Pre-underwriting result generation method and device based on large model, medium and computer

Abstract

The invention discloses a pre-underwriting result generation method and device based on a large model, a medium and a computer, and relates to the technical field of artificial intelligence and online medical treatment. The method comprises the following steps: acquiring context information of a user dialogue, and determining an underwriting key statement and an underwriting keyword based on the context information; through a knowledge base of a preset field, retrieving fragment similar statements of the underwriting key statements and keyword similar statements of the underwriting keywords; performing information fusion on the underwriting key statements, the fragment similar statements and the keyword similar statements to obtain a model input text; and inputting the model input text into a pre-trained large language model, so that the large language model performs reasoning based on information recorded by the model input text to obtain pre-underwriting result information. According to the method, the pre-underwriting accuracy of the pre-underwriting work can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and online medical technology, and in particular to a method, device, medium and computer for generating pre-underwriting results based on a large model. Background Art

[0002] With the rapid development of the overall social economic level and medical conditions, people pay more and more attention to their health problems, and purchasing medical insurance has become a common choice for the public. Currently, before users purchase medical insurance, insurance companies need to conduct pre-insurance review for users. Here, pre-insurance review refers to the process in which the insured provides his or her own health conditions, questions to be asked, and the type of insurance he or she wishes to purchase to the insurance agency before formally purchasing insurance, and the insurance agency conducts an advance review of the insured's actual situation and provides the insured with a suitable insurance product. This process can usually help consumers more clearly understand the insurance product that is most suitable for them at the moment, and the insurance company also has a basic assessment of the insured's health status and related risks.

[0003] Currently, major insurance institutions mostly generate pre-underwriting results through artificial intelligence. The specific method is to send the health information input by the user, or the chat records between the underwriter and the user, and other pre-underwriting information to the big model. The big model generates the pre-underwriting result of the user based on the received pre-underwriting information. However, in the above-mentioned method of generating pre-underwriting results, the quality of the pre-underwriting information input into the big model is difficult to guarantee. For example, there may be redundant information irrelevant to the pre-underwriting in the pre-underwriting information, which reduces the accuracy of the pre-underwriting results output by the big model, thereby making the pre-underwriting accuracy of the insurance institution low when performing pre-underwriting work. Summary of the invention

[0004] In view of this, the present application provides a method, device, medium and computer for generating pre-underwriting results based on a large model, the main purpose of which is to solve the technical problem of low pre-underwriting accuracy when insurance institutions perform pre-underwriting work.

[0005] According to a first aspect of the present invention, a method for generating a pre-underwriting result based on a large model is provided, the method comprising:

[0006] Acquire context information of the user conversation, and determine underwriting key sentences and underwriting keywords based on the context information;

[0007] Retrieving fragment-similar sentences of the underwriting key sentences and keyword-similar sentences of the underwriting key words through a knowledge base of a preset field;

[0008] Performing information fusion on the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences to obtain a model input text;

[0009] The model input text is input into a pre-trained large language model so that the large language model performs inference based on the information recorded in the model input text to obtain pre-underwriting result information.

[0010] According to a second aspect of the present invention, there is provided a device for generating a pre-underwriting result based on a large model, the device comprising:

[0011] A preprocessing module, used to obtain context information of the user dialogue and determine underwriting key sentences and underwriting keywords based on the context information;

[0012] An information screening module, used to retrieve similar sentences of the underwriting key sentences and similar sentences of the underwriting key sentences through a knowledge base of a preset field;

[0013] A text generation module, used for fusing information of the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences to obtain a model input text;

[0014] An information output module is used to input the model input text into a pre-trained large language model so that the large language model can perform inference based on the information recorded in the model input text to obtain pre-underwriting result information.

[0015] According to a third aspect of the present invention, there is provided a medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the above-mentioned method for generating pre-underwriting results based on a large model.

[0016] According to a fourth aspect of the present invention, there is provided a computer comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for generating pre-underwriting results based on a large model is implemented.

[0017] The present invention provides a method, device, medium and computer for generating pre-underwriting results based on a large model. First, context information including online communication with users is obtained, and the context information is pre-processed to analyze the underwriting key sentences of the context information, and the underwriting keywords related to the pre-underwriting are screened out in the context information to exclude redundant information irrelevant to the pre-underwriting in the context information. Then, fragment-similar sentences whose similarity with the underwriting key sentences meets the preset conditions and keyword-similar sentences whose similarity with the underwriting keywords meets the preset conditions are screened out in the knowledge base of the preset field. After that, the underwriting key sentences, fragment-similar sentences and keyword-similar sentences are information fused to obtain the model input text. Finally, the model input text is input into the pre-trained large language model so that the large language model performs reasoning based on the information recorded in the model input text to obtain the pre-underwriting result information. The technical solution of the present application can refine the context information, obtain the underwriting key sentences and underwriting keywords related to the pre-underwriting, and exclude other irrelevant information. Furthermore, based on the key underwriting sentences and underwriting keywords, the necessary information used by the large language model for reasoning is retrieved from the knowledge base of the preset field, thereby providing comprehensive information support for the large language model to reason about the pre-underwriting results through multi-level information retrieval, effectively improving the accuracy of the pre-underwriting results inferred by the large language model, and thus significantly improving the pre-underwriting accuracy of insurance institutions when performing pre-underwriting work.

[0018] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0020] Figure 1 A schematic diagram of a process of generating a pre-underwriting result based on a large model provided by an embodiment of the present invention is shown;

[0021] Figure 2 A schematic diagram of the structure of a device for generating pre-underwriting results based on a large model provided by an embodiment of the present invention is shown;

[0022] Figure 3 A schematic diagram of the structure of another device for generating pre-underwriting results based on a large model provided in an embodiment of the present invention is shown;

[0023] Figure 4 A structural diagram of a computer for executing a method for generating pre-underwriting results based on a large model provided in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0024] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0025] As people pay more and more attention to their health problems, buying medical insurance has become a common choice for the public. Before users buy medical insurance, insurance companies need to conduct pre-underwriting on users to help consumers more clearly understand the insurance products that are most suitable for them at present. At the same time, insurance companies also have a basic assessment of the health status and related risks of the insured. At present, major insurance institutions mostly generate pre-underwriting results through artificial intelligence. The specific method is to send the health information entered by the user, or the chat records of the communication between the underwriter and the user, and other pre-underwriting information to the big model. The big model generates the pre-underwriting result of the user based on the received pre-underwriting information. However, in the above-mentioned method of generating pre-underwriting results, the quality of the pre-underwriting information input into the big model is difficult to guarantee. For example, there may be redundant information in the pre-underwriting information that is not related to the pre-underwriting, which reduces the accuracy of the pre-underwriting results output by the big model, thereby making the pre-underwriting accuracy of the insurance institution low when performing pre-underwriting work.

[0026] In view of the above problems, in one embodiment, Figure 1 As shown, a method for generating pre-underwriting results based on a large model is provided, and the method is described by taking the application of the method to a computer as an example, and includes the following steps:

[0027] 101. Obtain context information of a user conversation, and determine underwriting key sentences and underwriting keywords based on the context information. The underwriting key sentences may be user question sentences containing questions that the user wants to ask.

[0028] Specifically, during the online communication between the underwriters of the insurance institution and the users, the computer can monitor the communication process between the underwriters and the users in real time, and collect the dialogue sentences in the chat records of both parties during the communication process.

[0029] As an example, if the communication process between the user and the underwriter is:

[0030] User: Can I purchase X insurance for XXX disease?

[0031] Underwriter: No.

[0032] User: Can I purchase XX insurance?

[0033] Then the above conversation sentences can be collected and organized into context information. Furthermore, the above context information can be input into a preset text summary model, the context information can be preprocessed, and the multiple conversation sentences contained in the context information can be converted into a key underwriting sentence. Here, a text summary model such as the BERT model can be used, which can capture context information and is suitable for a variety of natural language processing tasks, including text summarization. Here, the context information of the chat record is preprocessed to avoid directly retrieving multiple rounds of chat records, which results in the retrieval of a large amount of irrelevant interference knowledge. At the same time, it can also reduce the time consumption of the retrieval link by compressing the data to be retrieved.

[0034] Based on the above example, the key underwriting statement obtained may be: "Can I purchase XX insurance for XXX disease?"

[0035] Furthermore, the context information can be input into a pre-trained deep learning model so that the deep learning model can obtain underwriting keywords related to pre-underwriting work in the deep learning model based on the context information. The deep learning model can be a model such as RoBERTa, which performs well in multiple natural language processing tasks, including keyword extraction. Here, the number of underwriting keywords can be one or more. Furthermore, the above underwriting keywords can be organized into an entity list, with each underwriting keyword as an entity.

[0036] Furthermore, based on the context information, the method for determining the underwriting keywords can also be: first, obtain a keyword comparison table. The keyword comparison table includes multiple preset keywords, each of which can be a related word in the medical or insurance field related to the pre-underwriting work, such as a disease name, symptom description, and insurance description. Furthermore, the preset keyword can correspond to its own preset word vector.

[0037] Further, the context information is segmented to obtain multiple text words; here, the context information can be segmented based on an open source segmentation model to obtain multiple underwriting keywords after segmentation. Among them, the entity word list related to the industry field is pre-imported into the word segmentation model, and the word list is composed of product names and terms in the insurance field, so that the obtained underwriting keywords have a high correlation with the pre-underwriting. Finally, for each of the text words, the word similarity between the text word and all the preset keywords is determined, and when the word similarity between the text word and any one of the preset keywords in all the preset keywords meets the preset similarity threshold, the text word is determined as the underwriting keyword. Specifically, the word vector of the text word can be determined, and the vector similarity between the word vector of the text word and the preset word vector of each preset keyword can be determined. When the vector similarity between the word vector of the text word and any one of the preset word vectors in all the preset word vectors meets the preset vector similarity threshold, the text word is determined as the underwriting keyword.

[0038] 102. Retrieve similar sentences of the underwriting key sentences and similar sentences of the underwriting keywords through the knowledge base of the preset field. The preset field may be the field of health and insurance. Further, the knowledge base includes multiple sentence fragments, and the method of obtaining the sentence fragments is as follows:

[0039] The knowledge text of a preset field is obtained and the text segmentation is performed on the knowledge text to obtain a plurality of the fragment sentences. The knowledge text of the preset field may be medical literature and relevant insurance product clauses. Furthermore, a preset value of text overlap is retained between two consecutive fragment sentences obtained by text segmentation of the knowledge text to ensure the integrity of the context of the retrieval content, so that the returned knowledge fragments are more coherent and readable, wherein the preset value may be 10%, and the number of words in each fragment sentence does not exceed a preset fixed number of words, and the preset value and the value of the preset fixed number of words may be determined according to actual conditions.

[0040] Further, the similarity between the underwriting key sentence and each fragment sentence can be compared. When the similarity between the underwriting key sentence and a certain fragment sentence meets a preset similarity threshold, the fragment sentence is determined as a fragment similar sentence. Further, the number of fragment similar sentences can be one or more. Further, the similarity between all the underwriting keywords in the entity list and each fragment sentence can be compared. When the similarity between all the underwriting keywords and a certain fragment sentence meets a preset similarity threshold, the fragment sentence is determined as a keyword similar sentence. Here, the number of keyword similar sentences can be one or more.

[0041] 103. Perform information fusion on the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences to obtain a model input text.

[0042] Among them, the key underwriting statements, similar statements for each segment, and similar statements for each keyword can be configured into the preset Prompt template to obtain the model input text. Here, the Prompt template refers to a structured format for generating prompts, which are usually used to guide the model to generate specific types of outputs. The design of the Prompt template is crucial to obtaining high-quality generation results. Specifically, the key underwriting statements, similar statements for each segment, and similar statements for each keyword can be configured into the corresponding content in the Prompt template.

[0043] 104. Input the model input text into a pre-trained large language model, so that the large language model performs inference based on the information recorded in the model input text to obtain pre-underwriting result information.

[0044] Among them, the large language model is an efficient fine-tuning model obtained after a specific number of rounds of training based on the configured model input text and the standard answer text corresponding to the information in the model input text as the training set corpus.

[0045] Specifically, the model input text containing the key underwriting sentences, each fragment similar sentence and each keyword similar sentence can be input into the large language model, so that the large language model can perform inference based on the key underwriting sentences, each fragment similar sentence and each keyword similar sentence recorded in the model input text to obtain the pre-underwriting result information.

[0046] The method for generating pre-underwriting results based on a large model provided in this embodiment first obtains context information including online communication with users, pre-processes the context information, analyzes the underwriting key sentences of the context information, and screens out the underwriting keywords related to the pre-underwriting in the context information to exclude redundant information in the context information that is not related to the pre-underwriting. Then, in the knowledge base of the preset field, the fragment-similar sentences whose similarity with the underwriting key sentences meets the preset conditions and the keyword-similar sentences whose similarity with the underwriting keywords meets the preset conditions are screened out. Further, the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences are information fused to obtain the model input text. Finally, the model input text is input into the pre-trained large language model so that the large language model performs reasoning based on the information recorded in the model input text to obtain the pre-underwriting result information. The technical solution of the present application can refine the context information, obtain the underwriting key sentences and underwriting keywords related to the pre-underwriting, and exclude other irrelevant information. Furthermore, based on the key underwriting sentences and underwriting keywords, the necessary information used by the large language model for reasoning is retrieved from the knowledge base of the preset field, thereby providing comprehensive information support for the large language model to reason about the pre-underwriting results through multi-level information retrieval, effectively improving the accuracy of the pre-underwriting results inferred by the large language model, and thus significantly improving the pre-underwriting accuracy of insurance institutions when performing pre-underwriting work.

[0047] In an optional embodiment, the model input text includes output format information, wherein the output format information is used to define the text format of the pre-underwriting result information output by the large language model, and is a complete description statement when one of the pre-underwriting result information is obtained, including disease reasoning, standardized conclusions and matters needing attention, which can be used to define the text format of the pre-underwriting result information output by the large language model; as an example, the output format information can be in the form of:

[0048] ##Example

[0049] 【Conclusion: XXX

[0050] Reason: XXX

[0051] Others: XXX]

[0052] So that the pre-underwriting result information output by the large language model can be in the three-segment format mentioned above.

[0053] Furthermore, the output format information also includes system instructions, which will indicate all the candidate sets of conclusions of the large language model and the type of cases represented by each conclusion, so as to limit the output scope of the large language model and limit the scope of the pre-underwriting result information output by the large language model to the pre-underwriting field. At the same time, the system instructions will also indicate the judgment logic of the large language model for special situations, such as whether to reject insurance, etc. Furthermore, each fragment-like sentence and each keyword-like sentence can be set to the database reference information part of the model input text, and the underwriting key sentence can be set to the user question part of the model input text.

[0054] Furthermore, the large language model performs reasoning based on the information recorded in the model input text and outputs the pre-underwriting result information of the user, including: first, extracting the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences from the model input text. Then, based on the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences as reasoning basis, the pre-underwriting original conclusion information is generated. Finally, the format of the pre-underwriting original conclusion information is adjusted based on the output format information to obtain the pre-underwriting result information.

[0055] Furthermore, after obtaining the pre-underwriting result information, the pre-underwriting result information can be templated based on a pre-trained text sorting model, such as chatglm4_9b or qianwen14b, so that it conforms to the standardized business speech structure. The purpose of the template transformation here is to convert the speech directly inferred by the large language model into the commonly used speech of business personnel, filter and delete some content that does not need to be displayed, and ensure that the content of the answer is smooth and natural after deletion. At the same time, for situations where user questions cannot be answered, a backup speech is provided to guide users to obtain answers that meet their actual needs through supplementary questions. The embodiment provided in the present application can control the form, output range, etc. of the pre-underwriting result information in the process of obtaining the pre-underwriting result information based on the model input text by the large language model, thereby improving the availability of the pre-underwriting result information.

[0056] In an optional embodiment, retrieving similar sentences of the key underwriting sentences through a knowledge base of a preset field includes:

[0057] First, the key underwriting sentence is converted into a sentence vector, and the vector similarity between the sentence vector and the fragment sentence vector of each fragment sentence stored in the knowledge base is determined; wherein each fragment sentence in the knowledge base has pre-calculated its own fragment sentence vector, and the fragment sentence vector is stored in a vector database, which can be an open source vector tool such as Chroma, which can store and retrieve vectors.

[0058] Specifically, a vector model can be used to convert the key underwriting sentences into sentence vectors. Here, the vector model can be any open source vector model, such as nomic-embed-text-v1. For the word segmenter of the vector model, an industry-related entity vocabulary can be imported into it in advance. The entity vocabulary consists of product names and terms in the insurance field, etc., to improve the vector conversion ability of the vector model. Furthermore, the vector similarity between the sentence vector and the fragment sentence vector of each fragment sentence can be calculated.

[0059] Then, each of the vector similarities is arranged in descending order to obtain a segment similarity queue; specifically, the values ​​of the vector similarities between the sentence vector and each segment sentence vector can be arranged in descending order to obtain a segment similarity queue.

[0060] Finally, the segment sentence corresponding to the vector similarity of the first preset order arranged in front of the segment similarity queue is determined as the segment similar sentence. The value of the first preset order can be determined according to actual conditions.

[0061] Specifically, the vector similarities from the first position to the first preset order of the segment similarity queue can be obtained, the segment sentence vector corresponding to each acquired vector similarity can be determined, and the segment sentence corresponding to each segment sentence vector can be determined, so as to determine the segment sentence corresponding to each acquired vector similarity as a segment similar sentence. The embodiment provided by the present application can quickly and accurately retrieve the segment sentence with a strong correlation with the key underwriting sentence in the knowledge base, and determine it as a segment similar sentence, thereby improving the efficiency of generating pre-underwriting result information.

[0062] In an optional embodiment, each of the fragment sentences corresponds to a preset fragment keyword group consisting of multiple keywords; wherein the fragment keyword group is obtained by: performing word segmentation on the fragment sentence to obtain multiple fragment keywords including all words constituting the fragment sentence, and the fragment keyword group can be pre-stored locally in the form of a list of word segments. Here, the word segmentation can adopt Jieba or other open source word segmentation models.

[0063] Furthermore, the keyword similarity statements of the underwriting keyword are retrieved through the knowledge base of the preset field, including: first, determining the keyword similarity between the underwriting keyword and each of the fragment keyword groups; specifically, if there are multiple underwriting keywords, all of them can be set in a keyword entity list, and the keyword entity list as a whole is used as a comparison dimension to calculate the keyword similarity between the keyword entity list and each fragment keyword group, wherein the algorithm for calculating the similarity can be a keyword-based inverted index algorithm such as tfidf.

[0064] Then, each of the keyword similarities is arranged in descending order of value to obtain a keyword similarity queue. Finally, the segment keyword group corresponding to the keyword similarity arranged in the second preset order before the keyword similarity queue is determined as the target segment keyword group, and the segment sentence corresponding to the target segment keyword group is determined as the keyword similar sentence. The value of the second preset order can be determined according to actual conditions.

[0065] Specifically, the keyword similarities from the first to the second preset order in the keyword similarity queue are determined, and the segment keyword group corresponding to the above-determined keyword similarity is determined as the target segment keyword group. Further, the segment sentences corresponding to each target segment keyword group are determined as keyword similarity sentences. The embodiment provided by the present application can retrieve segment sentences with a strong correlation with the underwriting keywords in the knowledge base, and determine them as keyword similarity sentences, thereby improving the efficiency of generating pre-underwriting result information.

[0066] In an optional embodiment, the information fusion of the underwriting key sentence, the fragment-similar sentence and the keyword-similar sentence to obtain the model input text includes:

[0067] First, the key underwriting statement is configured into a preset input template to obtain an information configuration text; specifically, the key underwriting statement can be set into the user question part of the model input text as user question information.

[0068] Then, the identical sentences in the fragment-similar sentences and the keyword-similar sentences are determined; specifically, if there are multiple keyword-similar sentences and multiple fragment-similar sentences, and the same fragment sentences exist among the multiple keyword-similar sentences and the multiple fragment-similar sentences, in order to reduce redundant information in the model input text, each identical fragment sentence in the keyword-similar sentences and the fragment-similar sentences can be determined as the same sentence.

[0069] Then, determine the fragment arrangement order of the vector similarity corresponding to each of the identical sentences in the fragment similarity queue, and the keyword arrangement order of the keyword similarity corresponding to each of the identical sentences in the keyword similarity queue; specifically, for each identical sentence, it can be determined that when the fragment sentence corresponding to the identical sentence is filtered through the underwriting key sentence, the fragment arrangement order of the vector similarity between the underwriting key sentence and the fragment sentence in the fragment similarity queue can be determined to determine the fragment arrangement order corresponding to each identical sentence.

[0070] Similarly, for each identical sentence, when the fragment sentence corresponding to the identical sentence is filtered through the underwriting keyword, the keyword similarity between the underwriting keyword and the fragment sentence can be determined in the keyword arrangement order in the keyword similarity queue to determine the keyword arrangement order corresponding to each identical sentence.

[0071] Furthermore, based on the fragment arrangement order and the keyword arrangement order of each of the identical sentences, a similarity deviation score of each of the identical sentences is determined; specifically, the similarity deviation score of the identical sentence can be calculated by calculating the sum of the fragment arrangement order and the keyword arrangement order of the identical sentence, and then dividing the sum of the fragment arrangement order and the keyword arrangement order by two to obtain the similarity deviation value of the identical sentence.

[0072] Furthermore, all the same sentences are arranged in ascending order of similarity deviation scores to obtain a same sentence queue, and each of the same sentences in the same sentence queue is sequentially configured into the information configuration text. In the above manner, the same sentence with the highest degree of association with the context information can be prioritized in the front position of the database reference information according to the degree of association between the same sentence and the context information, thereby increasing the contribution of the same sentence to the reasoning process of the pre-underwriting result information of the large language model, thereby improving the accuracy of the pre-underwriting work.

[0073] Furthermore, all of the fragment-similar sentences and the keyword-similar sentences except the identical sentences are configured into the information configuration text, and the information configuration text is determined as the model input text. The embodiment provided by the present application can reduce redundant information in the information configuration text, and prioritize identical sentences with the highest degree of association with context information at the front position in the database reference information, thereby increasing the contribution of the identical sentences to the reasoning process of the pre-underwriting result information of the large language model, so as to improve the accuracy of the pre-underwriting work.

[0074] In an optional embodiment, before inputting the model input text into the pre-trained large language model, the method further includes:

[0075] First, determine whether the underwriting keyword is a descriptive word for a preset field. Specifically, the underwriting keyword can be retrieved through a graph database to obtain the attributes of the entity-specific relationship corresponding to the underwriting keyword, and transmitted in a structured form. Here, the graph database is preset with underwriting keywords for each preset field, as well as the associated information corresponding to the underwriting keywords for each preset field. Furthermore, the graph database can be an open source database such as neo4j. Before searching, the triple data needs to be imported into the neo4j database. The import method here is to first process the data into a relationship table of underwriting keywords and associated information for the preset field, and then import the relationship table into the database at one time.

[0076] Then, when the underwriting keyword is a descriptive word of the preset field, the preset associated information corresponding to the descriptive word of the preset field is obtained. As an example, if the preset field is the disease name field and the underwriting keyword is XX disease, the associated information corresponding to the underwriting keyword is other diseases, symptoms and diagnostic methods with the same symptoms as XX disease; if the preset field is the disease symptom field and the underwriting keyword is XXX disease, the associated information corresponding to the underwriting keyword is other diseases with the same symptoms as XXX disease and the diagnostic methods of the above other diseases.

[0077] Finally, the association information is configured into the model input text. Specifically, the association information can be configured into the database reference information portion in the model input text. The embodiment provided in the present application can use a graph database to retrieve relationship attributes of underwriting keywords, can provide structured entity relationship information, enhance the relevance and logic of the large language model reasoning objects, help the large language model to better understand and apply the retrieval results, and improve the reasoning accuracy of the large language model.

[0078] In an optional embodiment, the model input text is input into a pre-trained large language model so that the large language model performs inference based on the information recorded in the model input text to obtain pre-underwriting result information. The method further includes: determining the underwriting key sentences, the underwriting keywords and the pre-underwriting result information as historical underwriting key sentences, historical underwriting keywords and historical pre-underwriting result information, respectively, and mapping and storing the historical underwriting key sentences, the historical underwriting keywords and the historical pre-underwriting result information; specifically, after performing a pre-underwriting reasoning, the underwriting key sentences, underwriting keywords and pre-underwriting result information involved in the reasoning are respectively determined as historical underwriting key sentences, historical underwriting keywords and historical pre-underwriting result information, and the above-mentioned historical underwriting key sentences, historical underwriting keywords and historical pre-underwriting result information are determined as a historical information group, and the historical information group is stored locally.

[0079] Furthermore, after obtaining the context information of the user dialogue and determining the underwriting key sentences and underwriting keywords based on the context information, the method further includes:

[0080] First, it is determined whether the underwriting key statement and the underwriting keyword are the same as the historical underwriting key statement and the historical underwriting keyword that are mapped and stored; specifically, whether there is at least one historical underwriting key statement that is the same as the underwriting key statement among all the historical underwriting key statements stored locally. If there is a historical underwriting key statement that is the same as the underwriting key statement, it is determined whether the historical underwriting keyword that is mapped and stored with the historical underwriting key statement is the same as the underwriting keyword.

[0081] Furthermore, when the underwriting key statement and the underwriting keyword are identical to any set of historical underwriting key statements and historical underwriting keywords stored in a mapping, the historical pre-underwriting result information stored in a mapping with the historical underwriting key statement and the historical underwriting keyword is used as the pre-underwriting result information of the user. In the embodiment provided by the present application, when it is determined that the underwriting key statement and the underwriting keyword obtained by converting the context information are identical to a set of historical underwriting key statements and historical underwriting keywords stored in a mapping obtained when the pre-underwriting work was previously performed, the historical pre-underwriting result information corresponding to the historical underwriting key statement and the historical underwriting keyword is directly used as the pre-underwriting result information of the pre-underwriting work currently being performed, thereby improving the efficiency of the pre-underwriting work.

[0082] The method for generating pre-underwriting results based on a large model provided in this embodiment can extract context information, obtain key underwriting sentences and keywords related to pre-underwriting, exclude other irrelevant information, and can retrieve the necessary information used by the large language model for reasoning in the knowledge base of the preset field based on the key underwriting sentences and keywords, thereby realizing multi-level information retrieval to provide comprehensive information support for the pre-underwriting result reasoning of the large language model; further, weighted sorting and fusion of fragment-similar sentences and keyword-similar sentences are performed to ensure that the knowledge fragments finally returned have both high similarity and cover multi-level information, providing more comprehensive knowledge support for the large language model. Furthermore, the retrieval of relationship attributes of underwriting keywords can provide structured entity relationship information, enhance the relevance and logic of the reasoning object of the large language model, and thus significantly improve the accuracy of pre-underwriting when the insurance institution performs pre-underwriting work.

[0083] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. In addition, the numbers corresponding to the various steps in the above embodiments only serve as identification and are not used as a limitation on the order in which the steps are executed. The order in which the steps are executed in each embodiment can be set according to actual conditions.

[0084] Further, as Figure 1 The specific implementation of the method shown in the embodiment provides a device for generating pre-underwriting results based on a large model, such as Figure 2 As shown, the device includes: a preprocessing module 21, an information screening module 22, a text generating module 23, and an information output module 24.

[0085] A preprocessing module 21, configured to obtain context information of a user conversation and determine underwriting key sentences and underwriting keywords based on the context information;

[0086] The information screening module 22 is used to retrieve the fragment-similar sentences of the underwriting key sentences and the keyword-similar sentences of the underwriting key sentences through the knowledge base of the preset field;

[0087] A text generation module 23, configured to fuse information of the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences to obtain a model input text;

[0088] The information output module 24 is used to input the model input text into a pre-trained large language model so that the large language model can perform inference based on the information recorded in the model input text to obtain pre-underwriting result information.

[0089] In a specific application scenario, the model input text includes output format information, wherein the output format information is used to limit the text format of the pre-underwriting result information output by the large language model, and the information output module 24 can be specifically used to extract the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences from the model input text; generate pre-underwriting original conclusion information based on the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences as reasoning basis; and adjust the format of the pre-underwriting original conclusion information based on the output format information to obtain the pre-underwriting result information.

[0090] In a specific application scenario, the information screening module 22 can be specifically used to convert the key underwriting statement into a statement vector, and determine the vector similarity between the statement vector and the fragment statement vector of each fragment statement stored in the knowledge base; arrange each of the vector similarities in descending order of value to obtain a fragment similarity queue; and determine the fragment statement corresponding to the vector similarity arranged in the first preset order in front of the fragment similarity queue as the fragment similarity statement.

[0091] In a specific application scenario, each of the fragment sentences corresponds to a preset fragment keyword group composed of multiple keywords. The information screening module 22 can be specifically used to determine the keyword similarity between the underwriting keyword and each of the fragment keyword groups; arrange each of the keyword similarities in descending order of value to obtain a keyword similarity queue; determine the fragment keyword group corresponding to the keyword similarity arranged in the second preset order before the keyword similarity queue as the target fragment keyword group, and determine the fragment sentence corresponding to the target fragment keyword group as the keyword similarity sentence.

[0092] In a specific application scenario, the text generation module 23 can be specifically used to configure the underwriting key statements into a preset input template to obtain an information configuration text; determine the identical statements in the segment-similar statements and the keyword-similar statements; respectively determine the segment arrangement order of the vector similarity corresponding to each of the identical statements in the segment similarity queue, and the keyword arrangement order of the keyword similarity corresponding to each of the identical statements in the keyword similarity queue; determine the similarity deviation score of each of the identical statements based on the segment arrangement order and the keyword arrangement order of each of the identical statements; arrange all of the identical statements in ascending order of similarity deviation values ​​to obtain an identical statement queue, and configure each of the identical statements in the identical statement queue into the information configuration text in turn; configure all of the segment-similar statements and the keyword-similar statements except the identical statements into the information configuration text, and determine the information configuration text as the model input text.

[0093] Furthermore, the input data and output data of each module in the preprocessing module 21, the information screening module 22, the text generation module 23, and the information output module 24 can be mapped and stored as historical input data and historical output data. In other subsequent pre-underwriting work, if the input data of a module is the same as a historical input data, the historical output data corresponding to the historical input data is directly used as the output data of the module to increase the speed at which the device obtains the pre-underwriting result information.

[0094] In specific application scenarios, such as Figure 3 As shown, the device also includes an information expansion module 34, which can be specifically used to determine whether the underwriting keyword is a descriptive word of a preset field; when the underwriting keyword is a descriptive word of the preset field, obtain preset associated information corresponding to the descriptive word of the preset field; and configure the associated information into the model input text.

[0095] In specific application scenarios, such as Figure 3 As shown, the device also includes a data storage module 36, which can be specifically used to determine the underwriting key statements, the underwriting keywords and the pre-underwriting result information as historical underwriting key statements, historical underwriting keywords and historical pre-underwriting result information, respectively, and map and store the historical underwriting key statements, the historical underwriting keywords and the historical pre-underwriting result information.

[0096] In a specific application scenario, the data storage module 36 is also used to determine whether the underwriting key statement and the underwriting keyword are the same as the historical underwriting key statement and the historical underwriting keyword mapped and stored; when the underwriting key statement and the underwriting keyword are the same as any group of historical underwriting key statements and the historical underwriting keywords mapped and stored, the historical pre-underwriting result information mapped and stored with the historical underwriting key statement and the historical underwriting keyword will be used as the pre-underwriting result information of the user.

[0097] It should be noted that for other corresponding descriptions of the functional units involved in the large model-based pre-underwriting result generation device provided in this embodiment, please refer to Figure 1 The corresponding description in will not be repeated here.

[0098] Based on the above Figure 1 The method shown in the embodiment also provides a medium on which a computer program is stored. When the program is executed by a processor, the above-mentioned Figure 1 The method for generating pre-underwriting results based on a large model is shown.

[0099] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product. The software product to be identified can be stored in a non-volatile medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including a number of instructions for enabling a computer (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0100] Based on the above Figure 1 The method shown, and Figure 2 and Figure 3 In order to achieve the above-mentioned purpose, the embodiment of the pre-underwriting result generating device based on the large model is shown. Figure 4 As shown, this embodiment also provides a computer for generating pre-underwriting results based on a large model, which can be a personal computer, a server, a smart phone, a tablet computer, a smart watch, or other network devices, etc. The computer includes a medium and a processor; the medium is used to store a computer program and an operating system; the processor is used to execute the computer program to implement the above-mentioned Figure 1 The method shown.

[0101] Optionally, the computer may further include an internal memory, a communication interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, a display, an input device such as a keyboard, etc. Optionally, the communication interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0102] Those skilled in the art will appreciate that the computer structure provided in this embodiment does not constitute a limitation on the computer, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0103] The medium may also include an operating system and a network communication module. The operating system is a program that manages the above-mentioned computer hardware and software resources to be identified, and supports the operation of information processing programs and other software and / or programs to be identified. The network communication module is used to realize communication between components within the medium, and communication with other hardware and software in the information processing computer.

[0104] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus the necessary general hardware platform, or by hardware. By applying the technical solution of the present application, first, the context information of the user conversation is obtained, and based on the context information, the key underwriting sentences and the key underwriting keywords are determined; then, the fragment-similar sentences of the key underwriting sentences and the keyword-similar sentences of the key underwriting keywords are retrieved through the knowledge base of the preset field; then, the key underwriting sentences, the fragment-similar sentences and the keyword-similar sentences are information-fused to obtain the model input text; finally, the model input text is input into the pre-trained large language model, so that the large language model can perform reasoning based on the information recorded in the model input text to obtain the pre-underwriting result information. Compared with the prior art, the pre-underwriting accuracy of the pre-underwriting work can be significantly improved.

[0105] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0106] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A method for generating pre-underwriting results based on a large model, characterized in that: The method comprises: Acquire context information of the user conversation, and determine underwriting key sentences and underwriting keywords based on the context information; Retrieving fragment-similar sentences of the underwriting key sentences and keyword-similar sentences of the underwriting key words through a knowledge base of a preset field; Performing information fusion on the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences to obtain a model input text; The model input text is input into a pre-trained large language model so that the large language model performs inference based on the information recorded in the model input text to obtain pre-underwriting result information.

2. The method according to claim 1, characterized in that The model input text includes output format information, wherein the output format information is used to define the text format of the pre-underwriting result information output by the large language model; the large language model performs reasoning based on the information recorded in the model input text to obtain the pre-underwriting result information, including: Extracting the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences from the model input text; Generate pre-underwriting original conclusion information based on the underwriting key sentences, the fragment-similar sentences, and the keyword-similar sentences as reasoning basis; The format of the pre-underwriting original conclusion information is adjusted based on the output format information to obtain the pre-underwriting result information.

3. The method according to claim 1, characterized in that The retrieving similar sentences of the key underwriting sentences through the knowledge base of the preset field includes: Converting the underwriting key sentence into a sentence vector, and determining the vector similarity between the sentence vector and the segment sentence vector of each segment sentence stored in the knowledge base; Arrange each of the vector similarities in descending order of value to obtain a segment similarity queue; The segment sentence corresponding to the vector similarity arranged in a first preset order before the segment similarity queue is determined as the segment similar sentence.

4. The method according to claim 3, characterized in that Each of the segment sentences corresponds to a preset segment keyword group consisting of multiple keywords; the keyword-similar sentences of the underwriting keywords are retrieved through the knowledge base of the preset field, including: Determining the keyword similarity between the underwriting keyword and each of the segment keyword groups; Arrange each of the keyword similarities in descending order of value to obtain a keyword similarity queue; The segment keyword group corresponding to the keyword similarity arranged in a second preset order before the keyword similarity queue is determined as a target segment keyword group, and the segment sentence corresponding to the target segment keyword group is determined as the keyword similarity sentence.

5. The method according to claim 4, characterized in that The step of fusing information of the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences to obtain a model input text includes: Configuring the key underwriting statements into a preset input template to obtain an information configuration text; Determining identical sentences among the segment-similar sentences and the keyword-similar sentences; Respectively determining the segment arrangement order of the vector similarities corresponding to each of the same sentences in the segment similarity queue, and the keyword arrangement order of the keyword similarities corresponding to each of the same sentences in the keyword similarity queue; Determining a similarity deviation score of each of the same sentences based on the segment arrangement order and the keyword arrangement order of each of the same sentences; Arrange all the identical statements in ascending order of similarity deviation scores to obtain an identical statement queue, and sequentially configure each identical statement in the identical statement queue into the information configuration text; All of the segment-similar sentences and the keyword-similar sentences except the identical sentences are configured into the information configuration text, and the information configuration text is determined as the model input text.

6. The method according to claim 1, characterized in that Before inputting the model input text into the pre-trained large language model, the method further includes: Determining whether the underwriting keyword is a descriptive word of a preset field; When the underwriting keyword is a descriptive word of the preset field, obtaining preset association information corresponding to the descriptive word of the preset field; The association information is configured into the model input text.

7. The method according to claim 1, characterized in that After inputting the model input text into a pre-trained large language model so that the large language model performs inference based on the information recorded in the model input text and obtains pre-underwriting result information, the method further includes: Determining the underwriting key statement, the underwriting keyword, and the pre-underwriting result information as historical underwriting key statements, historical underwriting keywords, and historical pre-underwriting result information, respectively, and mapping and storing the historical underwriting key statements, the historical underwriting keywords, and the historical pre-underwriting result information; After obtaining the context information of the user dialogue and determining the underwriting key sentences and underwriting keywords based on the context information, the method further includes: Determining whether the underwriting key statement and the underwriting keyword are the same as the historical underwriting key statement and the historical underwriting keyword stored in the mapping; When the underwriting key statement and the underwriting keyword are the same as any group of historical underwriting key statements and historical underwriting keywords stored in the mapping, the historical pre-underwriting result information stored in the mapping with the historical underwriting key statement and the historical underwriting keyword will be used as the pre-underwriting result information of the user.

8. A device for generating pre-underwriting results based on a large model, characterized in that: The device comprises: A preprocessing module, used to obtain context information of the user dialogue and determine underwriting key sentences and underwriting keywords based on the context information; An information screening module, used to retrieve similar sentences of the underwriting key sentences and similar sentences of the underwriting key sentences through a knowledge base of a preset field; A text generation module, used for fusing information of the underwriting key sentences, the fragment-similar sentences and the keyword-similar sentences to obtain a model input text; An information output module is used to input the model input text into a pre-trained large language model so that the large language model can perform inference based on the information recorded in the model input text to obtain pre-underwriting result information.

9. A medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images