Large model-based enterprise dynamic intent tracking method and system

Abstract

The present application relates to the technical field of data processing, in particular to an enterprise dynamic intention tracking method and system based on a large model, comprising: reading enterprise operation data to generate a business anchor point vector, and combining historical dialogues to construct an intention anchor point set; using a large model to extract customer speech semantic evidence, and combining enterprise events to generate business evidence, recalibrating decay weights according to the correlation degree of the cache key vector and each anchor point, fusing dialogue channel and business channel evidence to obtain a comprehensive observation vector; establishing an intention state vector and a covariance matrix, and dynamically updating through Kalman gain; performing eigenvalue decomposition on the covariance matrix, identifying intention drift, splitting and merging, and adjusting business channel observation noise according to the consistency score of the large model output and the enterprise fact library, to realize enterprise dynamic intention tracking. The present application solves the problem of low reliability of intention tracking results.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method and system for tracking enterprise dynamic intent based on a large model. Background Technology

[0002] As digital transformation deepens, more and more enterprises are managing customer interactions through information systems such as Customer Relationship Management (CRM) systems, Enterprise Resource Planning (ERP) systems, ticketing systems, and online customer service platforms. In business scenarios involving marketing, sales conversion, customer service, and supply chain collaboration, customers typically interact continuously with enterprises through various channels such as telephone, email, instant messaging tools, and online customer service windows, expressing their purchasing needs, service requirements, consultation needs, and cooperation intentions. To improve customer conversion efficiency and business decision-making capabilities, enterprises usually need to identify and analyze customer intent to promptly grasp changing trends in customer needs and formulate corresponding business strategies.

[0003] In recent years, as large language models have demonstrated strong capabilities in natural language understanding, semantic reasoning, and contextual analysis, more and more enterprises have begun to utilize these models to identify customer intent within customer conversations. This enables functions such as customer needs mining, business opportunity prediction, risk warning, and customer profile construction. Currently, most intent recognition solutions based on large models rely primarily on semantic analysis of the customer's current utterance, outputting corresponding intent tags through classification models, cue word engineering, or large model inference mechanisms. Essentially, these are static intent recognition methods targeting a single conversation or short-term context. While some solutions incorporate historical conversation information or conversation state tracking mechanisms, they mainly focus on the dialogue content itself, rarely considering the significant impact of internal business data on customer intent judgment.

[0004] However, in practical applications, customer intent is often not static but evolves continuously with changes in communication content, business progress, and the company's operational status. Furthermore, existing technologies typically lack a mechanism for continuously verifying the reliability of intent recognition results. When the output of a large model is inconsistent with the actual business situation, it cannot dynamically correct and adjust the model output based on the business facts. This leads to the accumulation of erroneous judgments in subsequent analyses, further contributing to the low reliability of intent tracking results. Summary of the Invention

[0005] To address the problem of low reliability of intent tracking results mentioned in the background art, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for enterprise dynamic intent tracking based on a large model, comprising: reading structured fields from an enterprise's business database and concatenating them into a business state description string; feeding the business state description string into an embedded encoder to obtain a business anchor vector; and encoding an intent anchor set based on historical dialogues; extracting semantic evidence from customer utterances using a large model; generating business evidence for enterprise events according to mapping rules; calculating the correlation between the key vector of each slot in the large model cache and the business anchor vector and the intent anchor set; performing attention recalibration on the attenuation weights based on the correlation; fusing the dual-channel evidence within a sliding window to obtain a comprehensive observation vector, wherein the dual channels include a dialogue channel and a business channel; establishing an intent state vector and a covariance matrix; updating the intent state vector and covariance matrix using the comprehensive observation vector as an observation with Kalman gain; performing eigenvalue decomposition on the covariance matrix; determining intent drift, intent splitting, and intent merging based on the ratio of principal eigenvalues ​​to secondary eigenvalues ​​and the variance of each dimension; and adjusting the observation noise covariance of the business channel based on the consistency score between the large model output and the enterprise fact base.

[0007] The aforementioned technical solution integrates customer dialogue information with enterprise operational information for joint analysis and continuously tracks customer intent based on historical interaction processes. This avoids the bias in intent judgment caused by relying solely on single-turn dialogue content, improving the accuracy of identifying customers' true needs. Simultaneously, by dynamically adjusting the importance of historical information, key information closely related to business development is effectively preserved, reducing the risk of weakening or forgetting important semantic information during long-term interactions, thereby enhancing the continuity and stability of intent tracking. Furthermore, by incorporating a state update mechanism to record the changes in customer intent in real time, it can promptly identify trends in customer needs and the evolutionary relationships between various intents, improving the ability to perceive changes in intent. Finally, by dynamically verifying and correcting the analysis results using enterprise factual information, it reduces the impact of ambiguity in language expression, model inference bias, and noisy data on intent judgment results, thereby improving the reliability, consistency, and business adaptability of intent tracking results, providing more accurate support for enterprise customer management, opportunity discovery, and business decision-making.

[0008] Furthermore, the construction process of the business anchor vector and intent anchor set includes: extracting structured fields, including order stage, contract status, inventory warning, payment node, work order status, and customer relationship stage, from the enterprise operation database; concatenating the structured fields into a business status description string according to a predefined natural language template; feeding the business status description string into an embedding encoder of the same origin as the large model to obtain the business anchor vector; independently encoding the intent phrases marked or extracted in historical dialogues into intent anchors and sliding them according to a time window; removing intent anchors that exceed the time window; and registering new intents into the database to obtain the intent anchor set.

[0009] The above technical solution effectively integrates scattered business information and enhances the system's understanding of the actual operating status of enterprises by converting enterprise operating data into a unified semantic representation. At the same time, it dynamically maintains intent information by combining historical dialogues, promptly removes invalid intents and introduces new intents, ensuring that intent representation keeps pace with changes in customer needs and reducing interference from invalid historical information. This improves the correlation analysis capability between business information and customer needs, providing a more accurate and stable data foundation for subsequent intent recognition and dynamic tracking, and improving the accuracy and reliability of intent tracking results.

[0010] Furthermore, the process of attention recalibrating the time decay weight includes: calculating the cosine similarity between the key vector and the business anchor vector, and the maximum value of the cosine similarity between the key vector and each intent anchor in the intent anchor set, for the key vector of the target slot in the large model cache; weighting and summing the cosine similarity and the maximum value according to the balance coefficient of the business channel and the intent channel to obtain the relevance of the target slot; using the product of the relevance and the anchoring strength coefficient as a compensation factor, using the product of the decay weight of the target slot and 1 plus the compensation factor as the compensated decay weight, and writing the compensated decay weight back to the bias term before the attention normalization exponential function after normalization, where the target slot is any slot in the large model cache.

[0011] The above technical solution, by comprehensively considering the correlation between historical information and the current business status, dynamically adjusts the importance of historical information. This enables information highly relevant to current business progress and customer needs to receive more attention and retention, reducing the attenuation and loss of key semantic information during long-term interactions. At the same time, it avoids information with weak correlation from continuously occupying analysis resources and reduces the interference of irrelevant historical content on subsequent intent judgment. This enhances the model's ability to understand long-term context, improves the continuity, accuracy, and stability of customer intent recognition, and provides more reliable information support for subsequent intent evolution analysis.

[0012] Furthermore, the comprehensive observation vector includes: extracting semantic evidence from customer discourse using a large model, generating a dialogue channel evidence strength vector based on the strength of candidate intents; generating a business channel evidence strength vector for the enterprise event according to predefined mapping rules; and weighting and summing the dialogue channel evidence strength vector and the business channel evidence strength vector at each moment within the sliding window according to time decay weights to obtain the comprehensive observation vector, wherein the time decay weights have a negative exponential relationship with the time interval.

[0013] The aforementioned technical solution constructs observational data by simultaneously utilizing customer dialogue information and enterprise business information. This allows for a comprehensive reflection of customer needs from two dimensions: the content expressed by the customer and the actual progress of business operations, avoiding judgment biases caused by a single information source. Furthermore, by combining historical time-series information to differentiate data at different time points, recent information more accurately reflects changes in current customer needs, while historical information still provides necessary reference for intent analysis. This approach preserves the continuity of intent evolution while enhancing the ability to perceive the current intent state, thereby improving the completeness, accuracy, and timeliness of the observation results and providing more reliable data for subsequent intent tracking and status updates.

[0014] Furthermore, it also includes predicting the intention state vector and covariance matrix: Let the intention state vector at the previous time step be... The time interval between the two events is The intended natural decay coefficient is The sparse intention transition matrix is The process noise matrix is The time interval is raised to the power of the intention natural decay coefficient and multiplied by the sparse intention migration matrix, and then multiplied by the intention state vector of the previous time step to obtain the intention prediction vector; the covariance matrix is ​​transformed under the action of the sparse intention migration matrix and added to the process noise matrix to obtain the prediction covariance.

[0015] The aforementioned technical solution predicts future intent states by combining historical intent states, time-related factors, and the relationships between intents before state updates. This allows for the early characterization of evolving customer needs, enabling the intent tracking process to not only rely on current observations but also fully utilize historical patterns, thus improving the ability to predict the true direction of customer intent. Furthermore, the introduction of uncertainty descriptions during prediction effectively reflects potential fluctuations and random disturbances in changing customer needs, enhancing the adaptability of state estimation to complex business scenarios. This improves the continuity, stability, and accuracy of intent tracking results, providing a more reliable foundation for subsequent intent recognition and dynamic decision-making.

[0016] Further, the process of updating the intention state vector and covariance matrix includes: constructing a Kalman gain by merging the prediction covariance, the observation matrix, and the dual-channel observation noise covariance to obtain a total observation covariance; obtaining an observation residual by subtracting the predicted observation vector obtained by applying the combined observation vector and the observation matrix to the intention prediction vector; multiplying the Kalman gain and the observation residual and adding them to the intention prediction vector to obtain the updated intention state vector; obtaining the product of the identity matrix, the Kalman gain, and the observation matrix, and multiplying the product with the prediction covariance to obtain the updated covariance matrix, wherein the identity matrix is ​​a square matrix with 1s on the main diagonal, 0s on the remaining elements, and the same dimension as the prediction covariance matrix.

[0017] The above technical solution dynamically corrects the intent state by comprehensively utilizing prediction results and real-time observation information. It can promptly absorb new customer interaction information and business change information, reduce the bias caused by relying solely on historical predictions or current observations, and make the intent state estimation closer to the actual needs of customers. At the same time, it synchronously updates and corrects the uncertainty of the state, which can effectively reflect the credibility of the current intent judgment, improve the system's adaptability to noisy data, abnormal information and business fluctuations, enhance the stability and robustness of the intent tracking process, and thus improve the accuracy, continuity and reliability of customer intent recognition results, providing more accurate data support for subsequent intent evolution analysis and business decision-making.

[0018] Furthermore, it also includes expanding the dimension of the intent state vector, specifically: when the large model extracts an intent phrase that does not exist in the existing intent anchor set from the customer's utterance, the intent phrase is encoded as a new intent anchor and added to the intent anchor set; a new dimension is added to the intent state vector, and the covariance matrix is ​​expanded on the new dimension with a preset initial value; the rows and columns of the sparse intent transfer matrix corresponding to the new dimension are expanded, and transfer coefficients with the existing intent dimension are filled in according to business rules.

[0019] The above technical solution, by automatically introducing new intent dimensions when new customer needs are identified and simultaneously updating the relevant state description structure, can break through the limitations of a fixed intent category system. This enables the system to have continuous learning and dynamic adaptation capabilities, promptly covering newly emerging customer need types and business scenarios, and avoiding information omissions or identification errors due to missing intent categories. At the same time, it maintains the correlation between new and old intents, allowing new needs to participate in subsequent analysis and evolution tracking together with existing needs. This enhances the system's ability to express and adapt to complex and ever-changing customer needs, and improves the completeness, flexibility, and long-term stability of intent tracking.

[0020] Furthermore, the process of determining intent drift, intent splitting, and intent merging includes: performing eigenvalue decomposition on the updated covariance matrix to obtain principal and secondary eigenvalues, and obtaining the variance and mean sequence of each dimension of the intent state vector; when the variance of any dimension is higher than the drift threshold and its mean shifts by a preset magnitude, it is determined to be intent drift; when the ratio of the secondary eigenvalue to the principal eigenvalue is greater than the splitting threshold, it is determined to be intent splitting; when the Pearson correlation coefficient of two intent dimensions is higher than a preset upper limit and the corresponding intensity difference is lower than a preset lower limit, it is determined to be intent merging. The Pearson correlation coefficient is calculated based on the covariance and variance between the corresponding dimensions in the covariance matrix, and the intensity difference is obtained by subtracting the intensity corresponding to the two intent dimensions obtained from the intent state vector.

[0021] The aforementioned technical solution, through comprehensive analysis of the characteristics of intent state changes and the correlation between different intents, can promptly identify evolutionary phenomena such as changes in direction, refinement, and fusion of customer needs during their development, avoiding the simplistic treatment of continuously changing customer needs as static results. Furthermore, by combining the degree of correlation and activity between intents for joint judgment, it can more accurately distinguish between independent and converging needs, reducing misjudgments and omissions, and improving the perception and recognition accuracy of complex customer need changes. This enhances the adaptability and interpretability of intent tracking results to real-world business scenarios, enabling the system to more accurately reflect the dynamic evolution trend of customer needs and providing a more reliable basis for customer management, demand forecasting, and business decision-making.

[0022] Furthermore, the process of adjusting the observation noise covariance of the business channel based on the consistency score includes: comparing the intent labels and corresponding slots output by the large model with the corresponding fields in the enterprise fact base, and calculating the consistency score according to the field matching degree; constructing a monotonic adjustment factor based on the consistency score, wherein the monotonic adjustment factor is 1 minus the consistency score, and scaling the business channel observation noise covariance according to the monotonic adjustment factor to obtain the adjusted business channel observation noise covariance; using the adjusted business channel observation noise covariance for the next round of Kalman update, and recording the comprehensive observation vector, intent state vector, intent anchor set, and consistency score to the audit log.

[0023] The aforementioned technical solution continuously verifies the intent recognition results against the actual business information of the enterprise, enabling timely detection of discrepancies between model judgments and real business states. Based on the verification results, it dynamically adjusts the level of trust in business information during subsequent analysis, thereby reducing the impact of erroneous judgments and abnormal data on intent tracking results. Simultaneously, it makes subsequent state update processes more consistent with actual business conditions, improving the consistency and reliability of intent recognition results with the enterprise's operational facts. Furthermore, by recording key analysis results and evolution processes, it can fully preserve the trajectory of intent changes and decision-making basis, enhancing the traceability and auditability of the system's operation, thus improving the overall stability, accuracy, and business application value of intent tracking.

[0024] In a second aspect, the present invention provides a large-model-based enterprise dynamic intent tracking system, including a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the large-model-based enterprise dynamic intent tracking method described above is implemented.

[0025] The beneficial effects of this invention are as follows: This invention integrates customer dialogue information, enterprise operational data, and historical interaction information to continuously perceive, dynamically update, and track customer needs throughout the entire process. It overcomes the limitations of traditional intent recognition, which relies solely on single-turn dialogue content, making intent analysis results more closely aligned with actual business scenarios. By jointly modeling business status information and customer semantic information and dynamically adjusting the importance of historical information, it effectively retains key content closely related to current business progress, reduces interference from irrelevant information, and improves the continuity and accuracy of long-term intent tracking. Furthermore, by combining status prediction with real-time observation and update mechanisms, it can promptly reflect customer needs. By understanding changing trends and development directions, the system enhances its adaptability to complex and ever-changing business scenarios. Furthermore, by dynamically expanding intent categories, identifying relationships between changing intents, and continuously verifying and correcting analysis results using enterprise factual information, it can promptly identify new demands, accurately recognize the evolution of demands, and reduce the impact of model inference biases, thereby improving the consistency and reliability of intent recognition results with actual business conditions. In addition, by recording key analysis processes and results, the system enhances the traceability and auditability of its operation, thus comprehensively improving the intelligence level and application value of enterprise customer demand analysis, opportunity mining, customer management, and business decision-making processes. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an enterprise dynamic intent tracking method based on a large model according to an embodiment of the present invention. Detailed Implementation

[0027] Implementation example of a large-model-based enterprise dynamic intent tracking method.

[0028] like Figure 1 As shown in the flowchart of the enterprise dynamic intent tracking method based on a large model according to an embodiment of the present invention, the method includes the following steps: S1: Construct enterprise business anchor vectors and intent anchor sets.

[0029] In a preferred embodiment, the construction of enterprise business anchor vectors and intent anchor sets is performed first. This embodiment selects an enterprise-level intelligent customer service scenario as a typical application scenario. The enterprise's business database contains at least six core fields: order stage, contract status, inventory warning, payment node, work order status, and customer relationship stage. The order stage value set includes pending quotation, quotation received, contract signed, in production, shipped, and completed; the contract status value set includes draft, pending approval, pending execution, in execution, and settled; the inventory warning value set includes sufficient, tight, and out of stock; the payment node value set includes within the payment period, nearing due date, and overdue; the work order status value set includes unassigned, processing, resolved, and closed; and the customer relationship stage value set includes potential customer period, nurturing period, contract signing period, renewal preparation period, and churn risk period.

[0030] The server reads the current values ​​of the above six types of fields from the enterprise's operational database at preset periods. The specific value of the preset period is determined based on the enterprise's concurrency and event frequency. The read fields are concatenated according to a predefined natural language template to obtain a business status description string. The natural language template is in the form of order stage (e.g., contract status, inventory warning, payment node, work order status, customer relationship stage).

[0031] The business state description string is fed into an embedding encoder that is of the same origin as the large model. The encoded vector is denoted as the business anchor vector and stored in the business anchor cache.

[0032] At the end of each round of dialogue, the embedded encoder is independently invoked for each intent phrase extracted by the large model or manually labeled in that round to obtain intent anchor vectors. The intent anchor set is maintained by sliding a time window; in this embodiment, the time window is the most recent thirty rounds of dialogue. When the difference between the most recent occurrence time of an intent anchor and the current time is greater than the time window threshold, the anchor is removed from the intent anchor set. When the large model extracts a new intent phrase in the current round that is not in the intent anchor set, it is encoded as a new intent anchor and registered in the database.

[0033] The aforementioned sliding maintenance mechanism enables the enterprise's intent system to dynamically expand. All anchor vectors are timestamped when written to the cache for subsequent auditing and playback.

[0034] S2: Perform dual-channel evidence generation, attention recalibration, and dual-channel evidence fusion to obtain a comprehensive observation vector.

[0035] In a preferred embodiment, after obtaining the business anchor vector and intent anchor set, dual-channel evidence generation, attention recalibration, and dual-channel evidence fusion are performed. The dual channels include a dialogue channel and a business channel. The dialogue channel receives customer speech text and feeds it into a large model to extract semantic evidence; the business channel receives enterprise event streams, which come from change logs in customer relationship management, enterprise resource planning, and work order systems, such as customer downloading a quotation, contract signing completion, and work order status changing from "processing" to "resolved."

[0036] The dialogue channel evidence strength vector is generated according to the candidate intent dimension. For each customer utterance, the large model outputs a strength score of that utterance for each candidate intent dimension in the intent state vector while generating a response. The strength score ranges from 0 to 1. When a certain dimension is not matched, the strength score is set to 0.

[0037] The evidence strength vector for business channels is generated according to predefined event intent mapping rules. These rules are expressed using event type as the key and the intensity increment of the candidate intent dimension as the value. For example, a customer downloading a quotation triggers an increase of 0.8 in the intensity increment of the quotation inquiry intent dimension; a contract signing completion triggers an increase of 1 in the intensity increment of the contract confirmation intent dimension; and a resolved work order triggers a decrease of 0.5 in the intensity increment of the complaint handling intent dimension.

[0038] During the large model inference phase, the first [item] in its internal cache The key vector of each slot The correlation degree is calculated using the following formula: ,in, Indicates the first The correlation between slots Represents cosine similarity. This represents the balance coefficient between the service channel and the intent channel; in this embodiment, it is taken as... It is 0.5; Represents the business anchor vector. The first in the set of intentional anchor points One intention anchor point, among which The range of values ​​for is [1, ... ], This represents the total number of intent anchors in the intent anchor set. This is the function for finding the maximum value.

[0039] Furthermore, the attenuation weights of the large model under the original inference process are multiplicatively compensated using the following formula: ,in, Indicates the first The attenuation weight after slot compensation Indicates the anchoring strength coefficient; This indicates that the large model, under the original inference process, is in the first... The attenuation weight of each slot Indicates the first The correlation degree of each slot is calculated. The compensated attenuation weights are then normalized and written back to the bias term before the attention normalization exponential function.

[0040] After the dual-channel evidence is generated, the evidence within the sliding window is weighted and fused according to a time decay weight. The time decay weight is in the form of... ,in The attenuation coefficient is... For the natural constant The base is an exponential function, which is taken as 0.05 in this embodiment. The time interval between the evidence generation time and the current time is in seconds. In this embodiment, the sliding window length is the most recent 120 seconds. The evidence strength vectors of the dialogue channel and the business channel at each time point within the sliding window are weighted and summed according to the time decay weight, and then linearly superimposed according to the channel credibility to obtain the comprehensive observation vector for the current round.

[0041] S3: Establish the intention state vector and covariance matrix, and perform prediction and update.

[0042] In a preferred embodiment, after obtaining the comprehensive observation vector, the intent state vector is predicted and updated. An intent state vector with dynamically expandable dimensions is established. In this embodiment, the initial dimension of the intent state vector is twelve dimensions, corresponding to twelve registered candidate intents, including inquiry, comparison, order placement, order reminder, order modification, cancellation, complaint, consultation contract, consultation delivery, renewal, refund, and others. Simultaneously, a covariance matrix with the same dimension as the intent state vector is maintained. The initial covariance matrix is ​​set as a diagonal matrix, with diagonal elements set to 1 and off-diagonal elements set to 0.

[0043] At the beginning of each prediction step, record the intention state vector, the covariance matrix, and the time interval from the previous time step to the current time step. Let the intention natural decay coefficient be... In this embodiment, the value is 0.9; the sparse intention transition matrix is... The sparse intent migration matrix records only the sparse migration rules given by domain experts, such as migration items from inquiry to price comparison, price comparison to order placement, and order placement to order modification, with migration coefficients ranging from 0.1 to 0.3; the process noise matrix is... In this embodiment, a diagonal matrix is ​​used, with the diagonal element being 0.05.

[0044] After the prediction step is completed, a Kalman update is performed. An observation matrix H is constructed; in this embodiment, an identity matrix is ​​used to represent that the comprehensive observation vector directly corresponds to the candidate intent in each dimension. The observation noise covariance Rd of the dialogue channel is taken as a diagonal matrix with diagonal elements of 0.2; the observation noise covariance Rb of the service channel is also taken as a diagonal matrix with diagonal elements of 0.05. Rd and Rb are then superimposed according to the channel mapping to obtain the total observation noise covariance R.

[0045] When the large model extracts a new intent phrase from customer utterance that is not registered in the intent anchor set, it adds one dimension to the intent state vector and sets its initial strength to 0 and its initial variance to 1; simultaneously, it sets the diagonal elements of the new dimension of the covariance matrix to 1 and the off-diagonal elements to 0; and it fills the corresponding rows and columns of the new dimension of the sparse intent migration matrix with migration coefficients from the existing intent dimensions according to business rules. This process enables the dimension of the intent state vector to be expanded online, avoiding the problem of new intents being forcibly merged in a closed intent system.

[0046] S4: Execution intention topology change determination and consistency closed-loop feedback.

[0047] In a preferred embodiment, based on the updated covariance matrix, intent topology change determination and consistency closed-loop feedback are performed. First, real-time eigenvalue decomposition is performed on the covariance matrix to obtain principal and secondary eigenvalues, and then the variance and mean sequences of each dimension of the intent state vector are obtained. Intent topology change determination includes three types: intent drift, intent splitting, and intent merging.

[0048] The specific process for determining intent drift is as follows: when the variance of a certain dimension is higher than the drift threshold and its mean shifts by a preset magnitude, it is determined to be intent drift. In this embodiment, the drift threshold is set to 0.8, and the mean shift threshold is set to 0.3. After determining intent drift, a drift description is generated based on the intent anchor points associated with that dimension. For example, due to the contract status changing from negotiation to signing, the customer's intent drifts from inquiry to delivery confirmation.

[0049] The specific process for determining intent splitting is as follows: when the ratio of the secondary eigenvalue to the primary eigenvalue is greater than the splitting threshold, it is determined to be intent splitting. In this embodiment, the splitting threshold is set to 0.4. After determining intent splitting, a new intent dimension is derived according to the eigenvector direction corresponding to the secondary eigenvalue, and the intent state vector, the covariance matrix, and the sparse intent transition matrix are expanded on the new dimension according to the above steps.

[0050] The specific process for determining intent merging is as follows: when the Pearson correlation coefficient of any two intent dimensions is higher than a preset upper limit and the corresponding intensity difference is lower than a preset lower limit, it is determined that intents should be merged. In this embodiment, the preset upper limit is set to 0.95 and the preset lower limit is set to 0.05. After determining that intents should be merged, the two intent dimensions are collapsed into one dimension, the corresponding rows and columns in the covariance matrix are weighted and summed, and the corresponding intent anchors are merged.

[0051] A consistency closed-loop feedback is executed simultaneously with the topology change determination. The intent tags and corresponding slots output by the large model in the current round are compared with the corresponding fields in the enterprise fact base, and a consistency score is calculated based on the field matching degree. The fact base consists of an order system, a product database, and an enterprise knowledge graph. The matching degree is calculated with 1 for exact matching, 0.5 for fuzzy matching, and 0 for non-matching. The arithmetic mean of the matching degrees of all fields is the consistency score.

[0052] A monotonic adjustment factor is constructed based on the consistency score. This monotonic adjustment factor is negatively correlated with the consistency score; in this embodiment, the monotonic adjustment factor is set to 1 minus the consistency score. The adjusted service channel observation noise covariance is obtained by scaling the service channel observation noise covariance using the monotonic adjustment factor. When the consistency score is high, the monotonic adjustment factor is small, and the service channel has a higher weight in the next Kalman update; when the consistency score is low, the monotonic adjustment factor is large, and the service channel weight is correspondingly reduced to suppress the perturbation of the intention state by the illusion of a large model.

[0053] After each round of processing, the comprehensive observation vector, intent state vector, intent anchor set, and consistency score of that round are written into the auditable log. The auditable log is stored in partitions by customer identifier and timestamp, which facilitates future tracing of why customer intent migrates in a specific round and provides a complete chain of evidence for compliance audits.

[0054] For example, in the intelligent customer service system of an equipment manufacturing company, customer A was originally in the "quoted, pending contract" order stage. The company's business database showed the contract status as "pending approval," the inventory warning as "sufficient," the payment node as "within the payment period," and the customer relationship stage as "contract signing period." The system concatenates the above fields into a business status description string and encodes it to obtain a business anchor vector. Historical dialogues have already recorded intent anchors such as "inquiry," "price comparison," "contract confirmation," and "delivery consultation." Subsequently, the customer sent the message "I have reviewed the quotation. If I sign the contract this week, when can the goods be shipped?" The large model extracts the intent strength of "contract confirmation" as 0.85 and the intent strength of "delivery consultation" as 0.72, and the dialogue channel generates corresponding evidence strength vectors. At the same time, the system records the customer downloading the quotation and the contract approval completion event, and adds 0.8 and 1.0 business evidence strength to the "contract confirmation" intent dimension respectively according to the preset mapping rules. The system further calculates the correlation between the key vectors of each slot in the large model cache and the business anchor vectors and intent anchor sets. For slots with high correlation, the compensated attenuation weight is increased to preserve historical semantic information related to contract approval and delivery cycles in the attention calculation. Subsequently, time-attenuated weighted fusion is performed on dialogue channel evidence and business channel evidence within the last 120 seconds to form a comprehensive observation vector, and Kalman filtering is used to update the intent state vector. After the update, the strength of the "Contract Confirmation" dimension is 0.91, and the strength of the "Delivery Consultation" dimension is 0.83. At this point, the system finds that the Pearson correlation coefficient between the two dimensions, "Contract Confirmation" and "Delivery Consultation," reaches 0.97, and the difference in their strengths is only 0.04, which is less than the preset threshold. Therefore, it is determined that the two have formed a stable relationship and triggers intent merging, merging them into a new intent dimension, "Contract Delivery Confirmation." Simultaneously, the system performs a consistency check between the "Contract Signing and Delivery Confirmation" label output by the large model and the contract approval completion records in the enterprise fact base, achieving a consistency score of 0.95. This reduces the observation noise covariance in the business channel and improves the credibility of business event evidence in the next round of status updates. Finally, the comprehensive observation vector of this round, the updated intent state vector, changes in the intent anchor set, and the consistency score are all written to the audit log, enabling dynamic tracking and interpretable recording of customer intent from the simple inquiry stage to the contract signing and delivery focus stage.

[0055] This invention integrates enterprise operational data, customer dialogue data, and enterprise business event information to continuously perceive, dynamically model, and update customer intent in real time. This overcomes the limitations of traditional intent recognition, which relies solely on single-turn conversation content, enabling intent analysis results to be supported by both business and interactive contexts, thus improving the understanding of customers' true needs. By dynamically adjusting the importance of historical information, it effectively retains key historical information highly relevant to current business progress, reducing the weakening or forgetting of important context during long-term interactions, thereby enhancing the continuity and stability of intent tracking. Furthermore, by predicting and updating customer intent states, it more accurately depicts the development trend of customer needs over time, improving the ability to perceive potential changes in needs. By incorporating an automatic expansion mechanism for newly added intents, the system can continuously adapt to new business scenarios and new types of needs, avoiding the recognition limitations caused by a fixed intent system. Furthermore, by analyzing the evolution characteristics of intents and the relationships between intents, the system can promptly detect changes, refinements, and fusion processes in customer needs, improving recognition accuracy in complex intent scenarios. In addition, by utilizing enterprise factual information to continuously verify and dynamically correct the analysis results, the system can reduce the impact of model inference bias, semantic ambiguity, and abnormal data, improving the consistency and credibility of intent tracking results with actual business status, and enhancing the traceability, reliability, and business adaptability of the entire tracking process. This provides more accurate, comprehensive, and real-time data support for customer management, opportunity discovery, marketing decisions, and business operations.

[0056] Example of an enterprise dynamic intent tracking system based on a large model: The enterprise dynamic intent tracking system based on a large model, according to an embodiment of the present invention, includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the enterprise dynamic intent tracking method based on a large model as described above.

[0057] The large-model-based enterprise dynamic intent tracking system also includes other components well-known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0058] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for tracking enterprise dynamic intent based on a large model, characterized in that, include: The structured fields in the enterprise's business database are read and concatenated into a business status description string. The business status description string is then fed into an embedded encoder to obtain a business anchor vector, and a set of intent anchors is encoded based on historical dialogues. Semantic evidence is extracted from customer discourse by a large model, and business evidence is generated for enterprise events according to mapping rules. The correlation between the key vector of each slot in the large model cache and the business anchor vector and the intent anchor set is calculated. After attention recalibration of the attenuation weight based on the correlation, the dual-channel evidence in the sliding window is fused to obtain a comprehensive observation vector. The dual channels include the dialogue channel and the business channel. Establish an intention state vector and covariance matrix. Use the comprehensive observation vector as an observation to update the intention state vector and covariance matrix using Kalman gain. Perform eigenvalue decomposition on the covariance matrix. Determine intention drift, intention splitting, and intention merging based on the ratio of principal eigenvalues ​​to secondary eigenvalues ​​and the variance of each dimension. Adjust the observation noise covariance of the business channel based on the consistency score between the large model output and the enterprise fact base.

2. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The process of constructing business anchor vectors and intent anchor sets includes: Structured fields, including order stage, contract status, inventory warning, payment node, work order status, and customer relationship stage, are extracted from the enterprise's business database. These structured fields are then concatenated into a business status description string using a predefined natural language template. The business status description string is then fed into an embedding encoder that is of the same origin as the large model to obtain the business anchor vector. The intent phrases marked or extracted in the historical dialogue are independently encoded as intent anchors and slid according to the time window. Intent anchors that exceed the time window are removed, and new intents are registered in the database to obtain a set of intent anchors.

3. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The process of attention recalibrating the time decay weights includes: For the key vector of the target slot in the large model cache, calculate the cosine similarity between the key vector and the business anchor vector, and the maximum value of the cosine similarity between the key vector and each intent anchor in the intent anchor set; then, perform a weighted sum of the cosine similarity and the maximum value according to the balance coefficient between the business channel and the intent channel to obtain the correlation degree of the target slot. The product of the correlation degree and the anchoring strength coefficient is used as the compensation factor. The product of the attenuation weight of the target slot and 1 plus the compensation factor is used as the attenuation weight after compensation. The attenuation weight after compensation is normalized and written back to the bias term before the attention normalization exponential function. The target slot is any slot in the large model cache.

4. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The integrated observation vector includes: Semantic evidence is extracted from customer discourse by a large model, and a dialogue channel evidence strength vector is generated according to the strength of candidate intent; for the enterprise event, a business channel evidence strength vector is generated according to predefined mapping rules. The comprehensive observation vector is obtained by weighting and summing the evidence strength vectors of the dialogue channel and the business channel at each time point within the sliding window according to the time decay weight. The time decay weight has a negative exponential relationship with the time interval.

5. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, It also includes predicting the intention state vector and covariance matrix: Let the intentional state vector at the previous time step be... The time interval between the two events is The intended natural decay coefficient is The sparse intention transition matrix is The process noise matrix is ; The time interval is raised to the power of the natural decay coefficient of the intent and multiplied by the sparse intent migration matrix, and then multiplied by the intent state vector of the previous time step to obtain the intent prediction vector. The predicted covariance is obtained by transforming the covariance matrix under the action of the sparse intention migration matrix and adding it to the process noise matrix.

6. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The process of updating the intent state vector and covariance matrix includes: The total observation covariance obtained by combining the prediction covariance, the observation matrix, and the dual-channel observation noise covariance is used to construct the Kalman gain; the observation residual is obtained by subtracting the predicted observation vector obtained by applying the combined observation vector and the observation matrix to the intention prediction vector; the Kalman gain is multiplied by the observation residual and then added to the intention prediction vector to obtain the updated intention state vector; Obtain the product of the identity matrix, the Kalman gain, and the observation matrix. Multiply the product by the prediction covariance to obtain the updated covariance matrix. The identity matrix is ​​a square matrix with 1s on the main diagonal, 0s on the rest, and the same dimension as the prediction covariance matrix.

7. The enterprise dynamic intent tracking method based on a large model according to claim 1 or 5, characterized in that, It also includes expanding the dimension of the intent state vector, specifically: When the large model extracts an intent phrase that is not present in the existing intent anchor set from the customer's utterance, the intent phrase is encoded as a new intent anchor and added to the intent anchor set; and a new dimension is added to the intent state vector, and the covariance matrix is ​​expanded on the new dimension with a preset initial value; The sparse intent migration matrix is ​​expanded to include the rows and columns corresponding to the new dimension, and the migration coefficients with the existing intent dimensions are filled in according to business rules.

8. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The process of determining intent drift, intent splitting, and intent merging includes: performing eigenvalue decomposition on the updated covariance matrix to obtain principal eigenvalues ​​and secondary eigenvalues, and obtaining the variance and mean sequence of each dimension of the intent state vector; When the variance of any dimension is higher than the drift threshold and its mean shifts by a preset amount, it is determined to be intentional drift. When the ratio of the secondary eigenvalue to the primary eigenvalue is greater than the splitting threshold, it is determined to be an intentional split; When the Pearson correlation coefficient of two intent dimensions is higher than a preset upper limit and the corresponding intensity difference is lower than a preset lower limit, it is determined that the intent is merged. The Pearson correlation coefficient is calculated based on the covariance and variance between the corresponding dimensions in the covariance matrix, and the intensity difference is obtained by subtracting the intensity of the two intent dimensions obtained from the intent state vector.

9. The enterprise dynamic intent tracking method based on a large model according to claim 1, characterized in that, The process of adjusting the observation noise covariance of the service channel based on the consistency score includes: The intent tags and corresponding slots output by the large model are compared with the corresponding fields in the enterprise fact base, and a consistency score is calculated based on the field matching degree. Based on the consistency score, a monotonic adjustment factor is constructed. The monotonic adjustment factor is 1 minus the consistency score. The service channel observation noise covariance is scaled according to the monotonic adjustment factor to obtain the adjusted service channel observation noise covariance. The adjusted service channel observation noise covariance is used for the next round of Kalman update, and the integrated observation vector, intent state vector, intent anchor set, and consistency score are recorded in the audit log.

10. A large-scale enterprise dynamic intent tracking system, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the enterprise dynamic intent tracking method based on a large model as described in any one of claims 1 to 9 is implemented.

Images