Intelligent diagnostic methods, devices, equipment, and media for cross-border tourism sales dialogue

Abstract

This application provides an intelligent diagnostic method, apparatus, equipment, and medium for cross-border tourism sales dialogues, effectively solving the problem of the lack of intelligent diagnostic and early warning technologies with high accuracy, strong attribution capabilities, real-time performance, and cost controllability in the existing cross-border tourism service field. The method includes: collecting raw dialogue data and preprocessing it to obtain dialogue data; dynamically generating target prompt text based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget; extracting multi-dimensional scoring vectors and attribution information sets from the target prompt text, and performing dynamic scoring fusion and two-stage attribution analysis based on the multi-dimensional scoring vectors and attribution information sets to obtain the order cancellation probability and attribution analysis results of the order to be traded; triggering tiered early warnings based on the order cancellation probability and attribution analysis results, and taking corresponding early warning processing measures to intelligently diagnose the cross-border tourism sales dialogue.

Description

Technical Field

[0001] This application relates to, specifically, intelligent diagnostic methods, devices, equipment, and media for cross-border tourism sales dialogues. Background Technology

[0002] In the cross-border tourism service sector, sales dialogue is a core link connecting clients and consultants, facilitating order conversion. However, this dialogue faces several unique challenges: First, significant cross-language and cultural differences lead to substantial discrepancies in clients' expression habits and needs, increasing the difficulty of understanding their requirements. Second, communication is asynchronous across time zones, with an average dialogue cycle of up to 42 days, during which client needs may dynamically change, making it difficult for consultants to accurately control the follow-up pace. Third, the order decision-making process is complex, involving professional steps such as screening multiple POIs (Points of Interest) in various countries, calculating quotes in multiple currencies, and matching resources from multiple suppliers, demanding extremely high comprehensive abilities from consultants. These factors result in a persistently high loss rate for high-value cross-border tourism orders (average order value of 15,000 RMB), severely impacting the overall profitability of the industry. To address the issues of risk identification and order retention in sales dialogues, existing technologies have primarily developed three types of solutions: The first category is rule-based keyword matching methods. This method uses a pre-set keyword library (e.g., price-related terms like "expensive" and "cheap"), order status-related terms like "cancellation" and "refund"), and a rule tree to count the number of keyword hits in the dialogue and accumulate a risk score. An alert is triggered when the score exceeds a preset threshold. The second category is shallow machine learning classification methods. This approach uses TF-IDF or Word2Vec techniques to extract features from the dialogue text, converts them into fixed-dimensional feature vectors, and trains classifiers such as SVM or random forests. It learns the mapping relationship between dialogue status and order results through labeled data such as "completed," "cancelled," and "follow-up," thereby classifying and predicting new dialogues. The third category is solutions that directly call general-purpose large language models. This method inputs the full dialogue text and free text prompts, such as GPT-4, to determine whether a customer will cancel an order, relying on the model's natural language understanding capabilities to output the analysis results.

[0003] The three types of solutions share common core flaws: First, they lack attribution and localization capabilities, only able to determine whether an order will be cancelled, but unable to accurately pinpoint the reasons for cancellation, key issues, responsible parties, and other core information, making it difficult for consultants to take effective intervention measures; second, they lack real-time performance, with keyword matching methods being fast but having low accuracy, and machine learning and general large model methods having relatively high accuracy but taking a long time to analyze, failing to meet the real-time early warning needs of asynchronous communication scenarios in cross-border tourism; third, they lack dynamic optimization mechanisms, failing to establish a closed loop of analysis results, business feedback, and parameter adjustments, preventing the system performance from continuously improving with the accumulation of business data.

[0004] Therefore, there is an urgent need for an intelligent diagnostic and early warning technology that can adapt to the characteristics of cross-border tourism sales dialogue scenarios and has high accuracy, strong attribution ability, real-time performance, and cost controllability, in order to solve the pain points of existing technologies, reduce the loss rate of high-value orders, and improve the overall service quality and profitability of the industry. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide intelligent diagnostic methods, devices, equipment and media for cross-border tourism sales dialogues, which effectively solves the problem that the existing cross-border tourism service field lacks intelligent diagnostic and early warning technologies with high accuracy, strong attribution ability, real-time performance and cost controllability.

[0006] In a first aspect, embodiments of this application provide an intelligent diagnostic method for cross-border tourism sales dialogues, the method comprising: Raw dialogue data is collected from various data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and the raw dialogue data is preprocessed to obtain dialogue data. Based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, multiple paragraph texts are dynamically generated. The multiple paragraph texts are then concatenated according to the preset paragraph order and the corresponding token budget to generate the target prompt text. Extract the multi-dimensional rating vector and attribution information set from the target prompt text, and perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded; Based on the order cancellation probability and attribution analysis results, a tiered early warning is triggered, and corresponding early warning processing methods are adopted to intelligently diagnose the cross-border tourism sales dialogue.

[0007] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein dynamic score fusion and two-stage attribution analysis are performed based on the multi-dimensional score vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded, including: The order feature vector, customer profile vector, and dialogue stage vector are extracted from the dialogue data respectively, and then fused to form a dynamic weight vector. The overall score is calculated based on the dynamic weight vector and the multi-dimensional scoring vector. The overall score is mapped to the order cancellation probability, and the attribution analysis result is obtained based on the order cancellation probability.

[0008] In conjunction with the first aspect, this application provides a second possible implementation of the first aspect, wherein obtaining the attribution analysis result based on the order cancellation probability includes: A linear evaluation model is constructed from the dialogue completeness, frequency of dangerous words, response timeliness, and cancellation probability extracted from the dialogue data. Based on the linear evaluation model, the dialogue type corresponding to the dialogue data is quickly screened out, and it is determined whether the dialogue type is the target dialogue type. If so, key nodes in the time series are detected to perform in-depth attribution of the cross-border tourism sales dialogue, and the attribution analysis results are combined with the attribution information set to form a visually interpretable explanation.

[0009] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein deep attribution of the cross-border tourism sales dialogue based on key time series nodes includes: Calculate the sentiment score of each message in the customer message sequence to form a sentiment vector, calculate the sentiment vector to obtain multiple moving averages, and fit a target curve to mark the sentiment inflection point; Calculate the weighted deviation between the constructed initial customer demand vector and the solution attribute vector provided by the consultant, and mark the demand deviation point when the weighted deviation exceeds a preset deviation threshold; A feature function is constructed using a price-sensitive dictionary to calculate the price signal strength of each message, and price conflict points are marked based on the price signal strength. Calculate the time difference between adjacent messages and compare the time difference with a silence threshold set according to the regions where the customer and consultant are located, to mark the silence period.

[0010] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein multiple paragraph texts are dynamically generated based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of a preset large language model, including: The complexity score of the dialogue data is calculated to determine the scenario type of the cross-border tourism sales dialogue, so as to dynamically select the number of segments of the prompt template based on the scenario type; Traverse the entity set extracted from the dialogue data in the domain knowledge base to query the corresponding knowledge entries; Based on the knowledge entries and the total token budget, multiple paragraph texts are dynamically generated.

[0011] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein the step of dynamically generating multiple paragraph texts based on the knowledge entries and the total token budget includes: Calculate the cosine similarity between the dialogue context vector of the dialogue data and the knowledge entry vector of the knowledge entry, and add the corresponding knowledge entry to the injection set based on the similarity; The total token budget is allocated to the corresponding prompt paragraph based on the preset priority weight of each prompt template paragraph; The injected set is then merged into the corresponding paragraph template to generate multiple paragraph texts.

[0012] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein, after preprocessing the original dialogue data to obtain dialogue data, the following steps are taken: In a pre-established historical dialogue vector library, similar dialogue matching is performed on the dialogue data to obtain the matching results; Based on the matching results, the stored historical analysis results are retrieved to perform a two-stage attribution analysis.

[0013] Secondly, embodiments of this application provide an intelligent diagnostic device for cross-border tourism sales dialogues, the device comprising: The data acquisition module is used to collect raw dialogue data from various data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and to preprocess the raw dialogue data to obtain dialogue data. The splicing module is used to dynamically generate multiple paragraph texts based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, and splice the multiple paragraph texts according to the preset paragraph order and the corresponding token budget to generate the target prompt text; The analysis module is used to extract the multi-dimensional rating vector and attribution information set from the target prompt text, and to perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded. The diagnostic module is used to trigger tiered warnings based on the order cancellation probability and attribution analysis results, and to take corresponding warning processing measures to intelligently diagnose the cross-border tourism sales dialogue.

[0014] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of any of the intelligent diagnostic methods for cross-border tourism sales dialogues described in the present invention.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any of the intelligent diagnostic methods for cross-border tourism sales dialogues.

[0016] This application provides an intelligent diagnostic method for cross-border tourism sales dialogues. The method first collects raw dialogue data from various data sources of cross-border tourism sales dialogues generated from pending orders, and preprocesses the raw dialogue data to obtain dialogue data. Second, based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of a preset large language model, multiple paragraph texts are dynamically generated. These paragraph texts are then concatenated according to a preset paragraph order and corresponding token budgets to generate target prompt text. Next, multi-dimensional scoring vectors and attribution information sets are extracted from the target prompt text. Dynamic scoring fusion and two-stage attribution analysis are performed based on these multi-dimensional scoring vectors and attribution information sets to obtain the order cancellation probability and attribution analysis results for pending orders. Finally, based on the order cancellation probability and attribution analysis results, a tiered warning is triggered, and corresponding warning processing methods are adopted to intelligently diagnose the cross-border tourism sales dialogue. Based on the above methods, this application not only provides intelligent diagnostic and early warning technology that adapts to the characteristics of cross-border tourism sales dialogue scenarios while possessing high accuracy, strong attribution ability, real-time performance, and cost controllability, but also significantly improves diagnostic accuracy and risk identification capabilities, greatly enhances real-time performance and early warning timeliness in sales dialogues, successfully increases order recovery rate, and significantly reduces costs. It solves the limitations of existing technologies in terms of accuracy, attribution ability, real-time performance, and cost, and is feasible for large-scale promotion. It has strong adaptability and scalability, wide applicability, and broad application prospects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an intelligent diagnostic method for cross-border tourism sales dialogue provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating the call history analysis results provided in an embodiment of this application is shown; Figure 3 This illustration shows a flowchart of obtaining multiple paragraph texts according to an embodiment of this application; Figure 4 This diagram illustrates a structural block diagram of an intelligent diagnostic device for cross-border tourism sales dialogue provided in an embodiment of this application. Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0020] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0022] The existing technology field of cross-border tourism lacks an intelligent diagnostic and early warning technology that can adapt to the characteristics of cross-border tourism sales dialogue scenarios and has high accuracy, strong attribution ability, real-time performance, and cost controllability, in order to solve the pain points of existing technologies, reduce the loss rate of high-value orders, and improve the overall service quality and profitability of the industry.

[0023] Based on this, embodiments of this application provide intelligent diagnostic methods, devices, equipment, and media for cross-border tourism sales dialogues, which are described below through embodiments.

[0024] Example 1 To facilitate understanding of this embodiment, a detailed description of the intelligent diagnostic method for cross-border tourism sales dialogue disclosed in this application embodiment will be provided first. For example... Figure 1 The diagram shows a flowchart of an intelligent diagnostic method for cross-border tourism sales dialogues. This application provides an intelligent diagnostic method for cross-border tourism sales dialogues, the method comprising: S101. Collect raw dialogue data from multiple data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and preprocess the raw dialogue data to obtain dialogue data. S102. Based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, multiple paragraph texts are dynamically generated, and the multiple paragraph texts are concatenated according to the preset paragraph order and the corresponding token budget to generate the target prompt text. S103. Extract the multi-dimensional rating vector and attribution information set from the target prompt text, and perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded. S104. Based on the order cancellation probability and attribution analysis results, trigger a tiered warning and take corresponding warning processing measures to intelligently diagnose the cross-border tourism sales dialogue.

[0025] The method described in this application can be applied to an intelligent diagnostic system for cross-border tourism sales dialogues. The intelligent diagnostic system includes a data acquisition module, a prompt generation module, a large model invocation module, a result analysis module, a score fusion module, an attribution analysis module, an early warning intervention module, and a feedback optimization module. The output of the data acquisition module is connected to the input of the prompt generation module; the output of the prompt generation module is connected to the input of the large model invocation module; the output of the large model invocation module is connected to the input of the result analysis module; the output of the result analysis module is simultaneously connected to both the score fusion module and the attribution analysis module; the outputs of the score fusion module and the attribution analysis module are connected to the input of the early warning intervention module; the output of the early warning intervention module is connected to the feedback optimization module; and the output of the feedback optimization module is connected in reverse to the prompt generation module, the score fusion module, and the early warning intervention module, forming a closed-loop optimization.

[0026] In step S101, this application pre-sets multiple data sources for cross-border tourism sales dialogues in the data acquisition module. These multiple data sources include a message system, an order system, a user system, and a channel system. Raw dialogue data is collected from these multiple data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders. Specifically, message sequences M, including message content, direction, and timestamps, are extracted from the message system (e.g., an IM message library); order attributes O are extracted from the order system; customer profiles C are extracted from the user system; and channel attributes, promotional association information, channel interaction trajectories, and channel attribute configurations that identify the dialogue are extracted from the channel system. The raw dialogue data is then preprocessed to obtain dialogue data. Specific preprocessing includes data cleaning, merging, and desensitization operations. For example, data cleaning includes filtering system notification messages: M = {m i : p i≠System}, merge SMS messages: if t i+1 - t i <5 minutes and p i = p i+1 Then, the data is merged, and sensitive information is desensitized based on the voice / image to text description method, thereby obtaining the dialogue data D = (M, T, P, O, C).

[0027] This application defaults to processing Chinese translated by the preprocessing module, but it can be replaced with: end-to-end multilingual analysis: directly inputting the original multilingual dialogue into a large model that supports multiple languages, and having the model perform cross-language semantic understanding internally; hybrid translation strategy: only key entities and sales phrases are manually proofread and translated, while the rest are machine translated; localized language model: using a localized model finely tuned on the corpus of the language for specific purposes (such as Japanese or Thai). In the specific implementation of step S101, one embodiment is as follows: Figure 2 As shown, after preprocessing the original dialogue data to obtain the dialogue data, the process includes: S1011. Perform similar dialogue matching on the dialogue data in the pre-established historical dialogue vector library to obtain the matching result; S1012. Based on the matching results, call the stored historical analysis results to perform a two-stage attribution analysis.

[0028] In steps S1011-S1012, this application also establishes a historical dialogue vector library, which is constructed based on historical cross-border tourism sales dialogues. After obtaining the dialogue data, the dialogue text of the dialogue data is truncated and concatenated, and then transformed into high-dimensional vectors through an embedding model. Then the high-dimensional vector This refers to the vector fingerprint of the dialogue data. In the historical dialogue vector database, an ANN algorithm is used to quickly perform similar dialogue matching on the dialogue text to obtain the matching results. These matching results retrieve the Top-K candidate similar dialogues and calculate... Exact cosine similarity of candidate vectors to candidate similar dialogues , Dialogue vectors Candidate vectors If similarity exists Historical records that meet timeliness requirements (e.g., within 7 days) and scenario consistency constraints are considered cache hits. If the similarity threshold is not met, the historical analysis results stored in the cache library of the large model calling module are called, and the prompt generation module, large model calling module, result parsing module, and scoring fusion module are skipped. Instead, the historical analysis results are directly input into the attribution analysis module for two-stage attribution analysis. If no similarity is found... If the historical records meet the timeliness requirements (e.g., within 7 days) and scenario consistency constraints, then according to the method described in this application, the dialogue data is input to the prompt generation module to execute the entire method for full evaluation and write the corresponding evaluation results and their vector fingerprints into the cache library managed by the LRU strategy.

[0029] The cache hit rate of this application for similar dialogue cache reuse is approximately 20%, and the cost when a hit occurs is 0 yuan (no API call required). The overall cost is further reduced to 0.19 × 0.8 = 0.152 yuan per dialogue.

[0030] Domain knowledge (POI, price, supplier) is pre-encoded as a vector cache to avoid repeatedly injecting the same content, reducing token consumption by about 15%.

[0031] In step S102, this application calculates the complexity score of the dialogue data based on the prompt generation module to determine the dialogue scenario in which the dialogue data is located, and dynamically generates multiple paragraph texts based on the dialogue scenario in which the dialogue data is located and the total token budget of the pre-set Large Language Model (LLM) in the Large Model Calling module, wherein the total token budget... The system initializes with a fixed configuration based on the type of the model being called, and stores this configuration in the parameter configuration library of the prompt generation module. It does not require dynamic calculation for individual dialogues; different models correspond to different fixed configurations. The prompt generation module includes a five-segment template library. Domain knowledge base The system comprises a dynamic splicing engine and a total token budget allocator. The five-segment template library stores templates for role positioning, information injection, semantic decomposition, evaluation frameworks, and format constraints. The domain knowledge base stores POI information, price benchmarks, supplier capabilities, and time zone information. The dynamic splicing engine queries the knowledge base based on the dialogue content and generates complete prompts. The total token budget allocator allocates the total token budget for each segment according to the Large Language Model (LLM) constraints. It then splices multiple segment texts according to the preset segment order determined by the scenario type, and according to the segment order and their corresponding token budgets to generate the target prompt text. Different segment texts correspond to different priorities, and the total token budget of the LLM is allocated based on the segment text priorities. For segment texts with mandatory format constraints, the original length budget is allocated. For variable-length segment texts, such as chat logs, the remaining token budget is allocated according to their priority ratio to ensure the total length is maintained. .

[0032] This application binds the concatenation of the target paragraph text with the token resources provided by the Large Language Model (LLM), ensuring that the length of the input text strictly matches the upper limit of the model tokens. This avoids information loss and evaluation distortion caused by text truncation. It not only realizes the dynamic allocation of token resources by priority, ensuring the complete presentation of key content such as the evaluation framework and core dialogue information, thus improving the understanding and evaluation accuracy of the large language model, but also simplifies redundant information, reduces the consumption of invalid tokens, significantly reduces the cost of calling the large language model, and improves the system's inference speed and early warning timeliness.

[0033] This application also analyzes the target prompt text based on the large model call module, such as checking whether the six dimensions are complete, whether the score is within the range of [1,10], and whether the description is empty, thus obtaining the analysis result R in JSON format that meets the requirements. Parameters are also pre-set: temperature = 0.3 (to ensure output stability), response_format = "json" (to ensure format compliance), timeout = 30 seconds, and max_retries = 3. Simultaneously, it addresses engineering issues such as stability, cost control, and result reliability in API calls, providing high-quality input for subsequent result parsing and score fusion. This application adopts a six-dimensional evaluation framework, but the number and content of dimensions can be adjusted according to the business scenario. If the scenario is insurance sales, the corresponding dimensions are risk assessment, protection plan, trust establishment, premium communication, renewal follow-up, and objection handling; if the scenario is real estate brokerage, the corresponding dimensions are demand mining, property matching, professional consultation, price negotiation, property viewing follow-up, and closing ability; if the scenario is education consulting, the corresponding dimensions are learning diagnosis, curriculum design, teacher introduction, fee communication, learning follow-up, and renewal ability. The number of dimensions can be 4 (simplified version), 8 (refined version), or 10 (comprehensive version).

[0034] The large model calling module calls at a constant rate. To capacity Add tokens to the token bucket; current token count. When a call request is received, check the number of tokens in the current token bucket. :like The request is allowed to proceed, and 1 token is deducted from the bucket; if Refuse to execute immediately, calculate the waiting time. The request is added to a delayed queue, which uses a priority mechanism. When a new token is generated in the bucket, requests with high urgency (such as real-time alerts) are processed first. Token update: tokens = min(capacity, tokens + elapsed × rate); Request processing: if tokens >= 1, the call is allowed; otherwise, the waiting time is calculated.

[0035] In the above process, the rate controller in the large model call module uses the token bucket algorithm to limit the API call frequency to support the operation of the prompt generation module. The token bucket algorithm is replaced by the leaky bucket algorithm, which processes requests at a constant rate. The sliding window algorithm uses the number of requests in the most recent N seconds, and distributed rate limiting uses Redis for cross-server control. The large model call module includes a rate controller, a cache manager, a retry manager, and a result validator. The rate controller uses the token bucket algorithm to limit the API call frequency, the cache manager manages the analysis results of similar dialogues stored in the cache library, the retry manager handles timeouts and error retries, and the result validator verifies the integrity of the returned result format.

[0036] The large model in the large model calling module described in this application is OpenAI GPT-4, but it can be replaced with AnthropicClaude: similar capabilities, just adjust the API parameters; Google Gemini: supports longer contexts (1 million tokens); open source models (Llama 3, Qwen, etc.): private deployment, higher data security; multi-model combination: use open source for quick screening, and commercial for in-depth analysis.

[0037] In the specific implementation of step S102, one embodiment is as follows: Figure 3 As shown, multiple paragraph texts are dynamically generated based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, including: S1021. Calculate the complexity score of the dialogue data to determine the scenario type of the cross-border tourism sales dialogue, so as to dynamically select the number of segments of the prompt template based on the scenario type; S1022. Traverse the entity set extracted from the dialogue data in the domain knowledge base to query the corresponding knowledge entries. S1023. Based on the knowledge entries and the total token budget, dynamically generate multiple paragraph texts.

[0038] In steps S1021-S1023, the prompt generation module of this application pre-establishes a scene complexity scoring calculation model, which is expressed by formula (1): (1); in As a normalized message quantity factor, To normalize the dialogue duration factor, For the session stage factor, The complexity score of the dialogue data is calculated based on the scenario complexity scoring calculation model, using user interaction feature factors (such as historical interaction frequency and data retention). The scenario type of the cross-border tourism sales dialogue is determined, and the dialogue scenarios are divided into three types: simple, medium, and complex. The component paragraphs of the prompt template are dynamically selected according to the scenario type. If complexity score For simple scenarios, choose a three-part template that includes role positioning, evaluation framework, and format constraints; if the complexity score... This is considered a medium-level scenario, and a five-segment template including role positioning, information injection, semantic decomposition, evaluation framework, and format constraints should be selected; if the complexity score... For complex scenarios, a seven-segment complete template should be selected; among them , These are the lower and upper thresholds for the complexity score, respectively; where the weight adjustment coefficient in formula (1) can be adjusted in various ways to suit different business scenarios, such as linearly adjusting w. i ' = w i × (1 + order amount / 100,000); segmented adjustment w i ' = w i × {0.8, 1.0, 1.3, 1.5}, based on the amount range; multi-factor combined w i ' = w i × f(amount, historical value, conversation duration); Extract entity sets from dialogue content (Including destination, POI, service type, etc.), iterate through the collection. Each entity in In the domain knowledge base Search for the corresponding knowledge item in the middle The obtained knowledge entries are then combined with the total token budget to dynamically generate multiple paragraph texts.

[0039] In a specific implementation of step S1023, one embodiment is as follows: Based on the knowledge entries and the total token budget, multiple paragraph texts are dynamically generated, including: A1. Calculate the cosine similarity between the dialogue context vector of the dialogue data and the knowledge entry vector of the knowledge entry, and add the corresponding knowledge entry to the injection set based on the similarity; A2. Allocate the total token budget to the corresponding prompt paragraphs of each prompt template based on the preset priority weights of each prompt template paragraph; A3. Merge the injected set into the corresponding paragraph template to generate multiple paragraph texts.

[0040] In steps A1-A3, after obtaining the knowledge entries, this application calculates the dialogue context vector. With knowledge item vectors cosine similarity The process involves determining whether a knowledge item meets the requirements, and if the similarity is greater than a preset threshold. This knowledge entry Add to injection set Based on the preset priority weights of each template paragraph. For the total token budget The allocation is performed as follows: for paragraphs with formatting constraints that must be retained, the original length budget is allocated; for paragraphs with variable lengths, such as chat logs, the remaining budget is allocated according to priority proportions to ensure the total length is maintained. , merge the injection set To generate multiple paragraph texts with dialogue data in the corresponding paragraph template, specifically, the injection set is... Each knowledge entry in Standardization is performed, core fields are extracted and formatted to ensure semantic fluency and no redundancy after integration into paragraph templates. Preprocessed knowledge items are then integrated into corresponding paragraph templates using a fixed template description and dynamic knowledge item filling. After integration, the total token count for each paragraph is calculated. If the total token count > the total token budget Tmax, it is pruned according to the principle of retaining core knowledge and simplifying redundant descriptions, prioritizing knowledge items directly related to the dialogue. For knowledge items of the same type, core fields are retained, ensuring the total token count meets the total token budget. If the total token count ≤ the total token budget Tmax, the process proceeds directly to the subsequent paragraph concatenation step. If the field is empty (no highly similar knowledge entries), only the fixed expression of the paragraph template will be retained, and no knowledge will be forcibly injected to avoid meaningless redundancy.

[0041] This application uses an 85% similarity threshold, which can be adjusted to: strict mode (95%): reduce false matches and improve accuracy; lenient mode (75%): improve hit rate and reduce costs; dynamic threshold: automatically adjusted based on hit rate and accuracy.

[0042] In step S103, this application parses the JSON-formatted analysis result R using a result parsing module and extracts structured information. The result parsing module includes a format verification submodule, a score extraction submodule, an attribution extraction submodule, and an anomaly handling submodule. The format verification submodule verifies the integrity of the JSON structure; the score extraction submodule extracts a six-dimensional score vector; the attribution extraction submodule extracts attribution information for each dimension; and the anomaly handling submodule handles parsing anomalies. The result parsing module outputs the multi-dimensional score vector and attribution information set A. The multi-dimensional score vector is specifically a six-dimensional score vector S, including demand understanding ability, solution design ability, professional credibility, price communication ability, follow-up strategy ability, and retention effort. The dialogue data D and the multi-dimensional score vector S are input into the score fusion module, which extracts the order feature vector based on the included feature extraction submodule. Customer profile vector Dialogue stage vector Three vectors are processed to obtain a comprehensive total score Score_total, a cancellation probability P_cancel, and a weight vector W. The attribution information set A, the six-dimensional scoring vector S, the cancellation probability P_cancel, and the dialogue data D are input into the attribution analysis module for analysis and processing to obtain an attribution report Report = {main responsible party, reason for failure, list of key nodes, improvement suggestions}.

[0043] In a specific implementation of step S103, one embodiment is as follows: dynamic score fusion and two-stage attribution analysis are performed based on the multi-dimensional scoring vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded, including: S1031. Extract the order feature vector, customer profile vector and dialogue stage vector from the dialogue data respectively, and fuse them to form a dynamic weight vector; S1032. Calculate the overall score based on the dynamic weight vector and the multi-dimensional scoring vector; S1033. Map the overall total score to the order cancellation probability, and obtain the attribution analysis result based on the order cancellation probability.

[0044] In steps S1031-S1033, the scoring fusion module includes not only a feature extraction submodule, but also a weight calculation submodule, a score normalization submodule, and a probability mapping submodule. The weight calculation submodule is used to calculate dynamic weights based on a multi-factor model; the score normalization submodule is used to perform Z-score normalization on the original scores; the probability mapping submodule is used to map the overall total score to a cancellation probability; and the weight calculation submodule outputs a six-dimensional weight vector corresponding to the multi-dimensional score vector S. The six-dimensional weight vector It is dynamic, utilizing the average of historical data. and standard deviation The original scores of each dimension Based on formula (2), convert to standard scores and map to the 0-100 range: (2); The results are truncated to [0, 100], and the comprehensive score is calculated based on formula (3) using the dynamic weight vector: (3); The overall score is mapped to a cancellation probability score in the interval [0, 1] using the Sigmoid function: (4); in For decision boundary threshold, The slope coefficient is used to obtain the cancellation probability score of pending cross-border tourism orders.

[0045] The weight calculation submodule sets the initial weight vector. The weights of each dimension are evenly distributed, and a feature adjustment matrix is ​​constructed to enhance the weights of specific dimensions according to business rules. For example, based on the attribute features of the target object: numerical features (such as the value of the target object) are obtained from the order feature vector and compared with a preset grading threshold. When the numerical features are in a high-level range (i.e., high-complexity data processing objects), the weight coefficient of specific evaluation dimensions (such as dimensions involving credibility or complexity) is increased through a non-linear mapping function; if the decision cycle is short, the weight related to "strategy response" is increased. Another example is based on user interaction features: if it is a first-time user interaction, the weight related to "trust building" is increased; if it is a price-sensitive user (based on historical data analysis), the weight related to "value communication" is increased. Yet another example is based on the session progress: the weight is increased according to the session progress ratio. The approach emphasizes different capabilities at different stages (e.g., focusing on intent recognition in the early stages and on result achievement in the later stages). The revised formula is: ,in This indicates element-wise multiplication. The corrected weight vector is then normalized, as shown in formula (5), to ensure that the sum of all weights is 1. (5); Based on the above formula, the dynamic six-dimensional weight vector corresponding to the multi-dimensional scoring vector S is obtained. .

[0046] In a specific implementation of step S1033, one embodiment is as follows: obtaining the attribution analysis result based on the order cancellation probability includes: B1. Construct a linear evaluation model based on the dialogue completeness, frequency of dangerous words, response timeliness, and cancellation probability extracted from the dialogue data; B2. Based on the linear evaluation model, quickly screen out the dialogue type corresponding to the dialogue data, and determine whether the dialogue type is the target dialogue type; B3. If so, key nodes in the time series are detected to perform in-depth attribution of the cross-border tourism sales dialogue, and the attribution analysis results are combined with the attribution information set to form a visually interpretable explanation.

[0047] In steps B1-B3, after receiving the input attribution information set A, the six-dimensional rating vector S, the cancellation probability P_cancel, and the dialogue data D, the attribution analysis module performs two stages of processing. Stage 1: Lightweight Screening, which constructs a linear evaluation model based on lightweight features (dialogue completeness, frequency of dangerous words, response timeliness, and total score). =Σw i × f i Calculate rapid risk score ,according to The conversations are categorized into three types: high-risk, attention-based, and normal. Corresponding risk score thresholds are set for each category. If a rapid risk score is applied... Within a certain threshold range, the corresponding dialogue data represents the corresponding dialogue type. It is then determined whether this dialogue type is the target dialogue type, i.e., whether the dialogue type corresponding to the data is a high-risk dialogue type. If so, the second stage of deep attribution analysis is triggered; otherwise, the screening is complete. Stage Two: Deep Attribution, 1. Dimension Localization: Identify the dimension with the lowest standardized score. 1. As the primary failure point; 2. Causal chain analysis: dimensional correlation matrix based on historical data. Analyze whether it exists 3. Determining responsibility: Calculate the responsibility factor scores for the client side (e.g., changes in requirements), the consultant side (e.g., slow response), and the system side (e.g., insufficient inventory) respectively, and take the one with the highest score as the main responsible party; 4. Report generation: Combine key nodes and attribution results to generate a structured report containing improvement suggestions.

[0048] The attribution analysis module includes a rapid screener that uses a lightweight model for preliminary classification and a deep attributor that uses a high-performance model for in-depth analysis. The rapid screener includes: a feature calculation unit for calculating completeness, frequency of warning signals, and response quality; a risk scoring unit for calculating a rapid risk score; and a classification decision unit for classifying based on thresholds. The deep attributor includes: a dimensional analysis unit for analyzing key issues across dimensions; a correlation analysis unit for calculating the correlation matrix between dimensions; a responsibility determination unit for identifying the main responsible party; a node locator for identifying key turning points; an emotion detection unit for detecting emotional turning points; a demand deviation unit for detecting demand deviation points; a price conflict unit for detecting price conflict points; and a silence period detection unit for detecting abnormal silence periods.

[0049] In the specific implementation of step B3, one embodiment is as follows: Deep attribution of the cross-border tourism sales dialogue is performed by combining key time-series nodes, including: B31. Calculate the sentiment score of each message in the customer message sequence to form a sentiment vector, calculate the sentiment vector to obtain multiple moving averages, and fit a target curve to mark the sentiment inflection point. B32. Calculate the weighted deviation between the constructed initial customer demand vector and the solution attribute vector provided by the consultant, and mark the demand deviation point when the weighted deviation exceeds the preset deviation threshold; B33. Construct a feature function using a price-sensitive dictionary to calculate the price signal strength of each message, and mark price conflict points based on the price signal strength; B34. Calculate the time difference between adjacent messages and compare the time difference with the silence threshold set by the respective regions where the advisors are located to mark the silence period.

[0050] In steps B31-B34, this application calculates key nodes in the time series based on a deep attributor and calculates sentiment scores for each message in the customer message sequence based on a sentiment detection unit, forming a sentiment vector. , for emotion vector Application window size is The moving average is used to obtain multiple moving averages. : (6); And based on the obtained multiple moving averages, a target curve is fitted. The curve, when the target curve is detected, i.e. The curve crosses zero (from positive to negative) and the magnitude of the change exceeds the threshold. When, or when the target curve is detected When the curve changes abruptly and negatively, the position is marked as an emotional inflection point; and the method for detecting emotional inflection points in this application can also be replaced by: threshold method, which marks the point when the single decrease exceeds the threshold; inflection point detection, which uses the first / second derivative to detect the inflection point of the curve; and machine learning classifier, which trains the classifier to identify the inflection features. The demand deviation units respectively construct the customer's initial demand vector. Attribute vectors of the solutions provided by consultants In this context, the consultant represents the seller in the pending transaction order, and a heterogeneous distance function is predefined. The weighted deviation between the proposed solution and the requirements is calculated. , ,in w For the weight vector, For a certain component of the initial demand vector, For the solution attribute vector If a certain component is weighted by deviation Exceeding the preset deviation threshold If a customer's subsequent needs conflict with key dimensions in the initial needs (such as destination and number of people), it is marked as a deviation point in the needs. The price conflict unit constructs a feature function using a price-sensitive dictionary, and calculates the price signal strength of each message based on the feature function. ,in Where k represents all words in the i-th message that hit the price-sensitive dictionary, such as expensive, cheap, high price, discount, value for money, etc.; tf(k) is the word frequency of word k in the i-th message; and idf(k) is the inverse document frequency of word k in the full sales dialogue corpus. If the price signal strength... If the value is greater than 0, it is marked as a price conflict point and graded according to its intensity; The silence period detection unit calculates the time difference between adjacent messages. And compare the time difference with a preset silence threshold. Among them, the silent threshold This is set based on the time difference between the locations of both parties to the transaction, namely the client and the consultant, to avoid misjudging the silent period due to cross-time zone or asynchronous communication. If the silence period is marked as a period of silence, it will be categorized as "waiting for a response silence" or "decision silence" based on the type of the last message before the silence (question / quote). The threshold parameter is adjustable: Emotional transition threshold. : 0.3 ~ 0.7; Demand deviation threshold : 0.1 ~ 0.3; Silent period threshold : 12 hours ~ 72 hours.

[0051] In step S104, the warning and intervention module of the present application includes a state machine manager, a danger signal detector, a hierarchical warning device, and a conversation strategy recommender. The state machine manager is used to maintain the conversion of the conversation state; the danger signal detector is used to accumulate the danger signal count; the hierarchical warning device is used to trigger differential warnings according to the risk level; the conversation strategy recommender is used to generate targeted retention conversation strategies. The warning and intervention module triggers hierarchical warnings and adopts corresponding warning handling methods based on the order cancellation probability, the attribution analysis result, and the danger signal intensity output by the danger signal detector, including: Level 1 warning (P_cancel > 0.8 or danger signal > 8 times): Immediately push it to the team supervisor and the consultant himself, along with an attribution report and a retention conversation strategy template. The triggering methods are real-time push and phone reminder; Level 2 warning (0.5 < P_cancel ≤ 0.8): Push improvement suggestions to the consultant and provide reference to similar successful cases. The triggering method is message push; Level 3 warning (0.3 < P_cancel ≤ 0.5): Add it to the regular review queue and check the progress every 24 hours. The triggering method is a timed task to intelligently diagnose the cross-border tourism sales conversation, and input the final order result, warning notice, intervention measure, and conversation strategy template into the feedback and optimization module. The feedback and optimization module virtually stores the analysis result in the database, tracks the final order result (成交 / 取消), and calculates the warning accuracy rate: accuracy = |correct warning| / |total warning|; if the accuracy deviation > 10%, adjust the probability mapping parameters α and θ; statistically analyze the correlation between each dimension and the final result, and adjust the weight factor.

[0052] The feedback and optimization module includes a result storage sub-module, an accuracy rate statistics sub-module, a parameter tuning sub-module, a template optimization sub-module, and a knowledge base update sub-module. The result storage sub-module is used to store the analysis result in the database; the accuracy rate statistics sub-module is used to statistically analyze the warning accuracy rate; the parameter tuning sub-module is used to adjust the system parameters (such as μ, σ, α, θ) according to the feedback; the template optimization sub-module is used to optimize the prompt template; the knowledge base update sub-module is used to extract the latest quotes and POI features based on the successful cases to update the domain knowledge base.

[0053] This application implements a streaming real-time analysis mechanism for ongoing cross-border tourism sales dialogues in pending orders, specifically employing an incremental analysis mechanism: Regular triggering conditions: Analysis is triggered every 10 new messages; Emergency immediate triggering: When the following high-risk signals are detected, the waiting period is immediately interrupted, and analysis is executed: Rapidly deteriorating sentiment: Sentiment value of a single message <-0.8; High-intensity price conflict: I_price(i)>3×θ_price; Customer explicitly expresses cancellation intention: Hitting keywords {"cancel", "refund", "don't want it anymore"}; Consultant's prolonged silence: No response within 24 hours after a customer's question; Incremental processing: Only the association between new messages and historical context is analyzed; Status update: The dialogue state machine and cumulative score are updated; Continuous monitoring: The danger signal detector continuously accumulates signal counts.

[0054] This application also achieved verification with real business data: Dataset size: 215 canceled orders, 38,717 messages, averaging 180 messages / conversation; Labeling benchmark: independent scoring by 5 experts, Kappa coefficient 0.82; Verification results: Cancellation identification accuracy 92%, attribution accuracy 87%, and recoverable identification accuracy 89%.

[0055] This application also underwent comparative experiments with other existing methods, as shown in Table 1. It can be seen that this application achieves significant advantages. Table 2 shows the reasons for the improved accuracy: deep semantic understanding: a large language model captures contextual semantic relationships; long context modeling: segmented prompts support complete long sequence inputs; multi-dimensional evaluation framework: covering the entire interaction process. Table 3 shows that, based on real data from 215 canceled orders, the attribution capability mechanism of this application is reflected in the six-dimensional quantitative scoring to quickly locate the lowest dimension; time series algorithms to accurately locate turning points; and the determination of responsible parties to differentiate between customer / consultant / system factors. Table 4 shows that the real-time performance and timeliness of this application are significantly improved. The real-time implementation mechanism includes: incremental analysis: triggered every 10 messages, with an average time consumption of <5 seconds; layered invocation: 70% of dialogues require only 2 seconds for rapid screening; and second-level early warning: danger signals are detected and pushed out on average in 7.5 seconds. After applying this application, the order recovery rate increased by 10 percentage points (15%→25%), the average score of consultants across six dimensions increased by 16% (6.8→7.9), the supervisor's review time decreased by 83% (30 minutes→5 minutes), and the team's overall conversion rate increased by 8 percentage points (38%→46%).

[0056] Table 1. Comparative experimental data of this application with other existing methods.

[0057] Table 2. Performance Comparison of this Application with Other Existing Methods

[0058] Table 3 Comparison of Attribution Positioning Ability

[0059] Table 4 Comparison of Real-Time Performance and Timeliness

[0060] This application can also adopt a two-layer model architecture to reduce costs, as shown in Table 5.

[0061] Table 5 Comparison of Two-Layer Models

[0062] Average cost calculation: Cost_avg = Cost_fast × P_fast + Cost_deep × P_deep = 0.1 × 0.7 + 0.4 × 0.3 = 0.19 yuan / dialogue, which means that compared with using GPT-4 for the whole amount (1.5 yuan / dialogue), the cost is reduced by 87%.

[0063] This application also merges low-priority conversations (such as nighttime historical analysis), which can utilize API batch processing discounts to reduce costs by approximately 10%, as shown in Table 6. Table 7 explains the meaning and value range of some letters appearing in this application.

[0064] Table 6 Cost Comparison Table

[0065] Table 7 Symbol Table of this Application

[0066] Example 2 This application also provides an intelligent diagnostic device for cross-border tourism sales dialogue, such as... Figure 4 The diagram shows a block diagram of an intelligent diagnostic device for cross-border tourism sales dialogues. The functions implemented by this device correspond to the steps of the aforementioned intelligent diagnostic method for cross-border tourism sales dialogues executed on a terminal device. This device can be understood as a server component including a processor. The intelligent diagnostic device for cross-border tourism sales dialogues described in this application includes: The data acquisition module 401 is used to acquire raw dialogue data from multiple data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and to preprocess the raw dialogue data to obtain dialogue data. The splicing module 402 is used to dynamically generate multiple paragraph texts based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, and splice the multiple paragraph texts according to the preset paragraph order and the corresponding token budget to generate the target prompt text. Analysis module 403 is used to extract the multi-dimensional rating vector and attribution information set from the target prompt text, and perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded. The diagnostic module 404 is used to trigger a tiered warning based on the order cancellation probability and attribution analysis results, and to take corresponding warning processing measures to intelligently diagnose the cross-border tourism sales dialogue.

[0067] In one feasible implementation, the analysis module includes: The order feature vector, customer profile vector, and dialogue stage vector are extracted from the dialogue data respectively, and then fused to form a dynamic weight vector. The overall score is calculated based on the dynamic weight vector and the multi-dimensional scoring vector. The overall score is mapped to the order cancellation probability, and the attribution analysis result is obtained based on the order cancellation probability.

[0068] In one feasible implementation, an analysis module is also included, comprising: A linear evaluation model is constructed from the dialogue completeness, frequency of dangerous words, response timeliness, and cancellation probability extracted from the dialogue data. Based on the linear evaluation model, the dialogue type corresponding to the dialogue data is quickly screened out, and it is determined whether the dialogue type is the target dialogue type. If so, key nodes in the time series are detected to perform in-depth attribution of the cross-border tourism sales dialogue, and the attribution analysis results are combined with the attribution information set to form a visually interpretable explanation.

[0069] In one feasible implementation, the analysis module also includes: Calculate the sentiment score of each message in the customer message sequence to form a sentiment vector, calculate the sentiment vector to obtain multiple moving averages, and fit a target curve to mark the sentiment inflection point; Calculate the weighted deviation between the constructed initial customer demand vector and the solution attribute vector provided by the consultant, and mark the demand deviation point when the weighted deviation exceeds a preset deviation threshold; A feature function is constructed using a price-sensitive dictionary to calculate the price signal strength of each message, and price conflict points are marked based on the price signal strength. Calculate the time difference between adjacent messages and compare the time difference with a silence threshold set according to the regions where the customer and consultant are located, to mark the silence period.

[0070] In one feasible implementation, the splicing module includes: The complexity score of the dialogue data is calculated to determine the scenario type of the cross-border tourism sales dialogue, so as to dynamically select the number of segments of the prompt template based on the scenario type; Traverse the entity set extracted from the dialogue data in the domain knowledge base to query the corresponding knowledge entries; Based on the knowledge entries and the total token budget, multiple paragraph texts are dynamically generated.

[0071] In one feasible implementation, the splicing module further includes: Calculate the cosine similarity between the dialogue context vector of the dialogue data and the knowledge entry vector of the knowledge entry, and add the corresponding knowledge entry to the injection set based on the similarity; The total token budget is allocated to the corresponding prompt paragraph based on the preset priority weight of each prompt template paragraph; The injected set is then merged into the corresponding paragraph template to generate multiple paragraph texts.

[0072] In one feasible implementation, the acquisition module includes: In a pre-established historical dialogue vector library, similar dialogue matching is performed on the dialogue data to obtain the matching results; Based on the matching results, the stored historical analysis results are retrieved to perform a two-stage attribution analysis.

[0073] Example 3 This application also provides an electronic device, such as Figure 5 As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the intelligent diagnostic methods for cross-border tourism sales dialogues are performed.

[0074] Example 4 This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the intelligent diagnostic methods for cross-border tourism sales dialogues.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0076] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0078] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0079] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent diagnosis method for cross-border travel sales dialog, characterized in that, The method includes: Raw dialogue data is collected from various data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and the raw dialogue data is preprocessed to obtain dialogue data. Based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, multiple paragraph texts are dynamically generated. The multiple paragraph texts are then concatenated according to the preset paragraph order and the corresponding token budget to generate the target prompt text. Extract the multi-dimensional rating vector and attribution information set from the target prompt text, and perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded; Based on the order cancellation probability and attribution analysis results, a tiered early warning is triggered, and corresponding early warning processing methods are adopted to intelligently diagnose the cross-border tourism sales dialogue.

2. The method of claim 1, wherein, Based on the aforementioned multi-dimensional scoring vector and attribution information set, dynamic scoring fusion and two-stage attribution analysis are performed to obtain the order cancellation probability and attribution analysis results for orders to be traded, including: The order feature vector, customer profile vector, and dialogue stage vector are extracted from the dialogue data respectively, and then fused to form a dynamic weight vector. The overall score is calculated based on the dynamic weight vector and the multi-dimensional scoring vector. The overall score is mapped to the order cancellation probability, and the attribution analysis result is obtained based on the order cancellation probability.

3. The method of claim 2, wherein, The attribution analysis result obtained based on the order cancellation probability includes: A linear evaluation model is constructed from the dialogue completeness, frequency of dangerous words, response timeliness, and cancellation probability extracted from the dialogue data. Based on the linear evaluation model, the dialogue type corresponding to the dialogue data is quickly screened out, and it is determined whether the dialogue type is the target dialogue type. If so, key nodes in the time series are detected to perform in-depth attribution of the cross-border tourism sales dialogue, and the attribution analysis results are combined with the attribution information set to form a visually interpretable explanation.

4. The method of claim 3, wherein, A deep attribution analysis of the cross-border tourism sales dialogue was conducted by combining key time-series nodes, including: Calculate the sentiment score of each message in the customer message sequence to form a sentiment vector, calculate the sentiment vector to obtain multiple moving averages, and fit a target curve to mark the sentiment inflection point; Calculate the weighted deviation between the constructed initial customer demand vector and the solution attribute vector provided by the consultant, and mark the demand deviation point when the weighted deviation exceeds a preset deviation threshold; A feature function is constructed using a price-sensitive dictionary to calculate the price signal strength of each message, and price conflict points are marked based on the price signal strength. Calculate the time difference between adjacent messages and compare the time difference with a silence threshold set according to the regions where the customer and consultant are located, to mark the silence period.

5. The method of claim 1, wherein, Based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the pre-set large language model, multiple paragraph texts are dynamically generated, including: The complexity score of the dialogue data is calculated to determine the scenario type of the cross-border tourism sales dialogue, so as to dynamically select the number of segments of the prompt template based on the scenario type; Traverse the entity set extracted from the dialogue data in the domain knowledge base to query the corresponding knowledge entries; Based on the knowledge entries and the total token budget, multiple paragraph texts are dynamically generated.

6. The method of claim 5, wherein, The process of dynamically generating multiple paragraph texts based on the knowledge entries and the total token budget includes: Calculate the cosine similarity between the dialogue context vector of the dialogue data and the knowledge entry vector of the knowledge entry, and add the corresponding knowledge entry to the injection set based on the similarity; The total token budget is allocated to the corresponding prompt paragraph based on the preset priority weight of each prompt template paragraph; The injected set is then merged into the corresponding paragraph template to generate multiple paragraph texts.

7. The method of claim 1, wherein, After preprocessing the original dialogue data to obtain the dialogue data, the process includes: In a pre-established historical dialogue vector library, similar dialogue matching is performed on the dialogue data to obtain the matching results; Based on the matching results, the stored historical analysis results are retrieved to perform a two-stage attribution analysis.

8. An intelligent diagnosis device for cross-border travel sales dialog, characterized in that, The device includes: The data acquisition module is used to collect raw dialogue data from various data sources of cross-border tourism sales dialogues generated from pending cross-border tourism orders, and to preprocess the raw dialogue data to obtain dialogue data. The splicing module is used to dynamically generate multiple paragraph texts based on the dialogue data, the dialogue scenario in which the dialogue data is located, and the total token budget of the preset large language model, and splice the multiple paragraph texts according to the preset paragraph order and the corresponding token budget to generate the target prompt text. The analysis module is used to extract the multi-dimensional rating vector and attribution information set from the target prompt text, and to perform dynamic rating fusion and two-stage attribution analysis based on the multi-dimensional rating vector and attribution information set to obtain the order cancellation probability and attribution analysis results of the order to be traded. The diagnostic module is used to trigger tiered warnings based on the order cancellation probability and attribution analysis results, and to take corresponding warning processing measures to intelligently diagnose the cross-border tourism sales dialogue.

9. An electronic device, comprising: include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of an intelligent diagnostic method for cross-border tourism sales dialogue as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of an intelligent diagnostic method for cross-border tourism sales dialogue as described in any one of claims 1 to 7.

Images