Insurance clause knowledge base self-maintenance method, system and device based on large model

Abstract

The application discloses an insurance clause knowledge base self-maintenance method, system and device based on a large model, and relates to the technical field of data processing. The method comprises the following steps: obtaining an associated split clause fragment set; performing vector conversion and storing in a cloud database to obtain a vectorized cloud knowledge base; when the vectorized cloud knowledge base receives a user search request, adjusting the priority of the knowledge base clause through a two-level search mechanism to obtain an updated vector cloud knowledge base; responding to the user search request based on the updated vector cloud knowledge base to obtain a clause fragment search result; inputting the Prompt construction result and the clause fragment search result into a large language model constructed based on RAG enhancement generation technology for search enhancement generation to obtain a target search interpretation analysis result. The technical problems of low maintenance efficiency and poor search accuracy of the existing insurance clause knowledge base are solved, and the technical effects of improving the self-maintenance efficiency and search accuracy of the insurance clause knowledge base are achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a self-maintaining method, system, and device for an insurance clause knowledge base based on a large model. Background Technology

[0002] With the rapid development of the insurance business, the number and complexity of insurance clauses are constantly increasing. To improve customer service experience and business compliance, insurance companies have generally built text-based clause knowledge bases to support various application scenarios such as insurance product design, risk assessment, and customer consultation. However, existing insurance clause knowledge bases largely rely on manual input and updates, making it difficult to respond promptly to needs such as clause version iterations, regulatory changes, or new product launches. This results in high maintenance costs, long update cycles, and poor consistency. Furthermore, due to the complex structure and dense technical terminology of insurance clauses, traditional keyword matching retrieval methods struggle to accurately capture user intent, leading to poor relevance and readability of search results, and failing to effectively meet diverse and intelligent user query needs. Summary of the Invention

[0003] This application provides a self-maintenance method, system, and device for insurance clause knowledge base based on a large model, which solves the technical problems of low maintenance efficiency and poor retrieval accuracy of insurance clause knowledge base in the prior art.

[0004] The first aspect of this application provides a self-maintaining method for an insurance clause knowledge base based on a large model, the method comprising:

[0005] Long text data is segmented into fragments using semantic segmentation, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments. The set of associated segmented clause fragments is then vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base. When the vectorized cloud knowledge base receives a user search request, it adjusts the priority of knowledge base clauses through a two-level search mechanism to obtain an updated vectorized cloud knowledge base. Based on the updated vectorized cloud knowledge base, the user search request is responded to to obtain clause fragment search results. The Prompt construction results and clause fragment search results are input into a large language model built based on RAG enhanced generation technology for search enhancement generation to obtain target search interpretation and analysis results.

[0006] A second aspect of this application provides a self-maintaining system for an insurance clause knowledge base based on a large model, the system comprising:

[0007] Semantic Segmentation Module: Segments ultra-long text data into segments using semantic segmentation, and associates each segmented clause with the original clause data to obtain a set of associated segmented clauses; Vector Conversion Module: Performs vector conversion on the set of associated segmented clauses using a semantic embedding device, and stores it in a cloud database to obtain a vectorized cloud knowledge base; Update Module: When the vectorized cloud knowledge base receives a user search request, it adjusts the priority of knowledge base clauses through a two-level search mechanism to obtain an updated vectorized cloud knowledge base; Response Module: Responds to the user search request based on the updated vectorized cloud knowledge base to obtain clause segment search results; Search Enhancement Module: Inputs the Prompt construction results and clause segment search results into a large language model built based on RAG enhancement generation technology for search enhancement generation to obtain target search interpretation and analysis results.

[0008] A third aspect of this application provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the self-maintenance method for an insurance clause knowledge base based on a large model provided in this application when executing the executable instructions stored in the memory.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] First, the ultra-long text data is segmented into fragments using semantic segmentation, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments. Next, the set of associated segmented clause fragments is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base. When the vectorized cloud knowledge base receives a user search request, it adjusts the priority of knowledge base clauses through a two-level search mechanism to obtain an updated vectorized cloud knowledge base. Then, it responds to the user search request based on the updated vectorized cloud knowledge base, obtaining clause fragment search results. Finally, the Prompt construction results and clause fragment search results are input into a large language model built based on RAG augmented generation technology for enhanced search generation, obtaining target search interpretation and analysis results. This solves the technical problems of low maintenance efficiency and poor search accuracy of insurance clause knowledge bases in existing technologies, achieving the technical effect of improving the self-maintenance efficiency and search accuracy of insurance clause knowledge bases. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0012] Figure 1 A schematic diagram of the self-maintenance method for an insurance clause knowledge base based on a large model provided in this application embodiment;

[0013] Figure 2 A schematic diagram of the structure of a self-maintaining insurance clause knowledge base based on a large model, provided for embodiments of this application;

[0014] Figure 3 This is a schematic diagram of the structure of an exemplary electronic device of this application.

[0015] Explanation of reference numerals in the attached figures: semantic segmentation module 11, vector transformation module 12, update module 13, response module 14, retrieval enhancement module 15, processor 21, memory 22, input device 23, output device 24. Detailed Implementation

[0016] This application solves the technical problems of low maintenance efficiency and poor retrieval accuracy of insurance clause knowledge bases in the prior art by providing a self-maintenance method, system and equipment based on a large model.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0019] Example 1, as Figure 1 As shown in the embodiments of this application, a self-maintaining method for an insurance clause knowledge base based on a large model is provided, wherein the method includes:

[0020] The ultra-long text data is segmented into fragments by semantic segmentation, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments.

[0021] First, the extremely long text data is structurally divided according to preset structural rules, such as titles, chapters, clause numbers, paragraph marks, or semantic pause markers, transforming the original text data into a structured set of extremely long text data. Then, using a semantic paragraph segmentation algorithm trained on a Transformer architecture or rule fusion model, the semantic relationships between sentences and logical pauses in each clause of the structured text set are identified, thereby achieving automatic segmentation of semantically coherent and reasonably granular clause fragments. During the segmentation process, the association between each clause fragment and its original text data is maintained, and this association is recorded in the form of metadata, including the original document path, chapter number, clause number, context paragraph ID, and text index positions before and after segmentation, thus generating a well-structured and semantically complete set of associated segmented clause fragments.

[0022] Furthermore, semantic segmentation is used to segment ultra-long text data into fragments, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments, including:

[0023] The ultra-long text data is subjected to structured partitioning to obtain a structured ultra-long text data set; the structured ultra-long text data set is then segmented into segments using a semantic paragraph segmentation algorithm to obtain a set of related segmented clause segments; wherein, each related segmented clause segment includes the original text path, context ID, and document index.

[0024] The process involves structurally segmenting ultra-long text data. Specifically, by analyzing the logical structure and layout features of the original clause documents, structured markers such as chapter titles, clause numbers, bullet points, indentation levels, and line breaks are identified. Preliminary segmentation is then performed according to preset structural rules, transforming the original unstructured text into a structured ultra-long text dataset with clear hierarchical relationships. Based on a trained semantic paragraph segmentation algorithm, semantically driven content segmentation is then performed on the aforementioned structured ultra-long text dataset. This algorithm employs a contextual representation model built upon deep language models (such as BERT and RoBERTa), combined with sequence labeling or similarity thresholding mechanisms, to automatically identify semantic breakpoints in the text. Each structured clause is further segmented into semantically coherent and fully expressed clause fragments, ensuring the segmentation results possess logical independence and retrieval effectiveness in the semantic dimension.

[0025] While generating clause fragments, to ensure data traceability and subsequent information tracking capabilities, association information with the original data is added to each segmented clause fragment, thereby obtaining a set of associated segmented clause fragments. Each associated segmented clause fragment includes the following metadata information fields: original path (i.e., the storage path of the original file or a unique document identifier), context ID (indicating its contextual position in the original document structure, such as the chapter or paragraph number), and document index (indicating the start and end character positions or sentence number range of the fragment in the original text).

[0026] The set of related segmented clauses is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base.

[0027] Each clause fragment is encoded using a pre-trained semantic embedding model (such as BERT or Sentence-BERT) to generate a high-dimensional semantic vector representation. Subsequently, the generated semantic vectors are combined with the corresponding metadata information (including the original text path, context ID, document index, clause number, etc.) to construct a structured vector entry. Finally, the above structured vector entries are stored in batches in a cloud database to build a vectorized cloud knowledge base that can be efficiently used for semantic retrieval and dynamic maintenance, providing basic support for subsequent retrieval matching, weight updates, and generation enhancement.

[0028] Furthermore, the set of related segmented clause fragments is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base, including:

[0029] The semantic embedder is used to perform vector transformation on the set of related segmented clause fragments to obtain a set of semantic vectors for the related segmented clause fragments. The set of semantic vectors for the related segmented clause fragments is then traversed and identified according to a preset set of metadata field tags to obtain a set of semantic vectors for identifying the related segmented clause fragments. The preset set of metadata field tags includes the original clause ID, chapter name, clause number, update time, applicable insurance type, user access frequency, and click popularity. The set of semantic vectors for identifying the related segmented clause fragments is stored in a cloud database to obtain the vectorized cloud knowledge base.

[0030] First, a semantic embedding device is used to semantically encode each related segmented clause fragment. The semantic embedding device is preferably a deep language model based on the Transformer architecture (such as Sentence-BERT, SimCSE, E5, etc.), which has context-aware semantic expression capabilities and can embed natural language text into a fixed-dimensional vector space. Each clause fragment is input into the semantic embedding device model, and its semantic features are extracted to obtain the corresponding clause fragment semantic vector, thus forming a complete set of semantic vectors for related segmented clause fragments. Second, the set of semantic vectors is traversed and identified according to a preset set of metadata field tags. The metadata field tag set is used to annotate the business context and retrieval attributes of each semantic vector, preferably including but not limited to the following fields: original clause ID (for tracing), chapter name (locating structural position), clause number (specific entry), update time (reflecting timeliness), applicable insurance type (supporting product classification), user access frequency and click popularity (reflecting user interaction behavior). The above metadata field tag set can be appended to each semantic vector item as key-value pairs or nested structures, forming a set of semantic vectors for identifying related segmented clause fragments with indexing capabilities and configurability. Finally, the tagged semantic vector set is uploaded to a cloud-deployed database to build a vectorized cloud knowledge base that supports efficient semantic retrieval.

[0031] When the vectorized cloud knowledge base receives a user's search request, it adjusts the priority of knowledge base terms through a two-level search mechanism to obtain an updated vectorized cloud knowledge base.

[0032] When the vectorized cloud knowledge base receives a user's search request, the system first converts the request into a semantic vector and performs a first-level semantic coarse recall in the knowledge base based on the vector to obtain a preliminary set of candidate clause fragments. Subsequently, a cross-encoder is used to perform fine semantic relevance scoring on the user's request vector and candidate clauses, and the candidate results are reordered to achieve dynamic adjustment of clause priority. Finally, an updated vectorized cloud knowledge base is generated to improve the accuracy and relevance of search results.

[0033] Furthermore, when the vectorized cloud knowledge base receives a user's search request, it adjusts the priority of knowledge base terms through a two-level search mechanism to obtain an updated vectorized cloud knowledge base, including:

[0034] The user search request is vectorized to obtain a user search request vector; based on the vector cloud knowledge base, a first-level semantic coarse recall is performed on the user search request vector to obtain a first-level candidate clause fragment set; a cross encoder is used to perform semantic relevance scoring on the user search request vector and the first-level candidate clause fragment set, and the first-level candidate clause fragment set is reordered according to the scoring results to obtain an updated vector cloud knowledge base.

[0035] First, the user's search request is vectorized. Specifically, the same semantic embedding algorithm used to generate the term vectors in the knowledge base can be employed to transform the user's text search request into a fixed-dimensional user search request vector. This ensures that the request and the term vectors reside in the same semantic space, facilitating subsequent semantic matching. Second, a primary semantic coarse-grained recall is performed on the user's search request vector based on the vector cloud knowledge base. Specifically, a vector retrieval engine (e.g., an index structure based on an approximate nearest neighbor algorithm) is used to calculate the similarity between the request vector and the semantic vectors of all term fragments in the knowledge base, selecting the term fragments with the highest similarity to form a primary candidate term fragment set. Then, for this primary candidate term fragment set, a cross-encoder model is used to perform a more refined semantic relevance score between each term fragment and the user's search request vector. The cross-encoder employs a joint encoding mechanism, treating the search request and the term fragment as a single text pair as input, and using a deep neural network model to calculate their semantic matching score. Finally, based on the cross-encoder's scoring results, the primary candidate term fragment set is reordered, prioritizing term fragments with higher scores to form an updated vector cloud knowledge base with adjusted priorities.

[0036] Furthermore, this includes:

[0037] Obtain a set of user search behavior logs within a preset window; use search frequency as an index to perform a term fragment search frequency analysis on the set of user search behavior logs, assign weights to the term fragments in descending order of search frequency, and update and correct the weights of the updated vector cloud knowledge base based on the weight assignment results.

[0038] The weight update and correction process based on the updated vector cloud knowledge base includes: First, acquiring a set of user search behavior logs within a preset time window. This log set contains records of users' specific search operations on knowledge base clause fragments during that time period, such as search keywords, the number of times clause fragments were clicked, and the duration of dwell time. Second, using the search frequency of clause fragments as an index, statistical analysis is performed on the user search behavior log set. Specifically, the frequency of each clause fragment being searched and accessed by users within the preset time window is statistically analyzed, forming a clause fragment search frequency distribution. Then, based on the statistically analyzed search frequency data, corresponding weights are assigned to clause fragments in descending order of search frequency. Finally, based on the weight allocation results, the weights of each clause fragment in the updated vector cloud knowledge base are dynamically corrected.

[0039] The system responds to the user's search request based on the updated vector cloud knowledge base, and obtains the search results for the clause fragments.

[0040] The system receives the user's search request vector, represented in vectorized form, and inputs it into the updated vector cloud knowledge base as a search query. Using a vector similarity matching mechanism, it calculates the similarity between the user's search request vector and the semantic vectors of all clause fragments in the updated vector cloud knowledge base, selecting several clause fragments with the highest similarity scores to form a search candidate set. Based on this, the search candidate set is further filtered and reordered according to the metadata tag information of the clause fragments (such as clause number, chapter position, applicable insurance type, click popularity, etc.) and the ranking results after weight updates, ultimately forming the clause fragment search results. The clause fragment search result set can contain several clause content fragments most relevant to the user's query semantics, and can carry the associated original clause path, document index information, contextual scope, etc., ensuring the traceability and consistency of subsequent content interpretation and multi-round question-and-answer reasoning.

[0041] The Prompt construction results and clause fragment retrieval results are input into a large language model built based on RAG enhancement generation technology to perform retrieval enhancement generation and obtain target retrieval interpretation and analysis results.

[0042] The Prompt construction results and clause fragment retrieval results are input into a large language model built based on RAG enhanced generation technology for enhanced retrieval generation. Specifically, this includes mapping the query intent parsed from the user's search request into a structured Prompt construction result, which, together with the obtained clause fragment retrieval results, is submitted to a large language model that integrates retrieval information and generation capabilities. After receiving the input, the model automatically integrates relevant clause knowledge content with the Prompt context based on the RAG mechanism, guiding the generation of logically coherent, semantically relevant, and accurately interpreted target retrieval interpretation and analysis results, thereby improving the ability to understand the complex semantics of insurance clauses and the level of intelligence in user interaction.

[0043] Furthermore, the Prompt construction results and clause fragment retrieval results are input into a large language model built based on RAG enhancement generation technology for retrieval enhancement generation, obtaining target retrieval interpretation and analysis results, including:

[0044] Based on the user's search request, query intent is parsed, and the Prompt template library is matched and constructed according to the obtained query intent parsing features to obtain the Prompt construction result; the large language model is used to perform search enhancement generation on the Prompt construction result and the clause fragment search result to obtain the target search interpretation analysis result.

[0045] Based on user search requests, query intent is analyzed, extracting semantic features such as keywords, intent tags, query scope, and logical relationships contained in the user request to form query intent analysis features. Based on the obtained query intent analysis features, a pre-set Prompt template library is matched and constructed, selecting the Prompt template content that best fits the current semantic scenario. Combined with user historical behavior characteristics, search context, and clause fragment attribute information, a dynamically generated Prompt result is produced. The Prompt result includes guiding language descriptions, task objective statements, and format requirement prompts, guiding the large language model to generate more accurate interpretation text. The Prompt result and the clause fragment search results are used as joint inputs to a large language model built based on RAG enhanced generation technology. The large language model obtains embedded semantic background information of the clause fragments through the retrieval module and combines it with the generation module to generate conditional text for the Prompt, achieving a semantic response to user intent and intelligent interpretation of legal clause content. Finally, the target retrieval interpretation analysis results are output, providing users with a semantically clear, accurate, and highly readable intelligent interpretation service of insurance clauses.

[0046] Furthermore, based on the user's search request, query intent is parsed, and the Prompt template library is matched and constructed according to the obtained query intent parsing features to obtain the Prompt construction result, including:

[0047] The process involves: acquiring a set of historical query intent parsing features, a set of historical user feedback ratings, and a set of historical prompts; aggregating the historical query intent parsing features by similar categories to obtain multiple aggregated historical query intent parsing feature sets; mapping and aggregating the historical prompts based on the multiple aggregated historical query intent parsing feature sets to obtain multiple mapped aggregated prompts; performing representative filtering on the multiple mapped aggregated prompts based on the historical user feedback ratings to obtain multiple prompt templates; extracting prototypes from the multiple aggregated historical query intent parsing feature sets to obtain multiple query intent parsing feature prototypes; and mapping and associating the multiple prompt templates and multiple query intent parsing feature prototypes to construct the Prompt template library.

[0048] First, the system obtains a historical query intent parsing feature set, a historical user feedback rating set, and a historical Prompt hint set. The historical query intent parsing feature set contains structured semantic tags extracted from past user search requests; the historical Prompt hint set records the language templates used to generate the Prompt for the corresponding intent; and the historical user feedback rating set contains user evaluations of the generated results, used to optimize the quality of Prompt selection. Then, the historical query intent parsing feature set is aggregated based on similarity, using semantic similarity clustering or tag field-based grouping mechanisms to form multiple aggregated historical query intent parsing feature sets, thus forming similar intent paradigms. Based on these multiple aggregated historical query intent parsing feature sets, the historical Prompt hint set is mapped and aggregated to establish a correspondence between each semantic intent cluster and the corresponding Prompt content, resulting in multiple mapped aggregated Prompt hint sets. Further, the historical user feedback rating set is used to perform representative screening of the Prompt hints in each mapped aggregated Prompt hint set, eliminating low-quality hint templates and selecting the hint set with the best historical rating performance, ultimately obtaining multiple high-quality Prompt hint templates. Subsequently, prototype extraction processing is performed on each aggregated historical query intent parsing feature set. Specifically, by calculating the average similarity among members within the set, a representative set or single query intent parsing feature is selected as the query intent parsing feature prototype, thereby establishing a structured representation of query semantics. Finally, the multiple Prompt suggestion templates mentioned above are mapped one-to-one with multiple query intent parsing feature prototypes to construct a standardized Prompt template library. The Prompt template library supports the rapid matching of optimal suggestion content based on the user's real-time query intent and is used to guide the large language model to generate context-appropriate insurance clause explanation text.

[0049] Furthermore, prototype extraction is performed on the multiple aggregated historical query intent parsing feature sets to obtain multiple query intent parsing feature prototypes, including:

[0050] For each of the multiple aggregated historical query intent parsing feature sets, perform pairwise similarity recognition within the set to obtain multiple pairwise aggregated historical query intent parsing feature similarity sets; based on the multiple pairwise aggregated historical query intent parsing feature similarity sets, take the aggregated historical query intent parsing feature corresponding to the maximum value of the average similarity of the pairwise aggregated historical query intent parsing features associated in the multiple aggregated historical query intent parsing feature sets as a prototype to obtain multiple query intent parsing feature prototypes.

[0051] For each aggregated historical query intent parsing feature set, pairwise similarity recognition is performed within the set. Specifically, based on a preset semantic similarity calculation model (e.g., based on BERT, SBERT, or other semantic embedding models), the semantic similarity of any two historical query intent parsing features within the set is measured, forming multiple pairwise aggregated historical query intent parsing feature similarity sets. In each similarity set, each item records the similarity score between a pair of query intent parsing features, used to characterize their semantic consistency.

[0052] The pairwise similarity scores of all members in each aggregated historical query intent parsing feature set are statistically processed, and the mean similarity between each historical query intent parsing feature and other members in the set is calculated. Based on this mean, the feature with the largest mean value is selected as the representative of the aggregated set, i.e., the query intent parsing feature prototype. Through the above prototype extraction mechanism, highly representative prototype intents can be extracted from a large number of semantically similar but not completely consistent historical query intents, serving as the core semantic anchor points for Prompt template matching. Ultimately, the obtained multiple query intent parsing feature prototypes not only possess stability and representativeness in expression but can also be used for rapid classification and semantic alignment in new intent parsing tasks, providing accurate intent classification basis for Prompt template library matching and retrieval enhancement generation, thereby improving the intelligent response capability and semantic generalization capability of the entire large model-driven insurance clause knowledge interpretation system.

[0053] In summary, the embodiments of this application have at least the following technical effects:

[0054] First, the ultra-long text data is segmented into fragments using semantic segmentation, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments. Next, the set of associated segmented clause fragments is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base. When the vectorized cloud knowledge base receives a user search request, it adjusts the priority of knowledge base clauses through a two-level search mechanism to obtain an updated vectorized cloud knowledge base. Then, it responds to the user search request based on the updated vectorized cloud knowledge base, obtaining clause fragment search results. Finally, the Prompt construction results and clause fragment search results are input into a large language model built based on RAG augmented generation technology for enhanced search generation, obtaining target search interpretation and analysis results. This solves the technical problems of low maintenance efficiency and poor search accuracy of insurance clause knowledge bases in existing technologies, achieving the technical effect of improving the self-maintenance efficiency and search accuracy of insurance clause knowledge bases.

[0055] Example 2, based on the same inventive concept as the self-maintenance method of the insurance clause knowledge base based on a large model in the previous examples, such as... Figure 2As shown, this application provides a self-maintaining system for an insurance clause knowledge base based on a large model, wherein the system includes:

[0056] Semantic segmentation module 11: Segments the ultra-long text data into segments using semantic segmentation, and associates each segmented clause with the original clause data to obtain a set of associated segmented clauses; Vector conversion module 12: Performs vector conversion on the set of associated segmented clauses using a semantic embedding device, and stores it in a cloud database to obtain a vectorized cloud knowledge base; Update module 13: When the vectorized cloud knowledge base receives a user search request, it adjusts the priority of knowledge base clauses through a two-level search mechanism to obtain an updated vectorized cloud knowledge base; Response module 14: Responds to the user search request based on the updated vectorized cloud knowledge base to obtain clause segment search results; Search enhancement module 15: Inputs the Prompt construction results and clause segment search results into a large language model built based on RAG enhancement generation technology for search enhancement generation to obtain target search interpretation and analysis results.

[0057] Furthermore, the semantic segmentation module 11 is used to perform the following method:

[0058] The ultra-long text data is subjected to structured partitioning to obtain a structured ultra-long text data set; the structured ultra-long text data set is then segmented into segments using a semantic paragraph segmentation algorithm to obtain a set of related segmented clause segments; wherein, each related segmented clause segment includes the original text path, context ID, and document index.

[0059] Furthermore, the vector transformation module 12 is used to perform the following method:

[0060] The semantic embedder is used to perform vector transformation on the set of related segmented clause fragments to obtain a set of semantic vectors for the related segmented clause fragments. The set of semantic vectors for the related segmented clause fragments is then traversed and identified according to a preset set of metadata field tags to obtain a set of semantic vectors for identifying the related segmented clause fragments. The preset set of metadata field tags includes the original clause ID, chapter name, clause number, update time, applicable insurance type, user access frequency, and click popularity. The set of semantic vectors for identifying the related segmented clause fragments is stored in a cloud database to obtain the vectorized cloud knowledge base.

[0061] Furthermore, the update module 13 is used to perform the following method:

[0062] The user search request is vectorized to obtain a user search request vector; based on the vector cloud knowledge base, a first-level semantic coarse recall is performed on the user search request vector to obtain a first-level candidate clause fragment set; a cross encoder is used to perform semantic relevance scoring on the user search request vector and the first-level candidate clause fragment set, and the first-level candidate clause fragment set is reordered according to the scoring results to obtain an updated vector cloud knowledge base.

[0063] Furthermore, the update module 13 is used to perform the following method:

[0064] Obtain a set of user search behavior logs within a preset window; use search frequency as an index to perform a term fragment search frequency analysis on the set of user search behavior logs, assign weights to the term fragments in descending order of search frequency, and update and correct the weights of the updated vector cloud knowledge base based on the weight assignment results.

[0065] Furthermore, the retrieval enhancement module 15 is used to perform the following methods:

[0066] Based on the user's search request, query intent is parsed, and the Prompt template library is matched and constructed according to the obtained query intent parsing features to obtain the Prompt construction result; the large language model is used to perform search enhancement generation on the Prompt construction result and the clause fragment search result to obtain the target search interpretation analysis result.

[0067] Furthermore, the retrieval enhancement module 15 is used to perform the following methods:

[0068] The process involves: acquiring a set of historical query intent parsing features, a set of historical user feedback ratings, and a set of historical prompts; aggregating the historical query intent parsing features by similar categories to obtain multiple aggregated historical query intent parsing feature sets; mapping and aggregating the historical prompts based on the multiple aggregated historical query intent parsing feature sets to obtain multiple mapped aggregated prompts; performing representative filtering on the multiple mapped aggregated prompts based on the historical user feedback ratings to obtain multiple prompt templates; extracting prototypes from the multiple aggregated historical query intent parsing feature sets to obtain multiple query intent parsing feature prototypes; and mapping and associating the multiple prompt templates and multiple query intent parsing feature prototypes to construct the Prompt template library.

[0069] Furthermore, the retrieval enhancement module 15 is used to perform the following methods:

[0070] For each of the multiple aggregated historical query intent parsing feature sets, perform pairwise similarity recognition within the set to obtain multiple pairwise aggregated historical query intent parsing feature similarity sets; based on the multiple pairwise aggregated historical query intent parsing feature similarity sets, take the aggregated historical query intent parsing feature corresponding to the maximum value of the average similarity of the pairwise aggregated historical query intent parsing features associated in the multiple aggregated historical query intent parsing feature sets as a prototype to obtain multiple query intent parsing feature prototypes.

[0071] Example 3, Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention, showing a block diagram of an exemplary electronic device suitable for implementing the embodiments of the present invention. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the present invention. Figure 3 As shown, the electronic device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the electronic device can be one or more. Figure 3 Taking a processor 21 as an example, the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0072] The memory 22, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the self-maintenance method for the insurance clause knowledge base based on a large model in this embodiment of the invention. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 22, thereby realizing the aforementioned self-maintenance method for the insurance clause knowledge base based on a large model.

[0073] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0074] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0075] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A self-maintaining method for an insurance clause knowledge base based on a large model, characterized in that: The method includes: The ultra-long text data is segmented into fragments by semantic segmentation, and each segmented clause fragment is associated with the original clause data to obtain a set of associated segmented clause fragments; The set of related segmented clause fragments is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base. When the vectorized cloud knowledge base receives a user's search request, it adjusts the priority of knowledge base terms through a two-level search mechanism to obtain an updated vectorized cloud knowledge base. The system responds to the user's search request based on the updated vector cloud knowledge base, and obtains the search results for the clause fragments. The Prompt construction results and the clause fragment retrieval results are input into a large language model built based on RAG enhancement generation technology to perform retrieval enhancement generation and obtain target retrieval interpretation and analysis results. Specifically, the Prompt construction results and clause fragment retrieval results are input into a large language model built based on RAG enhancement generation technology for retrieval enhancement generation, obtaining target retrieval interpretation and analysis results, including: Based on the user's search request, the query intent is parsed, and the Prompt template library is matched and constructed according to the obtained query intent parsing features to obtain the Prompt construction result; The large language model is used to enhance the retrieval of the Prompt construction results and the clause fragment retrieval results, thereby obtaining the target retrieval interpretation and analysis results. Specifically, based on the user's search request, query intent is parsed, and the Prompt template library is matched and constructed according to the obtained query intent parsing features to obtain the Prompt construction result, including: Obtain the historical query intent parsing feature set, the historical user feedback rating set, and the historical prompt suggestion set; The historical query intent parsing feature set is aggregated to obtain multiple aggregated historical query intent parsing feature sets. Based on the multiple aggregated historical query intent parsing feature sets, the historical Prompt suggestion sets are mapped and aggregated to obtain multiple mapped aggregated Prompt suggestion sets; Based on the historical user feedback rating set, representative filtering is performed on the multiple mapping aggregated Prompt prompt sets to obtain multiple Prompt prompt templates; Prototype extraction is performed on the multiple aggregated historical query intent parsing feature sets to obtain multiple query intent parsing feature prototypes; The multiple Prompt prompt templates and multiple query intent parsing feature prototypes are mapped and associated to construct the Prompt template library; Specifically, prototype extraction is performed on the multiple aggregated historical query intent parsing feature sets to obtain multiple query intent parsing feature prototypes, including: The similarity of the aggregated historical query intent parsing features within each of the multiple aggregated historical query intent parsing feature sets is identified to obtain multiple pairs of aggregated historical query intent parsing feature similarity sets. Based on the multiple sets of similarity of aggregated historical query intent parsing features, the aggregated historical query intent parsing feature corresponding to the maximum value of the average similarity of the two aggregated historical query intent parsing features associated in the multiple sets of aggregated historical query intent parsing feature is used as a prototype to obtain multiple query intent parsing feature prototypes.

2. The self-maintenance method for an insurance clause knowledge base based on a large model as described in claim 1, characterized in that, Long text data is segmented into fragments using semantic segmentation, and each segmented clause is associated with the original clause data to obtain a set of associated segmented clause fragments, including: The extremely long text data is subjected to structured partitioning to obtain a set of structured extremely long text data; The structured ultra-long text data set is segmented using a semantic paragraph segmentation algorithm to obtain a set of related segmented clause fragments; Each associated segmented clause includes the original path, context ID, and document index.

3. The self-maintenance method for an insurance clause knowledge base based on a large model as described in claim 1, characterized in that, The set of related segmented clauses is vectorized using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base, including: The semantic embedder is used to perform vector transformation on the set of related segmented clause fragments to obtain a set of semantic vectors for the related segmented clause fragments. The semantic vector set of the associated segmented clause fragments is identified by traversing the set of preset metadata field tags to obtain the semantic vector set of the associated segmented clause fragments. The preset metadata field tag set includes the original clause ID, chapter name, clause number, update time, applicable insurance type, user access frequency, and click popularity. The semantic vector set of the identified segmented clause fragments is stored in a cloud database to obtain the vectorized cloud knowledge base.

4. The self-maintenance method for an insurance clause knowledge base based on a large model as described in claim 1, characterized in that, When the vectorized cloud knowledge base receives a user's search request, it adjusts the priority of knowledge base terms through a two-level search mechanism to obtain an updated vectorized cloud knowledge base, including: The user search request is vectorized to obtain the user search request vector; Based on the vector cloud knowledge base, a first-level semantic coarse recall is performed on the user's retrieval request vector to obtain a first-level candidate clause fragment set; A cross-encoder is used to perform semantic relevance scoring on the user retrieval request vector and the set of first-level candidate clause fragments. Based on the scoring results, the set of first-level candidate clause fragments is reordered to obtain an updated vector cloud knowledge base.

5. The self-maintenance method for an insurance clause knowledge base based on a large model as described in claim 4, characterized in that, include: Retrieve the set of user search behavior logs within a preset window; Using retrieval frequency as an index, the user retrieval behavior log set is analyzed for the frequency of retrieval of clause fragments. The clause fragments are weighted in descending order of retrieval frequency, and the weights of the updated vector cloud knowledge base are updated and corrected based on the weighting results.

6. A self-maintaining system for an insurance clause knowledge base based on a large model, characterized in that: The system is used to implement the self-maintaining method for an insurance clause knowledge base based on a large model as described in any one of claims 1-5, the system comprising: Semantic segmentation module: It segments ultra-long text data into fragments through semantic segmentation, and associates each segmented clause fragment with the original clause data to obtain a set of associated segmented clause fragments; Vector Conversion Module: The set of related segmented clause fragments is converted into vectors using a semantic embedding device and stored in a cloud database to obtain a vectorized cloud knowledge base; Update module: When the vectorized cloud knowledge base receives a user's search request, it adjusts the priority of knowledge base terms through a two-level search mechanism to obtain an updated vectorized cloud knowledge base; Response module: Responds to the user's search request based on the updated vector cloud knowledge base and obtains the search results for the clause fragments; Search Enhancement Module: Input the Prompt construction results and clause fragment search results into a large language model built based on RAG enhancement generation technology to perform search enhancement generation and obtain target search interpretation and analysis results.

7. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the self-maintaining method for the insurance clause knowledge base based on a large model as described in any one of claims 1-5.

Images