Text classification method, device and equipment for insurance underwriting scene

Abstract

The application is a text classification method, device and equipment for insurance underwriting scenarios, and relates to the technical field of data processing. The method comprises: acquiring an object description text; extracting modal data of at least two modalities from the object description text; encoding the modal data of the at least two modalities based on the encoding mode corresponding to each of the at least two modalities, to obtain modal feature values of the at least two modalities; acquiring feature vectors corresponding to the object description text and each underwriting type based on the modal feature values of the at least two modalities; and determining the underwriting type based on the feature vectors corresponding to the object description text and each underwriting type. In the insurance scenario, the above scheme does not require manual review of insurance information, and the feature vectors corresponding to the text and each type are obtained through the modal data of different modalities of the object description text, so as to determine the type of the object description text, thereby improving the efficiency of insurance underwriting and the accuracy of insurance underwriting.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a text classification method, apparatus, and device for insurance underwriting scenarios. Background Art

[0002] Insurance underwriting is the process by which an insurance company reviews, verifies, and selects risks based on a policyholder's insurance application. Underwriting is the prerequisite for insurance underwriting, the first step for an insurer in handling business, and a prerequisite for the stable operation of an insurance company.

[0003] In the related art, manual underwriting is currently the most commonly used method. Manual underwriting involves the review and approval of the policyholder's insurance application by an underwriter. This requires experienced underwriters to spend a considerable amount of time verifying the policyholder's insurance information in order to improve the accuracy of the insurance underwriting.

[0004] In the above technical solution, when the insurance application is reviewed through manual underwriting, the underwriting efficiency is low, and the manual underwriting method is limited by the influence of the accumulated experience of the underwriters, and the underwriting accuracy is also low. Summary of the Invention

[0005] The present application provides a method, apparatus, and device for text classification in insurance underwriting scenarios, which can improve the accuracy of classifying underwriting types in object description text. The technical solution is as follows:

[0006] In one aspect, a text classification method for insurance underwriting scenarios is provided, the method being used in a computer system, the method comprising:

[0007] Obtaining object description text; the object description text is used to describe insurance-related attributes of the target object;

[0008] extracting modal data of at least two modalities from the object description text;

[0009] Encoding the modal data of the at least two modalities based on the encoding modes corresponding to the at least two modalities to obtain modal eigenvalues ​​of the at least two modalities;

[0010] Based on the modal feature values ​​of the at least two modalities, obtaining feature vectors corresponding to the object description text and each underwriting type respectively;

[0011] The underwriting type of the object description text is determined based on the feature vectors corresponding to the object description text and the respective underwriting types.

[0012] In another aspect, a text classification method for insurance underwriting scenarios is provided, the method being applied to a computer device and comprising:

[0013] Obtain at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects;

[0014] Extracting sample object data from the sample object description text, wherein the sample features include sample modality data of at least two modalities;

[0015] Encoding the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities;

[0016] Based on the sample features corresponding to the at least two sample object description texts, the underwriting classification model is iteratively updated through a dynamic routing algorithm to obtain the updated underwriting classification model; the underwriting classification model is used to determine the underwriting type of the object description text based on the modal feature values ​​of at least two modalities of the object description text.

[0017] In another aspect, a text classification device for insurance underwriting scenarios is provided. The device is used in a computer device and includes:

[0018] A description text acquisition module is used to acquire an object description text; the object description text is used to describe insurance-related attributes of the target object;

[0019] A modality data extraction module, configured to extract modality data of at least two modalities from the object description text;

[0020] a modal data encoding module, configured to encode the modal data of the at least two modalities based on encoding modes corresponding to the at least two modalities, to obtain modal eigenvalues ​​of the at least two modalities;

[0021] a feature vector acquisition module, configured to acquire feature vectors corresponding to the object description text and each underwriting type based on the modal feature values ​​of the at least two modalities;

[0022] An underwriting type determination module is used to determine the underwriting type of the object description text based on the feature vectors corresponding to the object description text and each underwriting type.

[0023] In a possible implementation, the modal data of the at least two modalities include at least two of real number data, category data, and free text data;

[0024] The real number data is data for indicating the real value in the object description text;

[0025] The category data is data for indicating category information contained in the object description text;

[0026] The free text data is a natural language text used to indicate descriptive features in the object description text.

[0027] In a possible implementation, the modal data encoding module is further configured to:

[0028] In response to the modal data of the at least two modalities including the real number data, obtaining a numerical value corresponding to the real number data in the object description text;

[0029] Average normalization processing is performed on the numerical values ​​corresponding to the real number data in the object description text to obtain the modal eigenvalues ​​corresponding to the real number data in the object description text.

[0030] In a possible implementation, the modal data encoding module is further configured to:

[0031] In response to the modal data of the at least two modalities including the category data, obtaining a category corresponding to the category data;

[0032] Based on the category corresponding to the category data, the category data is encoded to obtain a modal feature value corresponding to the category data.

[0033] In a possible implementation, the modal data encoding module is further configured to:

[0034] In response to the modal data of the at least two modalities including the free text data, the free text data is processed based on a pre-trained language model to obtain a modal feature value corresponding to the free text data.

[0035] In a possible implementation, the method further includes:

[0036] An underwriting type display module, used to display the underwriting type of the object description text;

[0037] A related feature display module is used to display the related features of the underwriting type of the object description text; the related features are used to indicate the data in the object description text that has the highest correlation with the underwriting type corresponding to the object description text.

[0038] In a possible implementation, the feature vector acquisition module is further configured to:

[0039] Based on the modal eigenvalues ​​of the at least two modalities, data processing is performed through a feature transfer layer in an underwriting classification model to obtain feature vectors corresponding to the object description text and each underwriting type;

[0040] The underwriting type determination module is further configured to:

[0041] Based on the feature vectors corresponding to the object description text and the various underwriting types, data processing is performed through the feature classification layer in the underwriting classification model to obtain the underwriting type of the object description text.

[0042] In one possible implementation, the feature transfer layer includes feature transfer matrices corresponding to the at least two modalities and the respective underwriting types; the feature transfer matrix is ​​used to transfer features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix;

[0043] The feature vector acquisition module includes:

[0044] a feature transfer unit, configured to perform feature transfer on the modal data of the at least two modalities using a feature transfer matrix corresponding to a first underwriting type and the modal data of the at least two modalities, and a first weight corresponding to the feature transfer matrix, to obtain feature vectors corresponding to the at least two modalities and the first underwriting type, respectively; the first underwriting type is any one of the underwriting types; and the first weight is used to indicate a correlation between the modal data and different underwriting types;

[0045] A feature vector acquisition unit is used to acquire the feature vector corresponding to the object description text and the first underwriting type based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively.

[0046] In a possible implementation, the apparatus further includes:

[0047] A sample object description text acquisition module is used to acquire at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects;

[0048] A sample object data acquisition module, configured to extract sample object data from the sample object description text, wherein the sample features include sample modal data of at least two modalities;

[0049] a sample object data encoding module, configured to encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities, to obtain sample features corresponding to the at least two sample object description texts; the sample features comprising sample modal feature values ​​of the at least two modalities;

[0050] The underwriting classification module update module is used to iteratively update the underwriting classification model through a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain the updated underwriting classification model.

[0051] In a possible implementation, the underwriting classification model updating module is further configured to:

[0052] Obtain sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text; where i ≥ 2 and i is an integer;

[0053] Obtaining first weights corresponding to the i-th sample object description text and each underwriting type based on sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text and feature vectors corresponding to the i-1-th sample object description text and each underwriting type;

[0054] Obtaining a feature vector corresponding to the i-th sample object description text and each underwriting type based on the first weight corresponding to the i-th sample object description text and the sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text;

[0055] Based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, each feature transfer matrix in the underwriting classification model is updated.

[0056] In another aspect, a text classification device for insurance underwriting scenarios is provided, the device being used in a computer device and comprising:

[0057] A sample text acquisition module is used to acquire at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects;

[0058] A sample modality extraction module, configured to extract sample object data from the sample object description text, wherein the sample features include sample modality data of at least two modalities;

[0059] a sample feature acquisition module, configured to encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities, to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities;

[0060] An iterative training module is used to iteratively update the underwriting classification model through a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain the updated underwriting classification model; the underwriting classification model is used to determine the underwriting type of the object description text based on the modal feature values ​​of at least two modalities of the object description text.

[0061] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by a processor to implement the above-mentioned text classification method for insurance underwriting scenarios.

[0062] In yet another aspect, a computer program product or computer program is provided. The computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the above-described text classification method for insurance underwriting scenarios.

[0063] The beneficial effects of the technical solutions provided in the embodiments of the present application include at least:

[0064] In insurance underwriting scenarios, an object description text describing the insurance-related attributes of a target object is obtained, and data from at least two modalities is obtained and encoded within the text. The encoded features are then processed to obtain feature vectors corresponding to the object description text and each underwriting type, thereby determining the underwriting type of the object description text. In insurance scenarios, this approach eliminates the need for manual review of insurance information. Instead, the modal data from the different modalities of the object description text is used to obtain feature vectors corresponding to each type of text to determine the object description text type. This improves both underwriting efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0066] Figure 1 A schematic diagram of a computer system provided by an exemplary embodiment of the present application is shown;

[0067] Figure 2 This is a flowchart illustrating a text classification method for an insurance underwriting scenario according to an exemplary embodiment;

[0068] Figure 3This is a flowchart illustrating a text classification method for an insurance underwriting scenario according to an exemplary embodiment;

[0069] Figure 4 This is a flowchart of a text classification method for an insurance underwriting scenario according to an exemplary embodiment;

[0070] Figure 5 Shown Figure 4 A schematic diagram of a model training process involved in the illustrated embodiment;

[0071] Figure 6 A schematic diagram of an interface for obtaining a description text according to an embodiment of the present application is shown;

[0072] Figure 7 Shown Figure 4 A schematic diagram of feature transfer related to the illustrated embodiment;

[0073] Figure 8 Shown Figure 4 A characteristic explanatory diagram of the embodiment shown relates to;

[0074] Figure 9 is a schematic diagram of a text classification method for an insurance underwriting scenario according to an exemplary embodiment;

[0075] Figure 10 This is a structural block diagram of a text classification device for insurance underwriting scenarios according to an exemplary embodiment;

[0076] Figure 11 This is a structural block diagram of a text classification device for insurance underwriting scenarios according to an exemplary embodiment;

[0077] Figure 12 The figure is a schematic structural diagram of a computer device according to an exemplary embodiment. DETAILED DESCRIPTION

[0078] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0079] First, the terms involved in the embodiments of the present application are introduced.

[0080] 1) Artificial Intelligence (AI)

[0081] Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0082] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0083] 2) Machine Learning (ML)

[0084] Machine learning is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0085] 3) Modality

[0086] Each source or form of information can be called a modality. For example, information media include voice, images, and text; and various sensors, such as radar, infrared, and accelerometers, can each be considered a modality. In machine learning, multimodal data can be processed and understood through machine learning methods. Currently, a popular research area is multimodal machine learning across images, video, audio, and semantics.

[0087] 4) Intelligent underwriting

[0088] Insurance underwriting refers to the process by which an insurer reviews an insurance application, decides whether to accept the risk, and, if so, sets the premium. This means that based on comprehensive information about the insured subject matter and verification, the insurer assesses and categorizes the insurable risk, ultimately deciding whether to underwrite it and under what terms. During the underwriting process, underwriters assign different underwriting conditions based on the different risk categories of the subject matter to ensure business quality and the stability of insurance operations. Intelligent underwriting simplifies the underwriting process into a single, standardized health declaration, automating the review of insurance applications.

[0089] The text classification method for insurance underwriting scenarios provided by the embodiments of the present application can be applied to computer devices with strong data processing capabilities. In one possible implementation, the text classification method for insurance underwriting scenarios provided by the embodiments of the present application can be applied to personal computers, workstations or servers, that is, the underwriting classification model can be trained by personal computers, workstations or servers. In one possible implementation, the underwriting classification model trained by the text classification method for insurance underwriting scenarios provided by the embodiments of the present application can be used to classify object description texts to obtain the underwriting type corresponding to the object description texts.

[0090] Please refer to Figure 1 , which shows a schematic diagram of a computer system provided by an exemplary embodiment of the present application. The computer system 200 includes a terminal 110 and a server 120, wherein data communication is performed between the terminal 110 and the server 120 via a communication network. Optionally, the communication network can be a wired network or a wireless network, and the communication network can be at least one of a local area network, a metropolitan area network, and a wide area network.

[0091] An application with an insurance underwriting function is installed in the terminal 110. The application may be an insurance application or an artificial intelligence (AI) application with an insurance underwriting function, which is not limited in the embodiment of the present application.

[0092] Optionally, the terminal 110 may have a data transmission interface, and the data transmission interface is used to receive text data input from other computer devices.

[0093] Optionally, the terminal 110 may also have an image acquisition device, which is used to acquire image data corresponding to the object description text, and perform OCR recognition on the image data corresponding to the object description text through OCR (Optical Character Recognition) recognition technology to obtain the object description text.

[0094] Optionally, the computer device 110 can be a mobile terminal such as a smart phone, a tablet computer, a laptop computer, or a similar mobile terminal, or a terminal such as a desktop computer or a projection computer, or an intelligent terminal with a data processing component, which is not limited in the embodiments of the present application.

[0095] The server 120 may be implemented as a single server or a server cluster consisting of a group of servers. The server 120 may be a physical server or a cloud server. In one possible implementation, the server 120 is a background server of an application in the computer device 110 .

[0096] In one possible implementation of this embodiment, the server 120 trains the underwriting classification model using a pre-set training sample set (i.e., a first training sample set), wherein the training sample set may include sample object description text and the underwriting type corresponding to the sample object description text. After the server 120 completes the training process of the underwriting classification model, it sends the trained underwriting classification model to the terminal 110 via a wired or wireless connection. The terminal 110 receives the trained underwriting classification model and inputs the data information corresponding to the underwriting classification model into an application with an underwriting function, so that when the user uses the application to process the object description text, the object description text can be processed according to the trained underwriting classification model, thereby implementing all or part of the steps of the text classification method for insurance underwriting scenarios.

[0097] Figure 2 This is a flowchart of a text classification method for insurance underwriting scenarios according to an exemplary embodiment. The method can be executed by a computer device, wherein the computer device can be the above-mentioned Figure 1 The terminal 120 in the embodiment shown. Figure 2 As shown, the process of the text classification method for insurance underwriting scenarios may include the following steps:

[0098] Step 201: Obtain object description text; the object description text is used to describe insurance-related attributes of the target object.

[0099] Step 202: extract modal data of at least two modalities from the object description text.

[0100] Step 203 : Encode the modal data of the at least two modalities based on the encoding modes corresponding to the at least two modalities, and obtain modal eigenvalues ​​of the at least two modalities.

[0101] Step 204: Based on the modal feature values ​​of the at least two modalities, obtain feature vectors corresponding to the object description text and each underwriting type.

[0102] Step 205: Determine the underwriting type of the object description text based on the feature vectors corresponding to the object description text and the respective underwriting types.

[0103] In summary, the solution shown in the embodiment of the present application, in an insurance underwriting scenario, obtains an object description text that describes the insurance-related attributes of a target object, obtains data of at least two modalities in the text and encodes it, and then processes the features obtained after encoding to obtain a feature vector corresponding to the object description text and each underwriting type to determine the underwriting type of the object description text. The above solution, in an insurance scenario, does not require manual review of insurance information. The modal data of different modalities of the object description text is directly used to obtain the feature vector corresponding to the text and each type to determine the type of the object description text, thereby improving the efficiency of insurance underwriting while improving the accuracy of insurance underwriting.

[0104] Figure 3 This is a flowchart of a text classification method for insurance underwriting scenarios according to an exemplary embodiment. The method can be executed by a computer device, wherein the computer device can be the above-mentioned Figure 1 The server 140 in the embodiment shown. Figure 3 As shown, the process of the text classification method for insurance underwriting scenarios may include the following steps:

[0105] Step 301: Obtain at least two sample object description texts in a first sample training set; the sample object description texts include sample modality data of at least two modalities.

[0106] Step 302: extract sample object data from the sample object description text, where the sample features include sample modal data of at least two modalities.

[0107] Step 303: Encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities.

[0108] Step 304 : Based on the sample features corresponding to the at least two sample object description texts, the underwriting classification model is iteratively updated by a dynamic routing algorithm to obtain an updated underwriting classification model.

[0109] The underwriting classification model is used to determine the underwriting type of the object description text based on the modal feature values ​​of at least two modalities of the object description text.

[0110] In summary, the solution shown in the embodiment of the present application, in an insurance underwriting scenario, obtains an object description text that describes the insurance-related attributes of a target object, obtains data of at least two modalities in the text and encodes it, and then processes the features obtained after encoding to obtain a feature vector corresponding to the object description text and each underwriting type to determine the underwriting type of the object description text. The above solution, in an insurance scenario, does not require manual review of insurance information. The modal data of different modalities of the object description text is directly used to obtain the feature vector corresponding to the text and each type to determine the type of the object description text, thereby improving the efficiency of insurance underwriting while improving the accuracy of insurance underwriting.

[0111] Figure 4 This is a flowchart of a method for text classification in an insurance underwriting scenario according to an exemplary embodiment. The method can be performed by a model training device and a signal processing device, wherein the model training device can be the above-mentioned Figure 1 In the server 120 of the embodiment shown, the signal processing device may be the above-mentioned Figure 1 The terminal 120 in the embodiment shown. Figure 4 As shown, the process of the text classification method for the insurance underwriting scenario may include the following steps:

[0112] Step 401: Obtain at least two sample object description texts in a first sample training set.

[0113] The sample object description text is used to describe insurance-related attributes of the sample object, and the sample object description text includes sample modal data of at least two modalities.

[0114] In a possible implementation, the sample object description text in the first sample training set is a structured electronic medical examination report in an insurance underwriting scenario.

[0115] When the sample object description text is a structured electronic physical examination report in the insurance underwriting scenario, the insurance-related attribute of the sample object is the physical examination information of the sample object.

[0116] The structured electronic medical examination report may be a medical examination report of a sample subject, which is data formed after being structured and arranged according to specified rules.

[0117] In a possible implementation, the structured electronic medical examination report may be data generated by an operator after performing a structured arrangement based on the medical examination report of the sample subject according to specified rules, that is, the structured electronic medical examination report may be generated through manual operation.

[0118] In another possible implementation, the structured electronic medical examination report may be based on the image data of the sample subject's medical examination report. After OCR recognition, the text data in the medical examination report is obtained and structured according to specified rules to form data.

[0119] Step 402: extract sample object data from the sample object description text, where the sample features include sample modal data of at least two modalities.

[0120] In one possible implementation, the sample modal data of the at least two modalities includes at least two of real number class data, category class data, and free text class data; the real number class data is data used to indicate the real value contained in the sample object description text; the category class data is data used to indicate the category information contained in the sample object description text; and the free text class data is natural language text used to indicate descriptive features in the sample object description text.

[0121] In a possible implementation, the free text data is text data in the sample object description text except for other modal data.

[0122] In a possible implementation, the sample modality data of each modality in the sample object description text is pre-labeled by an operator.

[0123] That is, before training the model using sample object description text, it is necessary to manually annotate the text content in the sample object description text and classify the text content in the sample object description text into data of different modalities.

[0124] In another possible implementation, the modality type of the data in the sample object description text is determined based on the sample object description text.

[0125] When the data information corresponding to the sample object description text (i.e., the structured electronic physical examination report) is input into a computer device, the computer device can read the data information corresponding to the sample object description text and determine the type of each data information of the sample data based on the data information corresponding to the sample object description text. For example, the computer device can first determine the numerical information in the sample object description text and define the numerical information in the sample object description text as real number class data; after the real number class information in the sample object description text is defined, it can also search for keywords in the sample object description text according to the preset category. For example, based on the keyword "gender male", the "gender male" can be determined as the category class data, and the category "gender" corresponding to the category class data can be determined; after the real number class data and the category class data are determined, the other text data is marked as free text class data.

[0126] In a possible implementation, the free text data is text data with a character length greater than a character threshold.

[0127] When the length of the text characters is short, it may be an auxiliary text of real number data, or an auxiliary text of category data (that is, text used to modify the real number data and category data). At this time, the text data may not be treated as free text data to reduce the computing load of the computer device.

[0128] According to the insurance company's data definition, the input for this application is a structured electronic physical examination report, and the data format of the physical examination report is a hierarchical structure, for example, "Physical Examination Report_Other Medical Examinations_Electrocardiogram_Electrocardiogram_Value" and "Physical Examination Report_Imaging Examination_Ultrasound Examination_Prostate Ultrasound_Abnormality Summary". It can be seen that the underlined items are all specific examination items at a level. These examination items include all the insured's physical condition information, which is also a specific feature that the insurance company needs to consider. However, for the above two examples, "Electrocardiogram_Value" corresponds to a specific numerical value (real-number), such as "(HR) 54 beats / min", while "Prostate Ultrasound_Abnormality Summary" corresponds to a free text description (free-texts), such as "Prostate enlargement with calcification". The two types of data themselves represent different meanings, and evaluating these two types of information also requires different standards. For example, "Electrocardiogram_Value" needs to be below a certain threshold, while "Prostate Ultrasound_Abnormality Summary" requires semantic encoding of the text. In short, these data of different dimensions come from different modalities, and their corresponding processing methods also need to be adjusted specifically and cannot be generalized.

[0129] Therefore, this application first classifies and defines data of different modalities. Currently, three types of modal information are defined:

[0130] 1. Real-number data (real-number), such as "ECG_value";

[0131] 2. Category data (category), such as "basic information_gender_value" (only male and female options);

[0132] 3. Free-text data (free-texts), such as "Prostate Color Doppler Ultrasound_Abnormality Summary".

[0133] After quickly manually labeling all the data (only the category of this data needs to be labeled, so the labeling amount is much less), we can obtain three types of modal data. It is worth noting that in actual business scenarios, the number of predefined modal types can be increased, and the specific data included in each modality can be adjusted as needed, which greatly expands the practicality of this application.

[0134] Step 403: Encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities.

[0135] In one possible implementation, in response to the sample modal data of the at least two modalities including real number data, the numerical values ​​corresponding to all the real number data in the sample object description text are obtained; based on the numerical values ​​corresponding to all the real number data in the sample object description text, average normalization processing is performed to obtain the modal feature values ​​corresponding to all the real number data in the sample object description text.

[0136] For example, for real-number features, we can directly use their specific values. For example, in "(HR) 54 beats / min", we directly extract the value "54" as the feature vector with a dimension of 1. However, the numerical ranges of different real-number features are not the same, so we must first perform data normalization. Here, we use average normalization X′ = (X-μ) / (max(X)-min(X)), where X is the set of real-number features and u is the mean of the set X. X′ obtained after average normalization can map all values ​​to the interval [0, 1].

[0137] In one possible implementation, in response to the sample modal data of the at least two modalities including category data, a category corresponding to the category data is obtained; based on the category corresponding to the category data, the category data is encoded to obtain a modal feature value corresponding to the category data.

[0138] For example, for category features, one-hot vector encoding can be used. For example, in "Basic Information_Gender", we set feature 01 to male and feature 10 to female. Since this is set manually, the feature definition relationship needs to be stored to facilitate the use of subsequent interpretable models.

[0139] In one possible implementation, in response to the sample modal data of the at least two modalities including free text data, the free text data is processed based on a pre-trained language model to obtain modal feature values ​​corresponding to the free text data.

[0140] In one possible implementation, in response to the sample modal data of the at least two modalities including free text data, the free text data is processed based on the BERT model to obtain modal feature values ​​corresponding to the free text data.

[0141] For example, for text-like features (free-texts), a pre-trained language model can be used to semantically encode them, such as the BERT (Bidirectional Encoder Representation from Transformers) model, so that free-texts of different character lengths can generate text feature vectors of uniform length.

[0142] After encoding the multimodal data separately, we can obtain the characteristics of the three modes (i.e., the three modal eigenvalues), which can be defined as:

[0143]

[0144] Among them, f i are the characteristics of three different modes, They are the sets corresponding to the features of three different modes.

[0145] Step 404 : Based on the at least two sample modal feature values ​​corresponding to the at least two sample object description texts, the underwriting classification model is iteratively updated through a dynamic routing algorithm to obtain the updated underwriting classification model.

[0146] In one possible implementation, at least two sample modal feature values ​​corresponding to the i-th sample object description text are obtained; wherein, i≥2, and i is an integer; based on the at least two sample modal feature values ​​corresponding to the i-th sample object description text, and the feature vectors corresponding to the i-1-th sample object description text and each underwriting type, the first weight corresponding to the i-th sample object description text and each underwriting type is obtained; based on the first weight corresponding to the i-th sample object description text and each underwriting type, and at least two sample modal feature values ​​corresponding to the i-th sample object description text, the feature vectors corresponding to the i-th sample object description text and each underwriting type are obtained; based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, each feature transfer matrix in the underwriting classification model is updated.

[0147] In the embodiment of the present application, at least two sample object description texts can be obtained, and the underwriting classification model can be iteratively updated through a dynamic routing algorithm based on the at least two sample object description texts. Figure 5 , which shows a schematic diagram of a model training process involved in the embodiment of the present application. Figure 5 As shown, for the sample object description text 501, the sample object description text 501 contains modal data of three modes, namely real number class data 502, category class data 503 and free text class data 504; the real number class data 502 is encoded by the encoding method 505 corresponding to the real number class data to obtain the real number modal eigenvalue 506 corresponding to the real number class data; the category class data 503 is encoded by the encoding method 507 corresponding to the category class data to obtain the category modal eigenvalue 508 corresponding to the category class data; the free text data 504 is encoded by the encoding method 509 corresponding to the free text data to obtain the text modal eigenvalue 510 corresponding to the free text data.

[0148] Before model training, it is necessary to initialize the various parameters of the model, that is, initialize the various feature transfer matrices of the underwriting classification model, the first weights corresponding to each feature transfer matrix (that is, the modal feature values ​​of each modality in the sample object description text and the first weights corresponding to each underwriting type), and the initial feature vectors of each category (that is, the hidden layer vectors corresponding to each category).

[0149] During the first iteration, the eigenvector value of each modality transferred to the hidden layer corresponding to each label is determined based on the real modal eigenvalue 506, the category modal eigenvalue 508, the text modal eigenvalue 510 corresponding to the first sample object description text, and the feature transfer matrix corresponding to each modality and each label; based on the eigenvector value of the hidden layer transferred to each label by each modality, and the hidden layer vector corresponding to each category, the correlation between each modality and the hidden layer of each label can be determined, and the first weight between each modality and each label can be updated based on the correlation. Based on the first weight between each modality and each label, and the matrix of the hidden layer transferred from each modality to each label through the feature transfer matrix, the vector value corresponding to the first sample object description text and the hidden layer of each label can be obtained (that is, the first sample eigenvector after the initial vector is updated, in Figure 5 Where c1 and c2 are updated. Based on the first sample feature vector, the softmax function can be used to determine the probability value of the first sample object description text corresponding to label 1 and the probability value corresponding to label 2. Through the probability distribution corresponding to the first sample object description text and the underwriting category corresponding to the first sample object description text, each feature transfer matrix is ​​back-propagated and updated to complete the first iteration process.

[0150] During the i-th iteration, the feature vector values ​​of the hidden layers corresponding to the i-th sample object description text and the underwriting types can be determined through the above steps. The correlation between the hidden layers of the modalities and the labels can be determined based on the feature vector values ​​of the hidden layers of the modalities and the labels, and the first weights between the modalities and the labels can be updated based on the correlation. Based on the first weights between the modalities and the labels, and the matrix of the hidden layers of the modalities transmitted to the labels through the feature transfer matrix, the vector values ​​corresponding to the i-th sample object description text and the hidden layers of the labels can be obtained (that is, the i-1th sample feature vector is updated to the i-th sample feature vector, in Figure 5 Where c1 and c2 are updated). According to the i-th sample feature vector, the softmax function can be used to determine the probability value of the i-th sample object description text corresponding to label 1 and the probability value corresponding to label 2. Through the probability distribution corresponding to the i-th sample object description text and the underwriting category corresponding to the i-th sample object description text, each feature transfer matrix is ​​back-propagated and updated to complete the i-th iteration process.

[0151] For example, the above iterative training process (i.e., dynamic routing algorithm) may also include the following steps:

[0152] Define a class of intermediate hidden layers c j This intermediate hidden layer vector is 1-dimensional and is randomly initialized during training. Its function is to measure the correlation between a certain type of feature and a certain label. Therefore, the number of intermediate hidden layer features is consistent with the number of class labels. For example Figure 5 If there are two labels in , there are only two intermediate hidden features c1 and c2.

[0153] In addition to the intermediate hidden layer features, we also define r ij ∈[0, 1] is used to define the correlation value between a certain type of feature and a certain label. Therefore, in the training phase, the calculation formula is defined as:

[0154]

[0155] Where W ij Is a weight matrix, representing f i to c j As you can see, r ij In all c j The average is done, so its value can be directly used as the relevant weight of the i-th feature for the j-th classification label in the model prediction stage. Then it is necessary to update the intermediate hidden layer features, and the formula is as follows:

[0156] c i =∑ i r ij (f i w ij )(2)

[0157] You can see c j The calculation of is essentially a linear aggregation. It can also be seen that formula (1) and formula (2) are closely related. Therefore, in the model training phase, we need to iteratively use formula (1) (2) to update c j and r ij , the specific method is as follows:

[0158] 1. Initialize all model weights, including c j 、W ij 、r ij ;

[0159] 2. Update r with all features according to formula (1) ij ;

[0160] 3. Update c with all labels according to formula (2) j ;

[0161] 4. According to c j And the classification label corresponding to the i-th sample is updated W ij ;

[0162] 5. Return to step 2 and repeat the iteration until the specified training rounds are reached or the parameter change range is less than a certain threshold.

[0163] After i iterations of the update process, the updated hidden layer vectors and the feature transfer matrices can aggregate similar features from data of different modalities into multiple distinct vectors, thereby enabling feature classification of different modalities. When new features of different modalities are input, these features are linearly mapped through the feature transfer matrices and compared with the previously obtained hidden layer vectors. The similarity between the new features of different modalities and the hidden layer vectors is then determined, and this similarity is used as a weight to classify the features of the different modalities.

[0164] In one possible implementation, based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, the probability distribution corresponding to the i-th sample object description text is determined; the probability distribution corresponding to the i-th sample object description text is used to indicate the probability of the i-th sample object description text corresponding to various underwriting types; based on the probability distribution corresponding to the i-th sample object description text and the labeled underwriting type corresponding to the i-th sample object description text, each feature transfer matrix in the underwriting classification model is updated.

[0165] When the feature vectors corresponding to the i-th sample object description text and each underwriting type are determined, the probability distribution corresponding to the i-th sample object description text can be obtained through the softmax function based on the feature vectors corresponding to the i-th sample object description text and each underwriting type. For example, when the probability distribution is (0.7, 0.3), that is, the probability distribution corresponding to the i-th sample object description text indicates that the probability of the i-th sample object description text corresponding to the renewal type is 0.7, the probability of the i-th sample object description text corresponding to the rejection type is 0.3, and the underwriting type corresponding to the i-th sample object description text is renewal, the feature transfer matrices in the underwriting classification model can be updated based on the probability distribution corresponding to the i-th sample object description text through the back propagation algorithm.

[0166] In one possible implementation, based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, the predicted underwriting type corresponding to the i-th sample object description text is determined; based on the predicted underwriting type corresponding to the i-th sample object description text and the labeled underwriting type corresponding to the i-th sample object description text, each feature transfer matrix in the underwriting classification model is updated.

[0167] In one possible implementation, based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, the probability distribution corresponding to the i-th sample object description text is determined; based on the probability distribution corresponding to the i-th sample object description text, the predicted underwriting type corresponding to the i-th sample object description text is determined.

[0168] In a possible implementation, in the probability distribution corresponding to the i-th sample object description text, the one with the largest probability value is determined as the predicted underwriting type corresponding to the i-th sample object description text.

[0169] In one possible implementation, the predicted underwriting type corresponding to the maximum probability value greater than the probability threshold in the probability distribution corresponding to the i-th sample object description text is determined as the predicted underwriting type corresponding to the i-th sample object description text.

[0170] That is, when the feature vectors corresponding to the i-th sample object description text and each underwriting type are determined, the probability distribution corresponding to the i-th sample object description text can be obtained through the softmax function based on the feature vectors corresponding to the i-th sample object description text and each underwriting type. For example, when the probability distribution is (0.4, 0.2, 0.1, 0.1, 0.2), the predicted underwriting type corresponding to the maximum probability value in the probability distribution corresponding to the i-th sample object description text, that is, 0.4, can be determined as the predicted underwriting type corresponding to the i-th sample object description text, and the underwriting classification model can be updated through the back propagation algorithm based on the predicted probability 0.4 corresponding to the predicted underwriting type and the underwriting type corresponding to the i-th sample object description text.

[0171] Step 405: Obtain object description text.

[0172] The object description text is used to describe insurance-related attributes of the target object, and the object description text includes modal data of at least two modalities.

[0173] In a possible implementation, the object description text is a structured electronic medical examination report in an insurance underwriting scenario, and the structured electronic medical examination report is used to display the medical examination data of the target object according to specified data display rules.

[0174] In a possible implementation, the structured electronic physical examination report may be data formed by structurally arranging the physical examination reports of sample subjects according to specified rules.

[0175] In one possible implementation, the structured electronic physical examination report may be data formed by an operator after structuring and arranging the physical examination report of the sample object according to specified rules, that is, the structured electronic physical examination report may be formed through manual operation.

[0176] In another possible implementation, the structured electronic medical examination report may be based on the image data of the medical examination report of the sample subject. After OCR recognition, the text data in the medical examination report is obtained and the data is structured and arranged according to certain rules.

[0177] In a possible implementation, a description text acquisition interface is displayed; the description text acquisition interface includes at least two description text acquisition controls; and in response to receiving designated operations on the at least two description text acquisition controls respectively, the object description text is generated.

[0178] In which, the description text acquisition interface can be displayed in the image display component of the terminal, and the description text interface includes at least two description text acquisition controls; wherein, the at least two description text acquisition controls respectively correspond to modal data of at least two modalities, that is, modal data of different modalities can be obtained through different types of description text acquisition controls.

[0179] In a possible implementation, in response to receiving a designated operation on the description text acquisition control, modal data corresponding to the designated operation on the description text acquisition control is generated.

[0180] For example, when the description text acquisition control is a selection control, for example, the two options of the selection control are "Male" and "Female", when the user selects "Female" by clicking or touching, the terminal generates category-class modal data corresponding to the description text acquisition control, indicating that the user's gender is female. Alternatively, when the description text acquisition control is a text input control, when the user enters text data into the text input box corresponding to the text input control through text data, the description text acquisition control saves the input text data as modal data of the text description modal.

[0181] Figure 6 FIG1 shows an interface diagram of a description text acquisition interface involved in an embodiment of the present application. Figure 6 As shown, in the description text acquisition interface 600, there are different types of description text acquisition controls, and the different types of description text acquisition controls include category acquisition controls, real number acquisition controls, and text acquisition controls. Among them, in the description text acquisition interface, the category selection control includes a gender selection control 601, and the two options of the category selection control are "male" and "female", respectively. At this time, the user selects the "male" option by clicking and touching. In the description text acquisition interface, the real number selection control can also include an age input control 602, a blood pressure input control 603, and a heart rate input control 604, so that the user can enter the corresponding age, blood pressure, heart rate and other data in the input boxes corresponding to each control. In the description text acquisition interface, the text acquisition control includes a heart disease history acquisition control 605, and the user can freely enter text in the heart disease history acquisition control 605 to describe his or her own heart disease history.

[0182] Among them, when the user operates the above-mentioned selection controls in the description text acquisition interface, the terminal generates modal data corresponding to each selection control in response to the user's operation; or, when the user operates the above-mentioned selection controls in the description text acquisition interface, the user clicks or touches the confirmation control in the description text acquisition interface (not shown in the figure), when the terminal receives the user's specified operation on the confirmation control, it generates modal data corresponding to each control according to the user's operation on each description text acquisition control.

[0183] Figure 6 The description text acquisition interface and the description text control in the description text interface are schematic explanations. The description text acquisition interface can be an interface containing description text controls using other arrangements, and the description text acquisition control can also include other controls for obtaining object description text.

[0184] Step 406: extract modal data of at least two modalities from the object description text.

[0185] In one possible implementation, the modal data of the at least two modalities includes at least two of real number class data, category class data, and free text class data; the real number class data is data used to indicate the real value contained in the object description text; the category class data is data used to indicate the category information contained in the object description text; and the free text class data is natural language text used to indicate the descriptive features in the object description text.

[0186] Step 407 : Encode the modal data of the at least two modalities based on the encoding modes corresponding to the at least two modalities, and obtain modal eigenvalues ​​of the at least two modalities.

[0187] In one possible implementation, in response to the fact that the modal data of the at least two modalities include real number data, the numerical values ​​corresponding to all the real number data in the object description text are obtained; based on the numerical values ​​corresponding to all the real number data in the object description text, average normalization processing is performed to obtain the modal feature values ​​corresponding to all the real number data in the object description text.

[0188] In one possible implementation, the maximum and minimum values ​​of the numerical values ​​corresponding to all real number data in the object description text are obtained; based on the maximum and minimum values ​​of the numerical values ​​corresponding to the real number data in the object description text, the real number scaling coefficient corresponding to the object description text is obtained; based on the real number scaling coefficient corresponding to the object description text, each real number data is scaled to obtain the modal eigenvalue corresponding to each real number data in the object description text.

[0189] For example, when the real number data in the object description text is (1, 3, 4, 11), the maximum value of the real number data in the object description text is 11, and the minimum value is 1. At this time, the real number scaling coefficient corresponding to the object description text is the maximum value minus the minimum value, that is, 10. At this time, the real number data of the object description text is scaled by the real number scaling coefficient corresponding to the object description text to obtain the corresponding modal eigenvalues ​​(0.1, 0.3, 0.4, 1.1) in the object description text.

[0190] In one possible implementation, in response to the modal data of the at least two modalities including category data, a category corresponding to the category data is obtained; based on the category corresponding to the category data, the category data is encoded to obtain a modal feature value corresponding to the category data.

[0191] In one possible implementation, in response to the modal data of the at least two modalities including free text data, the free text data is processed based on the BERT model to obtain modal feature values ​​corresponding to the free text data.

[0192] In a possible implementation, the free text data is text data in the object description text except for other modal data.

[0193] In a possible implementation, the modality data of each modality in the object description text is pre-labeled by an operator.

[0194] That is, before training the model using object description text, the text content in the object description text needs to be manually annotated and classified into data of different modalities.

[0195] In another possible implementation, the modality type of the data in the object description text is determined based on the object description text.

[0196] When the data information corresponding to the object description text (i.e., the structured electronic physical examination report) is input into a computer device, the computer device can read the data information corresponding to the object description text and determine the type of each data information of the data based on the data information corresponding to the object description text. For example, the computer device can first determine the numerical information in the object description text and define the numerical information in the object description text as real number data; after the real number data in the object description text is defined, it can also search for keywords in the object description text according to the preset category. For example, based on the keyword "gender male", the "gender male" can be determined as the category data; after the real number data and the category data are determined, the other text data is marked as free text data.

[0197] In a possible implementation, the free text data is text data with a character length greater than a character threshold.

[0198] When the length of the text characters is short, it may be an auxiliary text of real number data, or an auxiliary text of category data (that is, text used to modify the real number data and category data). At this time, the text data may not be treated as free text data to reduce the computing load of the computer device.

[0199] Step 408: Based on the modal eigenvalues ​​of the at least two modalities, data processing is performed through the feature transfer layer in the underwriting classification model to obtain feature vectors corresponding to the object description text and each underwriting type.

[0200] In one possible implementation, the feature transfer layer includes at least two feature transfer matrices corresponding to the at least two modalities and each underwriting type respectively; the feature transfer matrix is ​​used to transfer the features of the modal data of the modality corresponding to the feature transfer matrix to the hidden layer of the underwriting type corresponding to the feature transfer matrix; based on the modal data of the at least two modalities, feature transfer is performed through the modal data of the at least two modalities and the feature transfer matrix corresponding to the first underwriting type, to obtain feature vectors corresponding to the at least two modalities and the first underwriting type respectively; the first underwriting type is one of the underwriting types; based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, the feature vector corresponding to the object description text and the first underwriting type is obtained.

[0201] Taking the first underwriting type of the at least two underwriting types as an example, when the object description text is input, the data of the different modalities in the object description text are first encoded according to the different modalities in the object description text to obtain the modal eigenvalues ​​corresponding to the different modalities in the object description text. The modal eigenvalues ​​of the different modalities of the object description text (taking the modal eigenvalues ​​of three different modalities as an example) are respectively transferred to the hidden layer of the first underwriting type through the three feature transfer matrices corresponding to the first underwriting type, so as to obtain the feature vectors corresponding to the object description text and the first underwriting type through the three different modalities.

[0202] In one possible implementation, based on the feature vectors corresponding to the first underwriting type respectively of the at least two modalities, weighted processing is performed using the first weights corresponding to the at least two modalities respectively of the first underwriting type to obtain the feature vectors corresponding to the object description text and the first underwriting type; the first weight is used to indicate the correlation between the modal data and different underwriting types.

[0203] After obtaining the features corresponding to the three different modalities and the first underwriting type respectively, the features of the three different modalities can also be weighted according to the first weight to obtain the feature vector corresponding to the object description text and the first underwriting type.

[0204] Figure 7 A schematic diagram of feature transfer involved in an embodiment of the present application is shown. Figure 7 As shown, the real modal eigenvalue 701 is transferred to the hidden layer 711 (i.e., c1) corresponding to label 1 and to the hidden layer 712 (i.e., c2) corresponding to label 2 through two feature transfer matrices respectively. There is a first weight between the real modal eigenvalue and the hidden layer 711 corresponding to label 1 and the hidden layer 712 corresponding to label 2. The sum of the first weight of the real modal eigenvalue and the hidden layer 711 and the first weight of the real modal eigenvalue and the hidden layer 712 is 1, that is, the first weight of the real modal eigenvalue is used to indicate the correlation between the real modal eigenvalue and each label type. When the real modal eigenvalue 701 is transferred to the hidden layer 711 and the hidden layer 712 respectively through different feature transfer matrices, it is necessary to multiply them by the corresponding first weights respectively to determine the feature vectors of the real modal eigenvalue 701 transferred to the hidden layers corresponding to the two label categories. Similarly, for the category modal eigenvalue 702 and the text modal eigenvalue 703, the features can also be transferred according to the feature transfer matrices corresponding to different underwriting categories and the first weights corresponding to different underwriting categories in the above manner. After the features of the three modes are transferred, the c1 vector and c2 vector corresponding to the object description text are obtained by summing them up.

[0205] Step 409 : Based on the feature vectors corresponding to the object description text and the various underwriting types, data processing is performed through the feature classification layer in the underwriting classification model to obtain the underwriting type of the object description text.

[0206] In the model prediction stage, our model is sensitive to the input feature f i ∈[1, 2, 3] makes the following predictions:

[0207]

[0208] Among them j It represents the correspondence between the intermediate hidden layer features (i.e., the hidden layer vector corresponding to the category label) and the category label, that is, the feature vector corresponding to each type is mapped to a real value. It can be seen that the model performs weighted aggregation on the three different modes and maps them to a category label in the multi-classification problem through the softmax activation function. For interpretability, we only need to output the visualization of the r in the above formula.ij (i.e. the first weight) and c i (The feature vectors corresponding to each underwriting type) can show the input mode that is most relevant to the output label. Moreover, if we parse the input mode features (according to the previously defined feature rules), we can also obtain the specific features that are most relevant to the output label.

[0209] Please refer to Figure 8 , which shows a characteristic interpretable schematic diagram involved in the embodiment of the present application. Figure 8 As shown, label 1 corresponds to c1, label 2 corresponds to c2, and the text modal feature value 801 has a greater correlation with c2, indicating that the text modal feature value 801 has a greater correlation with label 2. Therefore, according to the weight 802 of the text modal feature value 801 and label 2, the feature that has the greatest impact on the underwriting type classification during the data processing of the model can be read, thereby improving the credibility of the model in the underwriting type classification.

[0210] In one possible implementation, the underwriting type of the object description text is displayed; the relevant features of the underwriting type of the object description text are displayed; the relevant features are used to indicate the data of the object description text that is most relevant to the underwriting type corresponding to the object description text.

[0211] In the computer device, the underwriting classification model processes the input object description text, obtains the underwriting type of the object description text, and then displays the underwriting type of the object description text on the image display component of the computer device. For example, if the underwriting type corresponding to the input object description text is "renewal," the text or image corresponding to renewal may be displayed on the image display component of the computer device.

[0212] After displaying the underwriting type of the object description text, the computer device's image display component may also display the data within the object description text that is most relevant to the underwriting type corresponding to the object description text. The data with the highest relevance to the underwriting type corresponding to the object description text is the data with the largest first weight corresponding to the underwriting type of the object description text. For example, when the underwriting type corresponding to the input object description text is renewal, the model outputs the calculated data for the underwriting type, and determines the data corresponding to the largest first weight among the first weights corresponding to the "renewal" category. For example, when the data with the largest first weight corresponds to the text data "Cardiac Ultrasound_No Abnormality," the text data "Cardiac Ultrasound_No Abnormality" is displayed on the computer device's image display component. This improves the credibility of the underwriting type, makes it easier for users to understand the reason for the output of the underwriting type, and improves the efficiency of manual review when the underwriting type needs to be reviewed.

[0213] This application mainly proposes an intelligent underwriting prediction method, which aims to provide intelligent underwriting services with high automation, wide applicability, excellent performance, and strong explainability. After being deployed on a private cloud to empower commercial insurance companies, this application can improve the underwriting efficiency and accuracy of insurance companies, greatly reducing labor costs and time costs; at the same time, for policyholders, the intelligent underwriting system simplifies the insurance process, reduces the insured's running around, and responds to policyholders' needs in real time. In the future, the intelligent underwriting system can also help insurance companies quickly build a one-stop service of "underwriting + preservation", further improving the intelligence level of insurance business. In addition, this application can be applied to more application scenarios that require strong explainability, such as institutional insurance fraud.

[0214] The experiment used specific business data from a cooperative insurance company, with a total of 8,000 data items and 504 modal features, including 139 real number features, 219 categorical features, and 146 free text features. The classification task was a 5-category task. The specific experimental results are shown in Table 1:

[0215] Table 1

[0216]

[0217] The experimental structure above shows that the proposed multimodal classification model achieves significant performance improvements over the baseline model, with Macro-F1 improving by 0.15, a 44% improvement. Furthermore, we also analyzed the specific label performance for five categories, as shown in Table 2 (due to space limitations, only the F1-score for each category is shown).

[0218] Table 2

[0219]

[0220] As discussed above, the unimodal classification model did overfit label #1 due to the uneven distribution of classes, and even achieved a performance of 0 for label #4. Although the multimodal classification model's performance was not very high due to data distribution issues, it showed significant improvement over the unimodal classification model. For example, for label #4, the multimodal classification model achieved an F1-score of 0.18, demonstrating the effectiveness of the multimodal strategy.

[0221] In summary, the solution shown in the embodiment of the present application, in an insurance underwriting scenario, obtains an object description text that describes the insurance-related attributes of a target object, obtains data of at least two modalities in the text and encodes it, and then processes the features obtained after encoding to obtain a feature vector corresponding to the object description text and each underwriting type to determine the underwriting type of the object description text. The above solution, in an insurance scenario, does not require manual review of insurance information. The modal data of different modalities of the object description text is directly used to obtain the feature vector corresponding to the text and each type to determine the type of the object description text, thereby improving the efficiency of insurance underwriting while improving the accuracy of insurance underwriting.

[0222] Figure 9 This is a schematic diagram of a text classification method for an insurance underwriting scenario according to an exemplary embodiment. In this embodiment of the present application, the text classification method for an insurance underwriting scenario includes a model training process and a model application process. The model training and model application processes may include the following steps.

[0223] In the model training device 900, the first training sample set 901 includes at least two sample object description texts, and the modality of the data in each sample object description text is first determined based on the at least two sample object description texts in the first training sample set 901. In an embodiment of the present application, taking the example that the sample object description text includes real number class data, category class data and free text class data, the real number class data is encoded by the encoding method corresponding to the real number class data (i.e., encoding 1) to obtain the real number modal feature value corresponding to the real number class data; the category class data is encoded by the encoding method corresponding to the category class data (i.e., encoding 2) to obtain the category modal feature value; the free text class data is encoded by the encoding method corresponding to the free text class data (i.e., encoding 3) to obtain the text modal feature value. After the data of each sample object description text in the first training sample set 901 is divided into the features of each modality, it can be based on the following examples: Figure 5 The training is performed using the iterative training method shown in the figure, which will not be described here.

[0224] When the training reaches a certain number of times or the amount of data update after a single training is less than a threshold, it means that the model training device 900 has completed the training of the underwriting classification model. At this time, the model training device 900 transmits the trained underwriting classification model to the model application device 910, so that the model application device 910 performs data processing on the object description text 911 through the underwriting classification model 912. In an embodiment of the present application, the underwriting classification model includes a feature transfer layer composed of a feature transfer matrix between the modal eigenvalues ​​of each modality and different categories (i.e., different labels), and hidden layers corresponding to the different labels (i.e., the hidden layer corresponding to c1 and the hidden layer corresponding to c2), wherein the hidden layer corresponding to the different labels is also used to linearly transform the feature vector corresponding to the hidden layer into an output value corresponding to the label corresponding to the hidden layer (i.e., the hidden layer is also the feature classification layer in the underwriting classification model).

[0225] In the model application device 910, the object description text 911 also has the real number class data, category class data and the free text class data. Before inputting the object description text into the model application device, it is necessary to extract the features of different modes from the object description text 911, and input the features of different modes into the model application device 910 to obtain the underwriting type 913 output by the model application device 910.

[0226] Figure 10 This is a structural block diagram of a text classification device for insurance underwriting scenarios according to an exemplary embodiment. The text classification device for insurance underwriting scenarios can be implemented by Figure 2 or Figure 4 All or part of the steps in the method provided in the illustrated embodiment, the text classification device for insurance underwriting scenarios includes:

[0227] The description text acquisition module 1001 is used to acquire an object description text; the object description text is used to describe the insurance-related attributes of the target object;

[0228] A modality data extraction module 1002 is configured to extract modality data of at least two modalities from the object description text;

[0229] A modal data encoding module 1003 is configured to encode the modal data of the at least two modalities based on the encoding modes corresponding to the at least two modalities, to obtain modal eigenvalues ​​of the at least two modalities;

[0230] A feature vector acquisition module 1004 is configured to acquire feature vectors corresponding to the object description text and each underwriting type based on the modal feature values ​​of the at least two modalities;

[0231] The underwriting type determination module 1005 is configured to determine the underwriting type of the object description text based on the feature vectors corresponding to the object description text and the respective underwriting types.

[0232] In a possible implementation, the modal data of the at least two modalities include at least two of real number data, category data, and free text data;

[0233] The real number data is data for indicating the real value in the object description text;

[0234] The category data is data for indicating category information contained in the object description text;

[0235] The free text data is a natural language text used to indicate descriptive features in the object description text.

[0236] In a possible implementation, the modal data encoding module 1003 is further configured to:

[0237] In response to the modal data of the at least two modalities including the real number data, obtaining a numerical value corresponding to the real number data in the object description text;

[0238] Average normalization processing is performed on the numerical values ​​corresponding to the real number data in the object description text to obtain the modal eigenvalues ​​corresponding to the real number data in the object description text.

[0239] In a possible implementation, the modal data encoding module 1003 is further configured to:

[0240] In response to the modal data of the at least two modalities including the category data, obtaining a category corresponding to the category data;

[0241] Based on the category corresponding to the category data, the category data is encoded to obtain a modal feature value corresponding to the category data.

[0242] In a possible implementation, the modal data encoding module 1003 is further configured to:

[0243] In response to the modal data of the at least two modalities including the free text data, the free text data is processed based on a pre-trained language model to obtain a modal feature value corresponding to the free text data.

[0244] In a possible implementation, the method further includes:

[0245] An underwriting type display module, used to display the underwriting type of the object description text;

[0246] A related feature display module is used to display the related features of the underwriting type of the object description text; the related features are used to indicate the data in the object description text that has the highest correlation with the underwriting type corresponding to the object description text.

[0247] In a possible implementation, the feature vector acquisition module 1004 is further configured to:

[0248] Based on the modal eigenvalues ​​of the at least two modalities, data processing is performed through a feature transfer layer in an underwriting classification model to obtain feature vectors corresponding to the object description text and each underwriting type;

[0249] The underwriting type determination module 1005 is further configured to:

[0250] Based on the feature vectors corresponding to the object description text and the various underwriting types, data processing is performed through the feature classification layer in the underwriting classification model to obtain the underwriting type of the object description text.

[0251] In one possible implementation, the feature transfer layer includes feature transfer matrices corresponding to the at least two modalities and the respective underwriting types; the feature transfer matrix is ​​used to transfer features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix;

[0252] The feature vector acquisition module 1004 includes:

[0253] a feature transfer unit, configured to perform feature transfer on the modal data of the at least two modalities using a feature transfer matrix corresponding to a first underwriting type and the modal data of the at least two modalities, and a first weight corresponding to the feature transfer matrix, to obtain feature vectors corresponding to the at least two modalities and the first underwriting type, respectively; the first underwriting type is any one of the underwriting types; and the first weight is used to indicate a correlation between the modal data and different underwriting types;

[0254] A feature vector acquisition unit is used to acquire the feature vector corresponding to the object description text and the first underwriting type based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively.

[0255] In a possible implementation, the apparatus further includes:

[0256] A sample object description text acquisition module is used to acquire at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects;

[0257] A sample object data acquisition module, configured to extract sample object data from the sample object description text, wherein the sample features include sample modal data of at least two modalities;

[0258] a sample object data encoding module, configured to encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities, to obtain sample features corresponding to the at least two sample object description texts; the sample features comprising sample modal feature values ​​of the at least two modalities;

[0259] The underwriting classification module update module is used to iteratively update the underwriting classification model through a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain the updated underwriting classification model.

[0260] In a possible implementation, the underwriting classification model updating module is further configured to:

[0261] Obtain sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text; where i ≥ 2 and i is an integer;

[0262] Obtaining first weights corresponding to the i-th sample object description text and each underwriting type based on sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text and feature vectors corresponding to the i-1-th sample object description text and each underwriting type;

[0263] Obtaining a feature vector corresponding to the i-th sample object description text and each underwriting type based on the first weight corresponding to the i-th sample object description text and the sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text;

[0264] Based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, each feature transfer matrix in the underwriting classification model is updated.

[0265] In summary, the solution shown in the embodiment of the present application, in an insurance underwriting scenario, obtains an object description text that describes the insurance-related attributes of a target object, obtains data of at least two modalities in the text and encodes it, and then processes the features obtained after encoding to obtain a feature vector corresponding to the object description text and each underwriting type to determine the underwriting type of the object description text. The above solution, in an insurance scenario, does not require manual review of insurance information. The modal data of different modalities of the object description text is directly used to obtain the feature vector corresponding to the text and each type to determine the type of the object description text, thereby improving the efficiency of insurance underwriting while improving the accuracy of insurance underwriting.

[0266] Figure 11 This is a structural block diagram of a text classification device for insurance underwriting scenarios according to an exemplary embodiment. The text classification device for insurance underwriting scenarios can be implemented by Figure 3 or Figure 4 All or part of the steps in the method provided in the illustrated embodiment, the text classification device for insurance underwriting scenarios includes:

[0267] The sample text acquisition module 1101 is used to acquire at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects;

[0268] A sample modality extraction module 1102 is configured to extract sample object data from the sample object description text, wherein the sample features include sample modality data of at least two modalities;

[0269] A sample feature acquisition module 1103 is configured to encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities, to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities;

[0270] The iterative training module 1104 is used to iteratively update the underwriting classification model through a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain the updated underwriting classification model; the underwriting classification model is used to determine the underwriting type of the object description text based on the modal feature values ​​of at least two modalities of the object description text.

[0271] In summary, the solution shown in the embodiment of the present application, in an insurance underwriting scenario, obtains an object description text that describes the insurance-related attributes of a target object, obtains data of at least two modalities in the text and encodes it, and then processes the features obtained after encoding to obtain a feature vector corresponding to the object description text and each underwriting type to determine the underwriting type of the object description text. The above solution, in an insurance scenario, does not require manual review of insurance information. The modal data of different modalities of the object description text is directly used to obtain the feature vector corresponding to the text and each type to determine the type of the object description text, thereby improving the efficiency of insurance underwriting while improving the accuracy of insurance underwriting.

[0272] Figure 121 is a schematic diagram of the structure of a computer device according to an exemplary embodiment. The computer device can be implemented as a model training device and / or a signal processing device in each of the above-mentioned method embodiments. The computer device 1200 includes a central processing unit (CPU) 1201, a system memory 1204 including a random access memory (RAM) 1202 and a read-only memory (ROM) 1203, and a system bus 1205 connecting the system memory 1204 and the central processing unit 1201. The computer device 1200 also includes a basic input / output system 1206 that helps transmit information between various devices in the computer, and a large-capacity storage device 1207 for storing an operating system 1213, application programs 1214, and other program modules 1215.

[0273] The mass storage device 1207 is connected to the central processing unit 1201 via a mass storage controller (not shown) connected to the system bus 1205. The mass storage device 1207 and its associated computer-readable media provide non-volatile storage for the computer device 1200. In other words, the mass storage device 1207 may include a computer-readable medium (not shown) such as a hard disk or a Compact Disc Read-Only Memory (CD-ROM) drive.

[0274] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, flash memory or other solid-state storage technologies, CD-ROM, or other optical storage, tape cassettes, magnetic tape, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that the computer storage media is not limited to the aforementioned types. The above-mentioned system memory 1204 and mass storage device 1207 may be collectively referred to as memory.

[0275] The computer device 1200 can be connected to the Internet or other network devices through a network interface unit 1211 connected to the system bus 1205 .

[0276] The memory also includes one or more programs, which are stored in the memory. The CPU 1201 executes the one or more programs to implement Figure 2 、 Figure 3 or Figure 4 All or part of the steps of the method shown.

[0277] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is further provided, such as a memory including a computer program (instructions), and the program (instructions) can be executed by a processor of a computer device to perform the methods shown in various embodiments of the present application. For example, the non-transitory computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0278] In an exemplary embodiment, a computer program product or computer program is also provided. The computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods described in the various embodiments above.

[0279] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered merely as exemplary, and the true scope and spirit of the present application are indicated by the claims.

[0280] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A text classification method for insurance underwriting scenarios, characterized by: The method is executed by a computer device, and includes: Obtaining object description text; the object description text is used to describe insurance-related attributes of the target object; extracting modal data of at least two modalities from the object description text; Encoding the modal data of the at least two modalities based on the encoding modes corresponding to the at least two modalities to obtain modal eigenvalues ​​of the at least two modalities; The feature transfer layer in the underwriting classification model includes a feature transfer matrix corresponding to the first underwriting type and the modal data of the at least two modalities, and a first weight corresponding to the feature transfer matrix, and feature transfer is performed on the modal data of the at least two modalities to obtain feature vectors corresponding to the at least two modalities and the first underwriting type respectively; based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, the feature vector corresponding to the object description text and the first underwriting type is obtained; the first underwriting type is any one of the underwriting types; the first weight is used to indicate the correlation between the modal data and different underwriting types; the feature transfer layer includes a feature transfer matrix corresponding to the at least two modalities and each underwriting type respectively; the feature transfer matrix is ​​used to transfer the features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix; The underwriting type of the object description text is determined based on the feature vectors corresponding to the object description text and the respective underwriting types.

2. The method according to claim 1, characterized in that The modal data of the at least two modalities include at least two of real number data, category data, and free text data; The real number data is data for indicating the real value in the object description text; The category data is data for indicating category information contained in the object description text; The free text data is a natural language text used to indicate descriptive features in the object description text.

3. The method according to claim 2, characterized in that The modal data based on the at least two modalities are encoded respectively using encoding methods corresponding to the modalities of the modal data to obtain at least two modal eigenvalues, including: In response to the modal data of the at least two modalities including the real number data, obtaining a numerical value corresponding to the real number data in the object description text; Average normalization processing is performed on the numerical values ​​corresponding to the real number data in the object description text to obtain the modal eigenvalues ​​corresponding to the real number data in the object description text.

4. The method according to claim 2, characterized in that The modal data based on the at least two modalities are encoded respectively using encoding methods corresponding to the modalities of the modal data to obtain at least two modal eigenvalues, including: In response to the modal data of the at least two modalities including the category data, obtaining a category corresponding to the category data; Based on the category corresponding to the category data, the category data is encoded to obtain a modal feature value corresponding to the category data.

5. The method according to claim 2, characterized in that The modal data based on the at least two modalities are encoded respectively using encoding methods corresponding to the modalities of the modal data to obtain at least two modal eigenvalues, including: In response to the modal data of the at least two modalities including the free text data, the free text data is processed based on a pre-trained language model to obtain a modal feature value corresponding to the free text data.

6. The method according to claim 1, characterized in that The method further comprises: Display the underwriting type of the object description text; Display relevant features of the underwriting type of the object description text; the relevant features are used to indicate the data of the object description text that has the highest relevance to the underwriting type corresponding to the object description text.

7. The method according to any one of claims 1 to 6, characterized in that: The determining the underwriting type of the object description text based on the feature vectors corresponding to the object description text and the respective underwriting types includes: Based on the feature vectors corresponding to the object description text and the various underwriting types, data processing is performed through the feature classification layer in the underwriting classification model to obtain the underwriting type of the object description text.

8. The method according to claim 7, characterized in that The method further comprises: Obtain at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects; Extracting sample object data from the sample object description text, wherein the sample object data includes sample modality data of at least two modalities; Encoding the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities; Based on the sample features corresponding to the at least two sample object description texts, the underwriting classification model is iteratively updated through a dynamic routing algorithm to obtain the updated underwriting classification model.

9. The method according to claim 8, characterized in that The iterative updating of the underwriting classification model using a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain the updated underwriting classification model includes: Obtain sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text; where i ≥ 2 and i is an integer; Obtaining first weights corresponding to the i-th sample object description text and each underwriting type based on sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text and feature vectors corresponding to the i-1-th sample object description text and each underwriting type; Obtaining a feature vector corresponding to the i-th sample object description text and each underwriting type based on the first weight corresponding to the i-th sample object description text and the sample modal feature values ​​of at least two modalities corresponding to the i-th sample object description text; Based on the feature vectors corresponding to the i-th sample object description text and each underwriting type, each feature transfer matrix in the underwriting classification model is updated.

10. A text classification method for insurance underwriting scenarios, characterized by: The method is executed by a computer device, and includes: Obtain at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects; Extracting sample object data from the sample object description text, wherein the sample object data includes sample modality data of at least two modalities; Encoding the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities; Iteratively updating the underwriting classification model using a dynamic routing algorithm based on the sample features corresponding to the at least two sample object description texts to obtain an updated underwriting classification model; the underwriting classification model is used to determine the underwriting type of the object description text based on modal feature values ​​of at least two modalities of the object description text; The feature transfer layer in the underwriting classification model includes a feature transfer matrix corresponding to each underwriting type respectively of the at least two modalities; the feature transfer matrix is ​​used to transfer the features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix; wherein, the feature transfer layer is used to perform feature transfer on the modal data of the at least two modalities respectively through the feature transfer matrix corresponding to the first underwriting type and the modal data of the at least two modalities, and the first weight corresponding to the feature transfer matrix, to obtain the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, and based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, obtain the feature vector corresponding to the object description text and the first underwriting type; the first underwriting type is any one of the underwriting types; the first weight is used to indicate the correlation between the modal data and different underwriting types.

11. A text classification device for insurance underwriting scenarios, characterized in that: The device is used for a computer device, and the device includes: A description text acquisition module is used to acquire an object description text; the object description text is used to describe insurance-related attributes of the target object; A modality data extraction module, configured to extract modality data of at least two modalities from the object description text; a modal data encoding module, configured to encode the modal data of the at least two modalities based on encoding modes corresponding to the at least two modalities, to obtain modal eigenvalues ​​of the at least two modalities; A feature vector acquisition module is used to perform feature transfer on the modal data of the at least two modalities through the feature transfer matrix corresponding to the first underwriting type and the modal data of the at least two modalities, and the first weight corresponding to the feature transfer matrix, contained in the feature transfer layer in the underwriting classification model, to obtain the feature vectors corresponding to the at least two modalities and the first underwriting type respectively; based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, obtain the feature vector corresponding to the object description text and the first underwriting type; the first underwriting type is any one of the underwriting types; the first weight is used to indicate the correlation between the modal data and different underwriting types; the feature transfer layer contains the feature transfer matrix corresponding to the at least two modalities and each underwriting type respectively; the feature transfer matrix is ​​used to transfer the features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix; An underwriting type determination module is used to determine the underwriting type of the object description text based on the feature vectors corresponding to the object description text and each underwriting type.

12. A text classification device for insurance underwriting scenarios, characterized in that: The device is used for a computer device, and the device includes: A sample text acquisition module is used to acquire at least two sample object description texts in the first sample training set; the sample object description texts are used to describe insurance-related attributes of the sample objects; A sample modality extraction module, configured to extract sample object data from the sample object description text, wherein the sample object data includes sample modality data of at least two modalities; a sample feature acquisition module, configured to encode the sample object data extracted from the at least two sample object description texts based on the encoding methods corresponding to the at least two modalities, to obtain sample features corresponding to the at least two sample object description texts; the sample features include sample modal feature values ​​of the at least two modalities; an iterative training module, configured to iteratively update the underwriting classification model using a dynamic routing algorithm based on sample features corresponding to the at least two sample object description texts, to obtain an updated underwriting classification model; the underwriting classification model is configured to determine the underwriting type of the object description text based on modal feature values ​​of at least two modalities of the object description text; The feature transfer layer in the underwriting classification model includes a feature transfer matrix corresponding to each underwriting type respectively of the at least two modalities; the feature transfer matrix is ​​used to transfer the features of the corresponding modal data to the hidden layer of the underwriting type corresponding to the feature transfer matrix; wherein, the feature transfer layer is used to perform feature transfer on the modal data of the at least two modalities respectively through the feature transfer matrix corresponding to the first underwriting type and the modal data of the at least two modalities, and the first weight corresponding to the feature transfer matrix, to obtain the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, and based on the feature vectors corresponding to the at least two modalities and the first underwriting type respectively, obtain the feature vector corresponding to the object description text and the first underwriting type; the first underwriting type is any one of the underwriting types; the first weight is used to indicate the correlation between the modal data and different underwriting types.

13. A computer device, characterized in that: A computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the text classification method for insurance underwriting scenarios as described in any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the text classification method for insurance underwriting scenarios as described in any one of claims 1 to 10.

15. A computer program product, characterized in that The computer program product includes computer instructions, which are read and executed by a processor of a computer device, so that the computer device executes the text classification method for insurance underwriting scenarios according to any one of claims 1 to 10.

Images