An intelligent sales conversation navigation method and system based on a hierarchical state machine and portrait completion degree driving, and a medium

Abstract

The application relates to an intelligent sales dialogue navigation method and system based on a hierarchical state machine and portrait completion degree driving and a medium, and relates to the technical field of customer intelligent management. The intelligent sales dialogue navigation method comprises the following steps: collecting and cleaning multi-source data of a customer, constructing a portrait completion degree vector, combining an intention degree evaluation value, and generating a context vector; dividing a main process node according to a hierarchical finite state machine, combining the context vector, calculating a node transition probability, determining a target node, and triggering a sub-process; according to the target node and the context vector, generating personalized text dialogue according to dialogue retrieval scoring; monitoring the personalized text dialogue according to a preset semantic classification model, screening compliant text dialogue, and simultaneously outputting corresponding compliant voice dialogue; after completing the dialogue of all the compliant voice dialogue, collecting customer feedback signals, updating portrait dimension weight and dialogue effect scoring; and solving the problem of the separation of process management and dialogue recommendation in two independent systems.

Description

Technical Field

[0001] This application relates to the field of customer intelligent management technology, and in particular to an intelligent sales dialogue navigation method, system and medium based on hierarchical state machine and profile completion driven. Background Technology

[0002] In Customer Relationship Management (CRM) systems, Next-Best-Action (NBA) recommendations are a key technology for improving sales conversion rates. Existing NBA technology solutions mainly fall into the following categories, all of which have varying degrees of limitations: Option 1: The NBA system, represented by PegaCustomerDecisionHub, uses a four-factor scoring formula of P×C×V×L (propensity × context weight × business value × business leverage) to calculate the probability of a customer accepting a specific action and output action-level recommendations (such as "make a phone call", "send an email", "push offers").

[0003] The limitations of this solution are: 1. The recommendation granularity is action-level, only telling salespeople "what to do next," and cannot guide node-level progress and script content within a single conversation; 2. The decision model is stateless computation, with each recommendation independent of the preceding and following conversation context, and cannot model the sequential dependencies of the conversation; 3. It does not support nested sub-processes in the conversation, and cannot handle complex scenarios in sales conversations that require temporary entry into professional analysis (such as school comparison, professional evaluation) before returning to the main process.

[0004] Option 2: Real-time sales support systems, such as Gong and ClariCopilot, can detect keywords in the call (such as competitor names and price objections) and trigger preset BattleCards to be pushed to sales personnel.

[0005] The limitations of this solution are: 1. The triggering mechanism is keyword matching, which is a passive response and cannot actively plan the dialogue path; 2. The pushed content is a static preset template and cannot dynamically generate personalized scripts based on real-time changes in customer profiles; 3. The script cards are completely decoupled from the sales process, and it is unknown which sales stage is currently in, so it cannot provide process-level navigation.

[0006] Option 3: Agent inference engines, such as SalesforceAtlasReasoningEngine, break down user goals into step plans and select tools and actions through LLM inference.

[0007] The limitations of this approach are: 1. It is a general reasoning framework and lacks dedicated modeling for the sequential nature of sales dialogues; 2. Dialogue state management relies on the context window memory of LLM and lacks an explicit state transition model, making it easy to lose key context in long dialogues; 3. It cannot utilize structured knowledge from vertical domains (such as industry exam preparation cycles and subject difficulty gradients) for accurate gap analysis and script matching.

[0008] Existing patents disclose a marketing assistance method and system for enterprise micro-businesses based on a large language model, utilizing this model to improve marketing efficiency and effectiveness. The method integrates vector databases, embedding models, and contextual learning techniques to achieve real-time understanding of user intent, personalized responses, and rapid adaptation to business changes. Its core advantages lie in its accurate user needs analysis, personalized service provision, and efficient knowledge retrieval system. By analyzing user behavior and dialogue records, this invention can provide product recommendations and marketing strategies that match user preferences. Furthermore, the invention features real-time script updates, ensuring the timeliness and accuracy of information; automated order information extraction technology improves information processing efficiency and reduces the time and cost of manual operations. This invention not only improves marketing efficiency but also enhances the interactive experience between users and customer service, providing enterprises with a competitive advantage in a highly competitive market.

[0009] The existing technical solutions mentioned above have the following drawbacks: 1. The process navigation and dialogue generation are separated: no existing system has achieved joint recommendation of "dialogue path planning + dialogue content generation"; 2. Customer profile completeness is not being utilized: All existing systems treat the completeness of profile data as a passive prerequisite, rather than an active input to drive the direction of conversation and assess intent. 3. Lack of vertical domain knowledge-driven script generation: Existing script recommendations are based on general templates or general LLM world knowledge, and cannot dynamically generate targeted content based on the specific gap (GAP) between the client's current capabilities and target requirements. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this application is to provide an intelligent sales dialogue navigation method, system, and medium based on hierarchical state machines and profile completion-driven approaches. By explicitly modeling dialogue states, real-time assessment of intent based on profile completion, and script generation incorporating vertical domain knowledge, this approach solves the technical problems of coarse-grained recommendations, disconnect between process and script, and lack of dialogue sequence modeling in existing NBA systems.

[0011] This was achieved using the following technical solutions: Firstly, this application provides an intelligent sales dialogue navigation method based on hierarchical state machines and profile completion levels, including: Collect and clean multi-source customer data, construct a profile completion vector, and combine it with intent assessment values ​​to generate a context vector; The main process nodes are divided according to the hierarchical finite state machine, and the node transition probability is calculated by combining the context vector to determine the target node and trigger the sub-process. Based on the target node and context vector, personalized text scripts are generated according to the script retrieval score. Based on a preset semantic classification model, personalized text messages are monitored, compliant text messages are selected, and corresponding compliant voice messages are output. After completing all compliant voice scripts, customer feedback signals are collected to update the profile dimension weights and script effectiveness scores.

[0012] By adopting the above technical solution, the dialogue process is guided by a hierarchical finite state machine and a node transition probability algorithm. Personalized text is generated by combining the dialogue retrieval score, and the voice output is filtered by a semantic classification model to ensure compliance. At the same time, the profile weight and dialogue score are dynamically updated by using customer feedback signals. This achieves closed-loop adaptive optimization of intelligent customer service, which significantly improves the personalization and compliance efficiency of interaction.

[0013] This application further specifies: collecting and cleaning multi-source customer data, constructing a profile completion vector, and combining it with intent assessment values ​​to generate a context vector, including: Detect and classify multi-source customer data, and extract customer static data and customer dynamic data; The static customer data is parsed according to the preset profile dimensions, and the single-dimensional completion degree p_i(t) is calculated, i∈{1,2,3,4,5,6}, p_i(t)∈[0,1]: p_i(t)=count(Retrieved_i) / count(total_i); Where i is the dimension number, t is the total dialogue time, Retrieved is the number of fields obtained, and total is the total number of fields; Correlate all single-dimensional completion scores to construct a profile completion vector prog(t)=[p_1(t),p_2(t),p_3(t),p_4(t),p_5(t),p_6(t)]; The customer dynamic data is analyzed to extract the current behavioral event, and combined with the behavioral weight value signal_k, the behavioral signal score BS(t) is calculated: BS(t) = Σ_k (signal_k) × exp(-λ × (t-t_k)); Where t_k is the occurrence time of the current event, and λ is the time decay parameter; The current behavioral event is decomposed according to the customer's intent, the target date of the event D_target is determined, and the time urgency TU(t) is calculated in combination with the current date D_now: TU(t) = max(0, 1 - (D_target - D_now) / D_cycle); Where D_cycle is the event preparation cycle; Based on behavioral signal scores and time urgency, calculate the customer's intent assessment value Intent(t): Intent(t)=σ(Σi=16(w_i×p_i(t))+α×BS(t)+β×TU(t)); Where σ is the sigmoid function, w_i is the intention contribution weight of the i-th profile dimension, α is the behavior rating weight coefficient, and β is the time urgency weight coefficient. The profile completion vector, intention assessment value, behavioral signal score, and time urgency are concatenated to form the context vector C_t: C_t=[prog(t),BS(t),TU(t),Intent(t),cus_tb,pain_pb]; Where cus_tb is the customer type embedding vector and pain_pb is the pain point type embedding vector.

[0014] By adopting the above technical solution, a profile vector is calculated based on the completion of multi-source data fields. The intention assessment value is generated by combining behavioral signal scores and time urgency with sigmoid weighting. The context vector is constructed by splicing customer type and pain point embedding, which improves the timeliness and accuracy of customer intention recognition.

[0015] This application further specifies: dividing the main process nodes according to the hierarchical finite state machine, and calculating the node transition probability in conjunction with the context vector, determining the target node, and triggering the sub-process, including: Based on the main process state machine in the hierarchical finite state machine and the dialogue stage, the main process nodes are grouped to obtain the stage node group; Based on the main process node and the context vector, calculate the node transition probability P() of candidate node n_j: P(n_j|n_i,C_t)=softmax_j(W_trans·φ(n_i,C_t)); Where W_trans is the transition weight matrix with dimension M×dim(φ), φ() is the joint state-context feature function; ΔIntent(t) is the intention drift increment; ΔIntent(t) = Intent(t) - Intent(t-1); The joint state-context feature function φ(n_i,C_t) is: φ(n_i,C_t)=[emb(n_i),C_t,prog(t),ΔIntent(t)]; Where, emb(n_i) is the node embedding vector of the current main process node; emb(n_i)=concat(BERT(t_i),agg_pre({emb(n_j)∣j∈pre(i)}) ,encode_cond(T_i)); Wherein, BERT is the semantic encoder, pre(i) is the set of preceding nodes, T is the preceding conditional logic expression, agg_pre is the self-attention mechanism, and encode_cond is the rule encoder; Sort the node transition probabilities in descending order and select the node with the maximum transition probability as the target node n_next=argmax_j(P(n_j|n_i,C_t)); If the intentional contribution weight of any dimension in the portrait completion vector is greater than the preset contribution threshold, but the portrait completion is less than the preset completion threshold, then the current dimension is determined to be the highest collection priority, and the node transfer probability is corrected to obtain the priority transfer probability priority_dim(t): priority_dim(t)=argmax_i(w_i×(1-p_i(t))); If the maximum transition probability is less than the path determinism threshold and the vertical domain feature salience is greater than the domain expertise threshold, then the sub-process state machine is triggered to answer the question of the target node, and returns to the main process node after the sub-process completes the answer. The saliency of vertical domain features is: A comprehensive analysis of the completion rates of all single dimensions is performed to calculate the overall profile completion score, PrS; PrS = 1 / 6∑ i=1 6 p_i(t); The behavioral signal scores, time urgency assessment, and intention assessment values ​​are analyzed to calculate the comprehensive behavioral intention score (BeS). BeS=(BS(t)+TU(t)+Intent(t)) / 3; Based on the preset typical domain feature vector v_domain, the customer type embedding vector and the pain point type embedding vector are fused and mapped, and the embedding semantic similarity SemS is calculated: SemS=1 / 2(cos(cus_tb,v_domain)+cos(pain_pb,v_domain)); The comprehensive score of the portrait completion, the comprehensive score of behavior intent, and the embedded semantic similarity are weighted and fused to obtain the vertical domain feature saliency domain_specificity(C_t): domain_specificity(C_t)=w1*PrS+w2*BeS+w3*SemS; Among them, w1 is the profile completion weight coefficient, w2 is the behavioral intent weight coefficient, and w3 is the professional field weight coefficient.

[0016] By adopting the above technical solution, based on hierarchical state machine and softmax probabilistic reasoning algorithm, and combining intention drift increment and priority correction strategy to dynamically select target nodes, and trigger vertical domain sub-process interaction when there is low determinism, intelligent guidance of dialogue path and in-depth answers to professional questions are realized, which significantly improves the logical consistency, task completion rate and user experience of dialogue system.

[0017] This application is further configured to: generate personalized text scripts based on the target node and context vector, according to the script retrieval score, including: Based on the target node, the customer's static data is parsed, the level assessment scores of each subject are extracted, and the customer's test score V_obs(t) is calculated. Based on the customer's scores in two consecutive tests and the learning rate α, calculate the customer's current capability vector V_current(t), and combine it with the customer's target requirement vector V_target(t) to obtain the GAP vector GAP(t). V_current(t)=α*V_obs(t)+(1−α)*V_prev(t−1); Where V_prev(t−1) is the customer test score from the previous test; GAP(t)=V_target(t)-V_current(t); Based on each dialogue s_k in the preset dialogue knowledge base, semantic matching is performed on the target node n_i, and the dialogue retrieval score Score(s_k) is calculated. Score(s_k)=α×Rel(s_k,n_i)+β×GapCov(s_k,GAP(t))+γ×CVR(s_k,c_j)+δ×TempFit(s_k,t); Where α is the scene weight coefficient, Rel(s_k,n_i) is the scene relevance, β is the gap filling weight coefficient, GapCov(s_k,GAP(t)) is the gap filling degree, γ is the conversion weight coefficient, CVR(s_k,c_j) is the historical conversion rate, c_j is the customer cluster group, δ is the time series weight coefficient, and TempFit(s_k,t) is the time series fit degree. GapCov(s_k,GAP(t))=Σ_d(g_d>0∧s_k.covers_d)×|g_d| / ‖GAP(t)‖; Where d is the information dimension index, g_d is the information gap component of GAP(t) in the d-th dimension, and s_k.covers_d is whether the utterance s_k covers the information gap component in the d-th dimension. TempFit(s_k,t)=cos_sim(emb_season(s_k),emb_season(t)); Where, emb_season(s_k) is the temporal semantic embedding vector of the utterance s_k, and emb_season(t) is the embedding vector of the current time node t; Based on the context vector and customer profile completion, the script is retrieved and scored, and the script template is rewritten to generate personalized scripts.

[0018] By adopting the above technical solution, the customer capability GAP vector is calculated and multi-factor weighted scoring of scenario relevance, GAP filling degree, historical conversion rate and time sequence adaptability is integrated. The script knowledge base is dynamically retrieved and the template is personalized by combining context and profile completion degree, which significantly improves the targeting of the script and the efficiency of filling information gaps.

[0019] This application is further configured to: after completing all compliant voice scripts in the dialogue, collect customer feedback signals, update the profile dimension weights and script effectiveness scores, including: After completing all compliant voice scripts, the content of customer communication is analyzed, customer feedback signals are collected, and a comprehensive feedback score (outcome_score) is calculated. Based on the comprehensive feedback score and the baseline score, the weights of the portrait dimensions are updated, and the weight increment of the portrait dimension Δw_i is calculated; Δw_i=η×(outcome_score-baseline)×∂Intent(t) / ∂w_i; The original script effectiveness score EffS(s_k,t) is updated based on the comprehensive feedback score to obtain the new script effectiveness score EffS(s_k,t+1); EffS(s_k,t+1)=(1-μ)×EffS(s_k,t)+μ×outcome_score; Where μ is the smoothing coefficient.

[0020] By adopting the above technical solution, a comprehensive score is calculated based on customer feedback signals after the conversation ends. The profile dimension weights (∂Intent / ∂w_i) are dynamically adjusted through gradient increments, and the speech effect score is updated exponentially. This achieves closed-loop adaptive optimization of the speech library and customer profile, significantly improving the personalization and conversion efficiency of subsequent interactions.

[0021] Secondly, this application also provides an intelligent sales dialogue navigation system based on hierarchical state machine and profile completion level, employing the following technical solution: A hierarchical state machine and profile completion-driven intelligent sales dialogue navigation system, comprising the following intelligent sales dialogue navigation methods: The profile assessment module is used to collect and clean multi-source customer data, construct a profile completion vector, and combine it with the intent assessment value to generate a context vector. The state machine engine module is used to divide the main process nodes according to the hierarchical finite state machine, and in combination with the context vector, determine the target node and trigger the sub-process, including: The main process state machine unit is used to maintain the set of main process nodes and node embedding vectors, and to calculate the state transition probability. The subprocess state machine unit is used to maintain the set of subprocess nodes, which is activated when the triggering condition is met and returns to the main process after completion; The path decision unit is used to select the optimal next node based on the state transition probability and to prioritize the completion of customer profile information according to the profile completion rate. The script generation module is used to generate personalized text scripts based on the target node and context vector, according to the script retrieval score. The quality control module is used to monitor personalized text messages based on a preset semantic classification model, filter compliant text messages, and output corresponding compliant voice messages. The feedback optimization module is used to collect customer feedback signals and update the profile dimension weights and script effectiveness scores after all compliant voice scripts have been completed. The storage module is used to store multi-source customer data, a dialogue knowledge base, node embedding vectors, transition weight matrices, and dialogue performance scores based on customer identifiers.

[0022] By adopting the above technical solution, the input verification module parses the order to construct the specification data block, the calculation control module triggers the straight wire calculation model according to the cutting method, and the linkage control unit automatically triggers the core calculation unit to run when the status indicator cell is not empty. The output cleanup module finally calculates the total usage and records it to the material preparation list, realizing the automated and modular linkage control of cutting material calculation, which significantly improves the order processing efficiency and calculation accuracy.

[0023] Thirdly, this application also provides an electronic device, comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement any of the methods in the above scheme.

[0024] Fourthly, this application also provides a storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described above.

[0025] In summary, the beneficial technical effects of this application are as follows: By using hierarchical finite state machine modeling, the system provides node-level path navigation and specific dialogue content within a single conversation, rather than simply recommending actions like "make a phone call" or "send an email." Compared to existing action-level NBA systems (such as Pega P×C×V×L), the recommendation granularity is improved by an order of magnitude. By coupling the state machine engine module and the dialogue generation module (using the output of state transition probability calculation as the scenario input for dialogue retrieval), joint recommendation based on "at which node + what content to say" is achieved. This solves the problem of separation between process management and dialogue recommendation, which are two independent systems in existing technologies; By using the profile completion vector as an explicit input to the intention evaluation formula and state transition probability calculation, the system can proactively guide the direction of the dialogue based on the lack of information (prioritizing the collection of high-weight, low-completion dimensions). By calculating the gap vector between a customer's current capabilities and target requirements, and using this as a scoring dimension for script retrieval, the generated scripts accurately cover the customer's actual gaps. Compared to keyword-triggered static templates, the targeting and personalization of the scripts are significantly improved. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the intelligent sales dialogue navigation method in this application; Figure 2 This is a diagram of the hierarchical finite state machine structure in this application; Figure 3 This is a flowchart of the context-aware speech generation process in this application; Figure 4 This is a schematic diagram of the portrait completion-driven dialogue path in this application; Figure 5 This is a schematic diagram of the intelligent sales dialogue navigation system in this application. Detailed Implementation

[0027] Reference Figure 1 The present application discloses an intelligent sales dialogue navigation method based on hierarchical state machine and profile completion level, comprising: S1: Collect and clean multi-source customer data, construct a profile completion vector, and combine it with the intent evaluation value to generate a context vector; S2: Divide the main process nodes according to the hierarchical finite state machine, and calculate the node transition probability in combination with the context vector to determine the target node and trigger the sub-process; S3: Based on the target node and context vector, generate personalized text scripts according to the script retrieval score; S4: Monitor personalized text messages based on a preset semantic classification model, filter compliant text messages, and output the corresponding compliant voice messages. S5: After completing all compliant voice scripts, collect customer feedback signals and update the profile dimension weights and script effectiveness scores.

[0028] In this embodiment, in the intelligent sales dialogue navigation application for educational institutions promoting children's programming courses, the system first integrates multi-source data such as CRM backend, trial class behavior tracking, and parent questionnaires to clean and integrate the child's age group, past competition participation, parents' professional background, and historical interaction duration. This constructs a profile completion vector containing dimensions such as "completion of programming basics," "intensity of competition intention," and "parents' technical awareness." Combined with the intention evaluation value based on login frequency and data download behavior (such as the high-value tag "completed three trial classes"), a context vector is generated that integrates short-term needs and long-term profiles. Subsequently, the hierarchical finite state machine calculates the node transition probability based on the main process nodes of "icebreaking - needs diagnosis - course matching - price objection handling - enrollment facilitation" and the context vector (if the parent has inquired about the competition exit and the child has a graphical foundation, the transition from "needs diagnosis" to "course matching" is highly probable). The target node is determined to be "price objection handling," and a sub-process state machine containing competitor comparison and installment payment instructions is triggered.

[0029] Based on the "price-sensitive" feature in the current node and context vector, templates such as "interest-free installments in three installments" and "examples of scholarships for outstanding students" are retrieved and scored from the script library. Personalized text scripts are generated by combining the parents' preferred evening communication time: "Many students who won provincial awards in the Blue Bridge Cup, like your child, choose to pay in three installments, less than 500 yuan per month." A preset RoBERTa-based semantic classification model monitors the script in real time, eliminating absolute and illegal expressions such as "guaranteed to win awards." After screening compliant text, a natural and compliant voice script is output through emotional synthesis TTS for course consultants to adopt. After completing all compliant voice script dialogues (such as when parents say "need to discuss with their child"), the system collects customer feedback signals (post-call SMS open rate, willingness to follow up the next day, and page dwell time). The weight of "decision cycle sensitivity" in the profile dimension is updated through incremental learning. At the same time, the effectiveness score of "discount incentive" scripts is reduced based on the script click-through conversion rate. Ultimately, the strategy evolves after each sales dialogue, increasing the programming course enrollment conversion rate by about 25%.

[0030] Preferably, step S1 includes: Detect and classify multi-source customer data, and extract customer static data and customer dynamic data; The static customer data is parsed according to the preset profile dimensions, and the single-dimensional completion degree p_i(t) is calculated, i∈{1,2,3,4,5,6}, p_i(t)∈[0,1]: p_i(t)=count(Retrieved_i) / count(total_i); Where i is the dimension number, t is the total dialogue time, Retrieved is the number of fields obtained, and total is the total number of fields; Correlate all single-dimensional completion scores to construct a profile completion vector prog(t)=[p_1(t),p_2(t),p_3(t),p_4(t),p_5(t),p_6(t)]; The customer dynamic data is analyzed to extract the current behavioral event, and combined with the behavioral weight value signal_k, the behavioral signal score BS(t) is calculated: BS(t) = Σ_k (signal_k) × exp(-λ × (t-t_k)); Where t_k is the occurrence time of the current event, and λ is the time decay parameter; The current behavioral event is decomposed according to the customer's intent, the target date of the event D_target is determined, and the time urgency TU(t) is calculated in combination with the current date D_now: TU(t) = max(0, 1 - (D_target - D_now) / D_cycle); Where D_cycle is the event preparation cycle; Based on behavioral signal scores and time urgency, calculate the customer's intent assessment value Intent(t): Intent(t)=σ(Σi=16(w_i×p_i(t))+α×BS(t)+β×TU(t)); Where σ is the sigmoid function, w_i is the intention contribution weight of the i-th profile dimension, α is the behavior rating weight coefficient, and β is the time urgency weight coefficient. The profile completion vector, intention assessment value, behavioral signal score, and time urgency are concatenated to form the context vector C_t: C_t=[prog(t),BS(t),TU(t),Intent(t),cus_tb,pain_pb]; Where cus_tb is the customer type embedding vector and pain_pb is the pain point type embedding vector.

[0031] In this embodiment, customer static data (such as child's grade, past programming experience, parents' occupation, budget range, etc.) and dynamic data (such as the number of times recently logged into the online programming platform, the frequency of browsing competition information, and the time of contacting customer service) are first collected and classified through CRM, trial behavior tracking, and questionnaires. Based on the profile dimensions (basic information, academic background, goal planning, time availability, payment ability, and competitive environment), the completion rate of each dimension is calculated using the formula p_i(t)=count(Retrieved_i) / count(total_i). For example, the number of fields in basic information that have been completed is 5 / 5 (such as name, age, school, contact). The profile completion vector prog(t) = [1.0, 0.67, 0.5, 0.5, 0.5, 0.0].

[0032] Simultaneously, analyze customer dynamic behavior: Clicking on the "Competition Exit" article within the last day (signal_k=0.9, t_k=1 day ago), watching the trial lesson replay 3 days ago (signal_k=0.6, t_k=3 days ago), and adding customer service WeChat 7 days ago (signal_k=0.8, t_k=7 days ago). Taking the time decay parameter λ=0.5, calculate the behavioral signal score BS(t)=0.9×exp(-0.5×1)+0.6×exp(-0.5×3)+0.8×exp(-0.5×7)≈0.9×0.6065+0.6×0.2231+0.8×0.0067≈0.5459+0.1339+0.0054=0.6 852; Further breakdown based on customer intention: The customer wants their child to participate in a city-level competition 3 months from now (D_target), current date D_now, preparation period D_cycle = 90 days, calculate the time urgency TU(t) = max(0, 1 - (90 - 0) / 90) = 1.0 (urgent); Set the intention contribution weight for the profile dimension w_i = [0.2, 0.25, 0.3, 0.1, 0.1, 0.05], the behavior rating weight α = 0.3, and the time urgency weight β = 0.4, then the intention evaluation value Intent(t) = σ(0.2 × 1.0 + 0.25 × 0.67 + 0.3 × 0.5 + 0.1 × 0.5 + 0.1 × 0.5 + 0.05 × 0.0). +0.3×0.6852+0.4×1.0)=σ(0.2+0.1675+0.15+0.05+0.05+0+0.2056+0.4)=σ(1.2231)=0.773.

[0033] Finally, the profile completion vector, behavioral signal score (0.6852), time urgency (1.0), intention (0.773), customer type embedding vector ("upper elementary school students preparing for the intensive exam") and pain point embedding vector ("lack of competition guidance") are concatenated to form a context vector C_t. This vector drives the state machine to transition from the "needs diagnosis" node to the "competition planning and class type recommendation" node with a high probability, and generates personalized wording: "Based on your child's preparation urgency and existing foundation, we recommend choosing the 3-month intensive class. Classes start next week, which can ensure that simulation training is completed before the competition, and early bird discounts are available for those who register this month," thereby significantly improving the course conversion rate.

[0034] Preferably, step S2 includes: Based on the hierarchical finite state machine (refer to...) Figure 2 The main process state machine in the process is combined with the dialogue phase to group the main process nodes, resulting in phase node groups; Based on the main process node and the context vector, calculate the node transition probability P() of candidate node n_j: P(n_j|n_i,C_t)=softmax_j(W_trans·φ(n_i,C_t)); Where W_trans is the transition weight matrix with dimension M×dim(φ), φ() is the joint state-context feature function; ΔIntent(t) is the intention drift increment; ΔIntent(t) = Intent(t) - Intent(t-1); The joint state-context feature function φ(n_i,C_t) is: φ(n_i,C_t)=[emb(n_i),C_t,prog(t),ΔIntent(t)]; Where, emb(n_i) is the node embedding vector of the current main process node; emb(n_i)=concat(BERT(t_i),agg_pre({emb(n_j)∣j∈pre(i)}) ,encode_cond(T_i)); Wherein, BERT is the semantic encoder, pre(i) is the set of preceding nodes, T is the preceding conditional logic expression, agg_pre is the self-attention mechanism, and encode_cond is the rule encoder; Sort the node transition probabilities in descending order and select the node with the maximum transition probability as the target node n_next=argmax_j(P(n_j|n_i,C_t)); If the intentional contribution weight of any dimension in the portrait completion vector is greater than the preset contribution threshold, but the portrait completion is less than the preset completion threshold, then the current dimension is determined to be the highest collection priority, and the node transfer probability is adjusted to obtain the priority transfer probability priority_dim(t): priority_dim(t)=argmax_i(w_i×(1-p_i(t))); If the maximum transition probability is less than the path determinism threshold and the vertical domain feature salience is greater than the domain expertise threshold, then the sub-process state machine is triggered to answer the question of the target node, and returns to the main process node after the sub-process completes the answer. The saliency of vertical domain features is: A comprehensive analysis of the completion rates of all single dimensions is performed to calculate the overall profile completion score, PrS; PrS = 1 / 6∑ i=1 6 p_i(t); The behavioral signal scores, time urgency assessment, and intention assessment values ​​are analyzed to calculate the comprehensive behavioral intention score (BeS). BeS=(BS(t)+TU(t)+Intent(t)) / 3; Based on the preset typical domain feature vector v_domain, the customer type embedding vector and the pain point type embedding vector are fused and mapped, and the embedding semantic similarity SemS is calculated: SemS=1 / 2(cos(cus_tb,v_domain)+cos(pain_pb,v_domain)); The comprehensive score of the portrait, the comprehensive score of behavioral intent, and the embedded semantic similarity are weighted and fused to obtain the domain-specificity (C_t) of the vertical domain features: domain_specificity(C_t)=w1*PrS+w2*BeS+w3*SemS; Among them, w1 is the profile completion weight coefficient, w2 is the behavioral intent weight coefficient, and w3 is the professional field weight coefficient.

[0035] In this embodiment, the main process nodes are first grouped according to the main process state machine in the hierarchical finite state machine (which includes five stages: "initial communication - background assessment - school selection and positioning - document guidance - visa tutoring") to form stage node groups. Based on the current node n_i (e.g., "background assessment") and the context vector C_t (including profile completion prog(t)=[0.9,0.7,0.6,0.5,0.8,0.4], behavioral signal score BS(t)=0.75, time urgency TU(t)=0.9, intention Intent(t)=0.82, customer type embedding vector "ordinary 211 aiming for Ivy League school", pain point embedding vector "lack of research experience"), the state-context joint feature function φ(n_i,C_t)=[emb(n_i), C_t, prog(t)] is used. ΔIntent(t)] (where ΔIntent(t)=0.02), and calculate the transition probability P(n_j|n_i,C_t)=softmax_j(W_trans·φ) of the candidate node n_j through the transition weight matrix W_trans to obtain the node probability distribution; school selection positioning 0.65, document guidance 0.20, visa tutoring 0.15, therefore the target node n_next is "school selection positioning".

[0036] At this point, the intention contribution weight w_i=0.3 of the "academic background" dimension in the profile completion vector is greater than the contribution threshold of 0.2, while its completion degree p_i=0.7 is less than the completeness threshold of 0.85. Therefore, the highest collection priority is determined to be academic background, and the transition probability is adjusted to the priority transition probability priority_dim(t)=argmax_i(0.3×(1-0.7))=0.09, so that the system prioritizes inquiring about research details. Since the current maximum transition probability of 0.65 is still greater than the path determinism threshold of 0.6, and the vertical domain feature salience domain_specificity(C_t)=w1*PrS+ w2*BeS+ The formula is w3*SemS, where PrS=0.65, BeS=(0.75+0.9+0.82) / 3≈0.82, and the semantic similarity of the embedding is SemS=0.88. Taking w1=0.3, w2=0.3, and w3=0.4, the significance is 0.3*0.65+0.3*0.82+0.4*0.88≈0.195+0.246+0.352=0.793, which is greater than the domain expertise threshold of 0.7. Therefore, the sub-process is not triggered, and the process is directly transferred to the "school selection and positioning" node. Then, personalized text is generated according to the dialogue: "Based on your current GPA of 3.6 and lack of publications, we suggest you consider Columbia University's statistics program as your target school and supplement it with a summer research experience." This completes the high-quality dialogue navigation and drives a 30% increase in the signing conversion rate.

[0037] Reference Figures 3-4 Preferably, step S3 includes: Based on the target node, the customer's static data is parsed, the level assessment scores of each subject are extracted, and the customer's test score V_obs(t) is calculated. Based on the customer's scores in two consecutive tests and the learning rate α, calculate the customer's current capability vector V_current(t), and combine it with the customer's target requirement vector V_target(t) to obtain the GAP vector GAP(t). V_current(t)=α*V_obs(t)+(1−α)*V_prev(t−1); Where V_prev(t−1) is the customer test score from the previous test; GAP(t)=V_target(t)-V_current(t); Based on each dialogue s_k in the preset dialogue knowledge base, semantic matching is performed on the target node n_i, and the dialogue retrieval score Score(s_k) is calculated. Score(s_k)=α×Rel(s_k,n_i)+β×GapCov(s_k,GAP(t))+γ×CVR(s_k,c_j)+δ×TempFit(s_k,t); Where α is the scene weight coefficient, Rel(s_k,n_i) is the scene relevance, β is the gap filling weight coefficient, GapCov(s_k,GAP(t)) is the gap filling degree, γ is the conversion weight coefficient, CVR(s_k,c_j) is the historical conversion rate, c_j is the customer cluster group, δ is the time series weight coefficient, and TempFit(s_k,t) is the time series fit degree. GapCov(s_k,GAP(t))=Σ_d(g_d>0∧s_k.covers_d)×|g_d| / ‖GAP(t)‖; Where d is the information dimension index, g_d is the information gap component of GAP(t) in the d-th dimension, and s_k.covers_d is whether the utterance s_k covers the information gap component in the d-th dimension. TempFit(s_k,t)=cos_sim(emb_season(s_k),emb_season(t)); Where, emb_season(s_k) is the temporal semantic embedding vector of the utterance s_k, and emb_season(t) is the embedding vector of the current time node t; Based on the context vector and customer profile completion, the script is retrieved and scored, and the script template is rewritten to generate personalized scripts.

[0038] In this embodiment, customer static data (such as mock test scores) is parsed according to the target node (such as “ability diagnosis and course matching”), and listening 7.5, reading 6.0, writing 5.5, and speaking 6.0 are extracted. The customer test score V_obs(t)=[7.5,6.0,5.5,6.0] is calculated; combined with the learning rate α=0.7 and the previous score V_prev(t-1)=[7.0,6.0,5.0,5.5].

[0039] Update the customer's current capability vector V_current(t) = 0.7 × [7.5, 6.0, 5.5, 6.0] + 0.3 × [7.0, 6.0, 5.0, 5.5] = [7.35, 6.0, 5.35, 5.85]. The customer's target requirement vector V_target(t) = [8.0, 7.0, 6.5, 6.5] (IELTS overall score 6.5 with a minimum of 6.0 in each section). Therefore, GAP(t) = [0.65, 1.0, 1.15, 0.65]. The system starts from the call... Two dialogue phrases were retrieved from the knowledge base: s1 focuses on "writing logic improvement skills," and s2 focuses on "listening comprehension training." The dialogue phrase retrieval scoring formula, Score(s_k) = α × Rel(s_k, n_i) + β × GapCov(s_k, GAP(t)) + γ × CVR(s_k, c_j) + δ × TempFit(s_k, t), was used to calculate the scores: α = 0.2, β = 0.4, γ = 0.3, δ = 0.1. The relevance of s1 to the node "ability diagnosis," Rel = 0. 0.9, GapCov calculation shows that the writing gap of 1.15 and the speaking gap of 0.65 account for the total gap norm (0.65²+1.0²+1.15²+0.65²)^0.5≈1.73, the coverage (1.15+0.65) / 1.73≈1.04, and the normalized GapCov=0.8; CVR(s1, customer cluster "working students aiming for 6.5 points")=0.25; Temporal fit: the current season is summer vacation, s1 contains the semantics of "intensive summer breakthrough", TempFit=0. 0.9; therefore, Score(s1) = 0.2 × 0.9 + 0.4 × 0.8 + 0.3 × 0.25 + 0.1 × 0.9 = 0.18 + 0.32 + 0.075 + 0.09 = 0.665; s2 correlation Rel = 0.7, GapCov only covers the hearing gap of 0.65, so 0.38, CVR = 0.18, TempFit = 0.6, therefore Score(s2) = 0.14 + 0.152 + 0.054 + 0.06 = 0.406; therefore, s1 is selected.

[0040] Based on the context vector (including time urgency and behavioral scores) and profile completion, the original script template, "We suggest you prioritize improving your writing and speaking skills. Our summer intensive course can help you improve your score in each subject by 1 point within 6 weeks," was rewritten as a personalized script: "Based on your July mock exam scores of 5.5 in writing and 6.0 in speaking, there is still a significant gap to your target of 6.5. Our summer 'Logic Enhancement + Topic Review' course perfectly matches your needs. Enroll now to enjoy an early bird discount, and we expect to complete the sprint before the September exam season." This effectively guides customers to enroll.

[0041] Preferably, step S4 includes: Load a pre-trained semantic classification model (such as a BERT-based text classifier) ​​that can identify sensitive content and violation categories (such as pornography, violence, political sensitivity, fraud, etc.) in the text. Also load a compliance rule library, including keyword blacklists, regular expression patterns, and industry standard clauses.

[0042] It receives personalized text scripts (such as invitation scripts, Q&A scripts, etc.) from the script generation module or manual input, with each script being passed in as a string.

[0043] The dialogue text undergoes preprocessing, including removing extra spaces, standardizing capitalization, converting between simplified and traditional Chinese characters, masking sensitive words (preserving the original text but using it for analysis), and word segmentation. This ensures the text format is suitable for the model input.

[0044] The preprocessed text is input into a semantic classification model, which outputs probability distributions for each category (e.g., probability of "compliant", probability of "illegal-pornographic"). Simultaneously, it performs matching checks against a compliance rule base (e.g., keyword hit, regular expression matching).

[0045] Based on the combined model probability and rule matching results, a decision is made according to a preset judgment strategy (e.g., any rule match results in a violation, or the model violation probability is >0.7 and a rule match results in a violation). Statements judged as compliant are added to the "Compliant Text Statement List," while statements judged as violations are recorded for the reason for the violation and discarded or marked for manual review.

[0046] For each compliant text message, extract the necessary information for speech synthesis: the original text content, optional emotion tags (such as "friendly" or "serious"), speech rate / volume presets, etc. If the system supports personalized speech (such as cloning a specific voice), also load the vocoder parameters.

[0047] The compliant text script and its synthesis parameters are sent to a TTS engine (such as Microsoft Azure TTS, Amazon Polly, or locally deployed Tacotron2 / WaveGlow). The engine converts the text into an audio stream (such as WAV or MP3 format) and returns the audio file or stream address. The system supports multiple voices and multiple languages.

[0048] Post-processing is performed on the generated audio, such as volume normalization, noise reduction, and muting the beginning and end. Manual sampling or automatic scoring models (such as MOS scoring) can be added to evaluate speech naturalness and clarity. If the quality is unsatisfactory, the TTS engine is switched or parameters are adjusted for regeneration.

[0049] The compliant text script, its corresponding audio file address, review results (including model score and hit rules), generation time, and other information are stored in the database and associated with the user session ID or business scenario.

[0050] The final list of compliant text scripts and their corresponding audio (or URL) is returned to the caller (such as a conversation manager or outbound calling system) for use in the next round of customer interaction. It can also be pushed to a monitoring dashboard for operations personnel to view compliance statistics.

[0051] Preferably, step S5 includes: After completing all compliant voice scripts, the content of customer communication is analyzed, customer feedback signals are collected, and a comprehensive feedback score (outcome_score) is calculated. Based on the comprehensive feedback score and the baseline score, the weights of the portrait dimensions are updated, and the weight increment of the portrait dimension Δw_i is calculated; Δw_i=η×(outcome_score-baseline)×∂Intent(t) / ∂w_i; The original script effectiveness score EffS(s_k,t) is updated based on the comprehensive feedback score to obtain the new script effectiveness score EffS(s_k,t+1); EffS(s_k,t+1)=(1-μ)×EffS(s_k,t)+μ×outcome_score; Where μ is the smoothing coefficient.

[0052] In this embodiment, in an intelligent sales dialogue system for a children's programming course, after completing all compliant voice conversations, the system comprehensively scores the parents' feedback, including their emotional tone during the call, the depth of their questions, and whether they clicked the payment link the next day. The comprehensive feedback score is output_score = 0.85 (out of 1.0). The baseline is set to 0.7, and the learning rate η = 0.1. The increment of each dimension is calculated using the profile dimension weight update formula Δw_i = η × (outcome_score - baseline) × ∂Intent(t) / ∂w_i. The intention gradient ∂Intent / ∂w_pay for the "payment ability" dimension is 0.6, so the increment Δw_pay = 0.1 × (0.85 - 0.7) × 0.6 = 0.009, increasing the weight from 0.10 to 0.109. The gradient for the "basic information" dimension is only 0.1, with an increment of 0.0015 and the weight remaining basically unchanged, making the system place greater emphasis on data collection for the payment ability dimension in the future. Meanwhile, for the sales pitch s_k used in this instance (such as "competition exit case"), the original sales pitch effectiveness score EffS(s_k,t)=0.72 was updated with a smoothing coefficient μ=0.3 to a new score EffS(s_k,t+1)=(1-0.3)×0.72+0.3×0.85=0.504+0.255=0.759, thus increasing the priority of this sales pitch. After dozens of interactions, the system adaptively optimized the profile weight distribution and sales pitch ranking, resulting in a 25% increase in course sales conversion rate.

[0053] Reference Figure 5 A hierarchical state machine and profile completion-driven intelligent sales dialogue navigation system, applied to the intelligent sales dialogue navigation method, includes: The profile assessment module is used to collect and clean multi-source customer data, construct a profile completion vector, and combine it with the intent assessment value to generate a context vector. The state machine engine module is used to divide the main process nodes according to the hierarchical finite state machine, and in combination with the context vector, determine the target node and trigger the sub-process, including: The main process state machine unit is used to maintain the set of main process nodes and node embedding vectors, and to calculate the state transition probability. The subprocess state machine unit is used to maintain the set of subprocess nodes, which is activated when the triggering condition is met and returns to the main process after completion; The path decision unit is used to select the optimal next node based on the state transition probability and to prioritize the completion of customer profile information according to the profile completion rate. The script generation module is used to generate personalized text scripts based on the target node and context vector, according to the script retrieval score. The quality control module is used to monitor personalized text messages based on a preset semantic classification model, filter compliant text messages, and output corresponding compliant voice messages. The feedback optimization module is used to collect customer feedback signals and update the profile dimension weights and script effectiveness scores after all compliant voice scripts have been completed. The storage module is used to store multi-source customer data, a dialogue knowledge base, node embedding vectors, transition weight matrices, and dialogue performance scores based on customer identifiers.

[0054] In this embodiment, the script generation module includes: a GAP analysis unit: calculating the gap vector between the customer's current capabilities and target requirements; a script retrieval unit: ranking script retrieval based on a four-dimensional score of scenario relevance, GAP filling degree, historical conversion rate, and time-series adaptability; and a personalized rewriting unit: integrating vertical domain knowledge to rewrite the retrieved scripts in a personalized manner and output structured keyword content.

[0055] An electronic device, comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement any of the methods in the above scheme.

[0056] A storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the intelligent sales dialogue navigation method as described above.

[0057] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent sales dialogue navigation based on hierarchical state machine and profile completion, characterized in that, include: Collect and clean multi-source customer data, construct a profile completion vector, and combine it with intent assessment values ​​to generate a context vector; The main process nodes are divided according to the hierarchical finite state machine, and the node transition probability is calculated by combining the context vector to determine the target node and trigger the sub-process. Based on the target node and the context vector, personalized text scripts are generated according to the script retrieval score. The personalized text scripts are monitored according to the preset semantic classification model, compliant text scripts are selected, and corresponding compliant voice scripts are output. After completing all the compliant voice scripts, collect customer feedback signals and update the profile dimension weights and script effectiveness scores.

2. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 1, characterized in that, The collected and cleaned multi-source customer data is used to construct a profile completion vector, and combined with the intent evaluation value, a context vector is generated, including: Detect and classify multi-source customer data, and extract customer static data and customer dynamic data; The customer static data is parsed according to the preset profile dimensions, and the single-dimensional completion degree p_i(t) is calculated, i∈{1,2,3,4,5,6}, p_i(t)∈[0,1]: p_i(t)=count(Retrieved_i) / count(total_i); Where i is the dimension number, t is the total dialogue time, Retrieved is the number of fields obtained, and total is the total number of fields; The single-dimensional completion scores are correlated to construct a profile completion score vector prog(t)=[p_1(t),p_2(t),p_3(t),p_4(t),p_5(t),p_6(t)]; The customer dynamic data is parsed to extract the current behavioral event, and combined with the behavioral weight value signal_k, the behavioral signal score BS(t) is calculated: BS(t) = Σ_k (signal_k) × exp(-λ × (t-t_k)); Where t_k is the occurrence time of the current event, and λ is the time decay parameter; The current behavioral event is decomposed based on the customer's intent to determine the event target date D_target, and combined with the current date D_now, the time urgency TU(t) is calculated: TU(t)=max(0,1-(D_target-D_now) / D_cycle); Where D_cycle is the event preparation cycle; Based on the behavioral signal score and the time urgency, calculate the customer's intention assessment value Intent(t): Intent(t)=σ(Σ i=1 6 (w_i×p_i(t))+α×BS(t)+β×TU(t)); Where σ is the sigmoid function, w_i is the intention contribution weight of the i-th profile dimension, α is the behavior rating weight coefficient, and β is the time urgency weight coefficient. The profile completion vector, intention assessment value, behavioral signal score, and time urgency are concatenated to form the context vector C_t: C_t=[prog(t),BS(t),TU(t),Intent(t),cus_tb,pain_pb]; Where cus_tb is the customer type embedding vector and pain_pb is the pain point type embedding vector.

3. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 1, characterized in that, The step of dividing the main process nodes according to the hierarchical finite state machine, calculating the node transition probability in conjunction with the context vector, determining the target node, and triggering the sub-process includes: Based on the main process state machine in the hierarchical finite state machine and the dialogue stage, the main process nodes are grouped to obtain the stage node group; Based on the main process node and the context vector, calculate the node transition probability P() of candidate node n_j: P(n_j|n_i,C_t)=softmax_j(W_trans·φ(n_i,C_t)); Where W_trans is the transition weight matrix with dimension M×dim(φ), φ() is the joint state-context feature function; ΔIntent(t) is the intention drift increment; ΔIntent(t) = Intent(t) - Intent(t-1); Sort the node transition probabilities in descending order and select the node with the maximum transition probability as the target node n_next=argmax_j(P(n_j|n_i,C_t)); If the intentional contribution weight of any dimension in the portrait completion vector is greater than the preset contribution threshold, but the portrait completion is less than the preset completion threshold, then the current dimension is determined to be the highest collection priority, and the node transfer probability is corrected to obtain the priority transfer probability priority_dim(t): priority_dim(t)=argmax_i(w_i×(1-p_i(t))); If the maximum transition probability is less than the path determinism threshold and the vertical domain feature salience is greater than the domain expertise threshold, then the sub-process state machine is triggered to answer the question of the target node, and returns to the main process node after the sub-process completes the answer.

4. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 3, characterized in that, The state-context joint feature function φ(n_i,C_t) is: φ(n_i,C_t)=[emb(n_i),C_t,prog(t),ΔIntent(t)]; Where, emb(n_i) is the node embedding vector of the current main process node; emb(n_i)=concat(BERT(t_i),agg_pre({emb(n_j)∣j∈pre(i)}),encode_cond(T_i)); Wherein, BERT is the semantic encoder, pre(i) is the set of preceding nodes, T is the preceding conditional logical expression, agg_pre is the self-attention mechanism, and encode_cond is the rule encoder.

5. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 3, characterized in that, The saliency of the vertical domain features is: A comprehensive analysis of the completion rates of all single dimensions is performed to calculate the overall profile completion score, PrS; PrS = 1 / 6∑ i=1 6 p_i(t); The behavioral signal scores, time urgency assessment, and intention assessment values ​​are analyzed to calculate the comprehensive behavioral intention score (BeS). BeS=(BS(t)+TU(t)+Intent(t)) / 3; Based on the preset typical domain feature vector v_domain, the customer type embedding vector and the pain point type embedding vector are fused and mapped, and the embedding semantic similarity SemS is calculated: SemS=1 / 2(cos(cus_tb,v_domain)+cos(pain_pb,v_domain)); The comprehensive score of the portrait completion, the comprehensive score of behavior intent, and the embedded semantic similarity are weighted and fused to obtain the vertical domain feature saliency domain_specificity(C_t): domain_specificity(C_t)=w1*PrS+w2*BeS+w3*SemS; Among them, w1 is the profile completion weight coefficient, w2 is the behavioral intent weight coefficient, and w3 is the professional field weight coefficient.

6. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 1, characterized in that, The step of generating personalized text scripts based on the target node and the context vector, according to the script retrieval score, includes: Based on the target node, the customer's static data is parsed, the level assessment scores of each subject are extracted, and the customer's test score V_obs(t) is calculated. Based on the customer's scores in two consecutive tests and the learning rate α, calculate the customer's current capability vector V_current(t), and combine it with the customer's target requirement vector V_target(t) to obtain the GAP vector GAP(t). V_current(t)=α*V_obs(t)+(1−α)*V_prev(t−1); Where V_prev(t−1) is the customer test score from the previous test; GAP(t)=V_target(t)-V_current(t); Based on each dialogue s_k in the preset dialogue knowledge base, semantic matching is performed on the target node n_i, and the dialogue retrieval score Score(s_k) is calculated: Score(s_k)=α×Rel(s_k,n_i)+β×GapCov(s_k,GAP(t))+γ×CVR(s_k,c_j)+δ×TempFit(s_k,t); Where α is the scene weight coefficient, Rel(s_k,n_i) is the scene relevance, β is the gap filling weight coefficient, GapCov(s_k,GAP(t)) is the gap filling degree, γ is the conversion weight coefficient, CVR(s_k,c_j) is the historical conversion rate, c_j is the customer cluster group, δ is the time series weight coefficient, and TempFit(s_k,t) is the time series fit degree. GapCov(s_k,GAP(t))=Σ_d(g_d>0∧s_k.covers_d)×|g_d| / ‖GAP(t)‖; Where d is the information dimension index, g_d is the information gap component of GAP(t) in the d-th dimension, and s_k.covers_d is whether the utterance s_k covers the information gap component in the d-th dimension. TempFit(s_k,t)=cos_sim(emb_season(s_k),emb_season(t)); Where, emb_season(s_k) is the temporal semantic embedding vector of the utterance s_k, and emb_season(t) is the embedding vector of the current time node t; Based on the context vector and customer profile completion, the script is retrieved and scored, and the script template is rewritten to generate personalized scripts.

7. The intelligent sales dialogue navigation method based on hierarchical state machine and profile completion as described in claim 1, characterized in that, After completing all the compliant voice prompts in the dialogue, customer feedback signals are collected to update the profile dimension weights and the prompt effectiveness score, including: After completing all the compliant voice scripts in the dialogue, the customer communication content is analyzed, customer feedback signals are collected, and a comprehensive feedback score (outcome_score) is calculated. Based on the comprehensive feedback score and the baseline score, the weights of the portrait dimensions are updated, and the weight increment of the portrait dimension Δw_i is calculated; Δw_i=η×(outcome_score-baseline)×∂Intent(t) / ∂w_i; The original speech effect score EffS(s_k,t) is updated based on the comprehensive feedback score to obtain the new speech effect score EffS(s_k,t+1); EffS(s_k,t+1)=(1-μ)×EffS(s_k,t)+μ×outcome_score; Where μ is the smoothing coefficient.

8. A hierarchical state machine and profile completion-driven intelligent sales dialogue navigation system, used to implement the intelligent sales dialogue navigation method as described in any one of claims 1-7, characterized in that, include: The profile assessment module is used to collect and clean multi-source customer data, construct a profile completion vector, and combine it with the intent assessment value to generate a context vector. The state machine engine module is used to divide the main process nodes according to the hierarchical finite state machine, and determine the target node and trigger the sub-process by combining the context vector. The script generation module is used to generate personalized text scripts based on the target node and the context vector, according to the script retrieval score. The quality control module is used to monitor the personalized text scripts according to the preset semantic classification model, filter compliant text scripts, and output the corresponding compliant voice scripts. The feedback optimization module is used to collect customer feedback signals and update the profile dimension weights and script effectiveness scores after all the compliant voice scripts have been completed. The storage module is used to store multi-source customer data, a dialogue knowledge base, node embedding vectors, transition weight matrices, and dialogue performance scores based on customer identifiers.

9. The intelligent sales dialogue navigation system based on hierarchical state machine and profile completion as described in claim 1, characterized in that, The state machine engine module includes: The main process state machine unit is used to maintain the set of main process nodes and node embedding vectors, and to calculate the state transition probability. The subprocess state machine unit is used to maintain the set of subprocess nodes, which is activated when the triggering condition is met and returns to the main process after completion; The path decision unit is used to select the optimal next node based on the state transition probability and to prioritize the completion of customer profile information according to the profile completion rate.

10. A storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the intelligent sales dialogue navigation method as described in any one of claims 1 to 7.

Images