Insurance intention recognition method and device, electronic equipment and storage medium

Abstract

The invention provides an insurance intention recognition method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence and the field of financial science and technology. The method comprises the steps of obtaining a sample communication text; performing parameter adjustment on a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model comprises a first feature extraction layer, a second feature extraction layer and a third feature extraction layer; selecting the second feature extraction layer and the third feature extraction layer to obtain a target feature extraction layer; replacing a second feature extraction layer and a third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain a target intention recognition model; obtaining a target communication text; and performing intention recognition on the target communication text through the target intention recognition model to obtain a target insurance intention. According to the invention, the insurance intention identification efficiency for the insurance scene can be improved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence technology and financial technology, and in particular to an insurance intent recognition method and device, electronic equipment, and storage medium. Background Art

[0002] In the field of financial technology, insurance companies can identify intent based on the communication text (such as chat records between two people) in the communication window of the instant messaging application (such as a chat interface or dialog box between two or more parties), and recommend insurance (such as car insurance, pet insurance, etc.) to customers.

[0003] In related technologies, large language models can be used to identify the intent of communication texts. However, large language models are often complex in structure and have many parameters, and the efficiency of intent recognition in insurance scenarios still has a lot of room for improvement. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose an insurance intention recognition method and apparatus, equipment, and medium that can improve the efficiency of insurance intention recognition for insurance scenarios.

[0005] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present application proposes an insurance intention identification method, the method comprising:

[0006] Get sample communication texts;

[0007] Adjusting parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer and a third feature extraction layer;

[0008] Performing feature extraction on the sample communication text through the first feature extraction layer to obtain a first intention feature vector, performing feature extraction on the first intention feature vector through the second feature extraction layer to obtain a second intention feature vector, and performing feature extraction on the second intention feature vector through the third feature extraction layer to obtain a third intention feature vector;

[0009] determining a feature extraction score according to the second intent feature vector and the third intent feature vector;

[0010] Selecting the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer;

[0011] The target feature extraction layer replaces the second feature extraction layer and the third feature extraction layer in the temporary large language model to obtain a target intent recognition model;

[0012] Obtain target communication text;

[0013] The target communication text is subjected to intent recognition through the target intent recognition model to obtain the target insurance intent.

[0014] In some embodiments, the feature extraction score includes a first extraction score of the second feature extraction layer and a second extraction score of the second feature extraction layer; and determining the feature extraction score according to the second intent feature vector and the third intent feature vector includes:

[0015] Performing intent classification on the second intention feature vector to obtain a second insurance intention;

[0016] Calculate the similarity between the second insurance intent and a preset label insurance intent to obtain a first extraction score of the second feature extraction layer;

[0017] Performing intent classification on the third intention feature vector to obtain a third insurance intention;

[0018] A similarity calculation is performed based on the third intention feature vector and the label insurance intention to obtain a second extraction score of the third feature extraction layer.

[0019] In some embodiments, extracting features from the sample communication text vector through the first feature extraction layer to obtain a first intention feature vector includes:

[0020] Through the first feature extraction layer, the sample communication text is vector-embedded to obtain a sample communication text vector;

[0021] Generate a query vector for the sample communication text vector through the first feature extraction layer to obtain an intent query vector;

[0022] By using the first feature extraction layer, the sample communication text vector is segmented to obtain at least two sample communication text sub-vectors;

[0023] Generate a key-value pair vector for each of the sample communication text sub-vectors through the first feature extraction layer to obtain a candidate intent key-value pair vector;

[0024] Through the first feature extraction layer, vector selection is performed on each of the candidate intent key-value pair vectors to obtain at least one target selected intent key-value pair vector;

[0025] By using the first feature extraction layer, an attention score is calculated for the intent query vector and each of the target selected intent key-value pair vectors to obtain a target attention score;

[0026] Through the first feature extraction layer, the attention scores are concatenated to obtain the first intention feature vector.

[0027] In some embodiments, the step of performing vector selection on each of the candidate intent key-value pair vectors to obtain at least one target selected intent key-value pair vector includes:

[0028] Calculate the importance of the intent query vector and the candidate intent key-value pair vector to obtain the importance corresponding to each candidate intent key-value pair vector;

[0029] Vector selection is performed on each of the candidate intent key-value pair vectors according to the importance to obtain at least one of the target selected intent key-value pair vectors.

[0030] In some embodiments, before concatenating the attention scores to obtain the first intention feature vector, the method further includes:

[0031] Updating the first intention feature vector includes:

[0032] Perform feature fusion according to the sample communication text vector and the first intention feature vector to obtain a first fused intention feature vector;

[0033] Normalizing the first fusion intention feature vector to obtain a second fusion intention feature vector;

[0034] Inputting the second fusion intention feature vector into a feedforward neural network for feature mapping to obtain a third fusion intention feature vector;

[0035] Perform feature fusion according to the first fusion intention feature vector and the third fusion intention feature vector to obtain a fourth fusion intention feature vector;

[0036] Normalizing the fourth fusion intention feature vector to obtain a fifth fusion intention feature vector;

[0037] Performing a linear transformation on the fifth fusion intention feature vector to obtain a sixth fusion intention feature vector;

[0038] Nonlinearly activate the sixth fused intention feature vector to obtain an updated first intention feature vector.

[0039] In some embodiments, the initial large language model includes a text prefix layer, a first initial feature extraction layer, a second initial feature extraction layer, and a third initial feature extraction layer;

[0040] The step of adjusting parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model includes:

[0041] Embedding the intention guidance text into the sample communication text through the text prefix layer to obtain the sample intention communication text;

[0042] Performing feature extraction on the sample intention communication text through the first initial feature extraction layer to obtain a first initial intention feature vector, performing feature extraction on the first initial intention feature vector through the second initial feature extraction layer to obtain a second initial intention feature vector, and performing feature extraction on the second initial intention feature vector through the third initial feature extraction layer to obtain a third initial intention feature vector;

[0043] Performing intent classification on the third initial intent feature vector to obtain a sample insurance intent;

[0044] Calculate the loss based on the sample insurance intention and the preset label insurance intention to obtain target loss data;

[0045] According to the target loss data, the first initial feature extraction layer is parameter updated to obtain the first feature extraction layer, the second initial feature extraction layer is parameter updated to obtain the second feature extraction layer, and the third initial feature extraction layer is parameter updated to obtain the third feature extraction layer.

[0046] In some embodiments, the target insurance intention includes an insurance intention score and a reason for determining the intention score;

[0047] After performing intent recognition on the target communication text by using the target intent recognition model to obtain the target insurance intent, the method further includes:

[0048] Obtain the insurance intention score of the target potential subject;

[0049] Sorting the target intention objects according to the intention scores to obtain an object ranking order;

[0050] The target intended object whose object sorting order meets the predetermined conditions and the intention score determination reason corresponding to the target intended object are pushed to the target insurance responsible object.

[0051] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides an insurance intention recognition device, the device comprising:

[0052] A sample text acquisition module is used to acquire sample communication texts;

[0053] A model parameter adjustment module, used to adjust the parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer;

[0054] An intention feature extraction module, configured to perform feature extraction on the sample communication text through the first feature extraction layer to obtain a first intention feature vector, perform feature extraction on the first intention feature vector through the second feature extraction layer to obtain a second intention feature vector, and perform feature extraction on the second intention feature vector through the third feature extraction layer to obtain a third intention feature vector;

[0055] a score determination module, configured to determine a feature extraction score according to the second intention feature vector and the third intention feature vector;

[0056] A layer selection module, used for selecting the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer;

[0057] A model structure adjustment module, used to replace the second feature extraction layer and the third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain a target intent recognition model;

[0058] A target text acquisition module is used to acquire the target communication text;

[0059] The intention recognition module is used to perform intention recognition on the target communication text through the target intention recognition model to obtain the target insurance intention.

[0060] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the insurance intention identification method described in the first aspect when executing the computer program.

[0061] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium, and the storage medium stores a computer program. When the computer program is executed by a processor, it implements the insurance intention identification method described in the first aspect above.

[0062] The insurance intention recognition method, insurance intention recognition device, electronic device and computer-readable storage medium proposed in the present application do not directly use the obtained temporary large language model as the target intention recognition model after adjusting the parameters of the initial large language model based on the sample communication text, but further adjust the structure of the temporary large language model. In the process of structural adjustment, the feature extraction score is first determined according to the second intention feature vector output by the second feature extraction layer and the third intention feature vector output by the third feature extraction layer. Then the second feature extraction layer and the third feature extraction layer are selected according to the feature extraction score, which is equivalent to keeping the first feature extraction layer unchanged, selecting the second feature extraction layer as the target feature extraction layer according to the feature extraction score, or selecting the second feature extraction layer and the third feature extraction layer as the target feature extraction layer. Then replace the second feature extraction layer and the third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain the target intention recognition model. Finally, the target communication text is recognized by the target intention recognition model to obtain the target insurance intention. In summary, during the model training stage, this application adjusts the initial large language model from two aspects: model parameters and model structure, so as to obtain a target intent recognition model, which can improve the efficiency of insurance intent recognition while ensuring the accuracy of insurance intent recognition.

[0063] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a flow chart of the insurance intention identification method provided in an embodiment of the present application;

[0065] Figure 2 yes Figure 1 Flowchart of step 102 in FIG.

[0066] Figure 3 yes Figure 1 Flow chart of step 103 in;

[0067] Figure 4 yes Figure 3 Flow chart of step 305 in;

[0068] Figure 5 yes Figure 1 Flowchart of step 104 in;

[0069] Figure 6 It is a module structure block diagram of the insurance intention recognition device provided in an embodiment of the present application;

[0070] Figure 7 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0072] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0073] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0074] First, some nouns involved in this application are analyzed:

[0075] Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. AI is a branch of computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can respond in a similar way to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing and expert systems. AI can simulate the information process of human consciousness and thinking. AI is also a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0076] Natural language processing (NLP): NLP uses computers to process, understand and apply human languages ​​(such as Chinese, English, etc.). NLP is a branch of artificial intelligence and an interdisciplinary subject between computer science and linguistics. It is often referred to as computational linguistics. Natural language processing includes grammatical analysis, semantic analysis, and text comprehension. Natural language processing is commonly used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information image processing, information extraction and filtering, text classification and clustering, public opinion analysis and opinion mining. It involves data mining related to language processing, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computing.

[0077] In the tide of digital transformation, social media platforms have become an important channel for enterprises to connect with customers and gain insight into market needs. Instant messaging applications, in particular, offer huge potential for enterprises with their large user base and rich social functions. However, accurately identifying potential high-value customers, especially those who show a high interest in insurance services, in a massive amount of communication texts is an extremely challenging task.

[0078] Traditional manual screening methods are time-consuming, labor-intensive, inefficient, and difficult to cope with the ever-changing market environment and customer needs. Most automated customer identification tools on the market are based on keyword matching or simple statistical analysis, lacking the ability to deeply understand and perceive situations, and often fail to accurately capture the true intentions and needs of potential customers. Especially in the insurance industry, due to the complexity of products and the diversity of customer inquiries, it is difficult to effectively distinguish truly high-intent customers by simply relying on rules or shallow analysis methods.

[0079] Based on this, the embodiment of the present application proposes an insurance intention recognition method, an insurance intention recognition device, an electronic device and a computer-readable storage medium, which utilizes the deep understanding and generation capabilities of a large language model and combines the professional knowledge of the insurance industry to achieve accurate interpretation of communication texts, thereby quickly locating and locking high-intent customers in massive information, and providing sales staff with real-time high-intent clues. It can improve the efficiency of insurance intention recognition for insurance scenarios, and can significantly improve customer conversion rates and operating efficiency.

[0080] The insurance intention identification method provided in the embodiment of the present application can be applied to the terminal and the server side, and can also be software running on the server side. The server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the insurance intention identification method, but is not limited to the above forms.

[0081] The present application can be used in many general or special computer system environments or configurations. For example: server computers, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0082] The embodiments of the present application provide an insurance intention recognition method and an insurance intention recognition device, an electronic device, and a storage medium, which are specifically illustrated by the following embodiments. First, the insurance intention recognition method in the embodiments of the present application is described.

[0083] It should be noted that in each specific implementation method of the present application, when it comes to processing data related to the user's identity or characteristics such as the user's behavioral data, the user's permission or consent will be obtained first, and the collection, use and processing of such data will comply with relevant laws, regulations and standards.

[0084] Reference Figure 1 , Figure 1 It is an optional flowchart of the insurance intention identification method provided in an embodiment of the present application, which may include but is not limited to steps 101 to 108.

[0085] Step 101, obtaining a sample communication text;

[0086] Step 102, adjusting parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer and a third feature extraction layer;

[0087] Step 103, extracting features from the sample communication text through a first feature extraction layer to obtain a first intention feature vector, extracting features from the first intention feature vector through a second feature extraction layer to obtain a second intention feature vector, and extracting features from the second intention feature vector through a third feature extraction layer to obtain a third intention feature vector;

[0088] Step 104, determining a feature extraction score according to the second intention feature vector and the third intention feature vector;

[0089] Step 105, selecting the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer;

[0090] Step 106, replacing the second feature extraction layer and the third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain a target intent recognition model;

[0091] Step 107, obtaining the target communication text;

[0092] Step 108, performing intent recognition on the target communication text through a target intent recognition model to obtain a target insurance intent.

[0093] In steps 101 to 108 shown in the embodiment of the present application, after adjusting the parameters of the initial large language model based on the sample communication text, the obtained temporary large language model is not directly used as the target intent recognition model, but the temporary large language model is further structurally adjusted. In the process of structural adjustment, the feature extraction score is first determined based on the second intent feature vector output by the second feature extraction layer and the third intent feature vector output by the third feature extraction layer. Then the second feature extraction layer and the third feature extraction layer are selected according to the feature extraction score, which is equivalent to keeping the first feature extraction layer unchanged, selecting the second feature extraction layer as the target feature extraction layer according to the feature extraction score, or selecting the second feature extraction layer and the third feature extraction layer as the target feature extraction layer. Then the target feature extraction layer replaces the second feature extraction layer and the third feature extraction layer in the temporary large language model to obtain a target intent recognition model. Finally, the target intent recognition model is used to perform intent recognition on the target communication text to obtain the target insurance intent. In summary, during the model training stage, this application adjusts the initial large language model from two aspects: model parameters and model structure, so as to obtain a target intent recognition model, which can improve the efficiency of insurance intent recognition while ensuring the accuracy of insurance intent recognition.

[0094] In step 101 of some embodiments, a sample communication text is obtained. Communication text generally refers to the interactive text between the sender and the receiver, also known as the dialogue text. For example, the target intended object / customer has a dialogue with the insurance responsible object / insurance salesperson in the communication window, and the generated dialogue record can be used as the communication text. The sample communication text specifically refers to the communication text used as a sample, or refers to the communication text used for model training.

[0095] It should be noted that since the communication goals in insurance scenarios are clear and the communication text does not involve more complex semantics, the number of feature extraction layers can be reduced.

[0096] In one embodiment, step 101 includes: obtaining communication speech; performing speech recognition on the communication speech to obtain communication text; extracting insurance-related text from the communication text to obtain a sample communication text. When extracting insurance-related text, a keyword matching algorithm or a semantic detection algorithm may be used. When using the keyword matching algorithm, some keywords in the insurance field need to be set in advance. When using the semantic detection algorithm, some reference semantic vectors with insurance semantics need to be set in advance.

[0097] In one embodiment, insurance business communication records from an insurance website or an offline branch system are obtained, and a sample communication text is determined based on the insurance business communication records.

[0098] In step 102 of some embodiments, the parameters of the preset initial large language model are adjusted according to the sample communication text to obtain a temporary large language model. The initial large language model is a large language model. Large Language Model (LLM) refers to an artificial intelligence system that processes and generates natural language text based on deep learning technology, especially neural network models. Large language models are usually trained on a large amount of text data to learn the structure, grammar, semantics and contextual relationships of the language, so that they can perform a variety of language tasks. The following are some of the main features of large language models: Large scale: Large language models usually contain hundreds of millions to hundreds of billions of parameters, enabling them to capture complex language features. Lots of training data: Use large-scale text data for training, covering a variety of topics and styles to improve the model's language understanding and generation capabilities. Task diversity: Can perform a variety of natural language processing tasks, such as text generation, answering questions, translation, summarization, sentiment analysis, etc. Context understanding: Ability to understand and generate coherent text that conforms to the context, process longer text inputs, and help maintain the coherence of the conversation. Self-learning: Through pre-training and fine-tuning, it can be optimized on specific tasks or fields to improve performance.

[0099] In one embodiment, the initial large language model includes a text prefix layer, a first initial feature extraction layer, a second initial feature extraction layer, and a third initial feature extraction layer.

[0100] The text prefix layer is used to import intent-guided text in the insurance field. The intent-guided text is used to guide the initial large language model to output the insurance intent to which the text belongs based on the input text.

[0101] The first initial feature extraction layer is a processing layer based on the Transformer architecture for feature extraction. The typical Transformer architecture consists of two parts: the encoder and the decoder. The encoder converts the input sequence into an intermediate representation, and the decoder generates the output sequence based on this representation.

[0102] The second initial feature extraction layer is also a processing layer based on the Transformer architecture, which is used for feature extraction. The second initial feature extraction layer is also a processing layer based on the Transformer architecture, which is used for feature extraction.

[0103] The second initial feature extraction layer and the third initial feature extraction layer can be collectively referred to as an initial auxiliary feature extraction layer, that is, the initial large language model includes the first feature extraction layer and at least two cascaded initial auxiliary feature extraction layers. The latter initial auxiliary feature extraction layer uses the output vector obtained by the previous initial auxiliary feature extraction layer as an input vector, and obtains an output vector after feature extraction of the input vector.

[0104] It should be noted that the initial large language model is a model with particularly many parameters, so the initial large language model may have more than 2 initial auxiliary feature extraction layers, and may have 9, 15, etc. For example, the initial large language model may include a cascaded first initial feature extraction layer and 9 initial auxiliary feature extraction layers, specifically including a first initial feature extraction layer, a second initial feature extraction layer, a third initial feature extraction layer, a fourth initial feature extraction layer, a fifth initial feature extraction layer, a sixth initial feature extraction layer, a seventh initial feature extraction layer, an eighth initial feature extraction layer, a ninth initial feature extraction layer, and a tenth initial feature extraction layer.

[0105] In addition to the above-mentioned initial feature extraction layer, the initial large language model also includes a classification layer. The classification layer performs intent classification on the feature vector output by the last feature extraction layer. For example, if the last initial feature extraction layer is the third initial feature extraction layer, the classifier performs intent classification on the feature vector output by the third initial feature extraction layer to obtain the sample insurance intent. For another example, if the last initial feature extraction layer is the tenth initial feature extraction layer, the classifier performs intent classification on the feature vector output by the tenth initial feature extraction layer to obtain the sample insurance intent.

[0106] The temporary large language model and the initial large language model are the same in model structure, but different in model parameters. The temporary large language model includes a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer. The temporary large language model may also include a text prefix layer and a classification layer.

[0107] In one embodiment, referring to Figure 2 , step 102 may include:

[0108] Step 201, embedding the intention guidance text into the sample communication text through the text prefix layer to obtain the sample intention communication text;

[0109] Step 202, extracting features from the sample intention communication text through a first initial feature extraction layer to obtain a first initial intention feature vector, extracting features from the first initial intention feature vector through a second initial feature extraction layer to obtain a second initial intention feature vector, and extracting features from the second initial intention feature vector through a third initial feature extraction layer to obtain a third initial intention feature vector;

[0110] Step 203, performing intent classification on the third initial intent feature vector to obtain a sample insurance intent;

[0111] Step 204, performing loss calculation according to the sample insurance intention and the preset label insurance intention to obtain target loss data;

[0112] Step 205, according to the target loss data, update the parameters of the first initial feature extraction layer to obtain the first feature extraction layer, update the parameters of the second initial feature extraction layer to obtain the second feature extraction layer, and update the parameters of the third initial feature extraction layer to obtain the third feature extraction layer.

[0113] In step 201, the sample communication text may be input into the text prefix layer, which has a preset intention guiding text and can embed the intention guiding text into the sample communication text, such as splicing the intention guiding text at the beginning of the sample communication text with a splicing symbol to obtain the sample intention communication text. The intention guiding text may include insurance consultation, claim application, insurance inquiry, policy revision, pre-sales service, etc., and the intention guiding text is embedded at the beginning of the input text and participates in the model training.

[0114] In step 202, the data flow order is from small to large, specifically, the first initial feature extraction layer, followed by the second initial feature extraction layer, and then the third initial feature extraction layer. If the initial large language model also includes a fourth initial feature extraction layer and a fifth initial feature extraction layer, the data flow order is to pass through the first initial feature extraction layer, the second initial feature extraction layer, the third initial feature extraction layer, the fourth initial feature extraction layer, and the fifth initial feature extraction layer in sequence. For example, step 202 also includes: extracting features from the third initial intent feature vector through the fourth initial feature extraction layer to obtain a fourth initial intent feature vector; extracting features from the fourth initial intent feature vector through the fifth initial feature extraction layer to obtain a fifth initial intent feature vector; and so on.

[0115] In step 203, a classification layer may be used to perform intent classification on the third initial intent feature vector to obtain a sample insurance intent. The classification layer may include a linear mapping sublayer and a nonlinear activation sublayer.

[0116] In step 204, the error function sample insurance intent and the preset label insurance intent may be used to perform loss calculation to obtain target loss data. The error function may include a mean square error function, a Euclidean distance function, a cosine similarity function, and the like.

[0117] In step 205, during the parameter adjustment process, the target loss data is compared with a preset loss threshold. If the target loss data is greater than the loss threshold, the parameters of the first initial feature extraction layer, the second initial feature extraction layer, and the third initial feature extraction layer are continuously adjusted until the target loss data is less than or equal to the loss threshold. The parameter adjustment is then stopped to obtain a temporary large language model.

[0118] In steps 201 to 205 shown in the embodiment of the present application, a text prefix layer is introduced in the model parameter adjustment to further improve the accuracy of intent recognition by the temporary large language model.

[0119] It is understandable that if the initial large language model includes more than the first initial feature extraction layer, the second initial feature extraction layer and the third initial feature extraction layer, then in step 203, the intention feature vector output by the last initial feature extraction layer is specifically classified to obtain the sample insurance intention.

[0120] In step 103 of some embodiments, the first feature extraction layer is a processing layer having the same structure as the first initial feature extraction layer but different parameters. The second feature extraction layer is a processing layer having the same structure as the second initial feature extraction layer but different parameters. The third feature extraction layer is a processing layer having the same structure as the third initial feature extraction layer but different parameters.

[0121] The temporary large language model may further include a fourth feature extraction layer, a fifth feature extraction layer, a sixth feature extraction layer, a seventh feature extraction layer, an eighth feature extraction layer, a ninth feature extraction layer, and a tenth feature extraction layer. The introduction of these layers is similar to the introduction of the first feature extraction layer, and will not be repeated here.

[0122] In one embodiment, referring to Figure 3 In step 103, extracting features from the sample communication text through the first feature extraction layer to obtain a first intention feature vector may include:

[0123] Step 301, embedding the sample communication text into vectors through the first feature extraction layer to obtain the sample communication text vectors;

[0124] Step 302, generating a query vector for the sample communication text vector through the first feature extraction layer to obtain an intent query vector;

[0125] Step 303, performing vector segmentation on the sample communication text vector through the first feature extraction layer to obtain at least two sample communication text sub-vectors;

[0126] Step 304, generating a key-value pair vector for each sample communication text sub-vector through the first feature extraction layer to obtain a candidate intent key-value pair vector;

[0127] Step 305, performing vector selection on each candidate intent key-value pair vector through the first feature extraction layer to obtain at least one target selected intent key-value pair vector;

[0128] Step 306, calculating the attention score of the intent query vector and each target selected intent key-value pair vector through the first feature extraction layer to obtain the target attention score;

[0129] Step 307, concatenate the attention scores through the first feature extraction layer to obtain a first intention feature vector.

[0130] In step 301, the first feature extraction layer includes an embedding sublayer for performing vector embedding on the sample communication text.

[0131] In step 302, the first feature extraction layer includes a query sublayer, which is used to generate a query vector for the sample communication text vector through a query vector generation matrix.

[0132] In step 303, the first feature extraction layer includes a segmentation sublayer for segmenting the sample communication text vector, which can be random segmentation or average segmentation.

[0133] In step 304, the first feature extraction layer includes a key-value pair sublayer, which is used to generate key-value pair vectors for the sample communication text subvectors through a key vector generation matrix and a value vector generation matrix. Specifically, the sample communication text subvector is generated with a key vector generation matrix to obtain an intention key vector; the sample communication text subvector is generated with a value vector generation matrix to obtain an intention value vector; the intention key vector and the intention value vector are combined to obtain a candidate intention key-value pair vector.

[0134] In step 305, the first feature extraction layer includes a selection sublayer for performing vector selection on each candidate intent key-value pair vector.

[0135] In one embodiment, referring to Figure 4 , step 305 may include:

[0136] Step 401, calculating the importance of the intent query vector and the candidate intent key-value pair vector to obtain the importance corresponding to each candidate intent key-value pair vector;

[0137] Step 402 , selecting each candidate intent key-value pair vector according to the importance, and obtaining at least one target selected intent key-value pair vector.

[0138] In step 401, the importance of each candidate intent key-value pair vector can be obtained by selecting an importance function in a sub-layer to calculate the importance of the intent query vector and the candidate intent key-value pair vector. The importance function can be a cosine similarity function.

[0139] In step 402, each candidate intent key-value pair vector may be sorted according to importance to obtain a sorting order; and the candidate intent key-value pair vector whose sorting order meets a predetermined condition is determined as a target selected intent key-value pair vector.

[0140] The benefit of the embodiment of the above steps 401 to 402 is that the accuracy of vector selection can be improved.

[0141] In step 306, the first feature extraction layer includes an attention sublayer for calculating an attention score for the intent query vector and each target selected intent key-value pair vector using an attention function. The attention function represents a dot product of the intent query vector and the target selected intent key-value pair vector to obtain a target attention score.

[0142] In step 307, the first feature extraction layer includes a concatenation sublayer. The concatenation sublayer can concatenate each target attention score according to the order corresponding to each sample communication text subvector, thereby obtaining a first intention feature vector.

[0143] The benefit of the above-mentioned embodiment of step 301 to step 307 is that by performing multi-faceted processing on the sample communication text vector, the potential intention information can be fully mined, thereby improving the accuracy of the first intention feature vector.

[0144] In one embodiment, after step 307, the intention recognition method provided by this embodiment may further include: updating the first intention feature vector, specifically including:

[0145] Perform feature fusion according to the sample communication text vector and the first intention feature vector to obtain a first fused intention feature vector;

[0146] Normalizing the first fusion intention feature vector to obtain a second fusion intention feature vector;

[0147] Inputting the second fusion intention feature vector into a feedforward neural network for feature mapping to obtain a third fusion intention feature vector;

[0148] Perform feature fusion according to the first fusion intention feature vector and the third fusion intention feature vector to obtain a fourth fusion intention feature vector;

[0149] Normalizing the fourth fusion intent feature vector to obtain a fifth fusion intent feature vector;

[0150] Performing a linear transformation on the fifth fusion intent feature vector to obtain a sixth fusion intent feature vector;

[0151] Nonlinear activation is performed on the sixth fused intention feature vector to obtain an updated first intention feature vector.

[0152] The benefit of the above embodiment is that the accuracy of the first intention feature vector can be further improved.

[0153] In one embodiment, the process of extracting features from the first intention feature vector through the second feature extraction layer in step 103, and the process of extracting features from the second intention feature vector through the third feature extraction layer, are substantially the same as the process of extracting features from the sample communication text through the first feature extraction layer, except that there is no embedding sublayer in the first feature extraction layer. That is, the second feature extraction layer and the third feature extraction layer both include a query sublayer, a slicing sublayer, a key-value pair sublayer, a selection sublayer, an attention sublayer, and a splicing sublayer.

[0154] In step 104 of some embodiments, a feature extraction score is determined based on the second intent feature vector and the third intent feature vector. The feature extraction score is used to characterize the accuracy of feature extraction performed by the second feature extraction layer and the third feature extraction layer. The feature extraction score includes a first extraction score of the second feature extraction layer and a second extraction score of the second feature extraction layer. The first extraction score is used to characterize the accuracy of intent feature extraction performed by the second feature extraction layer on the first intent feature vector. The second extraction score is used to characterize the accuracy of intent feature extraction performed by the third feature extraction layer on the second intent feature vector.

[0155] In one embodiment, referring to Figure 5 , step 104 may include:

[0156] Step 501, performing intent classification on the second intention feature vector to obtain a second insurance intention;

[0157] Step 502, calculating the similarity between the second insurance intention and the preset label insurance intention to obtain a first extraction score of the second feature extraction layer;

[0158] Step 503, performing intent classification on the third intention feature vector to obtain a third insurance intention;

[0159] Step 504, performing similarity calculation based on the third intent feature vector and the label insurance intent to obtain a second extraction score of the third feature extraction layer.

[0160] In step 501, the second intention feature vector may be input into the classification layer for intent classification, thereby obtaining the second insurance intention. Specifically, step 501 may include: linearly mapping the second intention feature vector through a linear mapping function of the classification layer to obtain a second intention linear vector; and nonlinearly activating the second intention linear vector through a nonlinear activation function of the classification layer to obtain the second insurance intention.

[0161] In step 502, the label insurance intent is used to characterize the insurance intent to which the sample communication text belongs. The similarity calculation can select a cosine similarity function and a Euclidean distance function. The higher the similarity between the second insurance intent and the label insurance intent, the higher the first extraction score.

[0162] Step 503 is similar to step 501 , and step 504 is similar to step 502 , which will not be described in detail here.

[0163] Steps 501 to 504 shown in the embodiment of the present application introduce the label insurance intent to determine the extraction score, which improves the accuracy of the determination and is easy to implement.

[0164] In step 105 of some embodiments, the second feature extraction layer and the third feature extraction layer are selected according to the feature extraction score to obtain a target feature extraction layer. The target feature extraction layer is a feature extraction layer between the second feature extraction layer and the third feature extraction layer. The higher the feature extraction score, the more likely the feature extraction layer is to be used as the last feature extraction layer.

[0165] In one embodiment, the temporary large language model includes a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer, and step 105 may include:

[0166] Selecting the second feature extraction layer according to the first extraction score to obtain a first selection result;

[0167] If the first selection result indicates that the second feature extraction layer can be used as the last feature extraction layer, the second feature extraction layer is used as the target feature extraction layer;

[0168] If the first selection result indicates that the second feature extraction layer cannot be used as the last feature extraction layer, the third feature extraction layer is selected according to the second extraction score to obtain a second selection result;

[0169] If the second selection result indicates that the third feature extraction layer can be used as the last feature extraction layer, the second feature extraction layer and the third feature extraction layer are used as target feature extraction layers.

[0170] In another embodiment, the temporary large language model includes a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, a fourth feature extraction layer, and a fifth feature extraction layer, and step 105 may include:

[0171] Selecting the second feature extraction layer according to the first extraction score to obtain a first selection result;

[0172] If the first selection result indicates that the second feature extraction layer cannot be used as the last feature extraction layer, the second feature extraction layer is selected according to the second extraction score to obtain a second selection result;

[0173] If the second selection result indicates that the third feature extraction layer can be used as the last feature extraction layer, the second feature extraction layer and the third feature extraction layer are determined as target feature extraction layers.

[0174] If the second selection result indicates that the third feature extraction layer cannot be used as the last feature extraction layer, the fourth feature extraction layer is selected according to the third extraction score to obtain a third selection result;

[0175] If the third selection result indicates that the fourth feature extraction layer can be used as the last feature extraction layer, the second feature extraction layer, the third feature extraction layer, and the fourth feature extraction layer are determined as target feature extraction layers;

[0176] If the third selection result indicates that the fourth feature extraction layer cannot be used as the last feature extraction layer, the fifth feature extraction layer is selected according to the fourth extraction score to obtain a fourth selection result;

[0177] If the fourth selection result indicates that the fifth feature extraction layer can be used as the last feature extraction layer, the second feature extraction layer, the third feature extraction layer, the fourth feature extraction layer and the fifth feature extraction layer are determined as target feature extraction layers.

[0178] The benefit of the above embodiment is that it can ensure the accuracy of feature extraction while reducing the number of feature extraction layers.

[0179] In step 106 of some embodiments, the target feature extraction layer replaces the second feature extraction layer and the third feature extraction layer in the temporary large language model to obtain a target intent recognition model. For example, the target feature extraction layer includes the second feature extraction layer, then after the replacement, the target intent recognition model includes the first feature extraction layer and the second feature extraction layer, and no longer includes the third feature extraction layer. For another example, the temporary large language model includes the first feature extraction layer, the second feature extraction layer, the third feature extraction layer, the fourth feature extraction layer, and the fifth feature extraction layer, and the target feature extraction layer includes the second feature extraction layer and the third feature extraction layer. Then after the replacement, the target intent recognition model includes the first feature extraction layer, the second feature extraction layer, and the third feature extraction layer, and no longer includes the fourth feature extraction layer and the fifth feature extraction layer.

[0180] In step 107 of some embodiments, a target communication text is obtained. The target communication text has the same definition as the sample communication text described above, except that the target communication text is used in the model use phase, while the sample communication text is used in the model training phase.

[0181] In step 108 of some embodiments, the target communication text is subjected to intent recognition by a target intent recognition model to obtain a target insurance intent. The target intent recognition model is similar in structure to the temporary large language model, but the number of feature extraction layers of the target intent recognition model is less than the number of feature extraction layers of the temporary large language model. The target insurance intent includes an insurance intention score and a reason for determining the intention score.

[0182] In one embodiment, after step 108, the insurance intention identification method provided by this embodiment may further include:

[0183] Obtain insurance intention scores of target potential subjects;

[0184] Sorting the target intention objects according to the intention scores to obtain the object ranking order;

[0185] The target intended objects whose object sorting order meets the predetermined conditions and the reasons for the intention score determination corresponding to the target intended objects are pushed to the target insurance responsible object.

[0186] The benefit of the above embodiment is that it can push insurance-related information, which helps to improve the insurance business conversion rate.

[0187] In one example, after the target intention recognition model performs intent recognition, it can output the insurance intention score and the reason for the intention score determination. Within a certain time range, such as within a day, the list of high-intention objects is updated regularly according to the insurance intention score, and sorted from high to low by score, and pushed to the operating system of the target insurance responsible object (also known as the front-line salesperson) to provide the salesperson with a clear follow-up priority. In addition to pushing the list of high-intention objects, it will also explain why this list is marked as high-intention. For example, the target intention object said: I want to consult how much my compulsory traffic insurance is. I didn’t have any accidents last year, or is it the same as last year. The target intention recognition model outputs a score of 9 points, and the reason is vehicle insurance premium consultation.

[0188] In summary of the above embodiments, this application proposes a high-intention object intention recognition strategy based on a large language model, which aims to achieve accurate analysis of the content of the communication text through deep learning technology, so that high-intention objects can be identified and recommended to the target insurance responsible party in real time. Specifically, by utilizing the powerful language understanding and generation capabilities of the large model, as well as fine-tuning of industry-specific data, the model's understanding of insurance industry terminology, product features, and customer consultation scenarios is enhanced. Subsequently, the large language model identifies the insurance intention contained in the target communication text (such as the willingness to purchase auto insurance). Finally, the list of high-intention customers can be updated in real time and sorted by priority, providing the target insurance responsible party with clear follow-up goals to help conversion, and it is highly applicable.

[0189] See also Figure 6 The embodiment of the present application also provides an insurance intention recognition device, which can implement the above-mentioned insurance intention recognition method. Figure 6The module structure block diagram of the insurance intention recognition device provided in the embodiment of the present application includes: a sample text acquisition module 601, a model parameter adjustment module 602, an intention feature extraction module 603, a score determination module 604, a layer selection module 605, a model structure adjustment module 606, a target text acquisition module 607 and an intention recognition module 608. Among them, the sample text acquisition module 601 is used to obtain sample communication text; the model parameter adjustment module 602 is used to adjust the parameters of the preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer; the intention feature extraction module 603 is used to extract features from the sample communication text through the first feature extraction layer to obtain a first intention feature vector, extract features from the first intention feature vector through the second feature extraction layer to obtain a second intention feature vector, and extract features from the second intention feature vector through the third feature extraction layer to obtain a third Intent feature vector; a score determination module 604, used to determine the feature extraction score according to the second intent feature vector and the third intent feature vector; a layer selection module 605, used to select the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer; a model structure adjustment module 606, used to replace the second feature extraction layer and the third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain a target intent recognition model; a target text acquisition module 607, used to obtain a target communication text; an intent recognition module 608, used to perform intent recognition on the target communication text through the target intent recognition model to obtain a target insurance intent.

[0190] In one embodiment, the insurance intention recognition device further includes: a feature updating module, configured to update the first intention feature vector.

[0191] In one embodiment, the insurance intention identification device also includes: an information push module, which is used to: obtain the insurance intention score of the target intention object; sort the target intention objects according to the intention score to obtain the object sorting order; push the target intention objects whose object sorting order meets the predetermined conditions and the intention score determination reason corresponding to the target intention object to the target insurance responsible object.

[0192] It should be noted that the specific implementation of the insurance intention recognition device is basically the same as the specific implementation example of the above-mentioned insurance intention recognition method, and will not be repeated here.

[0193] The embodiment of the present application also provides an electronic device, which includes: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for realizing connection and communication between the processor and the memory, and the above-mentioned insurance intention recognition method is realized when the program is executed by the processor. The electronic device can be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.

[0194] See also Figure 7 , Figure 7 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0195] The processor 701 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0196] The memory 702 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 702, and the processor 701 calls and executes the insurance intention identification method of the embodiment of this application;

[0197] Input / output interface 703, used to implement information input and output;

[0198] Communication interface 704, used to realize communication interaction between the device and other devices, which can be realized through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0199] A bus 705 that transmits information between the various components of the device (e.g., the processor 701, the memory 702, the input / output interface 703, and the communication interface 704);

[0200] The processor 701 , the memory 702 , the input / output interface 703 and the communication interface 704 are connected to each other in communication within the device via a bus 705 .

[0201] An embodiment of the present application also provides a storage medium, which is a computer-readable storage medium used for computer-readable storage. The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned insurance intention identification method.

[0202] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0203] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0204] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0205] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0206] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0207] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0208] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0209] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0210] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0211] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0212] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable an electronic device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0213] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A method for identifying insurance intention, characterized in that: The method comprises: Get sample communication texts; Adjusting parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer and a third feature extraction layer; Performing feature extraction on the sample communication text through the first feature extraction layer to obtain a first intention feature vector, performing feature extraction on the first intention feature vector through the second feature extraction layer to obtain a second intention feature vector, and performing feature extraction on the second intention feature vector through the third feature extraction layer to obtain a third intention feature vector; determining a feature extraction score according to the second intent feature vector and the third intent feature vector; Selecting the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer; The target feature extraction layer replaces the second feature extraction layer and the third feature extraction layer in the temporary large language model to obtain a target intent recognition model; Obtain target communication text; The target communication text is subjected to intent recognition through the target intent recognition model to obtain the target insurance intent.

2. The method according to claim 1, characterized in that The feature extraction score includes a first extraction score of the second feature extraction layer and a second extraction score of the second feature extraction layer; The determining a feature extraction score according to the second intention feature vector and the third intention feature vector comprises: Performing intent classification on the second intention feature vector to obtain a second insurance intention; Calculate the similarity between the second insurance intent and a preset label insurance intent to obtain a first extraction score of the second feature extraction layer; Performing intent classification on the third intention feature vector to obtain a third insurance intention; A similarity calculation is performed based on the third intention feature vector and the label insurance intention to obtain a second extraction score of the third feature extraction layer.

3. The method according to claim 1, characterized in that The extracting features of the sample communication text by the first feature extraction layer to obtain a first intention feature vector includes: Through the first feature extraction layer, the sample communication text is vector-embedded to obtain a sample communication text vector; Generate a query vector for the sample communication text vector through the first feature extraction layer to obtain an intent query vector; By using the first feature extraction layer, the sample communication text vector is segmented to obtain at least two sample communication text sub-vectors; Generate a key-value pair vector for each of the sample communication text sub-vectors through the first feature extraction layer to obtain a candidate intent key-value pair vector; Through the first feature extraction layer, vector selection is performed on each of the candidate intent key-value pair vectors to obtain at least one target selected intent key-value pair vector; By using the first feature extraction layer, an attention score is calculated for the intent query vector and each of the target selected intent key-value pair vectors to obtain a target attention score; Through the first feature extraction layer, the attention scores are concatenated to obtain the first intention feature vector.

4. The method according to claim 3, characterized in that The step of selecting each of the candidate intent key-value pair vectors to obtain at least one target selected intent key-value pair vector includes: Calculate the importance of the intent query vector and the candidate intent key-value pair vector to obtain the importance corresponding to each candidate intent key-value pair vector; Vector selection is performed on each of the candidate intent key-value pair vectors according to the importance to obtain at least one of the target selected intent key-value pair vectors.

5. The method according to claim 3, characterized in that: In the step of concatenating the attention scores to obtain the first intention feature vector, the method further includes: Updating the first intention feature vector includes: Perform feature fusion according to the sample communication text vector and the first intention feature vector to obtain a first fused intention feature vector; Normalizing the first fusion intention feature vector to obtain a second fusion intention feature vector; Inputting the second fusion intention feature vector into a feedforward neural network for feature mapping to obtain a third fusion intention feature vector; Perform feature fusion according to the first fusion intention feature vector and the third fusion intention feature vector to obtain a fourth fusion intention feature vector; Normalizing the fourth fusion intention feature vector to obtain a fifth fusion intention feature vector; Performing a linear transformation on the fifth fusion intention feature vector to obtain a sixth fusion intention feature vector; Nonlinearly activate the sixth fused intention feature vector to obtain an updated first intention feature vector.

6. The method according to any one of claims 1 to 5, characterized in that: The initial large language model includes a text prefix layer, a first initial feature extraction layer, a second initial feature extraction layer and a third initial feature extraction layer; The step of adjusting parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model includes: Embedding the intention guidance text into the sample communication text through the text prefix layer to obtain the sample intention communication text; Performing feature extraction on the sample intention communication text through the first initial feature extraction layer to obtain a first initial intention feature vector, performing feature extraction on the first initial intention feature vector through the second initial feature extraction layer to obtain a second initial intention feature vector, and performing feature extraction on the second initial intention feature vector through the third initial feature extraction layer to obtain a third initial intention feature vector; Performing intent classification on the third initial intent feature vector to obtain a sample insurance intent; Calculate the loss based on the sample insurance intention and the preset label insurance intention to obtain target loss data; According to the target loss data, the first initial feature extraction layer is parameter updated to obtain the first feature extraction layer, the second initial feature extraction layer is parameter updated to obtain the second feature extraction layer, and the third initial feature extraction layer is parameter updated to obtain the third feature extraction layer.

7. The method according to any one of claims 1 to 5, characterized in that: The target insurance intention includes the insurance intention score and the reasons for determining the intention score; After performing intent recognition on the target communication text by using the target intent recognition model to obtain the target insurance intent, the method further includes: Obtain the insurance intention score of the target potential subject; Sorting the target intention objects according to the intention scores to obtain an object ranking order; The target intended object whose object sorting order meets the predetermined conditions and the intention score determination reason corresponding to the target intended object are pushed to the target insurance responsible object.

8. An insurance intention recognition device, characterized in that: The device comprises: A sample text acquisition module is used to acquire sample communication texts; A model parameter adjustment module, used to adjust the parameters of a preset initial large language model according to the sample communication text to obtain a temporary large language model; wherein the temporary large language model includes a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer; An intention feature extraction module, configured to perform feature extraction on the sample communication text through the first feature extraction layer to obtain a first intention feature vector, perform feature extraction on the first intention feature vector through the second feature extraction layer to obtain a second intention feature vector, and perform feature extraction on the second intention feature vector through the third feature extraction layer to obtain a third intention feature vector; a score determination module, configured to determine a feature extraction score according to the second intention feature vector and the third intention feature vector; A layer selection module, used for selecting the second feature extraction layer and the third feature extraction layer according to the feature extraction score to obtain a target feature extraction layer; A model structure adjustment module, used to replace the second feature extraction layer and the third feature extraction layer in the temporary large language model with the target feature extraction layer to obtain a target intent recognition model; A target text acquisition module is used to acquire the target communication text; The intention recognition module is used to perform intention recognition on the target communication text through the target intention recognition model to obtain the target insurance intention.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images