Insurance dialogue intent tracking method and system based on large model and electronic device

Abstract

The application discloses an insurance dialogue intention tracking method and system based on a large model and electronic equipment, relates to the technical field of data processing, and comprises the following steps: encoding and analyzing a clause knowledge graph and historical dialogue corpus through a bidirectional encoding structure, and constructing an insurance large model; performing dynamic intention state analysis on a target insurance dialogue to obtain a target dynamic intention vector; performing feedback analysis on the target dynamic intention vector to obtain a target single-round feedback record; introducing a clarification optimization mechanism to perform clarification optimization analysis on the target single-round feedback record to obtain a target clarification strategy; and performing intention tracking processing on the target insurance dialogue according to the target clarification strategy. The technical problems of inaccurate intention understanding of an insurance dialogue system, insufficient dynamic intention tracking capability and lack of ambiguity clarification mechanism in the prior art are solved, and the technical effects of enhancing dynamic intention state modeling capability, optimizing active clarification and feedback mechanism and improving multi-round dialogue intention recognition accuracy are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an insurance dialogue intention tracking method and system based on a large model and an electronic device. BACKGROUND

[0002] With the rapid development of artificial intelligence, large language models play an important role in intelligent dialogue systems. In the insurance industry, customer inquiries often involve a large number of professional terms, complex insurance clause knowledge, claim settlement rules, and personalized needs. Users may adjust or refine their needs in multiple rounds of dialogue, and it is necessary to capture the dynamic evolution of intentions in real time to avoid deviation of the dialogue due to understanding bias. When the user's expression is ambiguous or has ambiguity, a clarification strategy such as follow-up or confirmation needs to be generated automatically. Existing dialogue systems are difficult to accurately understand and identify user intentions, resulting in low dialogue efficiency and poor user experience.

[0003] Therefore, in the related art at the present stage, there are technical problems of inaccurate intention understanding of the insurance dialogue system, insufficient dynamic intention tracking capability, and lack of ambiguity clarification mechanism. SUMMARY

[0004] The present application provides an insurance dialogue intention tracking method and system based on a large model and an electronic device, which solves the technical problems of inaccurate intention understanding of the insurance dialogue system, insufficient dynamic intention tracking capability, and lack of ambiguity clarification mechanism in the prior art, and achieves the technical effects of enhancing dynamic intention state modeling capability, optimizing active clarification and feedback mechanism, and improving multi-round dialogue intention recognition accuracy.

[0005] The present application provides an insurance dialogue intention tracking method based on a large model, which comprises: encoding and analyzing a clause knowledge graph and historical dialogue corpus through a bidirectional encoding structure, and constructing an insurance large model; performing dynamic intention state analysis on a target insurance dialogue to obtain a target dynamic intention vector; performing feedback analysis on the target dynamic intention vector through the insurance large model to obtain a target single-round feedback record; introducing a clarification optimization mechanism to perform clarification optimization analysis on the target single-round feedback record to obtain a target clarification strategy; and performing intention tracking processing on the target insurance dialogue according to the target clarification strategy.

[0006] In a possible implementation, the insurance dialogue intent tracking method based on a large model further performs the following processing: the bidirectional encoding structure includes a first encoding layer and a second encoding layer, the first encoding layer is used to encode the clause knowledge graph to obtain a clause encoding unit, and the second encoding layer is used to encode the historical dialogue corpus to obtain a dialogue encoding unit; the clause encoding unit and the dialogue encoding unit are jointly optimized and analyzed to obtain the insurance large model; the first encoding layer is used to encode the first entity relationship in the clause knowledge graph to obtain a first encoding unit, the clause encoding unit is established based on the first encoding unit, the first entity relationship includes a first product liability, a first exemption situation and a first claim condition, the first dialogue segment in the historical dialogue corpus is identified based on a BiLSTM-CRF model, the first dialogue segment is subjected to intent identification to obtain a first intent label, the second encoding layer is used to encode the first dialogue segment and the first intent label to obtain a second encoding unit, and the dialogue encoding unit is established based on the second encoding unit.

[0007] In a possible implementation, the insurance dialogue intent tracking method based on a large model further performs the following processing: when the insurance clause version is changed, an incremental training plan is started; the incremental training plan refers to a plan for triggering incremental training of the insurance large model by comparing entity differences in the insurance clause version.

[0008] In a possible implementation, the insurance dialogue intent tracking method based on a large model further performs the following processing: any round dialogue in the target insurance dialogue is extracted, and the any round dialogue corresponds to any time identifier; dialogue information filtering processing is performed on the any round dialogue according to a gating mechanism to obtain any dialogue information; any entity set in the any dialogue information is established, and any dialogue intent is determined by analyzing the any entity set; the target dynamic intent vector is established according to a mapping relationship between the any time identifier and the any dialogue intent; the any dialogue information includes an any word set, the any word set includes a first word, a first entity corresponding to the first word is matched, a first intent coefficient of the first entity is obtained in combination with a first appearance frequency of the first word, a first intent coefficient in an intent coefficient sequence obtained in descending order of the first intent coefficient is acquired, and a first target entity corresponding to the first intent coefficient is taken as the any dialogue intent.

[0009] In a possible implementation, the insurance dialogue intention tracking method based on a large model further performs the following processing: obtaining a first depth of the first entity in the clause knowledge graph; and correcting the first intention coefficient by taking the normalized first depth as a weight.

[0010] In a possible implementation, the insurance dialogue intention tracking method based on a large model further performs the following processing: extracting a first entity category in a predetermined entity category; obtaining a first total frequency of the first entity category in the target dynamic intention vector; performing feedback sorting on the first entity category based on the first total frequency to obtain a target feedback list; performing expansion rendering on the target feedback list in combination with the insurance large model to obtain target single-round feedback information; obtaining dialogue tracking of the target single-round feedback information, denoted as the target single-round feedback record; and wherein the predetermined entity category includes product liability, exemption circumstances, and claim conditions.

[0011] In a possible implementation, the insurance dialogue intention tracking method based on a large model further performs the following processing: when the target single-round feedback record has a predetermined tracking situation, calling a dialogue tree database according to the clarification optimization mechanism; analyzing the dialogue tree database to determine an optimal feedback strategy, and taking the optimal feedback strategy as the target clarification strategy; and wherein the predetermined tracking situation includes appearance of negative emotions and / or determination reduction.

[0012] In a possible implementation, the insurance dialogue intention tracking method based on a large model further performs the following processing: traversing the target dynamic intention vector in a historical insurance dialogue database to obtain the dialogue tree database; obtaining a first dialogue tree in the dialogue tree database, the first dialogue tree corresponding to a first dialogue terminal state index; taking the maximum of the first dialogue terminal state index as a target, determining a target dialogue tree; and forming the optimal feedback strategy based on multiple rounds of dialogue in the target dialogue tree; wherein the forming includes: extracting a first historical dynamic intention vector in the historical insurance dialogue database; and when a first vector similarity of the first historical dynamic intention vector and the target dynamic intention vector reaches a predetermined limit value, adding first historical dialogue information corresponding to the first historical dynamic intention vector to the dialogue tree database.

[0013] The application also provides an insurance dialogue intention tracking system based on a large model, which comprises: an encoding analysis module, configured to perform encoding analysis on a clause knowledge graph and historical dialogue corpus through a bidirectional encoding structure, and construct an insurance large model; an intention state analysis module, configured to perform dynamic intention state analysis on a target insurance dialogue, and obtain a target dynamic intention vector; a feedback analysis module, configured to perform feedback analysis on the target dynamic intention vector through the insurance large model, and obtain a target single-round feedback record; a clarification optimization analysis module, configured to introduce a clarification optimization mechanism to perform clarification optimization analysis on the target single-round feedback record, and obtain a target clarification strategy; and an intention tracking processing module, configured to perform intention tracking processing on the target insurance dialogue according to the target clarification strategy.

[0014] The application also provides an electronic device, comprising: a memory configured to store executable instructions; and a processor configured to execute the executable instructions stored in the memory to implement the insurance dialogue intention tracking method based on a large model.

[0015] The insurance dialogue intention tracking method, system and electronic device based on a large model provided by the application perform encoding analysis on a clause knowledge graph and historical dialogue corpus through a bidirectional encoding structure, and construct an insurance large model; perform dynamic intention state analysis on a target insurance dialogue, and obtain a target dynamic intention vector; perform feedback analysis on the target dynamic intention vector, and obtain a target single-round feedback record; introduce a clarification optimization mechanism to perform clarification optimization analysis on the target single-round feedback record, and obtain a target clarification strategy; and perform intention tracking processing on the target insurance dialogue according to the target clarification strategy. The technical problems of inaccurate intention understanding, insufficient dynamic intention tracking capability and lack of ambiguity clarification mechanism of the prior art are solved, and the technical effects of enhancing dynamic intention state modeling capability, optimizing active clarification and feedback mechanism and improving multi-round dialogue intention recognition accuracy are achieved. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments of the present disclosure will be briefly introduced below. The flowcharts are used to illustrate the operations performed by the system according to the embodiments of the present disclosure. It should be understood that the foregoing or the following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously according to needs. Meanwhile, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 The insurance dialogue intention tracking method based on a large model provided by the embodiments of the present application is shown in the flowchart.

[0018] Figure 2A structure schematic diagram of an insurance dialogue intention tracking system based on a large model is provided for an embodiment of the present application.

[0019] Figure 3 A structure schematic diagram of an electronic device is provided for an embodiment of the present application.

[0020] Legend: coding analysis module 10, intention state analysis module 20, feedback analysis module 30, clarification optimization analysis module 40, intention tracking processing module 50, input device 401, processor 402, memory 403, output device 404. DETAILED DESCRIPTION

[0021] The above description is only a summary of the technical solutions of the present application. In order to make the technical solutions of the present application more clear, the following specific embodiments of the present application are described in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described.

[0022] In order to make the purposes, technical solutions and advantages of the present application more clear, the following will combine the drawings to make a further detailed description of the present application. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by those skilled in the art without making creative efforts belong to the scope of protection of the present application.

[0023] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subset of all possible embodiments, and can be combined with each other without conflict. The term "first\second" is only to distinguish similar objects, and does not represent the specific order of the objects. The terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art in the technical field to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application.

[0024] The embodiments of the present application provide a large model-based insurance dialogue intention tracking method, as shown in Figure 1 The method comprises the following steps:

[0025] In step S100, the clause knowledge graph and the historical dialogue corpus are analyzed by a bidirectional coding structure, and an insurance large model is constructed.

[0026] The step S100 further comprises: the bidirectional coding structure comprises a first coding layer and a second coding layer; the step S110 codes the clause knowledge graph through the first coding layer to obtain a clause coding unit; the step S120 codes the historical dialogue corpus through the second coding layer to obtain a dialogue coding unit; and the step S130 jointly optimizes and analyzes the clause coding unit and the dialogue coding unit to obtain the insurance large model. The step S110 of coding the clause knowledge graph through the first coding layer to obtain a clause coding unit comprises: the step A1 extracts a first entity relationship in the clause knowledge graph; the step A2 codes a first triple corresponding to the first entity relationship through the first coding layer to obtain a first coding unit; and the step A3 establishes the clause coding unit based on the first coding unit. The first entity relationship comprises a first product liability, a first exemption situation and a first claim condition. The step S120 of coding the historical dialogue corpus through the second coding layer to obtain a dialogue coding unit comprises: the step B1 identifies a first dialogue segment in the historical dialogue corpus based on a BiLSTM-CRF model; the step B2 identifies an intent of the first dialogue segment to obtain a first intent label; the step B3 codes the first dialogue segment and the first intent label through the second coding layer to obtain a second coding unit; and the step B4 establishes the dialogue coding unit based on the second coding unit.

[0027] Preferably, the bidirectional coding structure is composed of two independent coding layers, including a first coding layer and a second coding layer, which respectively process the insurance clause knowledge graph and the historical dialogue corpus, and finally fuse to generate the insurance large model. The clause knowledge graph refers to the structured data of insurance clauses, which may include multiple entity relationships. The first entity relationship includes a first product liability, such as "coverage liability" and "payout ratio", a first exemption situation, such as "exclusion liability" and "deductible", and a first claim condition, such as "requirement of hospitalization proof" and "reporting time limit". The first coding layer codes the clause knowledge graph to convert the details of insurance clauses, liability scope and other professional knowledge in the insurance field into vector representations. Specifically, a graph neural network GNN or a Transformer is used to establish the first coding layer, and then the first triple corresponding to the first entity relationship is extracted from the clause knowledge graph, such as medical insurance, coverage liability and hospitalization expenses. The first triple corresponding to the first entity relationship is coded into a vector through the first coding layer, and the semantic and logical relationship is retained through joint learning of entity embedding and relationship embedding to obtain a first clause coding unit. Then, the coding vectors of all entity relationship triples are weighted and aggregated to form a clause coding unit representing the semantic of insurance clauses.

[0028] Preferably, the historical dialogue corpus is encoded by the second encoding layer to learn the expression of user's real dialogue intention, such as colloquial questioning, multi-round context dependence, etc. Specifically, the second encoding layer is established based on a pre-training model such as BERT or RoBERTa, and the historical dialogue corpus refers to the historical insurance dialogue record of unstructured text, such as user consultation, complaint, etc. Then, a BiLSTM-CRF model is used to combine a bidirectional long short-term memory network (LSTM) and a conditional random field (CRF) to perform sequence labeling on the dialogue text to identify key intention segments in the dialogue, such as "I want to claim for hospitalization expenses", and then perform intention label annotation, i.e. label the identified intention segments with intention labels such as "claim consultation". The dialogue segments and intention labels are jointly encoded into vectors by the second encoding layer as the second encoding unit. Then, the encoding vectors of all dialogue segments are integrated to form a dialogue encoding unit representing the user's intention mode, representing the user's intention distribution and expression habit in real dialogue.

[0029] Preferably, the attention mechanism or joint training is used to calibrate the clause encoding unit and the dialogue encoding unit in the vector space, output the final insurance large model, combine the professionalism of the clause with the flexibility of the dialogue, and avoid the deviation caused by the model relying on a single data source. The insurance large model has a deep understanding of the insurance clause and a dynamic tracking ability of the user's intention, can process the intention change in the dialogue in real time, such as the user suddenly changing from asking about the product to the complaint process, and significantly improves the intention recognition accuracy and scene adaptability of the insurance dialogue system.

[0030] Further, the step S100 further includes starting an incremental training plan when the insurance clause version is changed; wherein the incremental training plan refers to a plan for triggering the incremental training of the insurance large model by comparing the entity differences in the insurance clause versions.

[0031] Preferably, the insurance clauses can be changed due to regulatory requirements, product updates or business adjustments, such as the addition of exemption clauses, the adjustment of the proportion of compensation, etc., thereby starting an incremental training plan and only updating the model for the changed part, avoiding the cost of full training. Specifically, when detecting the change of the insurance clause version, through NLP tools such as text difference analysis and knowledge graph comparison, the entity differences in the new and old insurance clause versions are compared to identify entity-level changes, including new entities, modified entities and deleted entities. Then, the influence range of the changed entity on the insurance big model is analyzed, i.e. whether the changed entity affects the existing intent recognition and whether the changed entity triggers the chain logic adjustment of other clauses, thereby triggering the incremental training of the insurance big model. Specifically, the samples involving the changed entity are selected from the historical dialogue, such as all dialogues mentioning "hospitalization compensation ratio", new training data is generated, such as simulating user asking "now how much can hospitalization be compensated?" corresponding to the new clause, and then local fine-tuning is performed, i.e. only updating the coding layer related to the changed entity, such as the coding vector corresponding to "compensation ratio" in the first coding layer. A light-weight model is trained with new clause data, and then fused into the insurance big model to ensure the accuracy of the model's response to the new clause. The incremental training plan is an "online patching" mechanism for the insurance big model, which realizes the rapid adaptation of clause changes through entity-level difference analysis and local parameter updating, balancing the model training efficiency and accuracy after the clause changes.

[0032] Step S200, dynamic intent state analysis is performed on the target insurance dialogue to obtain a target dynamic intent vector.

[0033] Step S200 further includes step S210 of extracting any round dialogue in the target insurance dialogue, and the any round dialogue corresponds to any time identifier; step S220 of performing dialogue information filtering processing on the any round dialogue according to a gating mechanism to obtain any dialogue information; step S230 of assembling any entity set in the any dialogue information and analyzing the any entity set to determine any dialogue intent; and step S240 of establishing the target dynamic intent vector according to a mapping relationship between the any time identifier and the any dialogue intent. Step S230 includes: step C1 of assembling any word set in the any dialogue information, wherein the any word set includes a first word; step C2 of matching a first entity corresponding to the first word and combining a first appearance frequency of the first word to obtain a first intent coefficient of the first entity; step C3 of obtaining a first intent coefficient in a descending intent coefficient sequence; and step C4 of taking a first target entity corresponding to the first intent coefficient as the any dialogue intent.

[0034] Preferably, in the insurance dialogue, the user's intention may change over time, such as from "product consultation" to "claim complaint", and the dynamic intention state analysis of the target insurance dialogue determines the target dynamic intention vector, that is, the intention of each round of dialogue is tracked in real time and encoded into a computable vector for the model to understand the context and intention drift. Specifically, any round of dialogue is extracted from the target insurance dialogue, wherein the any round of dialogue corresponds to any time identifier, that is, the time sequence of the dialogue round is recorded for analyzing the intention evolution; then the dialogue information of the any round of dialogue is processed by a gating mechanism, wherein the gating mechanism is similar to the forget gate in LSTM, which is used to filter irrelevant information in the any round of dialogue. Specifically, the dialogue information is preprocessed based on rules, including quickly removing adverbs, repeated content and irrelevant auxiliary words through regular expressions and keyword lists, and retaining core content related to insurance intention, such as constructing a stop word library and an insurance domain irrelevant word list for preliminary cleaning to determine core noun phrases and verb structures; the obtained insurance intention related core content is input into the fine-tuned large model in the insurance domain, and the large model is guided to complete semantic compression and intention focusing by designing structured prompt words, such as the prompt template "Please rewrite the user's question as an insurance intention, and only keep the insurance type, liability, condition and question type". According to the prompt template, the standardized question is output, for example, the original input "Uh… I want to ask, how much can I claim for hospitalization for bone fracture for the medical insurance I bought before?", and the processed result is "Medical insurance, how much can I claim for hospitalization for bone fracture?". Finally, the rule processing result and the large model output result are compared, if the entities are consistent, the large model processing result is used; if the entities conflict, an entity recognition model is introduced for secondary verification, and finally efficient and accurate dialogue information filtering is realized.

[0035] Preferably, an arbitrary entity set in any dialogue information is formed, that is, insurance related entities are identified from the filtered any dialogue information, and the any dialogue intention is determined by analyzing the any entity set, that is, the intention is inferred by entity combination. Specifically, an arbitrary word set in any dialogue information is formed, such as medical insurance, fracture, hospitalization or claim, wherein the arbitrary word set includes multiple words, each word corresponds to an entity, and the first word is any one of the words; then the first entity corresponding to the first word is matched, and the appearance frequency of the first word is counted, such as "fracture" in the insurance clause associated with "accidental medical liability", obtaining the first appearance frequency, and then according to the entity weight in the clause knowledge graph, the product of the entity weight and the first appearance frequency is calculated as the first intention coefficient of the first entity; then the intention coefficients are sorted in descending order, and the entity corresponding to the highest value is taken as the current round intention, and the first target entity corresponding to the first intention coefficient is determined as the any dialogue intention. Finally, the any time identifier and the any dialogue intention are associated to determine their mapping relationship, and a vector representation of the time identifier and the dialogue intention is generated to finally determine the target dynamic intention vector, which may include a timestamp, an intention type and an entity weight.

[0036] Further, step C2 further comprises step C21, obtaining a first depth of the first entity in the clause knowledge graph; and step C22, correcting the first intention coefficient by using the normalized first depth as a weight.

[0037] Preferably, the first depth of the first entity in the clause knowledge graph is obtained, i.e., the hierarchical position of the entity in the clause knowledge graph, which reflects the closeness of its association with the core clause. The root node has a depth of 0, and the child nodes increase layer by layer. In the normalization process, according to the structural characteristics of the insurance clause knowledge graph, the upper limit of the highest depth is set to 5. The structural characteristics of the insurance clause knowledge graph include the hierarchical characteristics of the insurance clause system and the complexity and timeliness of the calculation. Specifically, most insurance clause structures can cover the semantic granularity from product categories to specific clause details within 5 layers, and too deep hierarchical traversal may increase the complexity of graph traversal and vector calculation, thereby affecting the timeliness of the dialogue response. By setting the depth of 5 as the upper limit of the highest depth, the stability in a high-concurrency scenario can be ensured. For example, medical insurance covers liability, which has a depth of 1, and hospital expenses, which has a depth of 2. Then, the depth value is normalized, i.e., the first depth is converted into a weight in the range of [0, 1] to avoid numerical scale differences. Finally, the first intention coefficient is corrected, i.e., the corrected intention coefficient = intention coefficient x (1 + normalized depth weight). The greater the depth, the stronger the influence of the entity on the intention. For example, "hospital expenses" is more specific than "coverage liability" and should be given a higher weight. The intention coefficient correction improves the accuracy of the insurance dialogue system in complex clause scenarios.

[0038] Step S300, performing feedback analysis on the target dynamic intention vector by the insurance large model to obtain a target single-round feedback record.

[0039] Step S300 further comprises step S310, extracting a first entity category in a predetermined entity category; step S320, obtaining a first total frequency of the first entity category in the target dynamic intention vector; step S330, performing feedback sorting on the first entity category based on the first total frequency to obtain a target feedback list; step S340, combining the insurance large model to expand and render the target feedback list to obtain target single-round feedback information; step S350, obtaining dialogue tracking of the target single-round feedback information, denoted as the target single-round feedback record; wherein the predetermined entity category includes product liability, exemption circumstances, and claim conditions.

[0040] Preferably, the target dynamic intention vector is analyzed by the insurance large model, and structured feedback is quickly generated according to the key entity categories in the user dynamic intention, so as to ensure that the reply is consistent with the clauses and close to the user's needs. Specifically, a first entity category in the predetermined entity category is extracted, wherein the predetermined entity category includes product liability such as "hospitalization expense compensation" and "surgery guarantee", exemption cases such as "no compensation for previous illness" and "war excepted", and claim conditions such as "diagnosis proof required" and "reporting within 72 hours"; any one entity category is selected from the target dynamic intention vector as the first entity category, and a first total frequency of the first entity category in the target dynamic intention vector is obtained, that is, the number of occurrences of the same category entity in the dynamic intention vector is counted. For example, the target dynamic intention vector contains [fracture, hospitalization, compensation, diagnosis proof], and the frequency statistics are as follows: product liability, fracture and hospitalization, frequency = 2; claim condition, compensation and diagnosis proof, frequency = 2; exemption case, frequency = 0.

[0041] Preferably, the first entity category is sorted based on the first total frequency, that is, the entity categories are arranged in descending order of frequency, and the high-frequency categories are preferentially fed back to obtain a target feedback list, such as [product liability, claim condition, exemption case], which indicates that the user pays more attention to "whether to compensate" and "how to compensate" rather than "what not to compensate"; then the target feedback list is expanded and rendered by the insurance large model. Specifically, the entity categories of the sorted target feedback list are converted into natural language replies, and associated insurance clause details are supplemented, such as inputting product liability and claim condition, and the insurance large model outputs product liability "your medical insurance covers fracture and hospitalization expenses, with a compensation ratio of 80%", and claim condition "please provide hospital diagnosis proof and expense list, and apply within 30 days after discharge", and then the target single-round feedback information is obtained by combining the output of the insurance large model. Finally, the dialogue tracking of the target single-round feedback information is obtained, including saving the feedback information, user intention and entity category of the current round, which is recorded as the target single-round feedback record for dialogue state management. Thus, the insurance dialogue efficiency is improved, the insurance clauses are converted into colloquial expressions to optimize the user experience, and the accurate response of the insurance dialogue system is ensured.

[0042] Step S400, introducing a clarification optimization mechanism to clarify and optimize the target single-round feedback record to obtain a target clarification strategy.

[0043] Step S400 further includes step S410, when the target single-round feedback record has a predetermined tracking situation, the dialogue tree database is called according to the clarification optimization mechanism; step S420, the optimal feedback strategy is determined by analyzing the dialogue tree database, and the optimal feedback strategy is taken as the target clarification strategy; wherein the predetermined tracking situation includes negative emotions and / or reduced certainty.

[0044] Preferably, the target single-round feedback record is subjected to clarification optimization analysis by introducing a clarification optimization mechanism, which is an intelligent decision unit in the insurance dialogue system, for automatically generating an optimal response strategy to eliminate ambiguity, appease emotions or guide the dialogue to an effective solution path when detecting ambiguous user intent, negative emotions or dialogue logic contradictions, thereby improving dialogue efficiency and enhancing user experience. Specifically, when the target single-round feedback record has a predetermined tracking situation, the predetermined tracking situation includes the presence of negative emotions and / or a decrease in certainty, i.e., detecting and determining the tone and emotional polarity of the user's text through an emotional analysis model, such as "I've been waiting for half a month and still haven't been compensated. Is it possible?", and detecting a decrease in certainty through intent confidence score or dialogue logic contradiction judgment, such as the user asking "What about the outpatient department? Hospitalization? Surgery?", then the clarification optimization mechanism retrieves the dialogue tree database, wherein each node of the dialogue tree database contains emotional labels, intent types, and candidate strategies, according to the current dialogue state, such as "delayed compensation complaint", to retrieve the corresponding node in the dialogue tree to determine the optimal feedback strategy, which may include emotion priority, if the user is angry, prioritize appeasement or manual strategy; intent clarification, if the user is confused, choose to ask or illustrate the strategy; historical path reference, i.e., combine past dialogue records to output strategies, such as the user has refused the urgent option, then skip that branch; finally, the optimal feedback strategy is converted into a specific reply as the target clarification strategy, such as "Do you mean hospitalization expenses or compensation for outpatient surgery?" "I understand your feelings, and we will prioritize your case." "You can upload all materials through the APP to avoid repeated submissions." This achieves intelligent response to user emotions and improves satisfaction, while ensuring accurate intent clarification.

[0045] Further, step S420 further comprises step S421, traversing the target dynamic intent vector in the historical insurance dialogue database to obtain the dialogue tree database; step S422, obtaining a first dialogue tree in the dialogue tree database, the first dialogue tree corresponding to a first dialogue terminal state index; step S423, determining a target dialogue tree with the maximum first dialogue terminal state index as the target; step S424, forming the optimal feedback strategy based on the multi-round dialogue in the target dialogue tree; wherein step S421 comprises: step D1, extracting a first historical dynamic intent vector in the historical insurance dialogue database; step D2, when the first historical dynamic intent vector and the first vector of the target dynamic intent vector have a first vector similarity reaching a predetermined limit value, adding the first historical dialogue information corresponding to the first historical dynamic intent vector to the dialogue tree database.

[0046] Preferably, by comparing the dynamic intention vector of the current user with similar cases in the historical dialogue database, an optimized dialogue tree database is constructed, and the path that can bring the best dialogue outcome is selected as the optimal feedback strategy, that is, the dialogue terminal state index is the highest, the historical successful dialogue experience is used to guide the current interaction, the data-driven intelligent customer service optimization is realized, specifically, the target dynamic intention vector is used to search similar intention vectors in the historical insurance dialogue database, that is, the cosine similarity is used to calculate the similarity matching degree of the intention vector, if it exceeds the predetermined limit value (such as 0.8), the historical dialogue is included in the candidate pool, and the complete multi-round interaction of all matched historical dialogues is converted into a tree structure, each branch represents a possible dialogue path, and finally a dialogue tree database is constructed.

[0047] Preferably, the first historical dynamic intention vector is extracted from the historical insurance dialogue database, including intention type, key entity and weight, and time context such as dialogue round and interaction time, the cosine similarity or Euclidean distance is used to quantify the matching degree of the first historical dynamic intention vector and the target dynamic intention vector, to determine the first vector similarity, the predetermined limit value refers to the similarity limit value set according to historical data, which is used to judge whether it is similar enough, when the first vector similarity reaches the predetermined limit value, the first historical dialogue information corresponding to the first historical dynamic intention vector is added to the dialogue tree database, wherein the first historical dialogue information includes the complete multi-round record of the historical dialogue, the final solution and the user satisfaction, and finally the matched historical dialogue information is converted into a new dialogue tree branch, and the dialogue tree database is updated.

[0048] Preferably, a first dialogue tree is randomly obtained from the dialogue tree database, the first dialogue tree corresponds to the first dialogue terminal state index, wherein the dialogue terminal state index is used to quantify the solution effect of the historical dialogue, for example, successful solution, user satisfaction and closed loop, dialogue terminal state index = 1.0; partial solution, need to transfer to manual, dialogue terminal state index = 0.6; failure, user gives up, dialogue terminal state index = 0.2; wherein in some specific business scenarios, "transfer to manual" is regarded as the terminal state of dialogue success, such as in the stranger customer sales scenario, the core target is not to complete the transaction directly through automatic dialogue, but to guide the potential user to establish contact, if the user agrees to transfer to manual for dialogue or adds the contact method to establish contact through multi-round dialogue, it means that the target has been achieved, at this time the dialogue terminal state index is determined as successful. Then taking the maximum first dialogue terminal state index as the target, the target dialogue tree is determined, that is, the historical dialogue tree with the highest dialogue terminal state index is preferentially selected, such as the material list path with index 1.0; finally, based on the multi-round dialogue in the target dialogue tree, the key steps are extracted as the optimal feedback strategy.

[0049] Step S500, according to the target clarification strategy, the target insurance dialogue is tracked for intent processing.

[0050] Preferably, according to the target clarification strategy, the target insurance dialogue is tracked for intent processing, real-time monitoring of user feedback, dynamic adjustment of the dialogue path, dynamic maintenance of the consistency of the dialogue context, and real-time optimization of the response according to the user feedback, to ensure that the user intent is completely understood and finally solved. Specifically, after the dialogue system applies the target clarification strategy, the user's subsequent reply is analyzed in real time to determine whether the intent is clear, and the entity weight and intent label in the dynamic intent vector are adjusted according to the user's new input; the intent and entity of the historical round are recorded to prevent repeated questioning or logical contradiction, and if the intent is still unclear after clarification, such as the user continues to ask vague questions, a higher level strategy is triggered, such as transferring to manual processing. Further optimize the efficiency of the dialogue business, improve the accuracy of multi-round dialogue intent recognition, and improve the user experience, to ensure that the path from the problem to the solution is the shortest and most accurate.

[0051] In the foregoing, with reference to Figure 1 The method for tracking insurance dialogue intent based on a large model according to an embodiment of the application is described in detail. Next, with reference to Figure 2 The system for tracking insurance dialogue intent based on a large model according to an embodiment of the application will be described.

[0052] The system for tracking insurance dialogue intent based on a large model according to an embodiment of the application is used to solve the technical problems of inaccurate intent understanding, insufficient dynamic intent tracking capability, and lack of ambiguity clarification mechanism in the prior art, and achieves the technical effects of enhancing dynamic intent state modeling capability, optimizing active clarification and feedback mechanism, and improving multi-round dialogue intent recognition accuracy. As shown in Figure 2 The system for tracking insurance dialogue intent based on a large model includes an encoding analysis module 10, an intent state analysis module 20, a feedback analysis module 30, a clarification optimization analysis module 40, and an intent tracking processing module 50.

[0053] The encoding analysis module 10 is used to perform encoding analysis on the clause knowledge graph and historical dialogue corpus through a bidirectional encoding structure, and construct an insurance large model; the intent state analysis module 20 is used to perform dynamic intent state analysis on the target insurance dialogue to obtain a target dynamic intent vector; the feedback analysis module 30 is used to perform feedback analysis on the target dynamic intent vector through the insurance large model to obtain a target single-round feedback record; the clarification optimization analysis module 40 is used to introduce a clarification optimization mechanism to perform clarification optimization analysis on the target single-round feedback record to obtain a target clarification strategy; and the intent tracking processing module 50 is used to track the intent of the target insurance dialogue according to the target clarification strategy.

[0054] Hereinafter, the specific configuration of the encoding analysis module 10 will be described in detail. The encoding analysis module 10 further comprises: the bidirectional encoding structure comprises a first encoding layer and a second encoding layer, the first encoding layer is used for encoding processing of the clause knowledge graph, and a clause encoding unit is obtained; the second encoding layer is used for encoding processing of the historical dialogue corpus, and a dialogue encoding unit is obtained; the clause encoding unit and the dialogue encoding unit are jointly optimized and analyzed to obtain the insurance large model; wherein the first encoding layer is used for encoding processing of the clause knowledge graph to obtain the clause encoding unit, which comprises: extracting a first entity relationship in the clause knowledge graph; encoding a first triple corresponding to the first entity relationship through the first encoding layer to obtain a first encoding unit; and assembling the clause encoding unit based on the first encoding unit; wherein the first entity relationship comprises a first product liability, a first exemption situation and a first claim condition; wherein the second encoding layer is used for encoding processing of the historical dialogue corpus to obtain the dialogue encoding unit, which comprises: identifying a first dialogue segment in the historical dialogue corpus based on a BiLSTM-CRF model; performing intent identification on the first dialogue segment to obtain a first intent label; encoding the first dialogue segment and the first intent label through the second encoding layer to obtain a second encoding unit; and assembling the dialogue encoding unit based on the second encoding unit.

[0055] Hereinafter, the specific configuration of the encoding analysis module 10 will be described in detail. The encoding analysis module 10 further comprises: when the insurance clause version is changed, an incremental training plan is started; wherein the incremental training plan refers to a scheme for triggering incremental training of the insurance large model by comparing entity differences in the insurance clause version.

[0056] Hereinafter, the specific configuration of the intent state analysis module 20 will be described in detail. The intent state analysis module 20 further comprises: extracting an arbitrary round dialogue in the target insurance dialogue, and the arbitrary round dialogue corresponds to an arbitrary time identifier; performing dialogue information screening processing on the arbitrary round dialogue according to a gating mechanism to obtain arbitrary dialogue information; assembling an arbitrary entity set in the arbitrary dialogue information and analyzing the arbitrary entity set to determine an arbitrary dialogue intent; establishing a target dynamic intent vector according to a mapping relationship between the arbitrary time identifier and the arbitrary dialogue intent; wherein it comprises: assembling an arbitrary word set in the arbitrary dialogue information, wherein the arbitrary word set comprises a first word; matching a first entity corresponding to the first word and combining a first appearance frequency of the first word to obtain a first intent coefficient of the first entity; obtaining a first intent coefficient in an intent coefficient sequence obtained in descending order of the first intent coefficient; and taking a first target entity corresponding to the first intent coefficient as the arbitrary dialogue intent.

[0057] Next, the specific configuration of the intention state analysis module 20 will be described in detail. The intention state analysis module 20 further comprises: obtaining a first depth of the first entity in the clause knowledge graph; and correcting the first intention coefficient with the normalized first depth as a weight.

[0058] Next, the specific configuration of the feedback analysis module 30 will be described in detail. The feedback analysis module 30 further comprises: extracting a first entity category in a predetermined entity category; obtaining a first total frequency of the first entity category in the target dynamic intention vector; performing feedback sorting on the first entity category based on the first total frequency to obtain a target feedback list; combining the insurance large model to expand and render the target feedback list to obtain target single-round feedback information; obtaining a dialogue tracking of the target single-round feedback information, denoted as the target single-round feedback record; wherein the predetermined entity category includes product liability, exemption situation and claim condition.

[0059] Next, the specific configuration of the clarification optimization analysis module 40 will be described in detail. The clarification optimization analysis module 40 further comprises: when the target single-round feedback record has a predetermined tracking situation, calling a dialogue tree database according to the clarification optimization mechanism; analyzing the dialogue tree database to determine an optimal feedback strategy, and taking the optimal feedback strategy as the target clarification strategy; wherein the predetermined tracking situation includes the appearance of negative emotions and / or the determination degree being reduced.

[0060] Next, the specific configuration of the clarification optimization analysis module 40 will be described in detail. The clarification optimization analysis module 40 further comprises: traversing the target dynamic intention vector in a historical insurance dialogue database to obtain the dialogue tree database; obtaining a first dialogue tree in the dialogue tree database, the first dialogue tree corresponding to a first dialogue terminal state index; taking the maximum first dialogue terminal state index as a target, determining a target dialogue tree; forming the optimal feedback strategy based on multiple rounds of dialogue in the target dialogue tree; wherein it includes: extracting a first historical dynamic intention vector in the historical insurance dialogue database; when the first vector similarity between the first historical dynamic intention vector and the target dynamic intention vector reaches a predetermined limit value, adding first historical dialogue information corresponding to the first historical dynamic intention vector to the dialogue tree database.

[0061] The insurance dialogue intention tracking system based on a large model provided in the embodiments of the present application can execute the insurance dialogue intention tracking method based on a large model provided in any embodiment of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0062] Figure 3is a structural schematic diagram of an electronic device provided by an embodiment of the present application, which shows a block diagram of an exemplary electronic device suitable for implementing the embodiment of the present application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functions and use range of the embodiment of the present application. The electronic device is in the form of a general computing device, and its components can include but are not limited to an input device 401, a processor 402, a memory 403, and an output device 404. The processor 402 can be one or more; the memory 403 can include a computer readable medium and at least one program product, which has a set of (at least one) program modules configured to perform the functions of the embodiments of the present application.

[0063] The memory 403 shown in the embodiment of the present application can adopt any combination of one or more computer readable media; the computer readable storage medium can be but is not limited to an infrared ray, a semiconductor system, a device or a component, or any combination of the above, for storing software programs, computer executable programs and modules, such as the program instructions / modules of the insurance dialogue intent tracking method based on a large model in the embodiment of the present application. The processor 402 performs various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory 403, that is, implements the above-mentioned insurance dialogue intent tracking method based on a large model.

[0064] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be implemented; in addition, the specific names of the functional units are only for easy mutual differentiation, and do not limit the protection scope of the present application.

[0065] The above specific embodiments do not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for tracking insurance dialogue intent based on a large model, characterized in that, include: The insurance knowledge graph and historical dialogue data are encoded and analyzed using a two-way coding structure, and a large insurance model is constructed. Dynamic intent state analysis is performed on the target insurance dialogue to obtain the target dynamic intent vector; The target dynamic intent vector is analyzed by the insurance big data model to obtain the target single-round feedback record; A clarification and optimization mechanism is introduced to perform clarification and optimization analysis on the single-round feedback record of the target, thereby obtaining the target clarification strategy; The target insurance dialogue is processed for intent tracking based on the target clarification strategy. Among them, dynamic intent state analysis is performed on the target insurance dialogue to obtain the target dynamic intent vector, including: Extract any round of dialogue from the target insurance dialogue, where each round of dialogue corresponds to any time marker; The dialogue information of any round of dialogue is filtered and processed according to the gating mechanism to obtain arbitrary dialogue information; Construct an arbitrary set of entities from the arbitrary dialogue information, and analyze the arbitrary set of entities to determine the arbitrary dialogue intent; Based on the mapping relationship between the arbitrary time identifier and the arbitrary dialogue intent, the target dynamic intent vector is established; This includes: Construct an arbitrary set of words from the arbitrary dialogue information, wherein the arbitrary set of words includes a first word; Match the first entity corresponding to the first word, and combine the first occurrence frequency of the first word to obtain the first intent coefficient of the first entity; Obtain the first intent coefficient in the descending sequence of the first intent coefficients; The first target entity corresponding to the first intent coefficient is taken as the arbitrary dialogue intent; The process, after matching the first entity corresponding to the first word and combining it with the first frequency of occurrence of the first word to obtain the first intent coefficient of the first entity, further includes: Obtain the first depth of the first entity in the term knowledge graph; The first intent coefficient is corrected using the normalized first depth as the weight.

2. The insurance dialogue intent tracking method based on a large model as described in claim 1, characterized in that, The bidirectional coding structure includes a first coding layer and a second coding layer. It uses this structure to encode and analyze the clause knowledge graph and historical dialogue data, and constructs a large-scale insurance model, including: The clause knowledge graph is encoded using the first encoding layer to obtain clause encoding units; The historical dialogue corpus is encoded using the second encoding layer to obtain a dialogue encoding unit. By performing joint optimization analysis on the clause coding unit and the dialogue coding unit, the insurance big model is obtained; The clause knowledge graph is encoded through the first encoding layer to obtain clause encoding units, including: Extract the first entity relation from the aforementioned terms knowledge graph; The first coding unit is obtained by encoding the first triplet corresponding to the first entity relationship through the first coding layer; The clause coding unit is constructed based on the first coding unit; The first entity relationship includes the first product liability, the first exemption from liability, and the first claim conditions; The historical dialogue corpus is encoded through the second encoding layer to obtain a dialogue encoding unit, including: The first dialogue segment in the historical dialogue corpus was identified based on the BiLSTM-CRF model. The first dialogue segment is marked with intent to obtain a first intent label; The first dialogue segment and the first intent tag are encoded by the second encoding layer to obtain the second encoding unit; The dialogue coding unit is constructed based on the second coding unit.

3. The insurance dialogue intent tracking method based on a large model as described in claim 1, characterized in that, When the version of the insurance terms changes, the incremental training plan will be activated. The incremental training plan refers to a scheme that triggers incremental training of the insurance big data model by comparing the entity differences in the insurance terms versions.

4. The insurance dialogue intent tracking method based on a large model as described in claim 1, characterized in that, By performing feedback analysis on the target's dynamic intent vector using the aforementioned insurance big data model, a single-round feedback record of the target is obtained, including: Extract the first entity category from the predefined entity categories; Obtain the first total frequency of the first entity category in the target dynamic intent vector; Based on the first total frequency, the first entity category is sorted according to the feedback to obtain the target feedback list; The target feedback list is expanded and rendered by combining the aforementioned insurance big model to obtain single-round target feedback information; The dialogue tracking that obtains the target's single-turn feedback information is recorded as the target's single-turn feedback record; The categories of pre-determined entities include product liability, exemptions, and claims conditions.

5. The insurance dialogue intent tracking method based on a large model as described in claim 1, characterized in that, A clarification and optimization mechanism is introduced to perform clarification and optimization analysis on the single-round feedback records of the target, resulting in a target clarification strategy, including: When the target single-round feedback record has a predetermined tracking situation, the dialogue tree database is retrieved according to the clarification and optimization mechanism; The optimal feedback strategy is determined by analyzing the dialogue tree database, and the optimal feedback strategy is used as the target clarification strategy. The predetermined tracking scenarios include the occurrence of negative emotions and / or a decrease in certainty.

6. The insurance dialogue intent tracking method based on a large model as described in claim 5, characterized in that, Analyzing the dialogue tree database to determine the optimal feedback strategy includes: The target dynamic intent vector is traversed in the historical insurance dialogue database to obtain the dialogue tree database; Obtain the first dialogue tree from the dialogue tree database, where the first dialogue tree corresponds to the first dialogue end state index. The target dialogue tree is determined with the goal of maximizing the end state index of the first dialogue. The optimal feedback strategy is formed based on the multi-turn dialogue in the target dialogue tree; This includes: Extract the first historical dynamic intent vector from the historical insurance dialogue database; When the first vector similarity between the first historical dynamic intent vector and the first vector of the target dynamic intent vector reaches a predetermined limit, the first historical dialogue information corresponding to the first historical dynamic intent vector is added to the dialogue tree database.

7. An insurance dialogue intent tracking system based on a large model, characterized in that, The system is used to implement the insurance dialogue intent tracking method based on a large model as described in any one of claims 1 to 6, and the system comprises: The coding analysis module is used to perform coding analysis on the clause knowledge graph and historical dialogue data through a bidirectional coding structure, and to build a large insurance model. The intent state analysis module is used to perform dynamic intent state analysis on the target insurance dialogue to obtain the target dynamic intent vector. The feedback analysis module is used to perform feedback analysis on the target dynamic intent vector through the insurance big model to obtain the target single-round feedback record; The clarification and optimization analysis module is used to introduce a clarification and optimization mechanism to perform clarification and optimization analysis on the single-round feedback record of the target, and obtain the target clarification strategy. An intent tracking processing module is used to perform intent tracking processing on the target insurance dialogue according to the target clarification strategy.

8. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the insurance dialogue intent tracking method based on a large model as described in any one of claims 1 to 6.

Images