vector comparison technology: The cosine similarity algorithm is used to calculate the distance between the semantic vector output by the model and the knowledge graph entity vector to quantify the matching degree. First, the customer intention vector is aligned with the knowledge graph node vector dimension, and then the cosine value of the vector angle (-1 to 1) is calculated to determine the association strength. When the similarity exceeds the threshold, the corresponding knowledge node is automatically associated to supplement the business background. The following cosine similarity method is used to complete the vector comparison: Where: If the similarity is greater than the set threshold τ, it is determined to be a synonymous entity or a hyponymous concept relationship, thereby introducing the causal path in the graph to complete the original reasoning result and output the causal chain information; (3) Dynamic knowledge injection and logical chain construction: This module breaks through the traditional static knowledge model and realizes the semantic enhancement of "dynamic knowledge injection". When there is a semantic gap between the model output and the knowledge graph, the system maps and completes it according to the knowledge path rules to form a complete intention chain, aligning semantic understanding with business context and improving business relevance.

6. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The Insight Generation and Visualization Module: (1) Structured semantic insight generation mechanism: The structured semantic results output by the large model are converted into multi-dimensional insight conclusions, covering customer portraits, satisfaction trends, service risks, product preferences, and core dimensions. By analyzing the semantic features of "high-frequency returns" and "brand loyalty expressions" mentioned by customers in customer service conversations, customer portraits containing consumption habits and preference labels are generated; based on the keyword frequency of "logistics delay" and "product quality" in after-sales complaint data, service risk warning indicators are constructed, and finally visualized in the form of charts, label clusters, and text reports; (2) Transformer-encoded multidimensional clustering application: A multidimensional clustering algorithm based on Transformer encoding is used to construct customer portraits, breaking through the limitations of traditional K-Means. Customer feature vectors are first encoded with Transformer to capture semantic associations, and then combined with K-Means clustering to form segmented groups and support time series analysis. Perform K-Means clustering on customer group feature vectors: Where: x: customer represented by semantic vector, μ i : centroid of class i, C i : Clustering result of the i-th category; The results can be displayed in the form of charts, radar charts, and behavior trajectory charts for management and customer representatives to use; (3) Dynamic early warning and strategy linkage mechanism: Built-in custom strategy triggers can set early warning thresholds according to business needs, and automatically associate rules with semantic insights. For example, if a customer group frequently encounters "payment failures + cross-border transactions", a risk warning will be triggered and operational strategies will be pushed in conjunction, realizing the upgrade from static labels to dynamic predictions.

7. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The intelligent strategy output module includes: (1) Intelligent strategy generation mechanism: Based on language models, knowledge templates, and strategy mapping, this module automatically generates customer service recommendations, operational strategies, and marketing plans based on customer profiles and risk warning insights. Through semantic-level "strategy production," it breaks through the limitations of traditional rule engines and achieves intelligent transformation from raw insights to executable strategies. It also generates personalized push content based on customer consumption behavior trends, improving strategy generation efficiency. (2) Dynamic Adaptation and Flexible Application: This mechanism is highly adaptable and flexible, and can dynamically adjust the strategy generation logic in response to complex and changing business needs. Whether it is customer retention suggestions or response strategies, it can accurately match business scenarios, making the strategies more in line with actual needs, and providing strong support for enterprises to respond to market changes and optimize operational decisions. The strategy generation objective function is as follows: Where: f1(S): customer satisfaction improvement; f2(S): execution cost minimization; f3(S): conversion rate and response rate improvement; α i : Weight coefficient (dynamically adjustable); Ultimately, the language model generates a content template and matches it with the operational strategy library to determine an executable plan.

8. The customer data processing and insight system based on a large language model according to claim 1, characterized in that: The feedback learning and self-optimization module includes: (1) Construction of a closed-loop learning optimization system: Based on user operation feedback, customer response results, and business indicators, the system implements model fine-tuning, prompt template updates, and dynamic updates of the knowledge graph. By integrating the RLHF training mechanism, human feedback is used to guide the model optimization direction. At the same time, combined with prompt rewriting recommendations, a closed-loop learning mechanism is formed from data collection, strategy optimization to effect feedback, promoting the continuous improvement of large-scale model insight capabilities and local adaptive evolution; (2) Reinforcement learning-driven strategy optimization: Using reinforcement learning algorithms, we use business indicators such as customer response rate, satisfaction feedback, and number of complaints as reward signals to optimize strategy selection paths. Through continuous exploration and trial and error, we automatically adjust model parameters and prompt strategies, enabling the system to dynamically select the optimal strategy in complex business scenarios. This effectively addresses the problem of traditional systems lacking continuous optimization capabilities and enhances the model's business adaptability and decision-making accuracy. Use reinforcement learning algorithm to optimize strategy selection path: Where: s t : Current policy status, a t : Current strategy, r t : reward signal from the business system, γ: discount coefficient, α: learning rate; This module supports the RLHF mechanism and optimizes the adaptability of local use without affecting the capabilities of the original model.

Images