An online insurance application management method and system

Abstract

The application belongs to the field of insurance management, and discloses an online insurance management method and system. The method comprises the following steps: obtaining natural language interaction information input by a user in a current round; calling a pre-trained semantic analysis model to perform real-time intention analysis, and generating an insurance demand intention vector containing demand slots, values and confidence levels thereof; obtaining a conversation context memory containing historical intention vectors, time stamps and confidence levels of each historical slot; fusing the current intention vector and the conversation context memory, weighting and merging or conflict resolving historical values and current values of the same demand slot according to a time sequence decay factor, and generating a dynamic user demand portrait; and performing matching and searching in an insurance product library according to the dynamic user demand portrait, and outputting a candidate insurance product list and matching confidence levels of each product. The application improves the accuracy and interaction efficiency of online matching of insurance products.

Description

Technical Field

[0001] This invention relates to the field of insurance management, and more particularly to an online insurance management method and system. Background Technology

[0002] Currently, online insurance management systems typically employ user profile-based insurance product matching methods. During the insurance decision-making process, users often express their protection needs in natural language, such as "I want to buy critical illness insurance for my parents" or "Are there any accident insurance policies suitable for people who travel frequently?" To accurately capture user needs, existing systems attempt to introduce a conversational interaction module, guiding users to gradually provide their required information through multiple rounds of question-and-answer sessions.

[0003] However, existing systems suffer from a core technical deficiency in integrating natural language interaction with product matching: a lack of ability to perform real-time intent parsing of user natural language expressions and integrate multi-turn dialogue context. Specifically, this manifests as follows: When user input contains colloquial expressions, referential relationships, or implicit conditions, the system cannot accurately interpret the semantics of insurance needs, resulting in the loss of a large amount of information. In multi-turn dialogue scenarios, the system cannot maintain the memory of dialogue states across turns. When users supplement or correct their needs in different turns, the system cannot effectively integrate the information before and after, resulting in fragmented information about the needs. The update of user profiles relies on offline batch processing or fixed-period synchronization, which cannot immediately feed back the intent information generated in real time during the conversation to the profile, resulting in a time mismatch between the profile data used by the matching engine and the user's current expressed needs.

[0004] The essence of the above-mentioned defects lies in the fact that the existing system implements the three functional modules of intent parsing, context management, and profile update in a fragmented manner, lacking a unified technical framework to realize a real-time, continuous, context-aware demand acquisition and dynamic profile generation mechanism. Ultimately, this results in the inability of the accuracy of insurance product matching and interaction efficiency to meet the needs of actual applications.

[0005] Therefore, there is an urgent need for an online insurance management method and system that can achieve real-time intent perception and multi-turn dialogue context fusion to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to disclose an online insurance management method and system to solve the technical problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, the present invention provides an online insurance management method, including: S1, Obtain the natural language interaction information input by the user in the current round; S2, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information, and generate the insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level. S3, obtain the session context memory, which includes at least the historical intent vector obtained from the historical rounds parsing and the timestamps and confidence levels of each historical slot; S4, the insurance demand intent vector of the current round is fused with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, the historical value and the current value are weighted and merged or conflict resolved according to the time decay factor, and the value of the slot is updated according to the slot position confidence after fusion. S5. Based on the dynamic user demand profile, perform matching and retrieval in the insurance product database, and output a list of candidate insurance products and the matching confidence of each product.

[0008] Optionally, generating the insurance demand intent vector for the current round includes: S21, retrieve the value of the historical demand slot; S22, based on the value of the historical demand slot, perform referential resolution and omission completion on the natural language interaction information to obtain the completed natural language interaction information; S23, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the completed natural language interaction information to generate the insurance demand intent vector for the current round.

[0009] Optionally, following S4, the following also includes: The session context memory is updated based on the values ​​and confidence levels of each demand slot in the dynamic user demand profile.

[0010] Optionally, the weighted merging of historical and current values ​​based on the time-series decay factor includes: Get the first time interval between the timestamp of the historical values ​​of the same demand slot and the current time; The decay weight of the historical values ​​is calculated based on the first time interval and the preset time decay function. The historical values ​​and the current values ​​are weighted and merged according to the attenuation weight.

[0011] Optionally, the conflict resolution includes: Determine whether there is a conflict between the historical value and the current value for the same demand slot; If a conflict exists, and the confidence level of the current value is higher than the first preset threshold, then the current value replaces the historical value.

[0012] Optionally, the conflict resolution further includes: When there are multiple historical values ​​for the same demand slot that conflict with the current value, obtain the confidence level of each historical value after time-series decay. Sort the confidence scores of each historical value with the confidence score of the current value; The value with the highest confidence level is selected as the valid value for the required slot.

[0013] Optionally, S5 includes: The values ​​and confidence scores of each demand slot in the dynamic user demand profile are converted into retrieval feature vectors. Calculate the similarity between the retrieved feature vector and the product vectors of each insurance product in the insurance product database; Insurance products with similarity exceeding a second preset threshold are selected, a list of candidate insurance products is generated, and the similarity is used as the matching confidence level for each product.

[0014] Optionally, after S5, it also includes: Receive the user's selection operation on the candidate insurance product list and obtain the product vector of the selected insurance product; The product vector is associated with the insurance demand intent vector of the current round and stored as a training sample; When the cumulative number of training samples reaches a third preset threshold, the parameters of the pre-trained semantic parsing model are updated.

[0015] Optionally, prior to S3, the following are also included: Detect the session state of the current session; When a session end event is detected, or when the session idle time exceeds a fourth preset threshold, the session context memory is reset.

[0016] On the other hand, the present invention provides an online insurance management system, comprising: The acquisition module is used to acquire natural language interaction information input by the user in the current round; The parsing module is used to call a pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information and generate an insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level. The acquisition module is also used to acquire the session context memory, which includes at least the historical intent vector obtained from the historical rounds of parsing and the timestamps and confidence levels of each historical slot. The fusion module is used to fuse the insurance demand intent vector of the current round with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, weighting and merging the historical values ​​and the current values ​​or resolving conflicts according to the time decay factor, and updating the value of the slot according to the slot position confidence after fusion. The retrieval module is used to perform matching retrieval in the insurance product database based on the dynamic user demand profile, and output a list of candidate insurance products and the matching confidence of each product. Beneficial effects

[0017] Compared with existing technologies, this invention solves the problem of the fragmentation of intent parsing, context management, and profile updating in the background technology by constructing a unified processing framework that integrates real-time intent parsing, context memory management, and temporal decay fusion. Addressing the deficiency of existing systems in accurately parsing colloquial expressions and referential relationships, this invention extracts historical slot values ​​from the conversation context memory before intent parsing, performs referential resolution and omission completion on user input, and then parses it, significantly improving the accuracy and completeness of semantic parsing. Addressing the deficiency of fragmented information in multi-turn dialogues, this invention uses a temporal decay factor to weight and merge or resolve conflicts between historical and current values. After fusion, it updates the conversation context memory and resets the memory through conversation state detection, achieving continuous accumulation and intelligent management of cross-turn demand information. Addressing the temporal misalignment caused by offline profile updates, this invention uses temporal decay to decrease the weight of historical values ​​over time. In case of conflict, a value with higher confidence replaces the historical value. When multiple historical values ​​are available, the one with the highest confidence after decay is selected, ensuring that the user profile is updated in real-time with each round of dialogue, always reflecting the user's latest demand status. During matching and retrieval, dynamic profiles are converted into feature vectors, and similarity calculations are performed with product vectors, followed by threshold filtering. This ensures that the matching engine always performs retrievals based on real-time and accurate profiles. Furthermore, user selection feedback triggers updates to the semantic parsing model parameters, continuously optimizing the system's parsing capabilities with usage. This invention works synergistically across four levels: intent parsing accuracy, contextual continuity, profile real-time performance, and matching precision, effectively improving product matching accuracy and interaction efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of an online insurance management method according to the present invention.

[0020] Figure 2 This is a schematic diagram illustrating the method for generating the insurance demand intent vector for the current round according to the present invention.

[0021] Figure 3 This is a schematic diagram of an online insurance management system according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the present invention provides an online insurance management method, including: S1: Obtain the natural language interaction information input by the user in the current round.

[0024] The term "user" refers to the party initiating an online insurance consultation.

[0025] The current round refers to the current interaction round in which the online insurance management method is executed in a complete process.

[0026] The natural language interaction information refers to the information describing the insurance needs input by the user in natural language form, including text information and / or text information obtained by speech conversion.

[0027] Optionally, obtain natural language interaction information from the user's current input, including: Receive text information entered by the user in the interactive interface; Alternatively, the system can receive voice information input by the user on the interactive interface, and perform speech recognition on the voice information to obtain the corresponding text information. The text information is used as the natural language interaction information input in the current round.

[0028] S2, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information, and generate the insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level.

[0029] The semantic parsing model refers to a pre-trained neural network model used to perform semantic understanding on natural language text in order to extract insurance demand information contained therein.

[0030] The real-time intent parsing refers to the immediate semantic understanding of the natural language interaction information after the user inputs the natural language interaction information for the current round, in order to extract the insurance demand intent for the current round.

[0031] The insurance demand intent vector refers to structured data used to represent the insurance demand expressed by a user in the current round, including at least one demand slot and the corresponding value and confidence level of each demand slot.

[0032] The demand slot refers to the identifier of the demand dimension contained in the insurance demand intention. The demand dimension includes at least one of the following: insurance type, insured role, age range, coverage period, and premium budget.

[0033] The value refers to the specific content corresponding to the demand slot.

[0034] The confidence level refers to the probability assessment value of the semantic parsing model for the correctness of the value of the demand slot.

[0035] Optionally, the pre-trained semantic parsing model is a pre-trained language model based on the Transformer architecture, such as any one of the BERT model, RoBERTa model, and GPT model.

[0036] Optionally, the training process of the pre-trained semantic parsing model includes: Obtain insurance domain corpus, annotate the demand slots and values ​​of the insurance domain corpus, and construct an insurance domain annotated dataset; The pre-trained language model is fine-tuned using the labeled dataset in the insurance field. During the fine-tuning process, the cross-entropy loss between the probability distribution of the demand slot output by the pre-trained language model and the labeled data is used as the training objective. When the cross-entropy loss converges to a preset condition, the pre-trained semantic parsing model is obtained.

[0037] The insurance-related corpus refers to text data collected from insurance product brochures, policyholder notices, insurance terms and conditions, and insurance consultation dialogue records.

[0038] The labeled data refers to the standard answer data obtained by labeling the demand slots and values ​​of each text data in the insurance domain corpus. The labeled data is used to calculate the loss of the model output during the fine-tuning process.

[0039] Wherein, the value probability distribution refers to the distribution of probability values ​​of each candidate value output by the pre-trained language model for each demand slot, and the value probability distribution is expressed as:

[0040] in: Indicates the first The probability distribution vector of the values ​​of each demand slot; Indicates the first The weight matrix corresponding to each demand slot (if we take...) If viewed as a column vector, then usually , This represents the number of candidate values ​​for this slot. (Hidden layer dimension) This represents the hidden layer representation vector obtained by the pre-trained language model after encoding the input text; Indicates the first The bias vector corresponding to each demand slot.

[0041] Wherein, the cross-entropy loss refers to the cross-entropy loss function value between the probability distribution of the values ​​and the labeled data, and the calculation formula for the cross-entropy loss is:

[0042] in: This represents the cross-entropy loss value; The index representing the demand slot; Indicates the first The index of the candidate values ​​for each demand slot; Indicates the first in the labeled data The first demand slot The true labels for each candidate value (usually in one-hot or indicator function form): And for the same slot ); In the probability distribution of the stated value, the first... The first demand slot The predicted probability values ​​of each candidate value (obtained from the softmax output above).

[0043] The preset condition refers to the change in the cross-entropy loss value over multiple consecutive training rounds being less than a fifth preset threshold, or the training rounds reaching a preset maximum number of training rounds.

[0044] Optional, such as Figure 2 As shown, generating the insurance demand intent vector for the current round includes: S21, retrieve the value of the historical demand slot.

[0045] The historical demand slots refer to the demand slots that have been resolved in previous rounds before the current round.

[0046] Optionally, retrieve the values ​​of historical demand slots, including: Read the values ​​of each demand slot from the insurance demand intent vector generated in the previous round of parsing; Alternatively, retrieve the value of the demand slot most recently parsed before the current round from the stored session records.

[0047] S22, based on the value of the historical demand slot, perform referential resolution and omission completion on the natural language interaction information to obtain the completed natural language interaction information.

[0048] The term "referential resolution" refers to replacing the referential words in the natural language interaction information with the specific entities corresponding to the values ​​of the historical demand slots.

[0049] The omission completion refers to supplementing the missing requirement information in the natural language interaction information according to the value of the historical requirement slot.

[0050] Furthermore, based on the values ​​of the historical demand slots, the natural language interaction information is subjected to referential resolution and omission completion to obtain the completed natural language interaction information, including: Identify the pronouns in the natural language interaction information and replace the pronouns with the specific entities that match the values ​​of the historical demand slots; Identify the missing demand slots in the natural language interaction information, and supplement the natural language interaction information with the corresponding values ​​from the historical demand slots; The replaced and supplemented natural language interaction information is used as the completed natural language interaction information.

[0051] The pronoun refers to a word in the natural language interaction information used to refer to content already mentioned above, and the pronoun includes at least one of personal pronouns and demonstrative pronouns.

[0052] The specific entity refers to the specific content in the value of the historical demand slot that has a referential relationship with the pronoun. The specific entity includes at least one of a specific person, a specific type of insurance, and specific time information.

[0053] S23, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the completed natural language interaction information to generate the insurance demand intent vector for the current round.

[0054] Through the processes described in S21 to S23, before invoking the semantic parsing model, the values ​​of historical demand slots are used to resolve pronouns and complete omissions in the user's current input. This restores the pronouns in the user's colloquial expression to concrete entities and automatically completes the omitted demand information. For example, when the user inputs "Buy him a medical insurance policy" in the current round, by extracting the value of "insured person role" from the historical demand slot as "father," "he" can be resolved to "father," resulting in the completed input "Buy my father a medical insurance policy." Therefore, the input information received by the semantic parsing model is more complete and explicit, avoiding the loss of demand semantics due to colloquial expressions and significantly improving the accuracy of insurance demand intent parsing.

[0055] S3, obtain the session context memory, which includes at least the historical intent vector obtained from the historical rounds parsing and the timestamps and confidence levels of each historical slot.

[0056] The session context memory refers to a data structure used to store the parsing results of historical rounds already performed in the current session.

[0057] The historical rounds refer to the interaction rounds in which intent parsing has been completed before the current round.

[0058] The historical intent vector refers to the insurance demand intent vector generated in the historical rounds.

[0059] The historical slots refer to the demand slots contained in the historical intent vector.

[0060] Optionally, retrieve the session context memory, including: Determine if the current session is the first interaction round; If this is the first interaction round, initialize an empty session context memory; If it is not the first interaction round, the stored session context memory is read from the storage unit. The stored session context memory includes the historical intent vectors parsed from previous historical rounds and the timestamps and confidence levels of each historical slot. Use the read session context memory as the session context memory for the current round.

[0061] Optionally, prior to S3, the following are also included: Detect the session state of the current session; When a session end event is detected, or when the session idle time exceeds a fourth preset threshold, the session context memory is reset.

[0062] The session state refers to the current stage of the session, which includes an active state and an ended state.

[0063] The session end event refers to an event that triggers the termination of a session, and the session end event includes at least one of user-initiated session end operation and system-initiated session end operation.

[0064] The idle time of the session refers to the duration of the current session from the last time the user input natural language interaction information was received to the current moment.

[0065] The fourth preset threshold is a preset time threshold used to determine whether the session idle time has exceeded the limit.

[0066] The reset operation refers to the operation of clearing the historical intent vectors and timestamps and confidence scores of each historical slot stored in the session context memory.

[0067] Optionally, a reset operation is performed on the session context memory, including: Clear all historical intent vectors, timestamps, and confidence scores of each historical slot stored in the session context memory; Alternatively, delete the storage space corresponding to the session context memory and reinitialize an empty session context memory.

[0068] Through the aforementioned mechanism for acquiring and resetting the session context memory, the system can obtain complete dialogue state information, including historical intent vectors, timestamps of each historical slot, and confidence levels, in each round, providing a data foundation for subsequent fusion. Simultaneously, by resetting the session context memory when a session ends or when the session idle time exceeds a fourth preset threshold, the system ensures that demand information between different sessions is isolated from each other, preventing insurance demand information from the previous session from interfering with the current new session. This guarantees the intra-session continuity and inter-session independence of demand information accumulation across multiple rounds of dialogue.

[0069] S4, the insurance demand intent vector of the current round is fused with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, the historical value and the current value are weighted and merged or conflict is resolved according to the time decay factor, and the value of the slot is updated according to the slot position confidence after fusion.

[0070] The dynamic user demand profile refers to structured data generated by integrating the current round of insurance demand intent vector with the session context memory, which reflects the user's real-time insurance demand status in the current session.

[0071] The "same demand slot" refers to a demand slot that shares the same demand dimension with the insurance demand intent vector of the current round and the historical intent vector in the session context memory. The "same demand dimension" means that the type identifiers of the demand dimensions corresponding to the two demand slots are identical. The type identifiers of the demand dimension include at least one of the following: insurance type, insured role, age range, coverage period, and premium budget.

[0072] The time-series decay factor is a factor used to decay the weight of historical values ​​over time. The value of the time-series decay factor decreases as the time interval between the timestamp of the historical value and the current time increases.

[0073] Furthermore, the timing decay factor is calculated using the following formula:

[0074] in, Indicates the first Each demand slot in time The time decay factor, Indicates the first The attenuation coefficient corresponding to each demand slot This represents the time interval between the timestamp of a historical value and the current time.

[0075] The attenuation coefficient Based on the preset type of demand slot, the attenuation coefficient corresponding to the insurance type slot is less than the attenuation coefficient corresponding to the insured role slot, and the attenuation coefficient corresponding to the insured role slot is less than the attenuation coefficient corresponding to the premium budget slot.

[0076] The beneficial effect of this factor is that different slot types use different decay coefficients, allowing more stable demand information (such as insurance type) to maintain effective weight for a longer period during the dialogue process, while volatile demand information (such as premium budget) decays rapidly, avoiding interference from outdated information on the current fusion result. Compared with the existing unified decay function, this is more adaptable to the characteristics of the differences in the stability of different dimensions of information in the expression of insurance demand.

[0077] The weighted merging refers to the operation of weighted merging of historical values ​​with the current value after assigning decay weights according to the time-series decay factor.

[0078] The conflict resolution refers to the operation of determining the final value of the demand slot according to a preset resolution rule when there is a semantic inconsistency between the historical value and the current value.

[0079] The merged slot position confidence refers to the confidence level of the required slot after weighted merging or conflict resolution of historical and current values. The merged slot position confidence is calculated as follows: When weighted merging historical values ​​and current values, the confidence scores of the historical values ​​after time-series decay are weighted and summed with the confidence scores of the current values ​​to obtain the fused slot position confidence scores. When conflict resolution is performed between historical values ​​and current values, and the historical values ​​are replaced by the current values, the confidence level of the current values ​​is used as the confidence level of the fused slot positions. When conflict resolution is performed between historical values ​​and current values, and the historical values ​​are retained, the confidence level of the historical values ​​after time-series decay is used as the confidence level of the fused slot position.

[0080] Optionally, the historical and current values ​​can be weighted and merged or conflict resolved based on the time-series decay factor, including: Determine whether there is a semantic conflict between the historical value and the current value of the same demand slot; If there is no semantic conflict, the historical values ​​and the current values ​​are weighted and merged according to the time decay factor; If a semantic conflict exists, the conflict between the historical value and the current value is resolved.

[0081] Optionally, the value of the slot is updated based on the fused slot position confidence, including: Replace the current value of the required slot with the value obtained after weighted merging or conflict resolution; Replace the current confidence level of the required slot with the fused slot position confidence level.

[0082] Optionally, the weighted merging of historical and current values ​​based on the time-series decay factor includes: Get the first time interval between the timestamp of the historical values ​​of the same demand slot and the current time; The decay weight of the historical values ​​is calculated based on the first time interval and the preset time decay function. The historical values ​​and the current values ​​are weighted and merged according to the attenuation weight.

[0083] Optionally, the conflict resolution includes: Determine whether there is a conflict between the historical value and the current value for the same demand slot; If a conflict exists, and the confidence level of the current value is higher than the first preset threshold, then the current value replaces the historical value.

[0084] Optionally, the conflict resolution further includes: When there are multiple historical values ​​for the same demand slot that conflict with the current value, obtain the confidence level of each historical value after time-series decay. Sort the confidence scores of each historical value with the confidence score of the current value; The value with the highest confidence level is selected as the valid value for the required slot.

[0085] The system uses a time-decay factor to weight and merge historical and current values, causing the contribution of historical values ​​to gradually decrease as time intervals increase. For example, if a user initially expresses a budget of 5000 yuan and later expresses a budget of 8000 yuan, the weight of earlier historical values ​​decreases after time-decay, while the weight of the current value is relatively higher. The resulting profile more closely reflects the user's latest expressed needs. When there is a conflict between historical and current values ​​and the current confidence level is higher than a first preset threshold, the current value replaces the historical value. For example, if a user initially expresses "buying insurance for parents" and later corrects it to "buying insurance for myself," the system resolves the conflict by adopting the current expression with higher confidence, avoiding historical information hindering changes in user needs. When multiple historical values ​​exist, the one with the highest confidence level after time-decay is selected as the valid value, ensuring that the fusion result always prioritizes information with the highest credibility and best timeliness. The above mechanisms work together to ensure that the dynamic user demand profile is updated in real time with each round of dialogue, eliminating the time misalignment between the profile data and the user's current expressed needs caused by traditional offline batch processing updates.

[0086] Optionally, following S4, the following also includes: The session context memory is updated based on the values ​​and confidence levels of each demand slot in the dynamic user demand profile.

[0087] Optionally, the session context memory is updated based on the values ​​and confidence levels of each demand slot in the dynamic user demand profile, including: Write the values ​​and confidence levels of each demand slot in the dynamic user demand profile into the session context memory; Write the current time timestamp into the session context memory as the timestamp for each demand slot.

[0088] S5. Based on the dynamic user demand profile, perform matching and retrieval in the insurance product database, and output a list of candidate insurance products and the matching confidence of each product.

[0089] The insurance product database refers to a data set that stores product information for multiple insurance products, including product vectors and product attribute tags for each insurance product.

[0090] The matching retrieval refers to the process of retrieving insurance products that match the user's needs from the insurance product database based on the values ​​and confidence levels of each demand slot in the dynamic user demand profile.

[0091] The candidate insurance product list refers to an ordered set of insurance products that match the dynamic user demand profile and are selected through matching retrieval. The insurance products in the ordered set are arranged according to the matching confidence level.

[0092] The matching confidence level refers to a quantitative indicator of the degree of matching between each insurance product in the candidate insurance product list and the dynamic user demand profile.

[0093] Optionally, S5 includes: S51, convert the values ​​and confidence levels of each demand slot in the dynamic user demand profile into retrieval feature vectors.

[0094] The retrieval feature vector refers to a feature vector composed of the value encoding and confidence weight of each demand slot in the dynamic user demand profile, which is used to calculate the similarity with the product vector in the insurance product database.

[0095] Optionally, the values ​​and confidence scores of each demand slot in the dynamic user demand profile are converted into retrieval feature vectors, including: The values ​​of each demand slot in the dynamic user demand profile are encoded to obtain the value encoding vector corresponding to each demand slot. The value encoding vector is weighted by using the confidence level of each demand slot as the weight to obtain the weighted encoding vector of each demand slot. The weighted encoding vectors of each demand slot are concatenated to obtain the retrieval feature vector.

[0096] S52, calculate the similarity between the retrieved feature vector and the product vectors of each insurance product in the insurance product database.

[0097] The insurance products mentioned above refer to the insurance products stored in the insurance product database that are available for users to choose from.

[0098] The product vector refers to the feature vector obtained by converting the product attribute tags of the insurance product, and the product vector has the same vector dimension as the retrieval feature vector.

[0099] Optionally, calculating the similarity between the retrieved feature vector and the product vectors of each insurance product in the insurance product database includes: Obtain the product vector of each insurance product in the insurance product database; Calculate the cosine similarity between the retrieved feature vector and each product vector, and use the cosine similarity as the similarity.

[0100] S53, filter out insurance products whose similarity exceeds the second preset threshold, generate the candidate insurance product list, and use the similarity as the matching confidence level of each product.

[0101] The second preset threshold is a similarity threshold used to filter insurance products that match the dynamic user demand profile.

[0102] By converting the values ​​and confidence levels of each slot in the dynamic user demand profile into retrieval feature vectors, and then calculating the cosine similarity between these vectors and the product vectors in the insurance product database, the demand slots with higher confidence levels have a greater weight in the retrieval feature vectors. This makes the matching retrieval results more focused on the dimensions of the user's expressed demand with higher confidence. For example, when a user has a high confidence level in "insurance type" but a low confidence level in "premium budget," the retrieval results prioritize matching the insurance type demand explicitly expressed by the user, avoiding interference from low-confidence slots in the matching results, thereby improving the accuracy of matching the candidate insurance product list with the user's actual needs.

[0103] Optionally, generating the candidate insurance product list includes: Select insurance products from the insurance product database whose similarity exceeds a second preset threshold; The selected insurance products are sorted from high to low according to the similarity to generate the candidate insurance product list, and the similarity is used as the matching confidence of each product.

[0104] Optionally, after S5, it also includes: Receive the user's selection operation on the candidate insurance product list and obtain the product vector of the selected insurance product; The product vector is associated with the insurance demand intent vector of the current round and stored as a training sample; When the cumulative number of training samples reaches a third preset threshold, the parameters of the pre-trained semantic parsing model are updated.

[0105] The selection operation refers to the interactive operation in which a user selects an insurance product from the list of candidate insurance products.

[0106] The product vector refers to the feature vector obtained by converting the product attribute tags of the insurance product, and has the same vector dimension as the retrieval feature vector.

[0107] The training sample refers to a sample pair consisting of the product vector selected by the user and the insurance demand intent vector of the current round, which is used for parameter updates of the pre-trained semantic parsing model.

[0108] The cumulative number refers to the total number of training samples stored since the last time the parameters of the pre-trained semantic parsing model were updated.

[0109] The third preset threshold refers to the preset threshold of the cumulative number of training samples used to trigger parameter updates of the pre-trained semantic parsing model.

[0110] The parameter update refers to the operation of adjusting the model parameters of the pre-trained semantic parsing model based on the training samples.

[0111] Optionally, the parameter update process of the pre-trained semantic parsing model includes: Extract the product vector and insurance demand intent vector from each sample pair from the stored training samples; Using each product vector as input and the values ​​of each demand slot in the corresponding insurance demand intent vector as labels, the pre-trained semantic parsing model is incrementally trained. During incremental training, the cross-entropy loss between the value probability distribution output by the pre-trained semantic parsing model and the label is used as the training objective. When the cross-entropy loss converges to a preset condition, the parameter update of the pre-trained semantic parsing model is completed.

[0112] By associating the product vectors selected by users from the candidate insurance product list with the insurance demand intent vector of the current round and storing them as training samples, and triggering parameter updates of the semantic parsing model when the accumulated number of samples reaches a third preset threshold, the system can use users' actual selection behavior as feedback signals to continuously optimize the semantic parsing model's ability to parse insurance demand intent. This online incremental learning mechanism allows the model to gradually adapt to different users' expression habits and demand preferences. With continuous use of the system, the accuracy of intent parsing gradually improves, and the matching and ranking of candidate products better reflects users' actual selection tendencies.

[0113] like Figure 3 As shown, the present invention provides an online insurance management system, comprising: The acquisition module is used to acquire natural language interaction information input by the user in the current round; The parsing module is used to call a pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information and generate an insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level. The acquisition module is also used to acquire the session context memory, which includes at least the historical intent vector obtained from the historical rounds of parsing and the timestamps and confidence levels of each historical slot. The fusion module is used to fuse the insurance demand intent vector of the current round with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, weighting and merging the historical values ​​and the current values ​​or resolving conflicts according to the time decay factor, and updating the value of the slot according to the slot position confidence after fusion. The retrieval module is used to perform matching retrieval in the insurance product database based on the dynamic user demand profile, and output a list of candidate insurance products and the matching confidence of each product.

[0114] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An online insurance application management method, characterized in that, include: S1, Obtain the natural language interaction information input by the user in the current round; S2, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information, and generate the insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level. S3, obtain the session context memory, which includes at least the historical intent vector obtained from the historical rounds parsing and the timestamps and confidence levels of each historical slot; S4, the insurance demand intent vector of the current round is fused with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, the historical value and the current value are weighted and merged or conflict resolved according to the time decay factor, and the value of the slot is updated according to the slot position confidence after fusion. S5. Based on the dynamic user demand profile, perform matching and retrieval in the insurance product database, and output a list of candidate insurance products and the matching confidence of each product.

2. The online insurance management method according to claim 1, characterized in that, The process of generating the insurance demand intent vector for the current round includes: S21, retrieve the value of the historical demand slot; S22, based on the value of the historical demand slot, perform referential resolution and omission completion on the natural language interaction information to obtain the completed natural language interaction information; S23, invoke the pre-trained semantic parsing model to perform real-time intent parsing on the completed natural language interaction information to generate the insurance demand intent vector for the current round.

3. The online insurance management method according to claim 1, characterized in that, Following S4, it also includes: The session context memory is updated based on the values ​​and confidence levels of each demand slot in the dynamic user demand profile.

4. The online insurance management method according to claim 1, characterized in that, The weighted merging of historical and current values ​​based on the time-series decay factor includes: Get the first time interval between the timestamp of the historical values ​​of the same demand slot and the current time; The decay weight of the historical values ​​is calculated based on the first time interval and the preset time decay function. The historical values ​​and the current values ​​are weighted and merged according to the attenuation weight.

5. The online insurance management method according to claim 1, characterized in that, The conflict resolution includes: Determine whether there is a conflict between the historical value and the current value for the same demand slot; If a conflict exists, and the confidence level of the current value is higher than the first preset threshold, then the current value replaces the historical value.

6. The online insurance management method according to claim 5, characterized in that, The conflict resolution also includes: When there are multiple historical values ​​for the same demand slot that conflict with the current value, obtain the confidence level of each historical value after time-series decay. Sort the confidence scores of each historical value with the confidence score of the current value; The value with the highest confidence level is selected as the valid value for the required slot.

7. The online insurance management method according to claim 1, characterized in that, S5 include: The values ​​and confidence scores of each demand slot in the dynamic user demand profile are converted into retrieval feature vectors. Calculate the similarity between the retrieved feature vector and the product vectors of each insurance product in the insurance product database; Insurance products with similarity exceeding a second preset threshold are selected, a list of candidate insurance products is generated, and the similarity is used as the matching confidence level for each product.

8. The online insurance management method according to claim 1, characterized in that, Following S5, it also includes: Receive the user's selection operation on the candidate insurance product list and obtain the product vector of the selected insurance product; The product vector is associated with the insurance demand intent vector of the current round and stored as a training sample; When the cumulative number of training samples reaches a third preset threshold, the parameters of the pre-trained semantic parsing model are updated.

9. The online insurance management method according to claim 1, characterized in that, Before S3, it also included: Detect the session state of the current session; When a session end event is detected, or when the session idle time exceeds a fourth preset threshold, the session context memory is reset.

10. An online insurance management system, characterized in that, include: The acquisition module is used to acquire natural language interaction information input by the user in the current round; The parsing module is used to call a pre-trained semantic parsing model to perform real-time intent parsing on the natural language interaction information and generate an insurance demand intent vector for the current round. The insurance demand intent vector contains at least one demand slot and its value and confidence level. The acquisition module is also used to acquire the session context memory, which includes at least the historical intent vector obtained from the historical rounds of parsing and the timestamps and confidence levels of each historical slot. The fusion module is used to fuse the insurance demand intent vector of the current round with the session context memory to generate a dynamic user demand profile. The fusion includes: for the same demand slot, weighting and merging the historical values ​​and the current values ​​or resolving conflicts according to the time decay factor, and updating the value of the slot according to the slot position confidence after fusion. The retrieval module is used to perform matching retrieval in the insurance product database based on the dynamic user demand profile, and output a list of candidate insurance products and the matching confidence of each product.

Images