Insurance entity identifying and linking method fused with RAG technology

Abstract

The invention discloses an insurance entity identifying and linking method fused with an RAG technology. The method comprises the following steps: collecting standard names of insurance products and services; generalizing the standard name based on a cue project, and establishing a standard name-generalization abbreviation one-to-many mapping table; establishing a keyword retrieval index for the mapping table, and constructing a semantic vector index; receiving a user question, and calling the large-scale language model to extract original words in the user question based on the prompt word engineering; accurately matching the original word with the mapping table, and returning a corresponding standard name if the original word is hit; when the candidate short name is not hit, respectively triggering keyword retrieval and semantic vector retrieval based on the original words to obtain a candidate short name set; and assembling the question sentences, the original words and the candidate abbreviation set of the user into cue words according to a preset template, and inputting the cue words into a large-scale language model for semantic matching and rule verification to obtain a standard product name. According to the method, high recall and high accuracy are both considered, the maintenance cost of an artificial dictionary is reduced, and the field controllability and auditing performance are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of intelligent question answering technology, specifically relating to a method for identifying and linking insurance entities that integrates RAG technology. Background Technology

[0002] Information retrieval and natural language understanding in the insurance industry have received widespread attention in recent years. Businesses such as insurance Q&A, intelligent customer service, and product search heavily rely on the accurate identification and linking of insurance product names and related technical terms in text (i.e., mapping users' colloquial or non-standard expressions to the company's internal standard product identifiers). However, the insurance sector's characteristics, including diverse terminology, highly similar product names, and rapid business iteration, pose significant challenges to entity recognition and linking.

[0003] In early practice, the insurance industry generally adopted rule-based methods for entity recognition and linking: regular expression matching, keyword dictionaries, and multi-pattern matching based on Aho-Corasick (AC) automata were used to identify predefined entities in text, and manually maintained mapping tables were used to map the recognition results to standard product names or numbers. This method is simple to implement, highly interpretable, and performs stably within the controlled dictionary coverage, with an overall accuracy of approximately 80%–90% (typically estimated at around 85%). However, the main drawbacks of this approach are: the dictionary and rules require continuous manual maintenance to accommodate new products, new abbreviations, and colloquial expressions; it is difficult to cover long-tail products and diverse user expressions; and the maintenance costs are high, with response speed constrained by the maintenance schedule.

[0004] Subsequently, supervised learning methods have been widely adopted in the field of natural language processing to improve entity recognition accuracy. Sequence labeling models, represented by BiLSTM-CRF, have improved entity recognition rates by using contextual features and conditional random field decoding. Although sequence labeling has significant advantages in boundary detection and type classification, the entity linking stage still relies on a pre-built complete knowledge base or knowledge graph to accurately map the identified original words. For the insurance industry, building a knowledge graph covering all products, versions, aliases, and misspellings is costly, and the frequent launch of new products and the flexibility of naming rules make it difficult for static knowledge graphs to cover long-term, dynamic long-tail items, thus limiting the performance of the linking stage and further increasing maintenance costs. For example, in the prior art, Chinese patent CN112463944A discloses a retrieval-based intelligent question answering method and device based on multi-model fusion. The method includes: preprocessing the user's question text, performing word segmentation and named entity recognition to obtain the word segmentation result and entity tag of the question; passing the word segmentation result and entity tag through a rule classifier to match the rules in the rule base, and returning the standard question corresponding to the rule and the answer to the standard question to the user.

[0005] However, this method has the following limitations: 1. The rule / dictionary method relies on a manually constructed and continuously maintained rule base and alternative name table. When faced with frequent updates to insurance products, colloquial abbreviations used in marketing, and various typos, a large amount of manual work is required for additions and corrections, making it difficult to achieve automated, low-cost long-term maintenance. 2. Methods based on manual rules or static knowledge graphs can usually only cover high-frequency products and known alternative names, with low coverage of long-tail products, newly launched products, and rare expressions. This results in a large number of queries failing to hit or mishitting in actual business, affecting user experience and self-service rates.

[0006] In summary, existing technologies have significant shortcomings in terms of coverage, semantic fault tolerance, maintainability, controllability, and auditability, making it difficult to meet the entity recognition and linking needs of the insurance industry, which involves thousands of similarly named products, frequent iterations, and compliant auditability. Therefore, there is an urgent need for an entity recognition method that can balance high recall and high accuracy, and possess a closed loop of semantic retrieval and rule verification, to solve the aforementioned technical problems and improve the robustness and operability of industry applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an insurance entity identification and linking method that integrates RAG technology.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a method for identifying and linking insurance entities using RAG technology, comprising the following steps:

[0010] Collect and verify the standard names of insurance products and services, generalize the standard names based on prompt word engineering, generate multiple generalized abbreviations, and establish a one-to-many mapping table between standard names and generalized abbreviations;

[0011] A keyword retrieval index is established for the mapping table, and the generalized abbreviations and standard names are vectorized to construct a semantic vector index. At the same time, a list of easily confused products and corresponding differentiation rules are established.

[0012] Receive user questions and use a large-scale language model to extract the original words from the user questions based on the prompt words.

[0013] The original word is matched precisely with the mapping table. If a match is found, the corresponding standard name is returned and the process ends.

[0014] When no match is found, keyword retrieval and semantic vector retrieval are triggered based on the original word, and the two recall results are merged and deduplicated to obtain a set of candidate abbreviations;

[0015] The user's question, the original word, the set of candidate abbreviations, the list of easily confused products, and the corresponding differentiation rules are assembled into prompt words according to a predetermined template and input into a large-scale language model for semantic matching and rule verification, and the final matching standard product name is output.

[0016] Furthermore, the standard name of the insurance product and service refers to the product or service name officially released and uniquely identified by the insurance company or industry regulatory agency, used to distinguish different types of insurance, different versions and different coverage. The standard name includes at least one or a combination of the following: product series name, insurance category, version information, coverage period, insurance liability and additional conditions.

[0017] Furthermore, the standard name is generalized based on the prompt word engineering to generate multiple generalized abbreviations, and a one-to-many mapping table of standard name and generalized abbreviation is established, specifically including:

[0018] The standard names of the insurance products and services are parsed to extract at least one of the following: subject name E, insurance type T, version information V, coverage period P, insurance liability R, and additional conditions A.

[0019] Based on the prompt word project, a large-scale language model is invoked to combine the extracted fields according to predetermined rules, generating a generalized abbreviation set S containing all key information. g The set of generalized abbreviations includes:

[0020] S g ={E,E+T,E+V,E+V+T,E+T+P,N}

[0021] Where N represents the standard name, the main name refers to the name of the insurance company, insurance product series or sub-series, the insurance category refers to the type of insurance product, and the version information refers to the specific version number, year or generation identifier of the insurance product.

[0022] During the generation of generalized abbreviations, ensure that each abbreviation contains at least the main name;

[0023] Establish a set S of standard names and precise generalized abbreviations. g The mapping relationship.

[0024] Furthermore, the step of establishing a keyword retrieval index for the mapping table and vectorizing the generalized abbreviations and standard names to construct a semantic vector index specifically includes:

[0025] Set the standard name and its corresponding set of generalized abbreviations S g ={s1,s2,…,s n Extract the data as indexed fields, where each generalized abbreviation is s. i It must include at least the subject name E;

[0026] The index fields are keyword-encoded to generate an inverted index for fast keyword matching and candidate filtering;

[0027] For the set S of standard names and generalized abbreviations g Vectorization is performed, using a vectorized model f. emb (·) Map each abbreviation to a vector:

[0028]

[0029] Where d is the vector dimension;

[0030] The vectorized v i A mapping relationship is established with the corresponding standard name, and stored as a semantic vector index to support semantic similarity-based retrieval.

[0031] Furthermore, the list of easily confused products includes: a set of insurance products with highly similar standard names, where the product names in each set partially overlap or are similar in terms of subject name, version information, or insurance category;

[0032] Establish corresponding differentiation rules for each set of easily confused products, which are used to accurately match the original words and candidate abbreviations entered by the user during the entity linking stage;

[0033] The distinction rules include attribute matching priority, version information distinction, insurance type distinction and subject name verification, which are used to guide the large model to select the most matching standard product name when there is ambiguity in the candidate products.

[0034] In actual question-and-answer or retrieval processes, when there are multiple easily confused products in the candidate abbreviation set, the list of easily confused products and the corresponding differentiation rules are called for verification to ensure that the final output standard product name is unique and accurate.

[0035] Furthermore, the step of using a large-scale language model to extract original words from user questions based on prompt words specifically includes:

[0036] Based on the user's input question, the role, task, and constraints of the large model are clarified through prompt words, guiding the model to identify proper nouns contained in the question;

[0037] The identified proper nouns include insurance product names, value-added service names, insurance types, APP names, channel platforms, functional modules, benefit names, and activity names, etc.

[0038] Each identified proper noun is labeled with its type information, and a structured entity candidate list is output, including the original word name and its corresponding category;

[0039] The following rules are followed during the identification process:

[0040] a) Only extract the proper nouns that explicitly appear in the question, without making inferences or associations;

[0041] b) Perform special processing on easily confused fields such as version information and additional insurance to ensure accurate extraction of the main product name;

[0042] c) Supports the extraction of multiple proper nouns in a question, and processes connected words or abbreviations according to the maximum matching principle;

[0043] The output structured entity candidates include optimized questions and lists of proper nouns, which are used for subsequent keyword matching, semantic retrieval, and entity linking.

[0044] Furthermore, the precise matching of the original word with the mapping table specifically includes:

[0045] The original words extracted from the user's question are compared one by one with the standard names and their generalized abbreviations in the standard name-generalized abbreviation mapping table;

[0046] Matching rules include:

[0047] a) When the original term is exactly the same as the standard name or generalized abbreviation, it will be hit directly;

[0048] b) Supports handling common product aliases and typos, mapping them to corresponding standard names;

[0049] c) Strictly match key information such as version information, additional insurance, and entity name to distinguish easily confused products;

[0050] When the original word matches a unique standard name in the mapping table, the standard name is returned directly as the matching result, and subsequent retrieval steps are terminated; when the original word does not match the mapping table, subsequent keyword retrieval and semantic vector retrieval are triggered.

[0051] Furthermore, the keyword retrieval specifically includes:

[0052] Keyword retrieval based on the original words includes:

[0053] The original word q extracted from the user's question is matched with the generalized abbreviation s stored in the mapping table. i Perform BM25 matching and calculate the relevance score for each candidate document:

[0054]

[0055] Where, score(q,s) i ) represents the original word q and the candidate generalization abbreviation s. iThe relevance score, IDF(t) represents the inverse document frequency of term t, where t represents a single term in the original term q, and f(t, s) is the relevance score. i ) indicates that term t is abbreviated as s in candidate generalization. i The frequency of occurrence in, k1 represents the BM25 term frequency adjustment parameter, b represents the document length normalization parameter, |s i | indicates the length of the candidate generalized abbreviation, and avgdl indicates the average length of all generalized abbreviations in the mapping table;

[0056] The inverse document frequency (IDF)(t) is calculated using the following formula:

[0057]

[0058] Where N represents the total number of generalized abbreviations in the mapping table; n(t) represents the number of generalized abbreviations containing term t;

[0059] Based on relevance scores, several candidate abbreviations are selected to form a keyword recall set.

[0060] Furthermore, the semantic vector retrieval specifically includes:

[0061] KNN-based semantic vector retrieval based on original words includes:

[0062] The original word q extracted from the user's question is input into the vectorized model f. emb (·) can be converted to a vector representation:

[0063] v q =f emb (q)

[0064] Among them, v q This is the original word vector representation;

[0065] Call the semantic vector index to retrieve the expression related to v from the index. q The most similar generalized abbreviations s i ∈S g The similarity between the original word vector and the candidate abbreviation vector is calculated using cosine similarity:

[0066]

[0067] Where ||·|| represents the L2 norm of the vector; The original word vector represents v. q With candidate abbreviation vector Similarity;

[0068] Sort by similarity in descending order, and select the top few generalized abbreviations to form a semantic vector recall set.

[0069] Recall the semantic vector set C vec With keyword recall set C kw Merging and deduplicating yields the final set of candidate abbreviations, C. final .

[0070] Furthermore, the process of assembling user questions, original words, candidate abbreviation sets, easily confused product lists, and corresponding differentiation rules into prompt words according to a predetermined template, and inputting them into a large-scale language model for semantic matching and rule verification, outputting the final matched standard product name, specifically includes:

[0071] The user question Q, the set of original words q, and the set of candidate abbreviations C are used. final A list of easily confused products and corresponding differentiation rules are generated by piecing together a predefined template to create a prompt.

[0072] Input the prompt word Prompt into a large-scale language model LLM, and the model performs the following tasks:

[0073] 1) For the candidate abbreviation set C final Match the semantics of the user's question Q;

[0074] 2) Verify the rules based on the list of easily confused products and the differentiation rules to ensure that similar products can be accurately distinguished;

[0075] 3) Correct and map any abbreviations, alternative names, or spelling errors that may exist in the original word q;

[0076] 4) If there are multiple candidates for matching results, sort them according to the matching confidence and select the optimal match based on rule constraints;

[0077] Output the best match.

[0078] Compared with the prior art, the present invention has the following advantages:

[0079] (1) Existing rule-based regular expression and dictionary methods require a lot of manual maintenance. When faced with new products, new abbreviations and colloquial expressions, the coverage is insufficient, the maintenance cost is high, and the response is slow, resulting in a heavy long-term maintenance burden on the system and affecting the user experience. This invention achieves centralized management and automatic expansion of abbreviations by automatically generalizing standard names through a large language model and establishing a one-to-many mapping table between standard names and generalized abbreviations, and storing the mapping table as a searchable data structure. This technology reduces the dependence on manual maintenance and can quickly add new products and new abbreviations to the database through a rule-based generalization process and model generation, thereby significantly reducing the cost of manual maintenance and improving the timeliness and coverage of the mapping database, ensuring that the system can respond in a timely manner to product iterations and changes in colloquial expressions in the market.

[0080] (2) Traditional entity link dictionaries are often simple thesaurus or isolated alias sets, with limited generalization and abbreviations often lacking key information (e.g., omitting version or type of insurance), leading to missed links or incorrect matching; entity recognition often uses traditional sequence labeling models (such as CRF, BiLSTM) or large models driven by simple prompt words. These methods are prone to missing key information (such as version, type of insurance) when extracting original words of insurance products, leading to subsequent link deviations; this invention designs a complex prompt word engineering in the entity recognition stage to drive a large-scale language model, explicitly requiring that the extracted original words must contain core elements such as subject, version, and type of insurance, and performs structured annotation and question optimization on the output; this technology significantly improves the completeness and accuracy of original word extraction, avoids link errors caused by incomplete original words, provides reliable and structured input for subsequent retrieval and linking, thereby improving the accuracy of the end-to-end system.

[0081] (3) This invention combines the precision of statistical keyword recall with the generalization ability of vector semantic recall by employing a hybrid retrieval strategy (BM25+KNN) and merging and deduplicating the recall results. Simultaneously, after merging, candidates are initially screened based on a list of easily confused products and differentiation rules. This approach balances precise matching at the keyword level with relevance matching at the semantic level, resulting in more comprehensive and relevant candidate words. It reduces the omission of candidate words due to the limitations of a single recall method, providing a higher-quality candidate set for subsequent entity linking. This solution introduces a verification process, a constraint list (such as requiring the inclusion of insurance type), and an easily confused list (such as common confused product pairs) through complex prompt word design, guiding the large model to make accurate judgments. This significantly reduces the error rate of entity linking, especially solving the confusion problem in the insurance field caused by similar product names and subtle differences in versions / insurance types, improving the reliability and professionalism of the linking results.

[0082] (4) Although sequence labeling models improve the accuracy of the recognition end, entity links still heavily rely on a complete knowledge base or knowledge graph, which has high construction and maintenance costs and is difficult to update dynamically. This invention prioritizes precise matching between the identified original words and the mapping table, and triggers hybrid retrieval when no match is found, and uses a limited candidate pool for LLM judgment, thereby minimizing the dependence on static knowledge. This technical approach avoids complete dependence on a complete static knowledge graph, significantly reduces the cost of knowledge base construction and maintenance, and improves adaptability and link accuracy for newly launched or long-tail products.

[0083] (5) The end-to-end generation of general large models is prone to illusions and lacks domain-specific knowledge, and the generated results are uncontrollable in terms of compliance and auditing. Before candidate generation, this invention forms a controlled candidate set through precise matching and hybrid retrieval, and inputs the candidates, easily confused lists and distinguishing rules as prompt words into the LLM for semantic matching and rule verification, thereby limiting the output range of the model and adding business rule verification. This measure effectively suppresses model illusions, improves the interpretability and auditability of the output, and makes the final matching both semantically capable and business compliant. Attached Figure Description

[0084] Figure 1 This is a flowchart of the insurance entity identification and linking method according to an embodiment of the present invention;

[0085] Figure 2 This is a model diagram of the insurance entity identification and linking method according to an embodiment of the present invention. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0087] Example 1:

[0088] This embodiment specifically provides a method for insurance entity identification and linking that integrates RAG technology, such as... Figure 1 As shown, it includes the following steps:

[0089] Step S1: Collect and verify the standard names of insurance products and services;

[0090] The standard name of an insurance product or service refers to the product or service name that is officially released and uniquely identified by the insurance company or industry regulatory agency. It is used to distinguish different types of insurance, different versions, and different coverage. The standard name includes at least one or a combination of the following: product series name, insurance category, version information, coverage period, insurance liability, and additional conditions.

[0091] Step S2: Based on the prompt word engineering, generalize the standard name to generate multiple generalized abbreviations, and establish a one-to-many mapping table between the standard name and the generalized abbreviation, specifically including:

[0092] Parse the standard names of insurance products and services to extract at least one of the following: subject name E, insurance type T, version information V, coverage period P, insurance liability R, and additional conditions A.

[0093] Based on the prompt word project, a large-scale language model is invoked to combine the extracted fields according to predetermined rules, generating a generalized abbreviation set S containing all key information. g The generalized abbreviation set includes:

[0094] S g ={E,E+T,E+V,E+V+T,E+T+P,N}

[0095] Where N represents the standard name, the main name refers to the name of the insurance company, insurance product series or sub-series, the insurance category refers to the type of insurance product, and the version information refers to the specific version number, year or generation identifier of the insurance product.

[0096] During the generation of generalized abbreviations, ensure that each abbreviation contains at least the main name;

[0097] Establish a set S of standard names and precise generalized abbreviations. g The mapping relationship.

[0098] Step S3: Establish a keyword retrieval index for the mapping table, and vectorize the generalized abbreviations and standard names to construct a semantic vector index. Simultaneously, establish a list of easily confused products and corresponding differentiation rules, specifically including:

[0099] Set the standard name and its corresponding set of generalized abbreviations S g ={s1,s2,…,s n Extract the data as indexed fields, where each generalized abbreviation is s. i It must include at least the subject name E;

[0100] Keyword processing is performed on the indexed fields to generate inverted indexes for fast keyword matching and candidate filtering;

[0101] For the set S of standard names and generalized abbreviations g Vectorization is performed, using a vectorized model f. emb (·) Map each abbreviation to a vector:

[0102]

[0103] Where d is the vector dimension;

[0104] The vectorized v i A mapping relationship is established with the corresponding standard name, and stored as a semantic vector index to support semantic similarity-based retrieval.

[0105] The list of easily confused products includes: a set of insurance products with highly similar standard names, where the product names in each set partially overlap or are similar in terms of subject name, version information, or insurance category;

[0106] Establish corresponding differentiation rules for each set of easily confused products, which are used to accurately match the original words and candidate abbreviations entered by the user during the entity linking stage;

[0107] The differentiation rules include attribute matching priority, version information differentiation, insurance type differentiation, and entity name verification, which are used to guide the large model to select the most matching standard product name when there is ambiguity in the candidate products.

[0108] In actual question-and-answer or retrieval processes, when there are multiple easily confused products in the candidate abbreviation set, the list of easily confused products and the corresponding differentiation rules are called for verification to ensure that the final output standard product name is unique and accurate.

[0109] Step S4: Receive the user's question and, based on the prompt word engineering, call a large-scale language model to extract the original words from the user's question. Specifically, this includes:

[0110] Based on the user's input question, the role, task, and constraints of the large model are clarified through prompt words, guiding the model to identify proper nouns contained in the question;

[0111] The identified proper nouns include insurance product names, value-added service names, insurance types, APP names, channel platforms, functional modules, benefit names, and activity names, etc.

[0112] Each identified proper noun is labeled with its type information, and a structured entity candidate list is output, including the original word name and its corresponding category;

[0113] The following rules are followed during the identification process:

[0114] a) Only extract the proper nouns that explicitly appear in the question, without making inferences or associations;

[0115] b) Perform special processing on easily confused fields such as version information and additional insurance to ensure accurate extraction of the main product name;

[0116] c) Supports the extraction of multiple proper nouns in a question, and processes connected words or abbreviations according to the maximum matching principle;

[0117] The output structured entity candidates include optimized questions and lists of proper nouns, which are used for subsequent keyword matching, semantic retrieval, and entity linking.

[0118] Step S5: Perform an exact match between the original word and the mapping table. If a match is found, return the corresponding standard name and the process ends. This includes:

[0119] The original words extracted from the user's question are compared one by one with the standard names and their generalized abbreviations in the standard name-generalized abbreviation mapping table;

[0120] Matching rules include:

[0121] a) When the original term is exactly the same as the standard name or generalized abbreviation, it will be hit directly;

[0122] b) Supports handling common product aliases and typos, mapping them to corresponding standard names;

[0123] c) Strictly match key information such as version information, additional insurance, and entity name to distinguish easily confused products;

[0124] When the original word matches a unique standard name in the mapping table, the standard name is returned directly as the matching result, and subsequent retrieval steps are terminated; when the original word does not match the mapping table, subsequent keyword retrieval and semantic vector retrieval are triggered.

[0125] Step S6: When no match is found, trigger keyword retrieval and semantic vector retrieval based on the original word respectively, and merge and deduplicate the two recall results to obtain a set of candidate abbreviations;

[0126] Keyword search, specifically including:

[0127] Keyword retrieval based on the original words includes:

[0128] The original word q extracted from the user's question is matched with the generalized abbreviation s stored in the mapping table. i Perform BM25 matching and calculate the relevance score for each candidate document:

[0129]

[0130] Where, score(q,s) i ) represents the original word q and the candidate generalization abbreviation s. i The relevance score, IDF(t) represents the inverse document frequency of term t, where t represents a single term in the original term q, and f(t, s) is the relevance score. i ) indicates that term t is abbreviated as s in candidate generalization. i The frequency of occurrence in, k1 represents the BM25 term frequency adjustment parameter, b represents the document length normalization parameter, |s i | indicates the length of the candidate generalized abbreviation, and avgdl indicates the average length of all generalized abbreviations in the mapping table;

[0131] The inverse document frequency (IDF)(t) is calculated using the following formula:

[0132]

[0133] Where N represents the total number of generalized abbreviations in the mapping table; n(t) represents the number of generalized abbreviations containing term t;

[0134] Based on relevance scores, several candidate abbreviations are selected to form a keyword recall set.

[0135] Semantic vector retrieval, specifically including:

[0136] KNN-based semantic vector retrieval based on original words includes:

[0137] The original word q extracted from the user's question is input into the vectorized model f. emb (·) can be converted to a vector representation:

[0138] v q =f emb (q)

[0139] Among them, v q This is the original word vector representation;

[0140] Call the semantic vector index to retrieve the expression related to v from the index. q The most similar generalized abbreviations s i ∈S g The similarity between the original word vector and the candidate abbreviation vector is calculated using cosine similarity:

[0141]

[0142] Where ||·|| represents the L2 norm of the vector; The original word vector represents v. q With candidate abbreviation vector Similarity;

[0143] Sort by similarity in descending order, and select the top few generalized abbreviations to form a semantic vector recall set.

[0144] Recall the semantic vector set C vec With keyword recall set C kw Merging and deduplicating yields the final set of candidate abbreviations, C. final .

[0145] Step S7: Assemble the user question, original words, candidate abbreviation set, easily confused product list, and corresponding differentiation rules into prompt words according to a predetermined template, and input them into a large-scale language model for semantic matching and rule verification. Output the final matched standard product name, specifically including:

[0146] The user question Q, the set of original words q, and the set of candidate abbreviations C are used. final A list of easily confused products and corresponding differentiation rules are generated by piecing together a predefined template to create a prompt.

[0147] Input the prompt word Prompt into a large-scale language model LLM, and the model performs the following tasks:

[0148] 1) For the candidate abbreviation set C final Match the semantics of the user's question Q;

[0149] 2) Verify the rules based on the list of easily confused products and the differentiation rules to ensure that similar products can be accurately distinguished;

[0150] 3) Correct and map any abbreviations, alternative names, or spelling errors that may exist in the original word q;

[0151] 4) If there are multiple candidates for matching results, sort them according to the matching confidence and select the optimal match based on rule constraints;

[0152] Output the best match.

[0153] Example 2:

[0154] This invention primarily serves question-and-answer scenarios. The overall technical solution will be introduced from two parts: offline data preparation and online question processing, such as... Figure 2 As shown:

[0155] Offline data preparation primarily involves collecting and organizing relevant information about the products the system needs to support, including product names and insurance type numbers. This involves automatically generalizing product abbreviations using a large model, including both precise and indexed generalization. The generalized data will then be entered into the platform's Elasticsearch database for subsequent retrieval. To ensure the accuracy and completeness of the data, the collected information undergoes rigorous verification and review.

[0156] The online question processing section employs a three-stage cascaded architecture:

[0157] 1. Entity Recognition Stage: This stage involves original word extraction. Through specific prompt word engineering, a large model is used to directly extract the original words from the question. To address situations where insurance product names are confused with other types of proper nouns such as value-added services, activities, and benefits, the entity recognition stage extracts various proper nouns contained in the question.

[0158] 2. Search Enhancement Stage: First, the original words in the user's question are precisely matched, i.e., rapid matching is performed using precise and generalized product abbreviation data. If no match is found in the whitelist, the search stage begins. Then, the Elasticsearch multi-path recall method (KNN+BM25) is used to recall candidate product words related to the original words.

[0159] 3. Entity Linking Stage: Through specific prompt word engineering, a large model is used to match and analyze the original word and candidate words, outputting a standard product name. Even if the product name entered by the user has slight differences or variations, the system can accurately identify and match the corresponding standard product name.

[0160] The following will describe the process in detail from four stages: offline data preparation, entity recognition, retrieval enhancement, and entity linking.

[0161] (I) Offline Data Preparation

[0162] Based on a large-scale language model, we utilize prompt word engineering techniques to generalize standard product names and generate effective abbreviations. We abstracted standard insurance product names, defining them as consisting of one or more combinations of the following: subject + version + insurance type + category. Product name generalization needs to accurately capture the core features of the product name, ensuring that the generalized abbreviation concisely expresses the original meaning while avoiding confusion with other products. To this end, we adopted two different generalization strategies: precise generalization and indexed generalization.

[0163] 1. Precise generalization

[0164] The precise generalization strategy focuses on preserving key information from the product name to the greatest extent possible. It meticulously breaks down and reorganizes the product name, extracting crucial information such as the main body, version, insurance type, and category. Then, based on this information, a concise and clear abbreviation is generated. This set of abbreviations constitutes the product terminology database for precise matching. This database accurately reflects the core characteristics of the product, facilitating quick user identification and understanding, and highlighting the product's uniqueness and accuracy.

[0165] The core requirements include:

[0166] 1) All generalized abbreviations contain key information, such as subject, version, type of insurance, and category;

[0167] 2) The standard name is a single, generalized abbreviation;

[0168] 3) For example: The abbreviation for "Chang Xiang Ban (Chuan Shi 2024) Whole Life Insurance" can be generalized as "Chang Xiang Ban Chuan Shi 2024 Whole Life Insurance" or "Chang Xiang Ban Chuan Shi 2024".

[0169] 1) LLM model

[0170] Using the Qianwen 2.5-72B model, a precise generalized abbreviation of the insurance standard name is generated based on the standard name.

[0171] 2) Model Post-processing

[0172] An in-memory index is built from the precise generalized lexicon. When used in product Q&A scenarios, it enables fast and accurate matching of original product terms. At the same time, the precise generalized lexicon is merged into the index generalized data (merged and deduplicated) and stored in Elasticsearch for subsequent retrieval and recognition by the system.

[0173] 2. Index generalization

[0174] Our index generalization strategy focuses on comprehensive coverage and flexible matching of insurance product names. By building an index containing a large number of insurance product names and their abbreviations, we can automatically match and convert different names entered by users, thus providing them with more accurate and richer question-and-answer results. This strategy greatly improves the platform's applicability and flexibility, enabling it to better meet the diverse needs of users.

[0175] The core requirements include:

[0176] 1) All generalized abbreviations include the subject;

[0177] 2) Version attributes are retained: such as 2024, Supreme Edition, etc.

[0178] 3) The main body is a generalized abbreviation on its own.

[0179] For example, the abbreviation for "Chang Xiang Ban (Chuan Shi 2024) Whole Life Insurance" can be generalized as "Chang Xiang Ban" or "Chang Xiang Ban Chuan Shi".

[0180] 1) LLM model

[0181] Using the Qianwen 2.5-72B model, a generalized abbreviation of the insurance standard name is generated based on the standard name.

[0182] 2) Model Post-processing

[0183] The generalized data is stored uniformly using DB and ES, and when used in product Q&A scenarios, the original product terms are retrieved and matched.

[0184] 3. Product aliases and misspellings

[0185] 1) Product Alias

[0186] Some products have special abbreviations in actual business scenarios. These abbreviations are widely used in daily operations. The system collects and organizes common product aliases. Examples of aliases:

[0187] alternative name Product Standard Name Enjoy Million Family Edition Enjoy Million-Dollar 2025 Medical Insurance (Adjustable Rates) A pretty good life, personal development edition A Good Life (Career Excellence) Annuity Insurance (Participating)

[0188] 2) Handling typos

[0189] In practical applications, user-entered product names may contain typos or spelling errors. To improve the system's fault tolerance and user experience, we designed a typo handling mechanism. This mechanism can automatically identify and correct common typos in user input and convert them into correct product names, thereby ensuring that the platform can provide accurate answers.

[0190] A typo library for products was created, storing common product typos as generalized terms. When a user enters a product name with a typo, the standard product name can be retrieved normally.

[0191] Example of a misspelled word:

[0192] Misspellings Product Standard Name Anxiang Million Enjoy Million-Dollar 2025 Medical Insurance (Adjustable Rates) Anheng Million Enjoy Million-Dollar 2025 Medical Insurance (Adjustable Rates)

[0193] 4. Generalized data storage

[0194] Generalized data storage includes DB database storage and ES index (vector) storage, specifically including:

[0195] 4.1DB Database Storage

[0196] After generalization, the data forms a dataset with a correspondence of "product standard name: product generalization term set," which is stored in a dedicated table within the DB database. The core structure of the table is as follows:

[0197] property meaning product_name Product Standard Name simple_name Product generalization terms status state

[0198] 4.2 ES Index (Vector) Storage

[0199] We vectorize the generalized data, generate an index, and store it in Elasticsearch (ES). This allows the system to quickly retrieve and match the generalized product names from the ES database. We have also optimized the storage of the generalized data to ensure its efficiency and accuracy.

[0200] 1) Vector data generation

[0201] Using the vectorization model bge-base-zh-v1.5, text blocks are mapped into 768-dimensional vectors for vectorization; the original text content and the vectorized data are then stored in the Elasticsearch vector database.

[0202] 2) Indexed (vector) storage

[0203] The storage structure is shown in the table below. To facilitate storage and subsequent queries, we have set the keyword attribute for the key fields product_name and simple_name.

[0204] Field Name type illustrate Id DataType.text Automatically generate vector IDs product_name DataType.text Standard name of storage simple_name DataType.text Store a single generic name simple_name_vector DataType.dense_vector Store the result of generalized name vectorization

[0205] (II) Entity Recognition Stage

[0206] Original word extraction design and implementation

[0207] 1. Design Scheme

[0208] By employing large-scale language model-based prompt word engineering techniques, we can accurately identify various proper nouns contained in user questions, such as specific insurance product names, value-added services offered, different types of insurance, related app names, and application modules. Based on this identification, we can further clarify the type classification corresponding to each proper noun.

[0209] 2. Design concept of prompt words

[0210] 2.1 Overall Task Objectives

[0211] Main task: Identify and extract proper nouns (product names, service names, etc.) from questions, and optimize questions to clarify semantics when necessary.

[0212] Processing scope: Only questions with "product" or "value-added service" as the subject will be processed; if the subject is not a product / service (such as APP name, download operation, etc.), no extraction will be performed, and only the original question structure will be returned.

[0213] 2.2 Extraction Rule Classification and Key Point Summary

[0214] (1) Rules for extracting insurance product names

[0215] 1) It consists of the following combination of fields:

[0216] The format should be: Insurance Company Abbreviation + Product Name + Version (e.g., "Supreme Edition") + Insurance Type (e.g., "Critical Illness Insurance") + Type (e.g., "Participating"); Example formats include: "Additional + Name + Version + Insurance Type", etc.

[0217] 2) Only extract the explicitly mentioned names; do not make any associations or additions.

[0218] 3) Eliminate misidentified items: such as general descriptions that are not product names, such as "double version" or "universal insurance".

[0219] 4) When "additional responsibility" appears, only the main product name is extracted, and the responsibility name is not extracted.

[0220] (2) Rules for extracting value-added service names

[0221] Strictly limited to the list of services within the specified whitelist.

[0222] If there is no accurate match, no additional associations are allowed (e.g., "emergency rescue services" cannot be associated with "lifelong companionship customer club").

[0223] Mentioning specific keywords will be mapped to standard service names, such as:

[0224] "LOHAS Mall" category → Extracted as Centenarian Residence

[0225] "Health and Wellness Community" etc. → Extracted as Taibao Home

[0226] "Worry-Free Butler" → Output based on Taibao blueprint

[0227] Other category name extraction rules:

[0228] Extract insurance product categories (such as "critical illness insurance", "supplementary insurance", etc.), APP name, channel platform, functional module, benefit name, and activity name. Each category has a designated keyword whitelist.

[0229] Special Case Identification Rules:

[0230] Multi-word recognition: Supports the identification of multiple proper words in a question, which need to be extracted separately. Even if multiple words are not clearly separated (such as "Jinsheng Wuyou Baby is good"), they should still be extracted into two product names based on the maximum match.

[0231] Additional product identification: If the question contains the word "additional", determine whether it is "additional insurance" based on the structure. If so, it is also extracted as a product.

[0232] Abbreviation Standardization Recognition: For industry terms or abbreviations (such as "Geyi", "Duanjianxian", "Ruyika"), the full name needs to be output as the standard word (tword) for extraction.

[0233] Optimize question rules:

[0234] Question optimization goal: Output optimized questions for further model processing. Optimization must clearly define the insurance product while maintaining semantic clarity and not altering the original meaning.

[0235] Historical Dialogue Reference Mechanism

[0236] Triggering condition: When the current question does not contain any specific proper nouns.

[0237] Rule: If no proper nouns are recognized in the current question, use the most recent valid extraction result.

[0238] The current question already contains proper nouns → only the results extracted in this instance will be used.

[0239] (III) Search Enhancement Stage

[0240] 1. Process Description

[0241] 1) Based on the data after the [Entity Recognition] process, if the original word type is "product", first perform a quick matching filter on the original word to determine if it is in the exact matching terminology. If it is, directly match the corresponding full name and terminate the subsequent steps; otherwise, proceed to the search enhancement stage.

[0242] 2) Use the extracted original product keywords and ES product data for mixed search recall;

[0243] 3) Deduplicate the product abbreviations retrieved by the mixed search to obtain a list of candidate abbreviations;

[0244] 2. Generalized data recall

[0245] In the process of generalized data recall, we adopted a hybrid retrieval strategy. Specifically, based on the extracted original product terms, we conducted recall searches in the ES database using BM25 and KNN to retrieve all generalized data related to the product, and then merged the results of the two recalls to remove duplicates.

[0246] 1) BM25 recall

[0247] BM25 retrieval (BestMatching25) is a statistical keyword matching model based on a probabilistic retrieval model normalized by term frequency, inverse document frequency, and document length. It is used to calculate the relevance of a query to a generalized thesaurus. Specifically, based on the extracted original product terms, it performs a retrieval recall against the `simple_name` field in the index data, retrieving the 20 most relevant product terms.

[0248] 2) KNN recall

[0249] This method quickly finds the document vector most similar to the query vector, achieving approximate nearest neighbor search based on semantic similarity. Specifically, the first step involves mapping the extracted product keywords to a 768-dimensional vector using the bge-base-zh-v1.5 vector model, obtaining the product keyword embedding data. The second step uses this embedding data and the simple_name_vector field in the index data for retrieval and recall, utilizing Elasticsearch integration capabilities for querying. Cosine similarity is used to calculate the similarity; the 20 most similar product keywords are then retrieved.

[0250] (iv) Entity Linking Phase

[0251] This stage assembles the recalled candidate abbreviation list, user questions, and prompts, inputs them into the large model for final entity linking, and outputs a list of full names of insurance products that may match.

[0252] 1. Overcoming difficulties

[0253] The accuracy of matching cannot be guaranteed, such as matching primary and supplementary insurance, attribute matching principles, matching easily confused products, independent identification of special names, and identification of special semantics. Example:

[0254]

[0255]

[0256] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0257] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying and linking insurance entities integrating RAG technology, characterized in that, Includes the following steps: Collect and verify the standard names of insurance products and services; Based on the prompt word engineering, the standard name is generalized to generate multiple generalized abbreviations, and a one-to-many mapping table of standard name and generalized abbreviation is established. A keyword retrieval index is established for the mapping table, and the generalized abbreviations and standard names are vectorized to construct a semantic vector index. At the same time, a list of easily confused products and corresponding differentiation rules are established. Receive user questions and use a large-scale language model to extract the original words from the user questions based on the prompt words. The original word is matched precisely with the mapping table. If a match is found, the corresponding standard name is returned and the process ends. When no match is found, keyword retrieval and semantic vector retrieval are triggered based on the original word, and the two recall results are merged and deduplicated to obtain a set of candidate abbreviations; The user's question, the original word, the set of candidate abbreviations, the list of easily confused products, and the corresponding differentiation rules are assembled into prompt words according to a predetermined template and input into a large-scale language model for semantic matching and rule verification, and the final matching standard product name is output.

2. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The standard name of the insurance product and service refers to the product or service name officially released and uniquely identified by the insurance company or industry regulatory agency, used to distinguish different types of insurance, different versions and different coverage. The standard name includes at least one or a combination of the following: product series name, insurance category, version information, coverage period, insurance liability and additional conditions.

3. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The standard name is generalized based on prompt word engineering to generate multiple generalized abbreviations, and a one-to-many mapping table of standard name and generalized abbreviation is established, specifically including: The standard names of the insurance products and services are parsed to extract at least one of the following: subject name E, insurance type T, version information V, coverage period P, insurance liability R, and additional conditions A. Based on the prompt word project, a large-scale language model is invoked to combine the extracted fields according to predetermined rules, generating a generalized abbreviation set S containing all key information. g The set of generalized abbreviations includes: S g ={E,E+T,E+V,E+V+T,E+T+P,N} Where N represents the standard name, the main name refers to the name of the insurance company, insurance product series or sub-series, the insurance category refers to the type of insurance product, and the version information refers to the specific version number, year or generation identifier of the insurance product. During the generation of generalized abbreviations, ensure that each abbreviation contains at least the main name; Establish a set S of standard names and precise generalized abbreviations. g The mapping relationship.

4. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The step of establishing a keyword retrieval index for the mapping table and vectorizing the generalized abbreviations and standard names to construct a semantic vector index specifically includes: Set the standard name and its corresponding set of generalized abbreviations S g ={s1,s2,…,s n Extract the data as indexed fields, where each generalized abbreviation is s. i It must include at least the subject name E; The index fields are keyword-encoded to generate an inverted index for fast keyword matching and candidate filtering; For the set S of standard names and generalized abbreviations g Vectorization is performed, using a vectorized model f. emb (·) Map each abbreviation to a vector: Where d is the vector dimension; The vectorized v i A mapping relationship is established with the corresponding standard name, and stored as a semantic vector index to support semantic similarity-based retrieval.

5. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The list of easily confused products includes: a set of insurance products with highly similar standard names, where the product names in each set partially overlap or are similar in terms of main body name, version information, or insurance type; Establish corresponding differentiation rules for each set of easily confused products, which are used to accurately match the original words and candidate abbreviations entered by the user during the entity linking stage; The distinction rules include attribute matching priority, version information distinction, insurance type distinction and subject name verification, which are used to guide the large model to select the most matching standard product name when there is ambiguity in the candidate products. In actual question-and-answer or retrieval processes, when there are multiple easily confused products in the candidate abbreviation set, the list of easily confused products and the corresponding differentiation rules are called for verification to ensure that the final output standard product name is unique and accurate.

6. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The process of extracting original words from user questions using a large-scale language model based on prompt words specifically includes: Based on the user's input question, the role, task, and constraints of the large model are clarified through prompt words, guiding the model to identify proper nouns contained in the question; The identified proper nouns include insurance product names, value-added service names, insurance types, APP names, channel platforms, functional modules, benefit names, and activity names, etc. Each identified proper noun is labeled with its type information, and a structured entity candidate list is output, including the original word name and its corresponding category; The following rules are followed during the identification process: a) Only extract the proper nouns that explicitly appear in the question, without making inferences or associations; b) Perform special processing on easily confused fields such as version information and additional insurance to ensure accurate extraction of the main product name; c) Supports the extraction of multiple proper nouns in a question, and processes connected words or abbreviations according to the maximum matching principle; The output structured entity candidates include optimized questions and lists of proper nouns, which are used for subsequent keyword matching, semantic retrieval, and entity linking.

7. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The precise matching of the original words with the mapping table specifically includes: The original words extracted from the user's question are compared one by one with the standard names and their generalized abbreviations in the standard name-generalized abbreviation mapping table; Matching rules include: a) When the original term is exactly the same as the standard name or generalized abbreviation, it will be hit directly; b) Supports handling common product aliases and typos, mapping them to corresponding standard names; c) Strictly match key information such as version information, additional insurance, and entity name to distinguish easily confused products; When the original word matches a unique standard name in the mapping table, the standard name is returned directly as the matching result, and subsequent retrieval steps are terminated; when the original word does not match the mapping table, subsequent keyword retrieval and semantic vector retrieval are triggered.

8. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The keyword retrieval specifically includes: Keyword retrieval based on the original words includes: The original word q extracted from the user's question is matched with the generalized abbreviation s stored in the mapping table. i Perform BM25 matching and calculate the relevance score for each candidate document: Where, score(q,s) i ) represents the original word q and the candidate generalization abbreviation s. i The relevance score, IDF(t) represents the inverse document frequency of term t, where t represents a single term in the original term q, and f(t, s) is the relevance score. i ) indicates that term t is abbreviated as s in candidate generalization. i The frequency of occurrence in, k1 represents the BM25 term frequency adjustment parameter, b represents the document length normalization parameter, |s i | indicates the length of the candidate generalized abbreviation, and avgdl indicates the average length of all generalized abbreviations in the mapping table; The inverse document frequency (IDF)(t) is calculated using the following formula: Where N represents the total number of generalized abbreviations in the mapping table; n(t) represents the number of generalized abbreviations containing term t; Based on relevance scores, several candidate abbreviations are selected to form a keyword recall set.

9. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The semantic vector retrieval specifically includes: KNN-based semantic vector retrieval recall, including: The original word q extracted from the user's question is input into the vectorized model f. emb (·) can be converted to a vector representation: v q =f emb (q) Among them, v q This is the original word vector representation; Call the semantic vector index to retrieve the expression related to v from the index. q The most similar generalized abbreviations s i ∈S g The similarity between the original word vector and the candidate abbreviation vector is calculated using cosine similarity: Where ||·|| represents the L2 norm of the vector; The original word vector represents v. q With candidate abbreviation vector Similarity; Sort by similarity in descending order, and select the top few generalized abbreviations to form a semantic vector recall set. Recall the semantic vector set C vec With keyword recall set C kw Merging and deduplicating yields the final set of candidate abbreviations, C. final .

10. The insurance entity identification and linking method integrating RAG technology according to claim 1, characterized in that, The process involves assembling user questions, original words, a set of candidate abbreviations, a list of easily confused products, and corresponding differentiation rules according to a predetermined template into prompt words, which are then input into a large-scale language model for semantic matching and rule verification. The final matched standard product name is then output, specifically including: The user question Q, the set of original words q, and the set of candidate abbreviations C are used. final A list of easily confused products and corresponding differentiation rules are generated by piecing together a predefined template to create a prompt. Input the prompt word Prompt into a large-scale language model LLM, and the model performs the following tasks: 1) For the candidate abbreviation set C final Match the semantics of the user's question Q; 2) Verify the rules based on the list of easily confused products and the differentiation rules to ensure that similar products can be accurately distinguished; 3) Correct and map any abbreviations, alternative names, or spelling errors that may exist in the original word q; 4) If there are multiple candidates for matching results, sort them according to the matching confidence and select the optimal match based on rule constraints; Output the best match.

Images