Insurance clause accurate retrieval method and system based on large model incremental training

Abstract

The application discloses an insurance clause accurate retrieval method and system based on large model incremental training, relates to the technical field of natural language processing, and comprises the following steps: constructing an insurance clause database, and establishing a perception level through perception analysis; under the constraint of the perception level, establishing semantic tags through named entity recognition and semantic role labeling; using the two to construct a two-level retrieval vector; receiving user input, calling an intent recognition channel to establish a structured query vector; retrieving and matching to obtain results; recording feedback, periodically extracting query feedback history and newly added clause data, and incrementally updating the two-level retrieval vector and retrieval matching. The application solves the technical problem that the traditional data processing method cannot accurately capture semantic association and hierarchical structure when dealing with complex structure and numerous content insurance clauses, resulting in one-sided retrieval results, low matching degree, and difficulty in meeting the accurate retrieval requirements, and achieves the technical effects of accurate insurance clause retrieval and improved comprehensiveness and matching degree of retrieval results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an insurance clause accurate retrieval method and system based on large model incremental training. BACKGROUND

[0002] In the insurance service industry, accurate retrieval of insurance clauses is crucial for users to understand rights and interests and for enterprises to provide services. In the prior art, traditional retrieval methods such as keyword matching are used to process insurance clause data. These methods can work when the clause structure is simple and the content is less, but as insurance products become more complex, the content of the clauses increases dramatically and the structure levels are diverse. Traditional data processing methods cannot accurately capture semantic associations and hierarchical structures when dealing with complex clauses, resulting in one-sided retrieval results and low matching degrees, which makes it difficult to meet the needs of users to quickly obtain accurate clause information and affects the efficiency and quality of insurance services. SUMMARY

[0003] The present application provides an insurance clause accurate retrieval method and system based on large model incremental training, which solves the technical problem that traditional data processing methods cannot accurately capture semantic associations and hierarchical structures when dealing with complex structures and numerous contents of insurance clauses, resulting in one-sided retrieval results and low matching degrees, which makes it difficult to meet the needs of accurate retrieval.

[0004] In a first aspect, the present application provides an insurance clause accurate retrieval method based on large model incremental training, which comprises: after constructing an insurance clause database, performing perception analysis of the insurance clause database by a clause hierarchical parser to establish a perception hierarchy of general provisions, insurance liability, liability exemption, and claim settlement procedures; under the constraint of the perception hierarchy, establishing semantic tags through named entity recognition and semantic role labeling; using the perception hierarchy and the semantic tags to construct a two-level retrieval vector, the two-level retrieval vector including a first-level retrieval vector constructed based on structure positioning and a second-level retrieval vector constructed based on the first-level retrieval vector and sentence positioning; after receiving a user's input sentence segment, calling an intent recognition channel to perform intent recognition matching under a designated template to establish a structured query vector; using the structured query vector to perform retrieval matching based on the two-level retrieval vector to establish a retrieval matching result; after recording the retrieval matching result as a query feedback, extracting query feedback historical records and new clause data under a preset period, and using the extraction results as incremental data to perform incremental updates of the two-level retrieval vector and the retrieval matching.

[0005] In a second aspect of the present application, an insurance clause accurate retrieval system based on large model incremental training is provided, which comprises: a perception level construction module, which, after constructing an insurance clause database, performs perception analysis of the insurance clause database through a clause level parser to establish a perception level of general provisions, insurance liability, liability exemption, and claim settlement procedures; a semantic tag construction module, which, under the constraint of the perception level, establishes semantic tags through named entity recognition and semantic role labeling; a second-level retrieval vector construction module, which constructs a second-level retrieval vector using the perception level and the semantic tags, the second-level retrieval vector comprising a first-level retrieval vector constructed based on structure positioning and a second-level retrieval vector constructed based on the first-level retrieval vector and sentence positioning; a structured query vector construction module, which, after receiving an input sentence segment of a user, calls an intent recognition channel to perform intent recognition matching under a calibration template to establish a structured query vector; a retrieval matching result construction module, which performs retrieval matching based on the second-level retrieval vector using the structured query vector to establish a retrieval matching result; and an incremental update execution module, which, after recording the retrieval matching result as query feedback, extracts query feedback historical records and new clause data under a preset period, and executes incremental update of the second-level retrieval vector and retrieval matching using the extracted results as incremental data.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0007] After constructing an insurance clause database, the present application performs perception analysis through a clause level parser to establish a perception level, establishes semantic tags through named entity recognition and semantic role labeling under the constraint of the perception level, and constructs a second-level retrieval vector using the two; after receiving an input sentence segment of a user, the present application calls an intent recognition channel to establish a structured query vector, performs retrieval matching, and then extracts incremental data under a preset period to execute incremental update, thereby achieving accurate retrieval of insurance clauses, making the retrieval result more accurate and reliable, and achieving the technical effect of accurate retrieval of insurance clauses and improving the comprehensiveness and matching degree of the retrieval result. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0009] Figure 1 is a flowchart of the insurance clause accurate retrieval method based on large model incremental training provided by the embodiments of the present application.

[0010] Figure 2FIG. 1 is a structural schematic diagram of an insurance clause accurate retrieval system based on large model incremental training provided by an embodiment of the present application.

[0011] Legend: perception level construction module 1, semantic label construction module 2, second-level retrieval vector construction module 3, structured query vector construction module 4, retrieval matching result construction module 5, and incremental update execution module 6. DETAILED DESCRIPTION

[0012] The present application provides an insurance clause accurate retrieval method and system based on large model incremental training, which is used to solve the technical problem that the traditional data processing method cannot accurately capture semantic association and hierarchical structure when dealing with complex structure and numerous content insurance clauses, resulting in one-sided retrieval results and low matching degree, and it is difficult to meet the demand for accurate retrieval.

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0014] It should be noted that the terms "first", "second", etc. in the specification and the above drawings of the present application are used to distinguish similar objects, and do not necessarily mean a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to these processes, methods, products or devices.

[0015] Embodiment one, as shown in the figure, the insurance clause accurate retrieval method based on large model incremental training, wherein the method comprises: Figure 1

[0016] Step A100: After building the insurance clause database, the perception analysis of the insurance clause database is performed by the clause level parser to establish the perception level of general provisions, insurance liability, liability exemption, and claim settlement procedure.

[0017] In the embodiments of the present application, the clause level parser is a tool for performing perception analysis on the insurance clause database.

[0018] ​Specifically, when constructing the insurance clause database, first, collect various types of insurance clause raw data, covering different insurance types, such as life insurance, property insurance, health insurance, etc. The clause text includes electronic documents, scanned copies and other forms. It is assumed that a total of 100,000 different versions of insurance clauses are collected, covering the core content modules of general provisions, insurance liability, liability exemption, and claim settlement procedures.

[0019] Next, the collected raw data is processed for format unification. Scanned copies are converted into editable text through OCR technology, and text is extracted from PDF, Word and other format documents. After processing, all clause data is converted into pure text format, and redundant information such as repeated headers and footers, invalid symbols, etc. are removed, such as removing garbled characters and repeated paragraphs, to ensure the integrity of the text data.

[0020] Subsequently, data verification is performed to check the consistency of the clause text with the original file, correct recognition errors or omissions, correct the corresponding recognition bias, and ensure the accuracy of the data. Finally, the processed clause text is stored according to the type of insurance, release time, etc. dimensions, establishes a basic index, and forms a structured insurance clause database, providing a standardized data foundation for subsequent perception analysis.

[0021] Finally, the insurance clause database is cleaned up, the clause document tree structure is constructed according to the word matching and paragraph position characteristics of the organizational structure unit, the initial title level segmentation is performed using the structure, the hierarchical structure mapping table is generated based on the segmentation of the title level and the paragraph position, and the syntax dependency analysis of the clause paragraph is performed using the mapping table to obtain the syntax depth distribution. Finally, the perception level is established according to the initial title level segmentation, the hierarchical structure mapping table and the syntax depth distribution. The specific steps are described in detail in A110-A150.

[0022] Step A200: Under the constraint of the perception level, semantic tags are established through named entity recognition and semantic role labeling.

[0023] In the embodiments of the present application, the named entity includes name, organization name, equipment, parameter, and operation behavior. The semantic role is the semantic function assumed by each component in the sentence after the semantic block decomposition of the syntax tree.

[0024] Optionally, under the constraint of the perception level, first establish a semantic perception framework according to the perception level. Under the guidance of this framework, identify named entities including personal names, organization names, devices, parameters, and operation behaviors in the insurance clause database through a semantic salience scoring model based on the partition weight strategy. Disambiguate and update the identified named entities within a preset window, then decompose the semantics of the sentence into semantic chunks using a syntactic tree and extract semantic roles. Finally, establish semantic tags based on the updated named entities and semantic roles. The specific steps are described in detail in A210 - A250.

[0025] Step A300: Construct a secondary retrieval vector using the perception level and the semantic tags. The secondary retrieval vector includes a first - level retrieval vector constructed based on structural positioning and a second - level retrieval vector constructed based on the first - level retrieval vector and sentence positioning.

[0026] In an embodiment of the present application, divide the structural level units using the perception level, combine the structural position encoding and semantic tags to construct the first - level retrieval vector, and then generate the second - level retrieval vector through the analysis of the position and semantic features at the sentence granularity based on the structural positioning area pointed to by this vector, so as to construct the secondary retrieval vector. The specific steps are described in detail in A310 - A330.

[0027] Step A400: After receiving the input sentence segment from the user, call the intent recognition channel to perform intent recognition and matching under the calibration template to establish a structured query vector.

[0028] Specifically, after receiving the natural - language sentence segment input by the user, first pre - process the input content, including word segmentation, removing stop words (such as meaningless words like "de" and "le"), and standardization processing, such as converting "What materials are needed for claims settlement" to "What materials are required for claims settlement". After pre - processing, the effective input retention rate reaches more than 95%, ensuring the standardization and consistency of the input content.

[0029] Subsequently, call the intent recognition channel, which is constructed based on a language model fine - tuned for the insurance field and has a calibration template library for the insurance clause retrieval scenario built - in. The template library contains multiple types of core intent templates such as insurance liability query, liability exemption confirmation, and claims settlement process consultation. Each template is associated with 50 - 100 typical expressions, such as "Which diseases are covered" corresponding to the insurance liability template, and "What situations are not covered" corresponding to the liability exemption template, etc. Calculate the semantic similarity between the user - input sentence segment and each template using the cosine similarity algorithm to match the most likely user intent. For example, the similarity between an input "How much can be compensated for the accidental death of the insured" and the insurance liability - liability limit template reaches 0.85 and is determined to be in this intent category.

[0030] After completing the intent matching, key information in the input sentence segment is extracted based on the matching result, including named entities, semantic roles, and core parameters. These information is integrated according to a preset structured format, such as intent category + entity information + role relationship, and converted into a computable vector form, i.e. structured query vector. For example, the vector corresponding to the above input contains intent: insurance liability - liability limit, entity: insured, accidental death, parameter: payment amount, etc. dimensions, and the vector dimension is set to 64, of which 20 dimensions are used to represent the intent category and 44 dimensions are used to carry entity and parameter information.

[0031] By preprocessing user input, matching intent templates, extracting key information and converting it into a vector, a structured query vector is established, providing a clear query basis for precise matching based on secondary retrieval vectors.

[0032] Step A500: Using the structured query vector for retrieval matching based on secondary retrieval vectors, a retrieval matching result is established.

[0033] Specifically, when using the structured query vector for retrieval matching, the cosine similarity algorithm is first used to calculate the semantic similarity between it and the secondary retrieval vector. The secondary retrieval vector contains a first-level retrieval vector with structural positioning and a second-level retrieval vector with sentence positioning, so the first-level retrieval vector is matched first, and candidate units with a structural positioning area matching degree higher than 0.7 are selected, and then the second-level retrieval vector is matched within these units. The cosine similarity algorithm is used again to calculate the matching degree of the sentence-level vector, and it is assumed that the top 20 retrieval results are obtained.

[0034] On this basis, the account trust degree is configured in combination with the historical record of the retrieval account, and the specific steps are explained in detail in A510.

[0035] At the same time, real-time feedback from users on retrieval results is obtained, such as clicking, skipping, and marking as irrelevant. For results marked as irrelevant, the difference between them and the query vector is analyzed, and a first correction strategy, i.e. account retrieval correction strategy, is established. For example, if a certain account repeatedly marks results related to the exemption clause as irrelevant, the matching weight of the liability exemption unit in the account query is reduced. In combination with the common feedback from all users, a second correction strategy, i.e. general correction strategy, is established to adjust the weight coefficient of the corresponding semantic features. The specific steps are explained in detail in A510-A540.

[0036] After completing the above steps, the accuracy of the retrieval matching result is improved, and the adaptation degree to the individual needs of different users is improved, and finally a retrieval matching result is established that meets the general retrieval logic and adapts to individual habits.

[0037] Through vector similarity matching, account trust adjustment and user feedback correction, accurate retrieval matching results are established, and the accuracy and personalized adaptation ability of insurance clause retrieval are improved.

[0038] Step A600: After recording the retrieval matching results after query feedback, the query feedback history and new clause data are extracted in a preset period, and the extraction results are used as incremental data to perform incremental updating of the secondary retrieval vector and retrieval matching.

[0039] Specifically, the query feedback history and new clause data are extracted as incremental data in a preset period, and through feedback analysis, clustering, correction data construction, new clause processing, vector construction, conflict verification, etc., incremental updating of the secondary retrieval vector and retrieval matching is realized, and the specific steps are described in detail in A610-A670.

[0040] Further, the method provided in the embodiment of the application comprises the following steps:

[0041] A110: After data cleaning of the insurance clause database, a clause document tree structure is constructed according to the characteristics of the organizational structure unit, and the characteristics of the organizational structure unit include word matching features and paragraph position features.

[0042] A120: Initial title level segmentation of the insurance clause database after data cleaning using the clause document tree structure.

[0043] A130: Based on the initial title level segmentation, a clause level index of title level and paragraph position is established, and a hierarchical structure mapping table is generated.

[0044] A140: Using the hierarchical structure mapping table, clause paragraph syntax dependency analysis is performed to obtain syntax depth distribution.

[0045] A150: According to the initial title level segmentation, hierarchical structure mapping table and syntax depth distribution, a perception level is established.

[0046] In the embodiment of the application, the organizational structure unit is the basic component of the document structure in the insurance clause, which has word matching features and paragraph position features. These features are used to construct a clause document tree structure to support the establishment of a subsequent perception level.

[0047] Specifically, the construction of the clause level parser is based on natural language processing technology, which integrates structure recognition model, syntax analysis algorithm and feature extraction module. The construction of the clause level parser needs to be realized in modules, and the functions and construction logic of each module are closely related to the structural features of the insurance clause:

[0048] The structure recognition model is one of the core modules of the parser, which is constructed based on insurance clause samples with annotated structure levels as training data. These samples cover clause texts of different insurance types, and the annotated content includes title levels, such as the first-level title general provisions, the second-level title insured scope, corresponding paragraph positions, and core terms. Through machine learning algorithms, such as sequence labeling models based on Transformer, the samples are trained. The model learns the word matching rules of organizational structure units, such as the fixed occurrence position of core terms such as general provisions and insurance liability in the clause, the collocation pattern with other terms, such as the association of liability exemption with the following cases, and the corresponding relationship between paragraph position features, such as the first-line indentation and bold font of the title paragraph, and levels, such as the smaller indentation of the first-level title than the second-level title. Finally, a model capable of recognizing clause structure is formed.

[0049] The syntax analysis algorithm, as another key component of the parser, is constructed based on dependency syntax theory, which defines the grammatical relationship between words in a sentence through coding rules, such as subject-predicate, verb-object, and modifier relationships. For example, for sentences such as "the insured person will be responsible for compensation if an accident occurs during the insurance period", the algorithm will parse the subject-predicate relationship between the insured person and the occurrence, the subject-predicate relationship between the insurance company and the undertaking, and the verb-object relationship between the undertaking and the liability. By calculating the hierarchical depth of these relationships, such as the first level for subject-predicate relationships and the second level for verb-object relationships, an index is formed to measure the complexity of the sentence structure, providing algorithmic support for subsequent determination of the syntax depth distribution of the clause paragraphs.

[0050] The construction of the feature extraction module focuses on the quantitative extraction of organizational structure unit features. Through text preprocessing tools such as word segmentation and part-of-speech tagging, clause texts are processed to extract word matching features (such as the frequency of liability exemption appearing in the paragraph and the co-occurrence probability with words such as exclusion and non-undertaking) and paragraph position features (such as recording the page number of the paragraph in the document and the spacing from the previous paragraph), and these features are converted into calculable vector form for the structure recognition model and syntax analysis algorithm.

[0051] The integration of these three modules forms the clause level parser: the feature extraction module is responsible for outputting structured feature data, the structure recognition model identifies the structure units of the clause based on these features and constructs a tree structure, and the syntax analysis algorithm performs syntax analysis on the paragraphs to obtain the syntax depth, which collectively supports the perception parsing process.

[0052] Next, when performing cognitive analysis, the data in the insurance clause database needs to be systematically cleaned: for example, for the 100,000 collected clause data covering different types of insurance such as life insurance, property insurance, and health insurance, first, duplicate texts are identified and removed by comparing text hash values. Such repeated content is mostly the same general description in different versions of the clauses, which accounts for no more than 5% of the original data. Then, using Unicode encoding verification and character set conversion technology, the garbled characters generated due to format conversion or scanning recognition are screened out, such as unrecognized special symbols and broken characters. Such garbled characters account for about 1% of the original data. Invalid annotations in the clauses that do not belong to the main text are located and removed through keyword matching (such as "note:" and "editor's note"). Such annotations account for about 1% of the original data. After the above processing, all texts are in a unified format and complete content, providing a standardized and reliable data basis for subsequent construction of clause document tree structure based on organizational structure unit features.

[0053] Subsequently, the clause level parser calls the feature extraction module, and according to the word matching features (such as identifying keywords such as general provisions, insurance liability, and exemption from liability in the fixed expressions of the clauses) and paragraph position features (such as chapter titles being located at the beginning of the paragraph and in bold font) of the organizational structure units, the cleaned clause text is converted into a tree structure with chapters, clauses, and paragraphs as nodes. For example, a life insurance clause is divided into general provisions, insurance liability, and other first-level nodes, and each first-level node is further divided into second-level nodes such as insurance conditions and insurance period, forming a clause document tree structure.

[0054] Next, the clause level parser uses the constructed tree structure to perform initial title level segmentation on the clauses, and by identifying the hierarchical relationship of the nodes in the tree structure, such as first-level nodes corresponding to major titles and second-level nodes corresponding to minor titles, the clause text is segmented into different title levels. Assuming that an average of 6-8 levels can be segmented from each clause, covering the complete structure from the general title to the specific clause items.

[0055] After that, based on the initial title levels obtained by segmentation, the parser establishes an index of title levels and paragraph positions, records the paragraph range contained in each title level, such as the insurance liability level corresponding to paragraphs 10-25, generates a hierarchical structure mapping table, and realizes the precise association of titles and content positions.

[0056] On this basis, the parser calls the syntax analysis algorithm and combines the hierarchical structure mapping table to perform syntax dependency analysis on the clause paragraphs at each level, calculates the hierarchical depth of the grammar relationships such as subject-predicate and verb-object in the sentence, obtains the syntax depth distribution, and for example, the syntax depth of complex clause sentences is mostly 4-5 layers, and the syntax depth of simple statements is 2-3 layers.

[0057] Finally, the parser integrates the initial title level segmentation results, the position information in the hierarchical structure mapping table, and the syntactic depth distribution features to accurately classify the clause text into four perception levels: general provisions, insurance liability, liability exemption, and claims procedure, thereby constructing a complete perception hierarchy.

[0058] By using a clause-level parser to clean, structure, segment, index, analyze syntax, and integrate the insurance clause database, accurate perception and parsing of insurance clauses is achieved, and a standardized perception hierarchy is established.

[0059] Furthermore, step A200 in the method provided in this application embodiment includes:

[0060] A210: Establish a semantic perception framework based on the perception hierarchy.

[0061] A220: Guided by the semantic awareness framework, named entity recognition of data in the insurance terms database is performed based on a partition weight strategy and a semantic saliency scoring model. The named entities include personal names, organization names, equipment, parameters, and operational behaviors.

[0062] A230: Perform context disambiguation resolution within a preset window on the identified named entities and update the named entities.

[0063] A240: Use a syntax tree to decompose the semantic blocks of a statement and extract semantic roles based on the decomposition results.

[0064] A250: Create semantic tags based on updated named entities and semantic roles.

[0065] Optionally, based on the different semantic characteristics of the perception levels, a corresponding semantic perception framework can be established first. When establishing a semantic perception framework, it is necessary to first deeply analyze the core semantic characteristics of the four parts in the perception level: general provisions, insurance liability, exclusions, and claims procedures. The general provisions, as the foundation of the clauses, are mainly used to define basic concepts and applicable premises. Therefore, core semantic dimensions are extracted around the definitions of terms (such as the definition of "insured" and "insured object") and the scope of application (such as geographical restrictions and time ranges). Insurance liability is the core of the clauses, focusing on the coverage that the insurance company must bear. Therefore, dimensions such as coverage content (such as disease type and accident) and liability limits (such as maximum compensation amount and compensation ratio) are highlighted. Exclusions clarify the situations in which the insurance company does not bear liability. Therefore, exclusions (such as intentional acts and specific diseases) are the core dimensions. The claims procedures regulate the claims process, focusing on setting dimensions around process steps (such as reporting the incident, submitting materials, and review) and time limits (such as the reporting time limit and claims cycle).

[0066] Subsequently, these extracted semantic dimensions are bound to the corresponding levels to form a structured semantic awareness framework, which clearly defines the semantic direction that needs to be captured and analyzed under each level, and sets a clear semantic range for subsequent work such as named entity recognition and semantic role labeling under the guidance of the framework, ensuring that semantic analysis always aligns with the core functions of each level.

[0067] In the partition weight strategy under the guidance of the semantic awareness framework, weight distribution needs to be refined for the core content of the four awareness levels: the general level focuses on term definition and applicable scope, considering that it is the basis for clauses but not the core of retrieval; the insurance liability level is the most concerned part for users, with the core including guarantee content, liability limit, and payment conditions, and priority is given to identifying key guarantee information; the liability exemption level needs to highlight the application conditions of excluded situations and exemption clauses to ensure that users are clear about the scope of non-protection; the claim settlement procedure level focuses on process steps, time limit requirements, and material requirements to match users' query needs for claim settlement operations. Through this differentiated weight distribution of core content within each level, as shown in Table 1, semantic analysis resources are tilted towards key information, improving the relevance and accuracy of named entity recognition.

[0068] Table 1: Core content and weight distribution table of awareness levels

[0069] Content type General provisions Insurance liability Liability exemption Claims procedure Core content and weight Term definition (0.5); scope of application (0.4) Guarantee content (0.9); liability limit (0.8); payment conditions (0.7) Exclusion (0.8); exemption clause application conditions (0.6) Process steps (0.7); time limit requirements (0.6); material requirements (0.5)

[0070] At the same time, named entity recognition is performed on the data in the insurance clause database through a semantic salience scoring model. The model scores candidate entities such as names, organization names, devices, parameters, and operation behaviors in the text, and filters out entities with a salience score higher than a threshold (e.g., 0.6).

[0071] The construction of the semantic salience scoring model is centered on the named entity recognition needs in the insurance clause domain. First, a large-scale insurance clause corpus annotated with named entities such as names, organization names, devices, parameters, and operation behaviors is collected as training data. Each corpus needs to be annotated with entity categories and manually judged salience scores, ranging from 0 to 1, with higher scores indicating that the entity is more critical in the clause. The model uses the BERT pre-trained language model as the basic architecture, and connects a fully connected layer for salience scoring to the output layer, adapting to the insurance domain semantics through fine-tuning.

[0072] During training, the partition weight strategy is integrated, setting level weight parameters for different awareness levels such as general, insurance liability, and liability exemption. For example, according to expert recommendations, the weight coefficient of "guarantee content" entities in the insurance liability level can be set to 0.9, higher than the 0.5 of "term definition" entities in the general level, so that the model tilts towards key entities in the key level when scoring.

[0073] On the other hand, in addition to the fixed weight coefficient set based on the above expert recommendations, a dynamic weight prediction model can also be constructed to realize dynamic adjustment of the weight coefficients of each level label. The model is built using the XGBoost gradient boosting tree algorithm, and the input features include the user's historical search frequency of each perception level label within a preset time window (such as the last 90 days), the user feedback rate of the search results corresponding to each level label, including the click proportion and the proportion of relevant labels; the semantic association degree between the search term and each level label can be calculated by cosine similarity; and the search period distribution features, such as the difference in search frequency between weekdays and holidays. During the model training stage, the input features need to be preprocessed first to eliminate abnormal search records, such as invalid records with a single search duration of less than 1 second, and normalized. Then, the preprocessed feature data is divided into training set, validation set and test set in the ratio of 7:2:1, and the actual optimal value of the weight coefficient of each level label is used as the training target. The actual optimal value can be calculated by the historical search effect in reverse, i.e. the weight value that makes the search accuracy the highest. Then, the mean square error loss function is used to iteratively optimize the model parameters, and the learning rate is set to 0.01-0.1 and the tree depth is set to 3-8 during the process to avoid overfitting or underfitting of the model. Finally, the trained dynamic weight prediction model is obtained, which can receive the latest search data of the user in real time and update the input features, and output the dynamic weight coefficient of each perception level label. The dynamic weight coefficient and the fixed weight coefficient are fused according to the dynamic weight proportion of 0.6 and the fixed weight proportion of 0.4, and then input into the semantic saliency scoring model, so that the model's scoring of key level core entities is more inclined to the actual search needs of the user.

[0074] At the same time, the context features of the entity (such as the co-occurrence frequency of core terms such as insurance amount and claim settlement) and the position features (such as the probability of appearing at the beginning / end of a paragraph) are extracted as auxiliary inputs to optimize the scoring logic through multi-feature fusion. During the training process, the annotated saliency score is used as the target, and the mean square error loss function is used to iteratively optimize the model parameters. Finally, the optimal threshold value, such as 0.6, is determined through the validation set, so that the semantic saliency scoring model can output the saliency score of each candidate entity, and accurately select the core entities with actual meaning in the clause.

[0075] Then, the identified named entity is disambiguated and analyzed within a preset 5-sentence context window. For example, the beneficiary may refer to the statutory beneficiary or the designated beneficiary in different clauses, and through the expressions such as the policyholder can designate the beneficiary within the window, the specific reference is clarified, and the entity information is updated. After disambiguation, the entity recognition accuracy is significantly improved.

[0076] After that, the syntax tree is used to perform semantic block decomposition on the sentence, the sentence is split into semantic blocks such as subject-predicate, verb-object, and modification, and semantic roles such as agent (subject performing the action), patient (object of the action), and time adverbial are extracted according to the functions of the semantic blocks in the sentence. For example, in the sentence "the insured person occurs an accident during the insurance period, and the insurance company pays the insurance money", the semantic blocks of the insured person (agent), accident (action), insurance company (agent), and insurance money (verb-object) are decomposed, and the semantic roles of the insured person and insurance company as agents, the insurance money as the patient, and the insurance period as the time adverbial are extracted.

[0077] Finally, the updated named entity is associated and integrated with the extracted semantic role to form a semantic label. For example, insurance liability-parameter (insurance amount 500,000 yuan)-patient (insured person), and claim procedure-operation behavior (submit materials)-time adverbial (within 10 days after the accident occurs). These labels cover the key semantic information in the clauses, and provide accurate semantic basis for the construction of the secondary retrieval vector.

[0078] By establishing a semantic perception framework, recognizing and updating named entities, and extracting semantic roles, semantic labels covering key information are finally established, providing semantic support for accurate retrieval of insurance clauses.

[0079] Further, the method provided in the embodiment of the application comprises the following steps A300:

[0080] A310: performing structural hierarchical unit division on the data in the insurance clause database according to the perception level.

[0081] A320: After establishing the structural position code by structural hierarchical unit division, extracting the semantic labels corresponding to the structural hierarchical units, and establishing the first-level retrieval vector according to the structural hierarchical unit division and the corresponding semantic labels.

[0082] Specifically, first, the data in the insurance clause database is divided into structural hierarchical units according to the division of the general provisions, insurance liability, liability exemption, and claim procedure in the perception level. For example, 100,000 clauses are divided into four first-level structural units according to the perception level, corresponding to the four perception levels, and each first-level unit is further divided into two second-level units according to the core content: the general provisions unit is divided into two second-level units of term definition and applicable scope, the insurance liability unit is divided into three second-level units of guarantee content, liability limit, and payment condition, the liability exemption unit is divided into two second-level units of exclusion situation and exemption condition, and the claim procedure unit is divided into three second-level units of process steps, time limit requirement, and material requirement, finally forming 10 structural hierarchical units, each containing 5,000-15,000 clauses.

[0083] Then, the divided 10 structural level units are coded for structural position, using a multi-dimensional coordinate coding rule: the first-level units are distinguished by the first digit, such as 0 for general provisions, 1 for insurance liability, and the rest follow; the second-level units are distinguished by the second digit, such as 1.0 for protection content under insurance liability, 1.1 for liability limit, 1.3 for payment conditions, and the rest follow to generate unique codes such as 0.0, 1.2, 3.1, etc., and the codes are converted into 32-bit vector form to ensure that the position information of each unit can be recognized by the algorithm.

[0084] After that, the semantic labels corresponding to each structural level unit are extracted, which come from the labeling results established based on named entities and semantic roles in the foregoing step A250. For example, in the insurance liability-liability limit unit, the semantic labels such as parameter (insurance amount 500,000 yuan), parameter (payment ratio 80%), and operation behavior (calculate payment) are extracted; the semantic labels such as operation behavior (report claim), operation behavior (audit), and semantic role (agent: insurance company) are extracted in the claim settlement procedure-process step unit, and an average of 5-7 semantic labels are extracted for each structural level unit.

[0085] Finally, the coding vector of the structural level unit is fused with the embedding vector of the corresponding semantic label to generate a first-level retrieval vector. Specifically, the 32-bit position coding vector is spliced with the 32-bit embedding vector generated by the BERT model for the semantic label to form a 64-dimensional first-level retrieval vector. For example, the retrieval vector of the insurance liability-protection content unit fuses the position coding and the embedding information of the labels such as disease type and operation behavior, and finally 10 first-level retrieval vectors are generated corresponding to the 10 structural level units, each vector accurately associates the semantic information of a specific structural position.

[0086] By dividing the structural level units, establishing position coding, extracting semantic labels, and generating vectors, the first-level retrieval vector reflecting the association between structural position and semantics is constructed, laying a foundation for the structured retrieval of insurance clauses.

[0087] Further, the method provided in the embodiments of the present application comprises the following steps A300:

[0088] A330: based on the structural positioning area pointed by the first-level retrieval vector, performing sentence-level semantic analysis, which comprises: obtaining the relative position and absolute position index of each sentence in the paragraph; extracting semantic feature information of each sentence in the structural area, which comprises semantic role, named entity category, context entity association, and semantic weight; performing sentence-level semantic embedding generation according to the position feature and semantic feature information of the sentence to establish a second-level retrieval vector.

[0089] In the embodiments of the present application, the structure positioning area refers to a specific structure area pointed to by the first-level retrieval vector and determined by a structure level unit, such as a specific structure unit under the perception level of general principles, insurance liability, etc. The sentence granularity refers to the granularity of semantic analysis in units of sentences, that is, semantic analysis is separately performed on each sentence in the structure area, including obtaining position information, extracting semantic features, etc.

[0090] Specifically, based on the structure positioning area pointed to by the first-level retrieval vector, for example, the liability limit unit under the insurance liability, semantic analysis of the sentence granularity is performed on the clause text in this area. Assuming that this structure area contains 100 sentences, the position and semantic information of each sentence needs to be processed one by one.

[0091] First, the position features of each sentence are obtained: the relative position in the paragraph is recorded, that is, the sentence is the 1st, 2nd, etc. in the paragraph to which it belongs, such as a sentence being the 3rd in the paragraph; and the absolute position index is labeled, that is, the global serial number of the sentence in the entire insurance clause document, such as the 286th sentence in the document. Through these two kinds of position information, the specific position of the sentence in the structure area is accurately located, providing a spatial coordinate reference for subsequent retrieval.

[0092] Next, the semantic feature information of each sentence is extracted: semantic roles are identified, such as the insurance company (agent) and the insured (patient); named entity categories are divided, such as the parameter (50 million yuan of premium) and the operation behavior (calculate the compensation); the association between context entities is analyzed, such as the association strength between the premium and the compensation ratio in the same paragraph being set to 0.7; and the semantic weight is determined in combination with the zoning weight strategy, such as the semantic weight of the sentence in the liability limit unit being uniformly set to 0.8. Taking one of the sentences, “After the insured person of the insured, the insurance company pays the premium of 50 million yuan at a compensation ratio of 80%”, as an example, the semantic roles of the insured (patient) and the insurance company (agent) are extracted, the named entity categories of the parameter (compensation ratio 80%) and the parameter (premium 50 million yuan) are extracted, the association degree between the context of the accident and the compensation is 0.6, and the semantic weight is 0.8.

[0093] Then, according to the position feature and the semantic feature information of the sentence, sentence-level semantic embedding generation is performed. The relative position and the absolute position index are converted into a 32-dimensional position vector, the semantic feature information is converted into a 96-dimensional semantic embedding vector through the BERT pre-training language model, and the two vectors are spliced and fused to generate a 128-dimensional sentence-level vector, that is, the second-level retrieval vector. After processing the above 100 sentences, 100 second-level retrieval vectors are generated, which together with the first-level retrieval vector form the second-level retrieval vector covering the structure and the sentence level.

[0094] By positioning the sentence position, extracting the semantic features and generating the embedding vector, a second-level retrieval vector based on sentence positioning is constructed, which cooperates with the first-level retrieval vector to form a two-level retrieval vector, providing vector support for fine-grained and accurate retrieval of insurance clauses.

[0095] Further, the step A600 in the method provided by the embodiment of the application comprises:

[0096] A610: performing feedback analysis on the query feedback history record to establish a feedback analysis result.

[0097] A620: performing query feedback bias clustering of the feedback analysis result to generate a feedback bias clustering result.

[0098] A630: constructing incremental correction data based on the feedback bias clustering result and performing incremental learning update of retrieval matching according to the incremental correction data.

[0099] Specifically, first, the matching results and user interaction behaviors generated by each retrieval, such as clicks, marking irrelevant, and staying time, are recorded as query feedback, for example, 1000 retrieval feedbacks are recorded daily, covering different types of retrieval scenarios such as insurance liability query and claim process consultation. The preset period is set to be one month, and the query feedback history record of the past month is extracted within this period, and the newly added insurance clause data, such as the clauses added every month, are synchronously collected, and these data are combined as incremental data to provide a basis for subsequent update.

[0100] Next, the extracted query feedback history record is subjected to feedback analysis, and the core information of each feedback is identified through natural language processing technology: the feedback types are distinguished, such as irrelevant results, missing information, and positioning deviation, and the corresponding retrieval vector, user query intention and deviation point are extracted. For example, it is analyzed that 60% of the irrelevant result type feedback is due to the error of named entity recognition, such as mistakenly identifying accidental death as disease death, and 30% is due to structural positioning deviation, such as matching the exemption clause to the insurance liability area, to form a feedback analysis result.

[0101] Then, based on the feedback analysis result, the K-means clustering algorithm is used to perform query feedback bias clustering, and feedbacks with similar deviation characteristics are classified into one class. For example, feedbacks caused by parameter entity recognition deviation are classified into one class, and feedbacks caused by positioning errors of claim procedure steps are classified into another class, and finally multiple feedback bias clustering results are generated, each containing 100-300 similar feedbacks, clearly presenting the common problems of the retrieval system.

[0102] Afterwards, for each feedback-biased clustering result, common bias features are extracted to construct incremental correction data: for parameter identification bias class, supplement the training samples of this class of entities, for example, add 500 pieces of annotation data containing premium and claim ratio; for structure positioning bias class, adopt the method of rule preset + model auxiliary verification to perform positioning weight adjustment. First, common bias features are extracted from the structure positioning bias class clustering result, for example, the responsibility exemption unit clause is repeatedly matched to the insurance liability unit, and the basic adjustment rule is preset based on the partition weight strategy of the perception level. If the mis-matching rate of a certain structure unit is higher than 15% for two consecutive preset periods, the positioning weight of the unit is increased, and if the mis-matching rate is lower than 5%, the weight is maintained or slightly reduced.

[0103] Finally, the semantic significance scoring model constructed in the foregoing step A220 is called to assist in verifying the adjusted weight. The adjusted weight is substituted into the model to determine whether the semantic significance score of the clause corresponding to the structure unit is improved in the correct retrieval direction, for example, after the weight of the responsibility exemption unit is increased, the matching significance score of the responsibility exemption unit with the responsibility exemption query vector is also correspondingly improved. If the verification is passed, the final adjusted weight coefficient is determined. According to the incremental correction data, the retrieval matching model is updated by incremental learning. The adjusted structure level positioning weight and the newly added entity training sample are integrated into the model training process. The vector matching weight parameter of the model is fine-tuned (such as optimizing the weight proportion of the structure positioning dimension in the secondary retrieval vector) to optimize the vector matching logic, so as to ensure the accuracy of structure positioning in subsequent retrieval.

[0104] Among them, the construction of the retrieval matching model takes the secondary retrieval vector and the structured query vector as the core basis. First, the BERT pre-training language model is used as the underlying architecture, which is fine-tuned by integrating semantic features in the insurance clause field, so that it can understand professional terms and logical relationships in the clause. The model takes the first-level retrieval vector of structure positioning and the second-level retrieval vector of sentence positioning as the basic index library. The matching degree between the structured query vector and the secondary retrieval vector is calculated by the cosine similarity algorithm, and the candidate matching result is initially output.

[0105] At the same time, the model integrates account trust parameter and user feedback correction strategy to adjust the weight of the candidate result. The historical related result weight of high trust account is increased, and the sentence feature weight of high frequency recognized sentence in general feedback is enhanced, forming a dynamically optimized matching logic. In addition, the model reserves an incremental learning interface, which fine-tunes the parameters by continuously inputting incremental correction data such as newly added annotation samples and adjusted level weights, so that the accuracy of vector matching is continuously improved with data accumulation. Finally, a retrieval matching model is constructed which can capture both structure and sentence level semantics and adapt to user habits and new clauses.

[0106] By extracting incremental data, analyzing clustering feedback, constructing correction data, and updating the model, the incremental optimization of retrieval matching is realized, and the adaptability of the insurance clause retrieval system to user demand and new data is improved.

[0107] Further, the step A600 in the method provided by the embodiment of the application comprises:

[0108] A640: Under a preset period, the newly added clause data is collected, and after data cleaning of the newly added clause data is performed, structured analysis processing is performed.

[0109] A650: The structured analysis processing result is used to construct a newly added clause semantic vector.

[0110] A660: Index conflict verification is performed on the newly added clause semantic vector and the secondary retrieval vector, and according to the index conflict verification result, incremental updating of the secondary retrieval vector is performed.

[0111] A670: Retrieval matching updating is completed according to the updated secondary retrieval vector.

[0112] In an embodiment, in a preset period within a month, the existing insurance industry database or the insurance company website is connected through an API data interface to collect newly added insurance clause data, covering different types such as health insurance and property insurance. Data cleaning is performed on these newly added data to remove duplicate entries, incorrect codes, improper sentence breaks, and other format errors, standardized terms, such as uniformizing the payment amount to the compensation amount, to ensure that the data format is consistent with the existing insurance clause database, and the effective data retention rate is more than 90% after cleaning.

[0113] Next, structured analysis processing is performed on the cleaned newly added clause data, and the established perception level is referred to for structure division in four parts of general rules, insurance liability, liability exemption, and claim settlement procedure, for example, the protection scope part in a newly added health insurance clause is classified into the insurance liability-coverage content unit; at the same time, named entity recognition and semantic role labeling are performed, and an average of 6-8 structured information points are analyzed from each newly added clause to form a structured analysis result.

[0114] Then, based on the structured analysis result, a newly added clause semantic vector is constructed: the structure position code, such as the insurance liability-coverage content corresponding code 1.0 in the foregoing step A320, is converted into a 32-dimensional vector, the extracted semantic labels, i.e., named entities, semantic roles, etc., are converted into 32-dimensional embedding vectors through a BERT pre-training language model, and the two are spliced to form a 64-dimensional newly added clause semantic vector. The corresponding number of semantic vectors is generated for all newly added clauses, and the vector dimension is consistent with the existing secondary retrieval vector.

[0115] After that, the semantic vector of the newly added clause is verified for index conflict with the existing secondary retrieval vector. Whether there is semantic conflict is determined by calculating the cosine similarity. For example, if the semantic of the guarantee content and the exemption of responsibility in the same structural unit is contradictory, the similarity threshold is set to 0.15, and if the value is lower than the threshold, it is considered to be in conflict. After verification, the vectors in conflict in the newly added vector are obtained, and manual review is performed to confirm the conflict type.

[0116] Subsequently, according to the index conflict verification result, the incremental update of the secondary retrieval vector is performed: for the newly added vector without conflict, it is directly added to the secondary retrieval vector library; for the vector with conflict, its structure position code or semantic label weight is adjusted and then integrated, the total amount of the updated secondary retrieval vector library increases by the number of newly added insurance clause data, and the conflict rate is controlled within 3%.

[0117] Finally, according to the updated secondary retrieval vector, the retrieval matching model is retrained, the weight parameters of vector matching are adjusted, and the newly added clause can be accurately retrieved. For example, when the user queries the newly added guarantee scope of critical illness insurance in 2025, the updated system can list the CAR-T therapy included in the newly added clause as the first matching result, the retrieval response speed is maintained within 0.5 seconds, and the retrieval coverage rate of the newly added clause reaches 100%.

[0118] By collecting and cleaning the newly added data, structuring and analyzing, constructing semantic vectors, verifying conflicts and updating secondary retrieval vectors, the dynamic adaptation of the retrieval system to the newly added clauses is realized, ensuring that the retrieval results cover the latest content in time, and improving the timeliness and integrity of the insurance clause retrieval.

[0119] Further, the method provided in the embodiments of the present application comprises the following steps A500:

[0120] A510: Obtain the account historical retrieval record of the retrieval account, and configure the account trust degree according to the account historical retrieval record.

[0121] A520: Obtain the user retrieval feedback of the retrieval matching result, and establish a first correction strategy according to the user retrieval feedback, wherein the first correction strategy is an account retrieval correction strategy.

[0122] A530: Configure a second correction strategy according to the user retrieval feedback and the account trust degree, wherein the second correction strategy is a general correction strategy.

[0123] A540: Perform retrieval matching optimization management according to the first correction strategy and the second correction strategy.

[0124] Optionally, the account history search record of the search account is obtained, the search behavior data of the past 90 days is associated through the unique identification of the user account, including search words, corresponding click result serial number, user feedback of each result, such as relevant, irrelevant, no feedback, and average search time. Based on these data, the account trust degree is calculated: taking the feedback accuracy rate, i.e. the proportion of the click results marked as relevant in the total click amount, as the core index, supplemented by search stability, i.e. the consistency of the search theme for three consecutive times.

[0125] Further, the above calculation process can obtain data through the following specific steps:

[0126] Step a: calculation of feedback accuracy rate, count the number of search results clicked by the account within the past 90 days that are explicitly marked as relevant, and divide by the total click amount of the account, excluding unclicked or only browsed unmarked results. For example, account A has clicked 100 results within 90 days, of which 92 are marked as relevant, so the feedback accuracy rate = 92 ÷ 100 = 92%.

[0127] Step b: calculation of search stability, extract the search words of the account for three consecutive times, calculate the theme correlation of each of the three consecutive searches through a semantic similarity algorithm such as cosine similarity, with a threshold value of 0.7, and a theme consistency is considered if the value is higher than this value, and then count the proportion of the number of times of theme consistency in all three consecutive searches within 90 days. For example, account A has 20 groups of three consecutive searches within 90 days, of which 17 groups have a theme correlation degree ≥ 0.7, so the search stability = 17 ÷ 20 = 85%.

[0128] Step c: comprehensive calculation of trust degree, using a weighted formula, the feedback accuracy rate weight is 0.7, and the search stability weight is 0.3. The trust degree of account A = 92% × 0.7 + 85% × 0.3 = 0.644 + 0.255 = 0.899 ≈ 0.9; the feedback accuracy rate of account B is 65%, and the search stability is 50%, so the trust degree = 65% × 0.7 + 50% × 0.3 = 0.455 + 0.15 = 0.605 ≈ 0.6.

[0129] Next, the real-time feedback of the current search matching result of the user is obtained, including the position of the clicked result (such as the proportion of clicks in the top 3), the time length of staying (more than 30 seconds is considered as effective attention), and manually marked irrelevant reasons such as content error and position deviation. Based on these feedbacks, the first correction strategy, i.e., the account search correction strategy, is established, and the Prompt is optimized to refine the correction logic: for example, for a high-trust account A, which has repeatedly marked that the material requirement information in the claim settlement process result is missing, the Prompt is designed to preferentially match the sentence containing the material list and submission requirement entity, and the semantic weight of the claim settlement procedure-material requirement unit in the account search is increased to 0.8; for a low-trust account B, whose feedback is messy, the Prompt emphasizes strict matching according to the general semantic features, and reduces the personalized adjustment range.

[0130] Then, according to the search feedback data of all users and the account trust degree distribution, i.e., the proportion of different trust degree accounts, the second correction strategy, i.e., the general correction strategy, is configured. For example, through statistical analysis, it is found that 80% of high-trust accounts and 60% of low-trust accounts both feedback that the result ordering of insurance liability-coverage content is deviated. Combined with Prompt optimization, the general rule is formulated: the expression of priority analysis of user input such as protection range and insurance content is fixed in the Prompt of intent recognition, and the matching weight with the insurance liability-coverage content unit is strengthened to 0.9, while the rule strength is dynamically adjusted according to the trust degree, with a coefficient of 1.2 for high-trust accounts and a coefficient of 0.8 for low-trust accounts.

[0131] Finally, the search matching result is optimized and managed according to the first correction strategy and the second correction strategy, and the matching weight of the secondary search vector and the structured query vector is adjusted in real time to reorder the result.

[0132] By combining account trust degree and user feedback to build a hierarchical correction strategy, and integrating Prompt optimization to accurately guide intent recognition, the individualization and generalization optimization of search matching results are realized, and the accuracy of insurance clause search and user experience are improved.

[0133] Further, the step A520 in the method provided by the embodiment of the application includes:

[0134] A521: performing search exception level matching according to the user search feedback, and establishing a matching result.

[0135] A522: generating an exception warning based on the matching result, and performing exception reporting.

[0136] In one embodiment, first, collect user feedback information on retrieval matching results, including but not limited to specific types such as marking result errors, position deviations, information omissions, and the frequency and scope of feedback, such as more than 20% of users feeding back abnormalities for the same search term. Set retrieval abnormality level division criteria: slight abnormality is single user single feedback and does not involve core information error; moderate abnormality is single user multiple feedback or a small number of users within 5% feedback the same problem; serious abnormality is more than 5% of users feeding back the same core error.

[0137] Next, according to the above-mentioned criteria, the user retrieval feedback is matched with the abnormality level to establish a matching result. For example, among the 1200 feedback collected daily, 800 are slight abnormalities accounting for 67%, 300 are moderate abnormalities accounting for 25%, and 100 are serious abnormalities accounting for 8%, of which the serious abnormalities are mainly concentrated in the two problems of mis-matching the liability exemption clause to the insurance liability and identifying errors in the claim settlement time limit parameters.

[0138] Then, based on the abnormality level matching result, the corresponding warning is generated, the slight abnormality generates system log record as daily optimization reference; the moderate abnormality triggers department-level warning and is pushed to the retrieval optimization team for special analysis; the serious abnormality starts emergency warning, immediately notifies the technical person in charge and suspends the automatic matching of the related search term, and changes to manual review.

[0139] After that, the abnormality reporting process is executed: the slight abnormality log is automatically archived daily; the moderate abnormality is pushed through the internal management system and requires a response within 24 hours; the serious abnormality is notified through SMS and email and requires processing within 2 hours. The technical team traces the abnormality causes according to the warning information, such as the mis-matching of the liability exemption clause due to structural positioning weight setting deviation, and immediately adjusts the weight coefficient of the corresponding level, completes the repair and verification within 24 hours.

[0140] By dividing the abnormality level, matching the feedback information, generating the corresponding warning and executing the report, the graded response and rapid processing of the retrieval abnormality are realized, which guarantees the stability and accuracy of the insurance clause retrieval system.

[0141] In summary, the insurance clause accurate retrieval method based on large model incremental training provided by the embodiments of the present application has the following technical effects:

[0142] The present application establishes a perception level by constructing an insurance clause database, establishes a semantic tag under the constraint of the perception level, constructs a two-level retrieval vector using the perception level and the semantic tag, establishes a structured query vector after receiving user input and performs retrieval matching, and periodically updates incrementally after recording feedback, thereby accurately retrieving insurance clauses, making the insurance clause retrieval result more accurate and reliable, achieving accurate retrieval of insurance clauses, and improving the comprehensiveness and matching degree of the retrieval result.

[0143] Embodiment two, as Figure 2 shown, based on the same inventive concept as the preceding embodiment one, the present embodiment provides an insurance clause accurate retrieval system based on large model incremental training, comprising:

[0144] a perception level construction module 1, which is used to construct an insurance clause database, and then perform perception analysis of the insurance clause database through a clause level parser to establish the perception level of general provisions, insurance liability, liability exemption, and claim settlement procedures.

[0145] a semantic label construction module 2, which is used to establish semantic labels through named entity recognition and semantic role labeling under the constraint of the perception level.

[0146] a second-level retrieval vector construction module 3, which constructs a second-level retrieval vector using the perception level and the semantic labels, the second-level retrieval vector including a first-level retrieval vector constructed based on structural positioning and a second-level retrieval vector constructed based on the first-level retrieval vector and sentence positioning.

[0147] a structured query vector construction module 4, which is used to receive a user's input sentence segment, call an intent recognition channel to perform intent recognition matching under a calibrated template, and establish a structured query vector.

[0148] a retrieval matching result construction module 5, which is used to perform retrieval matching based on the second-level retrieval vector using the structured query vector, and establish a retrieval matching result.

[0149] an incremental update execution module 6, which is used to record the retrieval matching result as query feedback, extract query feedback historical records and new clause data under a preset period, and execute incremental update of the second-level retrieval vector and retrieval matching using the extracted results as incremental data.

[0150] Further, the perception level construction module 1 is used to perform the following steps:

[0151] After data cleaning of the insurance clause database, a clause document tree structure is constructed according to the characteristics of the organizational structure unit, the characteristics of the organizational structure unit including word matching features, paragraph position features; initial title level segmentation of the insurance clause database after data cleaning is performed using the clause document tree structure; a clause level index of title level and paragraph position is established based on the initial title level segmentation, and a hierarchical structure mapping table is generated; syntax dependency analysis of clause paragraphs is performed using the hierarchical structure mapping table, and a syntax depth distribution is obtained; and a perception level is established according to the initial title level segmentation, the hierarchical structure mapping table, and the syntax depth distribution.

[0152] Further, the semantic label construction module 2 is configured to perform the following steps:

[0153] A semantic perception framework is established according to the perception level; under the guidance of the semantic perception framework, named entity recognition of data in the insurance clause database is performed through a semantic saliency scoring model based on a partition weight strategy, the named entity including a person name, an organization name, an equipment, a parameter, and an operation behavior; context disambiguation analysis of the recognized named entity in a preset window is performed, and the named entity is updated; semantic role is extracted according to the semantic block decomposition result of the semantic block decomposition of the sentence using the syntax tree; and a semantic label is established according to the updated named entity and the semantic role.

[0154] Further, the second-level retrieval vector construction module 3 is configured to perform the following steps:

[0155] Structure level unit division of data in the insurance clause database is performed according to the perception level; after structure position coding is established using the structure level unit division, the semantic label corresponding to the structure level unit is extracted, and a first-level retrieval vector is established according to the structure level unit division and the corresponding semantic label.

[0156] Further, the second-level retrieval vector construction module 3 is configured to perform the following steps:

[0157] Based on the structure positioning area pointed to by the first-level retrieval vector, sentence granularity semantic analysis is performed, the sentence granularity semantic analysis including: obtaining the relative position and absolute position index of each sentence in the structure area; extracting semantic feature information of each sentence in the structure area, the semantic feature information including semantic role, named entity category, context entity association, and semantic weight; performing sentence-level semantic embedding generation according to the position features and semantic feature information of the sentence, and establishing a second-level retrieval vector.

[0158] Further, the incremental update execution module 6 is configured to perform the following steps:

[0159] feedback analysis is performed on the query feedback history record to establish a feedback analysis result; query feedback bias clustering of the feedback analysis result is performed to generate a feedback bias clustering result; incremental correction data is constructed based on the feedback bias clustering result, and incremental learning update of retrieval matching is performed according to the incremental correction data.

[0160] Further, the incremental update execution module 6 is configured to perform the following steps:

[0161] After the newly added clause data is collected in a preset period and the data cleaning of the newly added clause data is performed, the structured analysis processing is performed; the newly added clause semantic vector is constructed based on the structured analysis processing result; the index conflict verification is performed on the newly added clause semantic vector and the secondary retrieval vector, and the incremental update of the secondary retrieval vector is performed according to the index conflict verification result; and the retrieval matching update is completed according to the updated secondary retrieval vector.

[0162] Further, the retrieval matching result construction module 5 is configured to perform the following steps:

[0163] The account history retrieval record of the retrieval account is obtained, and the account trust degree is configured according to the account history retrieval record; the user retrieval feedback of the retrieval matching result is obtained, and the first correction strategy is established according to the user retrieval feedback, the first correction strategy being an account retrieval correction strategy; the second correction strategy is configured according to the user retrieval feedback and the account trust degree, the second correction strategy being a general correction strategy; and the retrieval matching optimization management is performed according to the first correction strategy and the second correction strategy.

[0164] Further, the retrieval matching result construction module 5 is configured to perform the following steps:

[0165] The retrieval abnormality level matching is performed according to the user retrieval feedback, and the matching result is established; the abnormality early warning is generated based on the matching result, and the abnormality report is performed.

[0166] The insurance clause accurate retrieval system based on the large model incremental training provided in the embodiment of the application can execute the insurance clause accurate retrieval method based on the large model incremental training provided in any embodiment of the application, has the function modules and beneficial effects corresponding to the execution method.

[0167] Although various references are made in this application to certain modules in the system according to the embodiments of the application, however, any number of different modules can be used and run on the user terminal and / or server, and each unit and module included is only divided according to the functional logic, but is not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of each functional unit are only for convenient mutual distinction, and do not limit the protection scope of the application.

[0168] The foregoing detailed description has set forth various embodiments of the devices and / or processes via the use of specific terminology. However, embodiments can be practiced without the specific details (e.g., efforts to maintain consistency or one technique over another can have been omitted) or with additional details that are within the scope and nature of the claims. By way of example, various embodiments can be practiced with additional layers, components, elements, devices, steps, operations, and / or the like. It is intended that each element recited in any of the claims is to be understood, where appropriate, as being additionally permissible under the patent statutes and judgments of equivalency. The disclosure can also include other embodiments not expressly shown or described herein. Additionally, it is intended that each symbolic representation of an element (including any reference signs) be understood as a generic term that can be non-limitingly replaced by any of the various embodiments described herein or equivalents.

Claims

1. A precise insurance clause retrieval method based on incremental training of a large model, characterized in that, The method includes: After constructing the insurance terms database, the perception parser of the insurance terms database is executed through the terms hierarchy parser to establish the perception hierarchy of general provisions, insurance liability, exclusions of liability, and claims procedures; Under the constraints of the aforementioned perception level, semantic tags are established through named entity recognition and semantic role labeling; A secondary retrieval vector is constructed using the perception level and the semantic tags. The secondary retrieval vector includes a first-level retrieval vector constructed based on structural localization and a second-level retrieval vector constructed based on the first-level retrieval vector and sentence localization. After receiving the user's input segment, the intent recognition channel is invoked to perform intent recognition matching under the calibration template to establish a structured query vector; The structured query vector is used to perform retrieval matching based on the secondary retrieval vector to establish retrieval matching results; After recording the search matching results as query feedback, the query feedback history and newly added clause data are extracted at a preset period. The extracted results are used as incremental data to perform secondary search vector and incremental update of search matching.

2. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 1, characterized in that, The process of performing perceptual parsing of the insurance terms database through a terms-level parser includes: After cleaning the insurance terms database, a tree structure of terms documents is constructed based on the characteristics of the organizational structure units. The characteristics of the organizational structure units include word matching features and paragraph position features. The initial header hierarchy segmentation of the insurance terms database after data cleaning is performed using the aforementioned tree structure of the terms document; Based on the initial heading level segmentation, a clause level index is established to determine the heading level and paragraph position, and a hierarchical structure mapping table is generated. The hierarchical structure mapping table is used to perform syntactic dependency analysis on clause paragraphs to obtain the syntactic depth distribution; A perception hierarchy is established based on the initial title hierarchy segmentation, hierarchy structure mapping table, and syntactic depth distribution.

3. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 1, characterized in that, The establishment of semantic labels through named entity recognition and semantic role labeling under the constraints of the perception level includes: A semantic perception framework is established based on the perception hierarchy; Guided by the semantic awareness framework, named entity recognition of data in the insurance terms database is performed based on a partition weight strategy and a semantic saliency scoring model. The named entities include personal names, organization names, equipment, parameters, and operational behaviors. Perform context disambiguation parsing within a preset window on the identified named entities and update the named entities; Semantic blocks of sentences are decomposed using syntactic trees, and semantic roles are extracted based on the decomposition results. Semantic tags are created based on the updated named entities and semantic roles.

4. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 1, characterized in that, The construction of a secondary retrieval vector using the perception level and the semantic tags includes: The data in the insurance terms database is divided into structural hierarchical units based on the perception level. After establishing structural position encoding by dividing the structural hierarchy units, semantic tags corresponding to the structural hierarchy units are extracted, and a first-level retrieval vector is established based on the structural hierarchy unit division and the corresponding semantic tags.

5. The method for accurate retrieval of insurance clauses based on incremental training of a large model as described in claim 4, characterized in that, The method of constructing a secondary retrieval vector using the perception level and the semantic tags also includes: Based on the structural localization region pointed to by the first-level retrieval vector, sentence-level semantic analysis is performed, which includes: Obtain the relative and absolute position indices of each sentence within the structural region. Extract semantic feature information of each sentence within the structural region, including semantic roles, named entity categories, contextual entity associations, and semantic weights; Sentence-level semantic embedding is generated based on the positional and semantic features of the sentence to establish a second-level retrieval vector.

6. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 1, characterized in that, The step of extracting query feedback history and newly added clause data within a preset period, and using the extracted results as incremental data to perform secondary retrieval vectors and incremental updates of retrieval matches includes: Perform feedback parsing on the query feedback history and establish feedback parsing results; Perform query feedback bias clustering based on the feedback parsing results to generate feedback bias clustering results; Incremental correction data is constructed based on the feedback bias clustering results, and incremental learning updates are performed based on the incremental correction data to retrieve matching results.

7. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 6, characterized in that, The step of extracting query feedback history and newly added clause data within a preset period, and using the extracted results as incremental data to perform secondary retrieval vectors and incremental updates of retrieval matching, further includes: Data on newly added clauses is collected at a preset period, and after data cleaning of the newly added clauses, structured parsing processing is performed. Construct semantic vectors for new clauses from the structured parsing results; The semantic vector and secondary retrieval vector of the newly added clauses are subjected to index conflict verification, and the secondary retrieval vector is incrementally updated based on the index conflict verification result. The search matching is updated based on the updated secondary search vector.

8. The method for accurate retrieval of insurance terms based on incremental training of a large model as described in claim 1, characterized in that, The step of using the structured query vector to perform retrieval matching based on the secondary retrieval vector and establishing retrieval matching results includes: Obtain the account's historical search records and configure the account trust level based on the historical search records; Obtain user search feedback on the search matching results, and establish a first correction strategy based on the user search feedback. The first correction strategy is an account search correction strategy. A second correction strategy is configured based on the user search feedback and the account trust level. The second correction strategy is a general correction strategy. Search matching optimization management is performed based on the first correction strategy and the second correction strategy.

9. The method for accurate retrieval of insurance clauses based on incremental training of a large model as described in claim 8, characterized in that, The user search feedback for obtaining search matching results includes: Based on the user's search feedback, anomaly levels are matched to establish matching results; An anomaly warning is generated based on the matching results, and an anomaly is reported.

10. A precise insurance clause retrieval system based on incremental training of a large model, characterized in that: The system is used to implement the precise insurance clause retrieval method based on large model incremental training as described in any one of claims 1-9, the system comprising: The perception hierarchy construction module is used to construct the insurance terms database and then perform perception parsing of the insurance terms database through the terms hierarchy parser to establish the perception hierarchy of general provisions, insurance liability, exclusions of liability, and claims procedures. A semantic tagging construction module is used to establish semantic tags through named entity recognition and semantic role labeling under the constraints of the perception level. The second-level retrieval vector construction module uses the perception level and the semantic tags to construct a second-level retrieval vector. The second-level retrieval vector includes a first-level retrieval vector constructed based on structural localization and a second-level retrieval vector constructed based on the first-level retrieval vector and sentence localization. The structured query vector construction module is used to call the intent recognition channel to perform intent recognition matching under the calibration template after receiving the user's input segment, and to build a structured query vector. The retrieval matching result construction module uses the structured query vector to perform retrieval matching based on the secondary retrieval vector and establishes retrieval matching results; The incremental update execution module is used to record the search matching results as query feedback, extract the query feedback history and new clause data at a preset period, and use the extracted results as incremental data to perform incremental updates of secondary search vectors and search matching.

Images