Text intention recognition method and device based on large language model, equipment and medium

Abstract

The invention relates to the technical field of intelligent decision making, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a text intention recognition method, device and equipment based on a large language model and a medium. When the intention category is a general insurance service, performing semantic analysis on the dialogue text to obtain a semantic feature vector; determining a general intention recognition result according to the semantic feature vector and a general intention label mapping library; when the intention category is a target insurance service, determining a target insurance type corresponding to the target insurance service through an insurance type label mapping library; loading a low-rank matrix corresponding to the target insurance type, and performing parameter adjustment on the attention layer according to the low-rank matrix to obtain an insurance type intention recognition model; and performing fusion semantic reasoning by using the insurance type intention recognition model to obtain a fine-grained intention recognition result of the dialogue text. The accuracy of a specific insurance type intention recognition result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology, and in particular to a text intent recognition method, device, equipment and medium based on a large language model. Background Art

[0002] Driven by the development of artificial intelligence (AI), intent recognition technology has evolved from early rule-based matching and statistical models to intelligent recognition centered on deep learning. Neural network-based models (such as CNN and RNN), trained on large amounts of labeled data, can efficiently capture semantic features of language and significantly improve recognition accuracy. This technology is now widely used in scenarios such as intelligent customer service, voice assistants, and chatbots, becoming a key technical support for achieving efficient human-computer interaction and understanding user needs in the digital transformation of various industries.

[0003] In the healthcare sector, intent recognition technology often fails to accurately identify specific insurance types due to independently trained models. For example, when a user inquires about "reimbursement for targeted cancer drugs," the independent health insurance model may misclassify it as a general hospitalization claim due to limited training data. Similarly, when faced with the "application process for rare disease specialty medications," the pediatric critical illness insurance model may be unable to accurately identify specific needs due to insufficient samples, requiring multiple user interactions. These issues reduce service efficiency and user experience, resulting in low accuracy in intent recognition results for specific insurance types.

[0004] In the fintech sector, intent recognition technology often suffers from discrepancies in the identification of specific insurance types due to independent model training. For example, when a user inquires about "gold futures hedging strategies," a futures trading model may misidentify it as general gold financial consulting due to insufficient data coverage. Similarly, when faced with the "cross-border trade financing approval process," a supply chain finance model may be unable to accurately interpret complex requirements due to a limited sample size, resulting in lengthy interactions. These scenarios increase user operational costs and hinder service efficiency, leading to low accuracy in intent recognition results for specific insurance types.

[0005] Existing intent recognition technology has several significant problems in the insurance field. First, models are usually trained separately for each type of insurance, which makes model maintenance complex and costly, and leads to serious repeated consumption of computing resources. If a universal model is used, it is difficult to accurately capture specific intents due to the large differences in business logic between different types of insurance, resulting in low accuracy in intent recognition results for specific types of insurance. Summary of the Invention

[0006] The present invention provides a text intent recognition method, device, equipment and medium based on a large language model to solve the technical problem of low accuracy of intent recognition results for specific insurance types.

[0007] In a first aspect, a method for identifying text intent based on a large language model is provided, comprising:

[0008] Obtaining a target user's conversation text, performing intent analysis on the conversation text using an intent classifier of a preset large language model, and obtaining an intent category corresponding to the conversation text;

[0009] When the intention category is general insurance business, performing semantic analysis on the conversation text to obtain a semantic feature vector;

[0010] Determining a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library;

[0011] When the intention category is a preset target insurance business, the target insurance type corresponding to the target insurance business is determined through a preset insurance type label mapping library;

[0012] Loading a low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, adjusting parameters of the attention layer in the large language model according to the low-rank matrix, and obtaining an insurance type intent recognition model;

[0013] The insurance type intention recognition model is used to perform fusion semantic reasoning on the dialogue text to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

[0014] In a second aspect, a text intent recognition device based on a large language model is provided, comprising:

[0015] An intent category acquisition module is used to acquire the target user's conversation text, perform intent analysis on the conversation text using an intent classifier based on a preset large language model, and obtain the intent category corresponding to the conversation text;

[0016] a semantic feature vector analysis module, configured to perform semantic analysis on the conversation text to obtain a semantic feature vector when the intent category is general insurance business;

[0017] A general intent recognition module, configured to determine a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library;

[0018] A target insurance type analysis module is configured to determine the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library when the intention category is a preset target insurance business;

[0019] An insurance type intention recognition model acquisition module is used to load the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, and adjust the parameters of the attention layer in the large language model according to the low-rank matrix to obtain the insurance type intention recognition model;

[0020] The fine-grained intention recognition result reasoning module is used to use the insurance type intention recognition model to perform fusion semantic reasoning on the dialogue text to obtain the fine-grained intention recognition result of the dialogue text for the target insurance type.

[0021] In a third aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for text intent recognition based on a large language model are implemented.

[0022] In a fourth aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned text intent recognition method based on a large language model are implemented.

[0023] In the scheme implemented by the above-mentioned text intent recognition method, device, equipment and medium based on the large language model, the conversation text of the target user can be obtained through the client, and the intent classifier of the preset large language model is used to perform intent analysis on the conversation text to obtain the intent category corresponding to the conversation text; when the intent category is general insurance business, the conversation text is semantically analyzed to obtain a semantic feature vector; the general intent recognition result of the conversation text is determined based on the semantic feature vector and the preset general intent label mapping library; when the intent category is the preset target insurance business, the target insurance type corresponding to the target insurance business is determined through the preset insurance type label mapping library; the low-rank matrix corresponding to the target insurance type is loaded according to the preset dynamic loading mechanism, and the parameters of the attention layer in the large language model are adjusted according to the low-rank matrix to obtain an insurance type intention recognition model; the insurance type intention recognition model is used to perform fusion semantic reasoning on the conversation text to obtain the fine-grained intent recognition result of the conversation text for the target insurance type. In the present invention, the large language model and low-rank matrix technology are innovatively combined to construct a unified infrastructure, and the low-rank matrix is ​​dynamically adapted to the needs of different insurance types. The system processes general intents based on a pre-trained large model. It then trains lightweight, low-rank matrices for specific insurance types, such as auto and health insurance. During inference, the corresponding matrices are dynamically loaded based on the conversation content, integrating general capabilities with the specific characteristics of insurance types. This solution significantly reduces model maintenance costs and improves computing resource utilization. Dynamic loading and combined inference optimize response speed. New insurance types only require generating the corresponding low-rank matrix model, offering strong scalability and improving recognition accuracy through continuous learning, balancing efficiency and flexibility. This solution can address the issue of low accuracy in intent recognition results for specific insurance types. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0025] Figure 1 1 is a schematic diagram of an application environment of a method for identifying text intent based on a large language model in one embodiment of the present invention;

[0026] Figure 2 1 is a flow chart of a method for identifying text intent based on a large language model in one embodiment of the present invention;

[0027] Figure 3 yes Figure 2 A schematic flow chart of a specific implementation of step S1;

[0028] Figure 4 yes Figure 2 A schematic flow chart of a specific implementation of step S3;

[0029] Figure 5 1 is a structural diagram of a text intent recognition device based on a large language model in one embodiment of the present invention;

[0030] Figure 6 is a structural diagram of a computer device in one embodiment of the present invention;

[0031] Figure 7 FIG. 2 is another structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] The text intention recognition method based on the large language model provided by the embodiment of the present invention can be applied in the following fields: Figure 1In an application environment, the client communicates with the server through a network. The server can obtain the target user's conversation text through the client, perform intent analysis on the conversation text using the intent classifier of the preset large language model, and obtain the intent category corresponding to the conversation text; when the intent category is general insurance business, perform semantic analysis on the conversation text to obtain a semantic feature vector; determine the general intent recognition result of the conversation text based on the semantic feature vector and the preset general intent label mapping library; when the intent category is the preset target insurance business, determine the target insurance type corresponding to the target insurance business through the preset insurance type label mapping library; load the low-rank matrix corresponding to the target insurance type according to the preset dynamic loading mechanism, adjust the parameters of the attention layer in the large language model according to the low-rank matrix, and obtain an insurance type intention recognition model; use the insurance type intention recognition model to perform fusion semantic reasoning on the conversation text to obtain a fine-grained intent recognition result of the conversation text for the target insurance type. In the present invention, the large language model and low-rank matrix technology are innovatively combined to construct a unified infrastructure, and dynamically adapt to the needs of different insurance types through the low-rank matrix. The system processes general intents based on a pre-trained large model, and then trains lightweight low-rank matrices for auto insurance, health insurance, etc. During reasoning, the corresponding matrices are dynamically loaded according to the conversation content to achieve a combination of general capabilities and insurance characteristics. This solution significantly reduces model maintenance costs and improves computing resource utilization. It optimizes response speed through dynamic loading and combinatorial reasoning. When adding new insurance types, only the corresponding low-rank matrix model needs to be generated. It has strong scalability and can improve recognition accuracy through a continuous learning mechanism, taking into account both efficiency and flexibility. Among them, the client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented with an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0034] See also Figure 2 As shown, Figure 2 A flowchart of a method for identifying text intent based on a large language model provided in an embodiment of the present invention includes the following steps:

[0035] S1. Obtain a conversation text of a target user, perform intent analysis on the conversation text using an intent classifier of a preset large language model, and obtain an intent category corresponding to the conversation text.

[0036] In an embodiment of the present invention, the target users refer to users who use the two applications "Good Car Owner" and "Good Life" to consult, purchase insurance products or seek insurance services. They interact with the system through the intelligent dialogue system in the application; the dialogue text is the text content entered by these users during the communication process with the intelligent dialogue system, such as specific query text asking about the car insurance claim process, health insurance premiums, etc.

[0037] In detail, the system obtains the target user's conversation text through the intelligent dialogue system embedded in the "Good Car Owner" and "Good Life" applications. When users use these two applications for insurance consultation, purchase or service seeking, the text content entered in the dialogue interaction interface (such as asking about the method of reporting car insurance, health insurance coverage conditions, etc.) will be collected by the system in real time as the original data for subsequent intent recognition.

[0038] For example, in a medical scenario, the target user inputs "How long is the waiting period for the critical illness insurance I bought" to the intelligent dialogue system through the APP, and the system collects the text information in real time as the dialogue text; or the user sends "How do I apply for compensation for a scratch accident" on the platform. The text content such as consultations, complaints, and inquiries actively input by such users are all obtained by the system as dialogue text data for intent recognition.

[0039] For example, in a financial scenario, the target user inputs "Will credit card overdue affect credit rating" into the intelligent dialogue system through the mobile banking APP, and the system obtains the text content as dialogue text in real time; or the user inquires "How to open a cross-border transfer service" in the online banking, and the text information such as account inquiries, business processing, policy consultation, etc. submitted by such users in the financial service scenario are all collected by the system as dialogue text data for intent recognition.

[0040] In an embodiment of the present invention, the preset large language model refers to a pre-trained model with the ability to understand and generate natural language, which can be used as a basic framework to process conversation text; the intent classifier is a component in the model used to identify the core purpose or requirements of the conversation text, and extracts and classifies text features through an algorithm; the intent category is a preset classification result of the conversation text intent, such as general insurance business and preset target insurance business.

[0041] In the embodiment of the present invention, referring to Figure 3 As shown, the intent classifier using the preset large language model performs intent analysis on the conversation text to obtain the intent category corresponding to the conversation text, including:

[0042] S31, performing semantic vectorization processing on the conversation text to obtain a text semantic vector;

[0043] S32, using the intent classifier of the preset large language model to perform feature screening on the text semantic vector to obtain target intent features;

[0044] S33: performing intent matching on the target intent feature and the intent tag feature in a preset intent tag library to obtain the intent category corresponding to the conversation text.

[0045] In detail, the text semantic vector is the conversion of the conversation text into a computer-understandable numerical vector through semantic vectorization, which is used to represent the semantic information of the text; the target intent feature is the key features related to the text intent extracted after the text semantic vector is screened using the intent classifier in the preset large language model. These features can effectively reflect the core intent expressed by the text.

[0046] Specifically, semantic vectorization of conversation text involves converting the text into a numerical vector form that can be understood by a computer through an algorithm, thereby obtaining a text semantic vector; the text semantic vector is processed using the intent classifier in the preset large language model, and key information related to the intent is filtered out from the vector based on the rules and patterns set during model training to form the target intent feature.

[0047] Furthermore, the preset intent tag library is a pre-built database containing various intent tags and their corresponding features, which is used to store standard feature templates of different intent categories; the intent tag feature is the representative semantic feature or pattern corresponding to each intent tag in the intent tag library, which is used to characterize the core attributes of the intent category.

[0048] Furthermore, the text semantic vector obtained by semantic vectorization of the conversation text is used to filter out the target intent features through the intent classifier in the preset large language model. These features are then compared with the features corresponding to each intent label in the pre-built intent label library. Based on matching rules such as feature similarity, it is determined which intent label feature the conversation text best matches, thereby obtaining the intent category corresponding to the conversation text.

[0049] For example, in a medical scenario, a patient asks "What should I do if I feel dizzy after taking diabetes medication?" The system first performs semantic vectorization on the dialogue text and converts it into a numerical vector containing semantic information such as medication, diabetes, and dizziness. Then, the intent classifier in the preset large language model is used to filter out target intent features such as "diabetes medication", "dizziness symptoms", and "seeking solutions" from the text semantic vector. These features are then compared with the intent label features such as "medication consultation" and "symptom consultation" in the preset intent label library, and it is found that they are highly matched with the label feature of "adverse reaction consultation after medication". Finally, the intent category corresponding to the dialogue text is determined to be "adverse reaction consultation of medication" so that the system can provide targeted medical advice.

[0050] For example, in a financial scenario, a user asks, "When will the principal of my structured deposit arrive after maturity?" The system first vectorizes the conversation text semantics, generating a text semantic vector containing elements such as "structured deposit," "maturity," and "principal arrival." The intent classifier within the large language model then filters out features related to the liquidation of wealth management products, forming a target intent feature. This is then matched with the label feature for "financial product maturity redemption query" in the intent label library, ultimately classifying the intent as "inquiry on the arrival time of structured deposit principal upon maturity." The corresponding liquidation rules and arrival timeline instructions are then pushed.

[0051] S2. When the intention category is general insurance business, perform semantic analysis on the conversation text to obtain a semantic feature vector.

[0052] In an embodiment of the present invention, the general insurance business refers to a pre-set, generally applicable insurance business type, such as common basic business categories such as life insurance, property insurance, and health insurance; the semantic feature vector is obtained by performing semantic analysis on the conversation text, converting the semantic information contained in the text into a computer-processable numerical vector composed of several semantic feature dimensions, which is used to represent the semantic content of the text.

[0053] Specifically, when the intent category belongs to general insurance business, the system uses the infrastructure of a unified large language model and its extensive natural language understanding capabilities after pre-training on large-scale datasets to directly perform semantic analysis on the conversation text. The Transformer architecture and self-attention mechanism within the model capture the text context relationship, thereby converting the text semantic information into a numerical vector containing general intent features, namely a semantic feature vector.

[0054] For example, in a financial scenario, when a user asks "how to check a credit card bill", the pre-trained large language model is called based on the unified large language model infrastructure to perform semantic analysis on the conversation text. The model uses the Transformer architecture and self-attention mechanism to capture the contextual semantic relationship of "credit card bill query", and converts the text into a numerical vector containing semantic features such as account query and bill cycle, and then quickly matches the general business processing logic.

[0055] S3. Determine a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library.

[0056] In an embodiment of the present invention, the preset general intent label mapping library is a pre-constructed database or rule set that stores the correspondence between various semantic feature vectors and general intent labels, which includes standard intent labels extracted for scenarios such as insurance business, as well as the mapping logic between these labels and semantic features; the general intent recognition result is determined by matching and comparing the semantic feature vectors obtained by semantic analysis of the conversation text with the content in the mapping library, and then determining the specific insurance business intent category label corresponding to the conversation text, such as clear intent classification results such as consulting insurance products, applying for insurance, and claim inquiry.

[0057] In the embodiment of the present invention, referring to Figure 4 As shown, the determining of the general intent recognition result of the conversation text based on the semantic feature vector and the preset general intent label mapping library includes:

[0058] S41, performing normalization processing on the semantic feature vector to obtain a normalized feature vector;

[0059] S42, calculating the matching degree between the standardized feature vector and the label features in the preset general intent label mapping library;

[0060] S43, selecting the intention tags with a matching degree higher than a preset matching threshold as a tag candidate set;

[0061] S44. Determine a general intent recognition result of the conversation text according to the matching degree of each intent tag in the tag candidate set and a preset intent determination threshold.

[0062] In detail, the standardized feature vector is a vector obtained by normalizing the semantic feature vector through normalization and regularization to unify its numerical range and distribution, and eliminate the influence of differences in different feature dimensions; the label feature is the feature vector corresponding to each label in the preset general intent label mapping library, representing the semantic characteristics and patterns of the label. The matching degree is calculated by an algorithm to calculate the semantic similarity or fit between the standardized feature vector and the label features in the preset general intent label mapping library. It represents the close association between the semantics of the conversation text and each intent label in numerical form (such as cosine similarity value, Euclidean distance value, etc.).

[0063] Specifically, when standardizing the semantic feature vector, its numerical range is adjusted to a uniform scale through specific algorithms (such as Z-score standardization, minimum-maximum standardization, etc.) to eliminate the influence of dimensional differences in different feature dimensions, thereby obtaining a standardized feature vector; then, the standardized feature vector is compared with the label features corresponding to each label in the preset general intent label mapping library, and the degree of proximity between the two in the semantic space is calculated using algorithms such as cosine similarity and Euclidean distance to determine the degree of matching.

[0064] For example, in a medical scenario, a patient describes "I've been feeling dizzy and nauseous lately." The system performs semantic analysis on the text to obtain a semantic feature vector; the vector is standardized to eliminate the dimensional differences in symptom descriptions such as "dizziness" and "nausea" to obtain a standardized feature vector; the matching degree is calculated with label features such as "symptom consultation" and "neurology" in the general intent label mapping library, and it is found that the cosine similarity with the label "dizziness and nausea cause consultation" is the highest. Finally, the patient's intention is quickly located and relevant department consultation recommendations and preliminary health guidance are pushed.

[0065] Furthermore, the preset matching threshold is a pre-set matching standard for screening intent labels. Only when the matching degree between the intent label and the dialogue text is higher than the threshold can it be selected into the candidate set; the intent label is a standardized label that summarizes the intent expressed in the dialogue text, such as "consultation", "complaint", etc.; the label candidate set is a set of intent labels with a matching degree higher than the preset matching threshold screened out from all intent labels; the preset intent determination threshold is a pre-set standard for determining the final general intent recognition result. The specific intent of the dialogue text is determined by comparing the matching degree of each intent label in the label candidate set with the threshold.

[0066] Furthermore, from the matching calculation between the conversation text and each intent label, the intent labels whose matching degree exceeds a preset value (i.e., the preset matching threshold) are screened out to form a label candidate set. Then, based on the comparison result of the matching degree of each intent label in the label candidate set with the preset intent judgment threshold, the general intent recognition result of the conversation text is determined. That is, when the matching degree of a certain intent label reaches or exceeds the intent judgment threshold, it is used as the final recognition result.

[0067] S4. When the intention category is a preset target insurance business, determine the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library.

[0068] In an embodiment of the present invention, the preset target insurance business refers to a specific insurance business type that is preset; the preset insurance type label mapping library is a pre-established database or rule set for matching insurance businesses with insurance type labels; and the target insurance type is a specific insurance type determined by the mapping library that corresponds to the preset target insurance business.

[0069] In the embodiment of the present invention, when the intention category is a preset target insurance business, determining the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library includes:

[0070] Perform semantic analysis on the text content of the preset target insurance business to obtain business keywords;

[0071] Querying the business keyword from a preset insurance type label mapping library to obtain a candidate insurance type label;

[0072] Match and verify the candidate insurance type label with the preset label rule library to obtain the target insurance type label;

[0073] The target insurance type corresponding to the target insurance business is determined according to the mapping relationship of the target insurance type label in the insurance type label mapping library.

[0074] In detail, the text content refers to the specific textual expression corresponding to the preset target insurance business; the business keywords are words that can reflect the core characteristics of the business extracted from the text content through semantic analysis; the candidate insurance type labels are a set of possible matching insurance type labels obtained after searching the preset insurance type label mapping library based on the business keywords.

[0075] Specifically, semantic analysis is performed on the text content involved in the preset target insurance business (such as business descriptions, terms, etc.), and business keywords that can reflect the core characteristics of the business are extracted through natural language processing technology. These keywords are then used as indexes to perform matching queries in the preset insurance type label mapping library to obtain candidate insurance type labels associated with the keywords.

[0076] Furthermore, the preset label rule library is a set of pre-set rules for verifying candidate insurance type labels; the target insurance type label is a label that complies with the rules after the candidate insurance type label is matched and verified with the label rule library; the mapping relationship is the corresponding association relationship between the target insurance type label and the insurance type corresponding to the target insurance business in the insurance type label mapping library.

[0077] Furthermore, the candidate insurance type labels are matched and verified with the preset rules in the preset label rule library, and the labels that meet the rules are screened out as the target insurance type labels. Then, based on the corresponding relationship established between the target insurance type labels and the insurance type label mapping library, the target insurance type corresponding to the mapped target insurance business is found.

[0078] S5. Load the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, adjust the parameters of the attention layer in the large language model according to the low-rank matrix, and obtain an insurance type intention recognition model.

[0079] In an embodiment of the present invention, the preset dynamic loading mechanism refers to a technical design in which the system automatically identifies the type of insurance involved according to the content of the user conversation through an API interface or memory mapping method during the reasoning process, based on a lightweight loading framework (such as the model service component of PyTorch or TensorFlow), and loads the low-rank matrix model of the corresponding insurance type in real time for combinatorial reasoning; the low-rank matrix is ​​the core structure in low-rank matrix technology for adaptive fine-tuning of specific layers of the basic model. Its essence is to approximate the update of the original weight matrix as the product of two low-rank matrices through low-rank decomposition, so that while retaining the general capabilities of the model, only millions to tens of millions of parameters need to be trained to achieve efficient recognition of specific intentions of insurance types.

[0080] In the embodiment of the present invention, the step of loading the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism includes:

[0081] Determine the low-rank matrix identifier corresponding to the target insurance type according to the target insurance type and the preset insurance type matrix mapping rule;

[0082] Using a preset loading framework to perform index matching retrieval on the low-rank matrix identifier to obtain low-rank matrix data;

[0083] Performing structured analysis and calibration processing on the low-rank matrix data to obtain a standardized matrix loading package;

[0084] Dynamically memory-map the parameter address in the standardized matrix loading package to a preset memory alignment format to obtain a low-rank matrix corresponding to the target insurance type.

[0085] In detail, the preset insurance type matrix mapping rules are pre-set rules for establishing a corresponding relationship between the target insurance type and the low-rank matrix identifier; the low-rank matrix identifier is a symbol or number obtained after mapping the insurance type for identifying a specific low-rank matrix, which can be used as index key information; the preset loading framework is a pre-constructed architecture or mechanism for index matching and retrieval of low-rank matrix identifiers, which specifies the retrieval method and process; the low-rank matrix data is data related to the insurance type stored in the form of a low-rank matrix after retrieving the low-rank matrix identifier through the loading framework.

[0086] Specifically, according to the target insurance type and the preset insurance type matrix mapping rules, the low-rank matrix identifier corresponding to the target insurance type is determined by matching the characteristics or attributes of the target insurance type with the corresponding relationship preset in the mapping rules; then, using the preset loading framework, based on the index information contained in the low-rank matrix identifier, accurate matching retrieval is performed in the corresponding data storage structure or index system, and the low-rank matrix data corresponding to the identifier is obtained, thereby completing the determination and acquisition process from the target insurance type to the low-rank matrix data.

[0087] Furthermore, the standardized matrix loading package is a standardized data set obtained after structured parsing and calibration of low-rank matrix data, which contains matrix-related information in a standardized format and structure; the parameter address is the storage location identifier of each parameter in the standardized matrix loading package in the memory, which is used to locate and access specific data; the preset memory alignment format is a pre-set memory allocation rule, which specifies the starting address alignment method when data is stored in the memory.

[0088] Furthermore, structured parsing is performed on the low-rank matrix data, and the data structure is combed and calibrated according to preset rules to remove abnormal or erroneous information, forming a standardized matrix loading package with a unified format and standardized structure; then the parameter addresses in the standardized matrix loading package are dynamically memory mapped according to the preset memory alignment format (that is, the starting address alignment rule for data stored in the memory), so that the parameter addresses correspond to the memory space in a specific way, thereby generating the low-rank matrix required for the target insurance type.

[0089] In an embodiment of the present invention, the attention layer is the core component in the large language model responsible for calculating the association weights of each element of the input sequence and capturing semantic dependencies, and the attention mechanism is realized through matrix operations; the insurance type intention recognition model is a special model obtained by adjusting the parameters of the attention layer of the large language model using a low-rank matrix (such as injecting insurance type-specific parameters based on low-rank matrix technology), which can accurately identify user intentions, needs and semantic information related to insurance types from the text.

[0090] In an embodiment of the present invention, the parameter adjustment of the attention layer in the large language model according to the low-rank matrix to obtain the insurance type intention recognition model includes:

[0091] Performing dimension matching processing on the low-rank matrix according to the dimension parameters in the attention layer in the large language model to obtain a dimension matching matrix;

[0092] Superimposing and fusing the dimension matching matrix with the original parameter matrix in the attention layer of the large language model to obtain a parameter fusion matrix;

[0093] Utilizing a preset insurance type dialogue data set to configure parameters of the parameter fusion matrix, a target parameter matrix is ​​obtained;

[0094] The parameters of the attention layer in the large language model are reconstructed according to the target parameter matrix to obtain an insurance type intention recognition model.

[0095] In detail, the dimension parameters are the parameters used in the attention layer of the large language model to characterize the input and output dimensions and determine the dimension matching rules of matrix operations, such as the hidden layer dimension and the number of attention heads; the dimension matching matrix is ​​the matrix that adjusts and adapts the low-rank matrix according to the dimension parameters of the attention layer so that its dimension is consistent with the original parameter matrix; the original parameter matrix is ​​the initial weight parameter matrix in the attention layer of the large language model without fine-tuning by the low-rank matrix; the parameter fusion matrix is ​​a new matrix formed by superimposing and fusing the dimension matching matrix and the original parameter matrix in a preset manner, which not only retains the general capabilities of the basic model, but also incorporates the insurance type-specific intent recognition parameters.

[0096] Specifically, based on the dimensional parameters in the attention layer of the large language model, the row and column dimensions of the low-rank matrix are adjusted or completed to make them consistent with the dimensional structure of the original parameter matrix of the attention layer, forming a dimension matching matrix; then the dimension matching matrix and the original parameter matrix are superimposed according to preset rules, so that the insurance type-specific parameters contained in the low-rank matrix are fused with the general parameters of the basic model to generate a parameter fusion matrix that has both general language understanding capabilities and insurance type intention recognition capabilities.

[0097] Furthermore, the preset insurance type dialogue dataset is a pre-prepared dataset containing a large number of dialogue texts related to insurance types and corresponding intent labels, which is used for targeted training and parameter configuration of the parameter fusion matrix; the target parameter matrix is ​​the final parameter matrix obtained by using the insurance type dialogue dataset to train and optimize the parameter fusion matrix and adjust the parameter values. Its parameter values ​​are adapted to the requirements of the insurance type intent recognition task and can be used to reconstruct the attention layer of the large language model to form a dedicated model.

[0098] Furthermore, a preset insurance type dialogue dataset containing a large number of insurance type-related dialogue texts and corresponding intent labels is used to train and optimize the parameter fusion matrix. By adjusting the parameter values ​​in the matrix to adapt it to the insurance type intent recognition task, a target parameter matrix is ​​obtained. Based on the parameter values ​​in the target parameter matrix, the original parameters of the attention layer of the large language model are replaced or corrected to complete the parameter reconstruction, forming an insurance type intent recognition model with the ability to accurately identify the specific intent of an insurance type.

[0099] S6. Use the insurance type intention recognition model to perform fusion semantic reasoning on the dialogue text to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

[0100] In an embodiment of the present invention, the fine-grained intent recognition result refers to the system's ability to accurately capture the specific and detailed intentions regarding the target insurance type in the conversation text through the fusion of semantic reasoning of a large language model and low-rank matrix technology. For example, in a car insurance scenario, it can accurately identify specific operational inquiries such as "how to report car insurance", or in a health insurance scenario, it can accurately locate segmented needs such as "health insurance claims process", which can more deeply explore the details of user intentions than traditional models.

[0101] In the embodiment of the present invention, the method of performing fusion semantic reasoning on the dialogue text using the insurance type intention recognition model to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type includes:

[0102] Performing contextual semantic encoding on the conversation text according to the insurance intention recognition model to obtain a general insurance semantic feature vector;

[0103] Using a preset mixed-precision inference algorithm, performing insurance-type-specific enhancement calculation on the general insurance semantic feature vector to obtain a combined semantic feature;

[0104] Semantically matching the combined semantic features with a preset fine-grained intent rule library to obtain a fine-grained intent label for the target insurance type, and determining a confidence score corresponding to the fine-grained intent label;

[0105] The fine-grained intent label with the confidence score higher than the preset confidence threshold is used as the fine-grained intent recognition result of the target insurance type.

[0106] In detail, the general insurance semantic feature vector is obtained by contextual semantic encoding of the conversation text through the insurance type intention recognition model, and converting the text information into a numerical vector representation that can reflect the semantics of the general insurance field; the preset mixed precision inference algorithm is a pre-set algorithm that combines different precision calculation strategies, mainly through the combination of FP16 and FP32 to improve computing efficiency while ensuring the accuracy of the reasoning results; the combined semantic feature is the result obtained by using the mixed precision inference algorithm to perform insurance type specific enhancement calculation on the general insurance semantic feature vector. It strengthens the semantic features related to specific insurance types on the basis of general semantic features, making the features more targeted and discriminative, and facilitating the subsequent accurate identification of insurance type intentions.

[0107] Specifically, the insurance type intention recognition model is used to perform contextual semantic encoding on the conversation text, converting the text into a numerical vector containing semantic information in the general insurance field, namely the general insurance semantic feature vector. Then, with the help of a preset mixed-precision inference algorithm, the algorithm performs inference calculations on the general insurance semantic feature vector, enhancing the insurance type-related semantic features while ensuring computational efficiency, thereby obtaining a combined semantic feature that integrates general semantics and insurance type-specific semantics.

[0108] Furthermore, the preset fine-grained intent rule base is a pre-set database containing fine-grained intent judgment rules and semantic patterns in the insurance field, which is used to provide standards for semantic matching; the fine-grained intent label is a detailed identification of the specific intent of the target insurance type, representing different types of insurance business intentions; the confidence score is calculated through semantic matching, and is a numerical indicator that measures the degree of matching between the fine-grained intent label and the combined semantic features; the preset confidence threshold is a pre-set judgment standard, which is used to screen out fine-grained intent labels with sufficiently high credibility.

[0109] Furthermore, by semantically matching the combined semantic features with the fine-grained intent judgment rules and semantic patterns in the insurance field stored in the preset fine-grained intent rule library, the fine-grained intent label corresponding to the target insurance type is obtained, and the degree of matching between the label and the combined semantic features is calculated to determine the confidence score. The confidence score is then compared with a pre-set confidence threshold, and the fine-grained intent label above the threshold is used as the fine-grained intent recognition result of the target insurance type.

[0110] In addition, the system uses cloud-based model management systems (such as Kubernetes and Ping An Cloud) to implement version control, update deployment, and dynamic adjustment of the basic model and low-rank matrix model. It also uses online learning mechanisms to collect user interaction data, incrementally update the low-rank matrix model, and gradually fine-tune the basic model. The framework can automatically deploy model updates and monitor performance, triggering retraining when recognition effectiveness declines or user needs change. This achieves unified management and continuous optimization of the model, allowing the system to maintain efficiency and adaptability in a dynamic environment, quickly respond to new insurance types or business changes, improve scalability and stability, and continue to provide users with high-quality services.

[0111] As can be seen, the above solution innovatively combines large language models with low-rank matrix technology to build a unified infrastructure, dynamically adapting to the needs of different insurance types through low-rank matrices. The system processes general intents based on pre-trained large models, and then trains lightweight low-rank matrices for auto insurance, health insurance, etc. During inference, the corresponding matrices are dynamically loaded according to the content of the conversation, realizing the combination of general capabilities and insurance type characteristics. This solution significantly reduces model maintenance costs and improves computing resource utilization. It optimizes response speed through dynamic loading and combined reasoning. When adding new insurance types, only the corresponding low-rank matrix model needs to be generated. It has strong scalability and can also improve recognition accuracy through a continuous learning mechanism, taking into account both efficiency and flexibility.

[0112] It should be understood that the order of execution of the steps in the above embodiments does not necessarily mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0113] In one embodiment, a text intention recognition device based on a large language model is provided, and the text intention recognition device based on a large language model corresponds one-to-one to the text intention recognition method based on a large language model in the above embodiment. Figure 5 As shown, the text intent recognition device 100 based on a large language model includes an intent category acquisition module 101, a semantic feature vector analysis module 102, a general intent recognition module 103, a target insurance type analysis module 104, an insurance type intent recognition model acquisition module 105, and a fine-grained intent recognition result inference module 106. The functional modules are described in detail as follows:

[0114] Intent category acquisition module 101 is used to acquire a target user's conversation text, perform intent analysis on the conversation text using an intent classifier of a preset large language model, and obtain an intent category corresponding to the conversation text;

[0115] A semantic feature vector analysis module 102 is configured to perform semantic analysis on the conversation text to obtain a semantic feature vector when the intent category is general insurance business;

[0116] A general intent recognition module 103 is configured to determine a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library;

[0117] The target insurance type analysis module 104 is configured to determine the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library when the intention category is a preset target insurance business;

[0118] The insurance type intention recognition model acquisition module 105 is used to load the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, and adjust the parameters of the attention layer in the large language model according to the low-rank matrix to obtain the insurance type intention recognition model;

[0119] The fine-grained intention recognition result reasoning module 106 is used to perform fusion semantic reasoning on the dialogue text using the insurance type intention recognition model to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

[0120] In one embodiment, the intent category acquisition module 101, when performing intent analysis on the conversation text using an intent classifier using a preset large language model to obtain an intent category corresponding to the conversation text, is configured to:

[0121] Performing semantic vectorization processing on the conversation text to obtain a text semantic vector;

[0122] Using the intent classifier of the preset large language model to perform feature screening on the text semantic vector to obtain the target intent feature;

[0123] The target intent feature is matched with the intent tag feature in a preset intent tag library to obtain the intent category corresponding to the conversation text.

[0124] In one embodiment, the general intent recognition module 103, when determining the general intent recognition result of the conversation text based on the semantic feature vector and the preset general intent label mapping library, is configured to:

[0125] performing normalization processing on the semantic feature vector to obtain a normalized feature vector;

[0126] Calculating the matching degree between the standardized feature vector and the label features in the preset general intent label mapping library;

[0127] Selecting the intent tags whose matching degree is higher than a preset matching threshold as the tag candidate set;

[0128] The general intent recognition result of the conversation text is determined according to the matching degree of each intent tag in the tag candidate set and a preset intent determination threshold.

[0129] In one embodiment, when the intention category is a preset target insurance business, the target insurance type analysis module 104, when determining the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library, is configured to:

[0130] Perform semantic analysis on the text content of the preset target insurance business to obtain business keywords;

[0131] Querying the business keyword from a preset insurance type label mapping library to obtain a candidate insurance type label;

[0132] Match and verify the candidate insurance type label with the preset label rule library to obtain the target insurance type label;

[0133] The target insurance type corresponding to the target insurance business is determined according to the mapping relationship of the target insurance type label in the insurance type label mapping library.

[0134] In one embodiment, the insurance type intention recognition model acquisition module 105, when loading the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, is configured to:

[0135] Determine the low-rank matrix identifier corresponding to the target insurance type according to the target insurance type and the preset insurance type matrix mapping rule;

[0136] Using a preset loading framework to perform index matching retrieval on the low-rank matrix identifier to obtain low-rank matrix data;

[0137] Performing structured analysis and calibration processing on the low-rank matrix data to obtain a standardized matrix loading package;

[0138] Dynamically memory-map the parameter address in the standardized matrix loading package to a preset memory alignment format to obtain a low-rank matrix corresponding to the target insurance type.

[0139] In one embodiment, the insurance type intention recognition model acquisition module 105, when performing parameter adjustment on the attention layer in the large language model according to the low-rank matrix to obtain the insurance type intention recognition model, is further configured to:

[0140] Performing dimension matching processing on the low-rank matrix according to the dimension parameters in the attention layer in the large language model to obtain a dimension matching matrix;

[0141] Superimposing and fusing the dimension matching matrix with the original parameter matrix in the attention layer of the large language model to obtain a parameter fusion matrix;

[0142] Utilizing a preset insurance type dialogue data set to configure parameters of the parameter fusion matrix, a target parameter matrix is ​​obtained;

[0143] The parameters of the attention layer in the large language model are reconstructed according to the target parameter matrix to obtain an insurance type intention recognition model.

[0144] In one embodiment, the fine-grained intent recognition result reasoning module 106, when performing fusion semantic reasoning on the dialogue text using the insurance type intent recognition model to obtain a fine-grained intent recognition result of the dialogue text for the target insurance type, is configured to:

[0145] Performing contextual semantic encoding on the conversation text according to the insurance intention recognition model to obtain a general insurance semantic feature vector;

[0146] Using a preset mixed-precision inference algorithm, performing insurance-type-specific enhancement calculation on the general insurance semantic feature vector to obtain a combined semantic feature;

[0147] Semantically matching the combined semantic features with a preset fine-grained intent rule library to obtain a fine-grained intent label for the target insurance type, and determining a confidence score corresponding to the fine-grained intent label;

[0148] The fine-grained intent label with the confidence score higher than the preset confidence threshold is used as the fine-grained intent recognition result of the target insurance type.

[0149] The present invention provides a text intent recognition device based on a large language model, which innovatively combines the large language model with low-rank matrix technology to build a unified infrastructure, and dynamically adapts to the needs of different insurance types through low-rank matrices. The system processes general intents based on a pre-trained large model, and then trains lightweight low-rank matrices for auto insurance, health insurance, etc. During reasoning, the corresponding matrix is ​​dynamically loaded according to the content of the conversation to achieve a combination of general capabilities and insurance type characteristics. This solution significantly reduces model maintenance costs, improves computing resource utilization, optimizes response speed through dynamic loading and combinatorial reasoning, and only needs to generate the corresponding low-rank matrix model when adding a new insurance type. It has strong scalability and can also improve recognition accuracy through a continuous learning mechanism, taking into account both efficiency and flexibility.

[0150] For the specific limitations of the text intent recognition device based on a large language model, please refer to the limitations of the text intent recognition method based on a large language model above, which will not be repeated here. The various modules in the above-mentioned text intent recognition device based on a large language model can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0151] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external client via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the server side of a text intent recognition method based on a large language model.

[0152] In one embodiment, a computer device is provided. The computer device may be a client, and its internal structure diagram may be as follows: Figure 7As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps on the client side of a text intent recognition method based on a large language model.

[0153] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following steps are performed:

[0154] Obtaining a target user's conversation text, performing intent analysis on the conversation text using an intent classifier of a preset large language model, and obtaining an intent category corresponding to the conversation text;

[0155] When the intention category is general insurance business, performing semantic analysis on the conversation text to obtain a semantic feature vector;

[0156] Determining a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library;

[0157] When the intention category is a preset target insurance business, the target insurance type corresponding to the target insurance business is determined through a preset insurance type label mapping library;

[0158] Loading a low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, adjusting parameters of the attention layer in the large language model according to the low-rank matrix, and obtaining an insurance type intent recognition model;

[0159] The insurance type intention recognition model is used to perform fusion semantic reasoning on the dialogue text to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

[0160] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0161] Obtaining a target user's conversation text, performing intent analysis on the conversation text using an intent classifier of a preset large language model, and obtaining an intent category corresponding to the conversation text;

[0162] When the intention category is general insurance business, performing semantic analysis on the conversation text to obtain a semantic feature vector;

[0163] Determining a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library;

[0164] When the intention category is a preset target insurance business, the target insurance type corresponding to the target insurance business is determined through a preset insurance type label mapping library;

[0165] Loading a low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, adjusting parameters of the attention layer in the large language model according to the low-rank matrix, and obtaining an insurance type intent recognition model;

[0166] The insurance type intention recognition model is used to perform fusion semantic reasoning on the dialogue text to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

[0167] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can be found in the relevant descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0168] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0169] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0170] It should be noted that if software tools or components other than those of our company appear in the embodiments of this application, they are only used for illustration and do not represent actual use.

[0171] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A text intent recognition method based on a large language model, characterized in that: include: Obtaining a target user's conversation text, performing intent analysis on the conversation text using an intent classifier of a preset large language model, and obtaining an intent category corresponding to the conversation text; When the intention category is general insurance business, performing semantic analysis on the conversation text to obtain a semantic feature vector; Determining a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library; When the intention category is a preset target insurance business, determining the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library; Loading the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, adjusting the parameters of the attention layer of the large language model according to the low-rank matrix, and obtaining an insurance type intention recognition model; The insurance type intention recognition model is used to perform fusion semantic reasoning on the dialogue text to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type.

2. The text intent recognition method based on a large language model according to claim 1, characterized in that The intent classifier using the preset large language model performs intent analysis on the conversation text to obtain the intent category corresponding to the conversation text, including: Performing semantic vectorization processing on the conversation text to obtain a text semantic vector; Using the intent classifier of the preset large language model to perform feature screening on the text semantic vector to obtain the target intent feature; The target intent feature is matched with the intent tag feature in the preset intent tag library to obtain the intent category corresponding to the conversation text.

3. The text intent recognition method based on a large language model according to claim 1, characterized in that Determining the general intent recognition result of the conversation text according to the semantic feature vector and a preset general intent label mapping library includes: performing normalization processing on the semantic feature vector to obtain a normalized feature vector; Calculating the matching degree between the standardized feature vector and the label features in the preset general intent label mapping library; Selecting the intent tags with matching degrees higher than a preset matching threshold as the tag candidate set; The general intent recognition result of the conversation text is determined according to the matching degree of each intent tag in the tag candidate set and a preset intent determination threshold.

4. The text intent recognition method based on a large language model according to claim 1, wherein: When the intention category is a preset target insurance business, determining the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library includes: Perform semantic analysis on the text content of the preset target insurance business to obtain business keywords; Querying the business keyword from a preset insurance type label mapping library to obtain a candidate insurance type label; Match and verify the candidate insurance type label with the preset label rule library to obtain the target insurance type label; The target insurance type corresponding to the target insurance business is determined according to the mapping relationship of the target insurance type label in the insurance type label mapping library.

5. The text intent recognition method based on a large language model according to claim 1, wherein: The loading of the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism includes: Determine the low-rank matrix identifier corresponding to the target insurance type according to the target insurance type and the preset insurance type matrix mapping rule; Using a preset loading framework to perform index matching retrieval on the low-rank matrix identifier to obtain low-rank matrix data; Performing structured analysis and calibration processing on the low-rank matrix data to obtain a standardized matrix loading package; Dynamically memory-map the parameter address in the standardized matrix loading package to a preset memory alignment format to obtain a low-rank matrix corresponding to the target insurance type.

6. The text intent recognition method based on a large language model according to claim 1, characterized in that The parameter adjustment of the attention layer of the large language model according to the low-rank matrix to obtain the insurance type intention recognition model includes: Performing dimension matching processing on the low-rank matrix according to the dimension parameters in the attention layer of the large language model to obtain a dimension matching matrix; Superimposing and fusing the dimension matching matrix with the original parameter matrix in the attention layer to obtain a parameter fusion matrix; Utilizing a preset insurance type dialogue data set to configure parameters of the parameter fusion matrix, a target parameter matrix is ​​obtained; The parameters of the attention layer are reconstructed according to the target parameter matrix to obtain an insurance type intention recognition model.

7. The text intent recognition method based on a large language model according to claim 1, characterized in that The method of performing fusion semantic reasoning on the dialogue text using the insurance type intention recognition model to obtain a fine-grained intention recognition result of the dialogue text for the target insurance type includes: Performing contextual semantic encoding on the conversation text according to the insurance intention recognition model to obtain a general insurance semantic feature vector; Performing insurance-type-specific enhancement calculations on the general insurance semantic feature vector using a preset mixed-precision inference algorithm to obtain a combined semantic feature; Semantically matching the combined semantic features with a preset fine-grained intent rule base to obtain a fine-grained intent label for the target insurance type, and determining a confidence score corresponding to the fine-grained intent label; The fine-grained intent label with the confidence score higher than the preset confidence threshold is used as the fine-grained intent recognition result of the target insurance type.

8. A text intent recognition device based on a large language model, characterized in that: include: An intent category acquisition module is used to acquire the target user's conversation text, perform intent analysis on the conversation text using an intent classifier based on a preset large language model, and obtain the intent category corresponding to the conversation text; a semantic feature vector analysis module, configured to perform semantic analysis on the conversation text to obtain a semantic feature vector when the intent category is general insurance business; A general intent recognition module, configured to determine a general intent recognition result of the conversation text based on the semantic feature vector and a preset general intent label mapping library; A target insurance type analysis module is configured to determine the target insurance type corresponding to the target insurance business through a preset insurance type label mapping library when the intention category is a preset target insurance business; An insurance type intention recognition model acquisition module is used to load the low-rank matrix corresponding to the target insurance type according to a preset dynamic loading mechanism, and adjust the parameters of the attention layer in the large language model according to the low-rank matrix to obtain the insurance type intention recognition model; The fine-grained intention recognition result reasoning module is used to use the insurance type intention recognition model to perform fusion semantic reasoning on the dialogue text to obtain the fine-grained intention recognition result of the dialogue text for the target insurance type.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the text intent recognition method based on a large language model is implemented as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for text intent recognition based on a large language model as described in any one of claims 1 to 7 is implemented.

Images