Insurance claim auditing method based on multi-modal feature extraction and related equipment thereof

Abstract

The invention discloses an insurance claim auditing method based on multi-modal feature extraction and related equipment thereof, and belongs to the technical field of artificial intelligence. The method comprises the steps of obtaining an accident description text and an accident scene picture; performing text feature coding on the accident description text; carrying out image feature extraction on the accident scene picture; calculating feature weight values of the accident text features and the accident image features based on a preset cross attention mechanism; performing feature weighted summation on the accident text features and the accident image features by using the feature weight values to obtain accident multi-modal features; and inputting the accident multi-modal features into a pre-trained insurance claim audit model, and outputting an accident claim audit result. The invention further relates to the technical field of block chains, and claim settlement application materials are stored on a block chain network. Through comprehensive processing of the multi-modal features, the risk of information omission and misjudgment is reduced, the period of claim settlement auditing is shortened, the fairness and transparency of the claim settlement process are improved, and the user experience is improved.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to an insurance claim review method based on multimodal feature extraction and related equipment. Background Art

[0002] Insurance claims review refers to the process of a series of reviews and verifications conducted by the insurance company on the claims filed by the insured or beneficiary after the insurance accident occurs. The claims adjuster will carefully review the insurance accident and the claims filed by the insured or beneficiary according to the insurance contract and relevant laws and regulations to determine whether they meet the scope of insurance liability, and make a decision on whether to pay compensation and how much to pay accordingly.

[0003] In the traditional insurance claim review scheme, the claimant submits text materials, pictures, and videos in various formats as evidence. These materials are often large in quantity and complex in content. Subsequently, the claims adjuster needs to spend a lot of time manually reviewing these relevant materials and extracting and verifying the information item by item for review. This manual processing of claims review is not only inefficient, resulting in a longer claim cycle and increased waiting time for applicants, but also prone to omissions during the long review process, such as missing information and misjudgment. In addition, due to the strong subjectivity of manual review, the review standards may also vary due to differences in experience and ability of claims adjusters, affecting the fairness and accuracy of claims. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose an insurance claim review method, apparatus, computer equipment and storage medium based on multimodal feature extraction to solve the technical problems of low efficiency and easy omissions in traditional insurance claim review solutions.

[0005] In order to solve the above technical problems, the embodiment of the present application provides an insurance claim review method based on multimodal feature extraction, which adopts the following technical solution:

[0006] An insurance claim review method based on multimodal feature extraction, comprising:

[0007] Obtaining the claim application materials submitted by the claimant, wherein the claim application materials include the accident description text and accident scene pictures;

[0008] Encode the text features of the accident description text to obtain the accident text features;

[0009] Extract image features from the accident scene pictures to obtain accident image features;

[0010] Calculate the feature weight values ​​of accident text features and accident image features based on the preset cross-attention mechanism;

[0011] The accident text features and the accident image features are aligned in feature dimensions, and feature weight values ​​are used to perform feature weighted summation on the dimensionally aligned accident text features and accident image features to obtain the accident multimodal features;

[0012] Input the multimodal features of the accident into the pre-trained insurance claim review model and output the accident claim review results.

[0013] Furthermore, the step of encoding the text features of the accident description text to obtain the accident text features specifically includes:

[0014] Perform word segmentation on the accident description text to obtain the accident text segmentation;

[0015] Vectorize the accident text segmentation to obtain the text segmentation vector;

[0016] Perform feature extraction and aggregation on the text word segmentation vector to obtain the accident text features.

[0017] Furthermore, the step of extracting image features from the accident scene picture to obtain accident image features specifically includes:

[0018] Input the accident scene picture into a preset image feature extraction model, wherein the image feature extraction model is constructed based on a residual network, and the image feature extraction model includes a number of convolutional layers of different specifications, each of which is connected to a residual block;

[0019] The accident scene picture is convolved through several convolution layers of different specifications to obtain several initial image features;

[0020] Importing several initial image features into corresponding residual blocks to obtain several accident image feature maps;

[0021] The accident image feature map is used as the accident image feature, and the accident image feature is output.

[0022] Furthermore, the step of calculating the feature weight values ​​of the accident text features and the accident image features based on the preset cross attention mechanism specifically includes:

[0023] Construct the initial query matrix and the initial key-value pair matrix based on the cross-attention mechanism;

[0024] Based on the accident text features, the initial query matrix is ​​updated to obtain a first query matrix, and based on the accident image features, the initial key-value pair matrix is ​​updated to obtain a first key-value pair matrix;

[0025] Calculate a first feature weight value based on the first query matrix and the first key-value pair matrix;

[0026] Based on the accident image features, the initial query matrix is ​​updated to obtain a second query matrix, and based on the accident text features, the initial key-value pair matrix is ​​updated to obtain a second key-value pair matrix;

[0027] A second feature weight value is calculated based on the second query matrix and the first key-value pair matrix.

[0028] Furthermore, the first feature weight value is calculated based on the following formula:

[0029]

[0030] Attention(Q T ,K I ,V I ) is the first feature weight value, Q T is the first query matrix, K I is the key matrix in the first key-value pair matrix, V I is the value matrix in the first key-value pair matrix, d k is the dimension of the key matrix, and softmax is a normalization function, which means normalizing the attention scores of the query matrix on different keys;

[0031] The second feature weight value is calculated based on the following formula:

[0032]

[0033] In the formula, Attention(Q I ,K T ,V T ) is the second feature weight value, Q I is the second query matrix, K T is the key matrix in the second key-value pair matrix, V T is the value matrix in the second key-value pair matrix.

[0034] Further, the step of aligning the feature dimensions of the accident text features and the accident image features, and performing feature weighted summation on the dimensionally aligned accident text features and the accident image features using feature weight values ​​to obtain the multimodal features of the accident specifically includes:

[0035] Mapping accident text features and accident image features into feature spaces of the same dimension;

[0036] Using the first feature weight value to perform feature weighting on the dimensionally aligned accident image features to obtain a first weighted feature;

[0037] Using the second feature weight value to perform feature weighting on the accident text features after dimension alignment, to obtain a second weighted feature;

[0038] The first weighted feature and the second weighted feature are fused to obtain the multimodal feature of the accident.

[0039] Furthermore, before the step of inputting the accident multimodal features into the pre-trained insurance claim review model and outputting the accident claim review result, the method further includes:

[0040] Inputting the accident text features and the accident multimodal features into a preset residual network to obtain a first residual feature;

[0041] Inputting the accident image features and the accident multimodal features into a preset residual network to obtain a second residual feature;

[0042] The feature combination of the first residual feature, the second residual feature and the accident multimodal feature is used as the input of the insurance claim review model.

[0043] In order to solve the above technical problems, the embodiment of the present application further provides an insurance claim review device based on multimodal feature extraction, which adopts the following technical solution:

[0044] An insurance claim review device based on multimodal feature extraction, comprising:

[0045] A material submission module is used to obtain the claim application materials submitted by the claim applicant, wherein the claim application materials include the accident description text and accident scene pictures;

[0046] A text encoding module is used to encode the text features of the accident description text to obtain accident text features;

[0047] An image processing module is used to extract image features from accident scene pictures to obtain accident image features;

[0048] A weight calculation module, used to calculate feature weight values ​​of accident text features and accident image features based on a preset cross-attention mechanism;

[0049] A feature weighting module is used to align the feature dimensions of the accident text features and the accident image features, and use the feature weight value to perform feature weighted summation on the accident text features and the accident image features after dimension alignment to obtain the accident multimodal features;

[0050] The claims review module is used to input the multimodal features of the accident into the pre-trained insurance claims review model and output the accident claims review results.

[0051] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0052] A computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the insurance claim review method based on multimodal feature extraction as described in any one of the above are implemented.

[0053] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0054] A computer-readable storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by a processor, implement the steps of the insurance claim review method based on multimodal feature extraction as described in any one of the above.

[0055] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0056] The present application discloses an insurance claim review method based on multimodal feature extraction and related equipment, which belongs to the field of artificial intelligence technology. The present application obtains the accident description text and the accident scene picture; encodes the text feature of the accident description text; extracts the image feature of the accident scene picture; calculates the feature weight value of the accident text feature and the accident image feature based on the preset cross-attention mechanism; uses the feature weight value to perform feature weighted summation on the accident text feature and the accident image feature to obtain the accident multimodal feature; inputs the accident multimodal feature into the pre-trained insurance claim review model, and outputs the accident claim review result. Compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operation and avoid the problem of inconsistent review standards caused by differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omission and misjudgment, shortens the cycle of claim review, improves the fairness and transparency of the claim process, and improves the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 An exemplary system architecture diagram is shown in which the present application can be applied;

[0059] Figure 2 A flowchart of an embodiment of an insurance claim review method based on multimodal feature extraction according to the present application is shown;

[0060] Figure 3 A structural schematic diagram of an embodiment of an insurance claim review device based on multimodal feature extraction according to the present application is shown;

[0061] Figure 4 A schematic structural diagram of an embodiment of a computer device according to the present application is shown. DETAILED DESCRIPTION

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0063] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0064] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0065] like Figure 1 As shown, the system architecture 100 may include a terminal device 101, a network 102 and a server 103. The terminal device 101 may be a laptop 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is used to provide a medium for a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links or optical fiber cables.

[0066] The user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0067] The terminal device 101 can be any electronic device with a display screen and supporting web browsing. In addition to a laptop computer 1011, a tablet computer 1012 or a mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV), a laptop computer, a desktop computer, etc.

[0068] The server 103 may be a server that provides various services, such as a background server that provides support for a web page displayed on the terminal device 101 .

[0069] It should be noted that the insurance claim review method based on multimodal feature extraction provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the insurance claim review device based on multimodal feature extraction is generally set in the server / terminal device.

[0070] It should be understood that Figure 1 The numbers of terminal devices, networks and servers in the embodiment are merely illustrative. The above system may have any number of terminal devices, networks and servers according to implementation requirements.

[0071] Continue to refer Figure 2 , shows a flow chart of an embodiment of a method for insurance claim review based on multimodal feature extraction according to the present application. The insurance claim review method based on multimodal feature extraction comprises the following steps:

[0072] S201, obtaining claim application materials submitted by the claim applicant, wherein the claim application materials include accident description text and accident scene pictures.

[0073] Specifically, when submitting a claim application, the claimant usually provides a series of materials, including a detailed accident description text and accident scene pictures. The accident description text is the applicant's description of the accident process, loss situation, responsibility sharing, etc., containing detailed information such as time, place, and people. The accident description text is a semantic interpretation of the accident, mainly expressing the cause and effect of the accident through language. The accident scene pictures provide visual evidence and can intuitively show the physical conditions at the scene, such as the degree of damage to the vehicle, the accident scene environment, weather conditions, etc. These two sources of information each provide descriptions of different aspects of the accident, so combining the information from both can fully and accurately understand the accident.

[0074] Acquiring and processing these materials in an automated manner can significantly improve efficiency and reduce errors and omissions caused by manual sorting, summarizing and preliminary analysis.

[0075] S202, performing text feature encoding on the accident description text to obtain accident text features.

[0076] Specifically, the accident description text usually contains a large amount of natural language information, and in order for the system to understand and process this information, the text needs to be converted into a processable feature vector. First, the text can be segmented, the sentences can be decomposed into words or phrases, and the text can be preprocessed by removing stop words, restoring stems, etc. Next, natural language processing technology (such as pre-trained language models such as Word2Vec, BERT, and GPT) is used to encode the text features. These models can convert the semantic information of the text into multi-dimensional feature vectors, capturing the core content and implicit relationships in the text. For example, the BERT model can understand the subtle semantic differences in the text through contextual bidirectional modeling, thereby generating a highly semantic text feature vector. The accident text features not only contain the basic information of the accident, but may also reveal implicit information such as the responsibility for the accident and the applicant's demands.

[0077] S203, extracting image features from the accident scene picture to obtain accident image features.

[0078] Specifically, accident scene pictures usually provide rich visual information in the form of images, such as vehicle damage, weather conditions, road conditions, etc. In order to convert this image information into features that can be used for analysis, it is necessary to use image processing models in deep learning, the most common of which are convolutional neural networks (CNN) or visual transformer (ViT) models. First, the image is extracted through the convolution layer to extract local low-level features (such as edges, textures, etc.), and then aggregated layer by layer into more complex high-level features (such as object shape, scene information). These feature representations can capture the spatial structure and contextual information in the image. In addition, in some special cases, target detection or segmentation techniques can be combined to identify key objects in the image, such as damaged vehicles, personal injuries, etc. Through these models, accident scene pictures are converted into a set of high-dimensional image feature vectors, which can represent the main visual content in the picture in the vector space.

[0079] S204, calculating feature weight values ​​of accident text features and accident image features based on a preset cross-attention mechanism.

[0080] Specifically, in the multimodal processing process, text features and image features often belong to different modes, and direct fusion will lead to information loss or misunderstanding. In order to better mine the relationship between text and image, the cross-attention mechanism is used to calculate the feature weight values ​​corresponding to the accident text features and accident image features.

[0081] In this step, the text features can be used as queries, the image features as keys and values, and the feature weights of the image features can be calculated through the attention mechanism. Then, the image features can be used as queries, the text features as keys and values, and the feature weights of the text features can be calculated through the attention mechanism.

[0082] The core idea of ​​the cross-attention mechanism is to measure the correlation between text features and image features by calculating the dot product between Query and Key, and then derive the corresponding attention weights. This process can help the model identify which image areas have a strong connection with the text content, and then focus on those highly correlated information. This method can avoid ignoring important information during multimodal fusion. For example, in the part of "front vehicle damage" described in the accident description text, the model can find the area related to the "front vehicle" in the picture through the cross-attention mechanism and assign a higher weight. Finally, the feature weight values ​​generated by the cross-attention mechanism can be used for feature fusion, ensuring that the model can make full use of key information in different modalities when processing multimodal data.

[0083] S205, aligning the feature dimensions of the accident text features and the accident image features, and using feature weight values ​​to perform feature weighted summation on the dimensionally aligned accident text features and accident image features to obtain multimodal features of the accident.

[0084] Specifically, after their respective encoding and extraction processes, text features and image features are usually in different vector spaces, and the dimensional sizes may also be different, which will bring difficulties to the subsequent feature fusion. Therefore, the system first needs to align the feature dimensions. Commonly used alignment methods include mapping text and image features to the same feature dimension space through linear mapping, fully connected layers, or projection layers. The aligned features can be calculated in the same dimensional space. This step ensures that the features of different modalities have the same feature dimensions. Next, the feature weight values ​​calculated by the cross-attention mechanism in the previous step are weighted and summed for the features of each modality. In this way, text and image features can be weightedly fused according to their importance in the current scene to obtain a comprehensive multimodal feature. This feature retains the semantic information of the text and the visual information of the image, and ensures that important information is fully expressed through weighted summation.

[0085] S206, input the accident multimodal features into the pre-trained insurance claim review model, and output the accident claim review result.

[0086] Specifically, the fused multimodal features contain comprehensive information from text and images, which can better describe the overall picture of the accident. In this step, these multimodal features are input into a pre-trained insurance claim review model. This review model is usually based on deep learning and has been trained with a large amount of historical claims data. It can automatically judge the rationality of the claim application based on the input multimodal features and output the claim review results. The model involves tasks such as classification, regression or sequence prediction. The specific goals include judging the authenticity of the accident, the rationality of the division of responsibilities, and the reasonable range of the claim amount. The pre-trained model has strong generalization and decision-making capabilities by learning a large amount of case data in advance, and can automatically and quickly complete the review process. Compared with manual review, this automated review method is efficient and consistent, which can avoid the subjectivity and experience bias in manual review, greatly improve the efficiency of claim review, and ensure the consistency and fairness of the review standards.

[0087] In a specific embodiment of the present application, the model architecture of MLP (multilayer perceptron) + CRF (conditional random field) can be used. As a feedforward neural network, MLP can learn complex feature representations through multi-layer nonlinear transformations. Through appropriate activation functions, MLP can capture the nonlinear relationship between input features, and is particularly suitable for processing high-dimensional multimodal features (such as fusion features of text and images). CRF is good at processing sequence data and structured output, and can model the dependencies between input features. For claims review, there may be multiple related judgments (such as accident liability, degree of loss, etc.), and CRF can effectively capture the correlation between these variables and improve the accuracy of prediction.

[0088] The multimodal features of accidents can be input into MLP to learn more abstract and advanced feature representations (i.e., multimodal features of accidents), and finally the feature representations can be input into the CRF module to capture the contextual relationship of the sequence, make full use of global information, and improve the accuracy and consistency of claims review.

[0089] At the same time, a global labeling method is defined in CRF. Through the decoding algorithm, the optimal label sequence can be found under a given observation sequence, realizing the joint optimization of labels. The specific implementation method in the model is as follows:

[0090]

[0091] In the formula, H c is the input feature sequence, Y is the label sequence in the model, Z(Hc ) is the input feature sequence H c The normalization factor of , n is the total number of labels in the label sequence Y, and k is the input feature sequence H c The length, f j is the characteristic function of the jth input feature, used to capture the input feature sequence H c The jth input feature corresponds to the previous label y in the label sequence Y i-1 and the current label y i The relationship between j is the function f j The weight of P(Y|H c ) is the input sequence H c The conditional probability of the label sequence Y.

[0092] In the above embodiment, compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operations and avoid the problem of inconsistent review standards due to differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omissions and misjudgment, shortens the claim review cycle, improves the fairness and transparency of the claims process, and improves the user experience.

[0093] Furthermore, the step of encoding the text features of the accident description text to obtain the accident text features specifically includes:

[0094] Perform word segmentation on the accident description text to obtain the accident text segmentation;

[0095] Vectorize the accident text segmentation to obtain the text segmentation vector;

[0096] Perform feature extraction and aggregation on the text word segmentation vector to obtain the accident text features.

[0097] In this embodiment, the accident description text is segmented, and the long text is divided into words or phrases that can be analyzed independently. In the process of segmentation, the segmentation algorithm in natural language processing (NLP), such as a rule-based segmentation method or a machine learning-based model, can effectively identify the key information in the text. Next, the accident text segmentation is vectorized, and each segmentation is converted into a high-dimensional vector using word embedding technology, such as Word2Vec or GloVe, or a pre-trained model such as BERT. These vectors not only contain the semantic information of the word, but also consider the contextual relationship. After vectorization, the text segmentation vector needs to be feature extracted and aggregated, which can be achieved through different levels of neural networks, such as using convolutional neural networks (CNN) or recurrent neural networks (RNN) to capture important features in the text. Aggregation processing helps to reduce the redundancy of features and improve the expressiveness of features. Finally, through these steps, the accident text features obtained can accurately reflect the core content of the text.

[0098] Through the above steps, the processing of accident texts realizes the effective extraction and expression of information, so that the text content can be understood by the machine in a structured form, which significantly improves the accuracy and efficiency of text analysis.

[0099] Furthermore, the step of extracting image features from the accident scene picture to obtain accident image features specifically includes:

[0100] Input the accident scene picture into a preset image feature extraction model, wherein the image feature extraction model is constructed based on a residual network, and the image feature extraction model includes a number of convolutional layers of different specifications, each of which is connected to a residual block;

[0101] The accident scene picture is convolved through several convolution layers of different specifications to obtain several initial image features;

[0102] Importing several initial image features into corresponding residual blocks to obtain several accident image feature maps;

[0103] The accident image feature map is used as the accident image feature, and the accident image feature is output.

[0104] In this embodiment, the accident scene picture is input into a preset image feature extraction model, which adopts the residual network (ResNet) architecture. The ResNet architecture effectively solves the gradient vanishing problem in deep networks by introducing residual blocks. The model consists of multiple convolutional layers of different specifications, and each convolutional layer is connected to a residual block. Each convolutional layer is responsible for extracting features at different levels. The low-level convolutional layers capture basic features (such as edges and textures), while the high-level convolutional layers aggregate more complex shapes and object information.

[0105] After inputting the accident scene picture, the model performs convolution processing through these convolution layers, extracting the initial image features layer by layer, and then these initial image features are imported into the corresponding residual blocks for processing. The role of the residual block is to add the input features with the features processed by the convolution layer to enhance the model's learning ability and information transfer. With this structure, the model can efficiently learn and retain important image features, reduce information loss and instability during training. Finally, the processed accident image feature map is selected as the accident image feature for output. This method can not only capture the key visual information of the accident scene, but also ensure the effectiveness and accuracy of the model in the feature extraction process.

[0106] Through the above steps, the feature extraction of accident scene pictures achieves deep and detailed visual information capture, significantly improves the efficiency and accuracy of feature extraction, effectively integrates the advantages of convolutional layers and residual blocks, and enhances the expressiveness of image features.

[0107] Furthermore, the step of calculating the feature weight values ​​of the accident text features and the accident image features based on the preset cross attention mechanism specifically includes:

[0108] Construct the initial query matrix and the initial key-value pair matrix based on the cross-attention mechanism;

[0109] Based on the accident text features, the initial query matrix is ​​updated to obtain a first query matrix, and based on the accident image features, the initial key-value pair matrix is ​​updated to obtain a first key-value pair matrix;

[0110] Calculate a first feature weight value based on the first query matrix and the first key-value pair matrix;

[0111] Based on the accident image features, the initial query matrix is ​​updated to obtain a second query matrix, and based on the accident text features, the initial key-value pair matrix is ​​updated to obtain a second key-value pair matrix;

[0112] A second feature weight value is calculated based on the second query matrix and the first key-value pair matrix.

[0113] In this embodiment, the initial query matrix and the initial key-value pair matrix are first constructed. In the cross-attention mechanism, the query, key, and value matrices come from data of different modalities. For example, the initial query matrix is ​​updated based on the accident text features, and the initial key-value pair matrix is ​​updated based on the accident image features. At this time, the feature weight values ​​of the accident image features can be calculated; the initial query matrix is ​​updated based on the accident image features, and the initial key-value pair matrix is ​​updated based on the accident text features. At this time, the feature weight values ​​of the accident text features can be calculated.

[0114] The matrix obtained by the cross-attention calculation is a weight value, which represents the mutual attention between the two modalities (text and image), that is, the relevance of each text feature to each image feature, and vice versa. The larger the attention weight, the higher the attention of the feature to the other modality.

[0115] Based on the accident text features, the initial query matrix is ​​updated to obtain the first query matrix. This update process can be achieved through matrix operations or deep learning networks to enhance the expressive power of text features for queries. At the same time, based on the accident image features, the initial key-value pair matrix is ​​updated to obtain the first key-value pair matrix, thereby ensuring that the image features also have appropriate weights in the interaction. Based on these two matrices, the first feature weight value is calculated, which reflects the importance of the image features compared to the text features in affecting the results.

[0116] In the second round of calculation of the cross-attention mechanism, the initial query matrix is ​​updated based on the accident image features to obtain the second query matrix; at the same time, the initial key-value pair matrix is ​​updated based on the accident text features to obtain the second key-value pair matrix. The core of this process is to let the text features play a leading role in the update process, reflecting the importance of text features in affecting the results compared to image features.

[0117] Finally, the second feature weight value is calculated based on the second query matrix and the first key-value pair matrix to further improve the accuracy and effectiveness of feature fusion.

[0118] Through the above steps, the cross-attention mechanism realizes the effective interaction between text and image features, greatly enhancing the accuracy of multimodal feature fusion. This method not only improves the meticulousness of information capture, but also helps to improve the rationality and fairness of the audit results.

[0119] Furthermore, the first feature weight value is calculated based on the following formula:

[0120]

[0121] Attention(Q T ,K I ,V I ) is the first feature weight value, Q T is the first query matrix, K I is the key matrix in the first key-value pair matrix, V I is the value matrix in the first key-value pair matrix, d k is the dimension of the key matrix, and softmax is a normalization function, which means normalizing the attention scores of the query matrix on different keys;

[0122] The second feature weight value is calculated based on the following formula:

[0123]

[0124] In the formula, Attention(Q I ,K T ,V T ) is the second feature weight value, Q I is the second query matrix, K T is the key matrix in the second key-value pair matrix, V T is the value matrix in the second key-value pair matrix.

[0125] Further, the step of aligning the feature dimensions of the accident text features and the accident image features, and performing feature weighted summation on the dimensionally aligned accident text features and the accident image features using feature weight values ​​to obtain the multimodal features of the accident specifically includes:

[0126] Mapping accident text features and accident image features into feature spaces of the same dimension;

[0127] Using the first feature weight value to perform feature weighting on the dimensionally aligned accident image features to obtain a first weighted feature;

[0128] Using the second feature weight value to perform feature weighting on the accident text features after dimension alignment, to obtain a second weighted feature;

[0129] The first weighted feature and the second weighted feature are fused to obtain the multimodal feature of the accident.

[0130] In this embodiment, mapping the accident text features and the accident image features to the feature space of the same dimension is to achieve multimodal feature fusion. This process involves linear transformation or nonlinear mapping of the features to ensure that the text and image features can be effectively compared and calculated in the same feature space. By using a fully connected layer or dimensionality reduction technology (such as principal component analysis PCA), feature vectors of different dimensions are adjusted to the same dimension.

[0131] Next, the first feature weight value is used to weight the accident image features after dimension alignment to generate the first weighted feature. This process can highlight the important parts of the image related to the text content, so that the model can focus on the visual information that best fits the accident description. Correspondingly, the second feature weight value is used to weight the accident text features after dimension alignment to obtain the second weighted feature. This step is also designed to emphasize the key parts of the text information to ensure that the semantic information of the text is not ignored during feature fusion.

[0132] Finally, the first weighted feature is fused with the second weighted feature to obtain the multimodal feature of the accident, which can be achieved by simple weighted summation, concatenation or more complex fusion methods (such as self-attention mechanism). Through these steps, the final multimodal feature integrates the text information and image information of the accident.

[0133] Through the above steps, effective alignment and weighted fusion of accident text features and image features are achieved, making the model more accurate in processing multimodal information, improving the quality of feature fusion, and enhancing the understanding of accident situations.

[0134] Furthermore, before the step of inputting the accident multimodal features into the pre-trained insurance claim review model and outputting the accident claim review result, the method further includes:

[0135] Inputting the accident text features and the accident multimodal features into a preset residual network to obtain a first residual feature;

[0136] Inputting the accident image features and the accident multimodal features into a preset residual network to obtain a second residual feature;

[0137] The feature combination of the first residual feature, the second residual feature and the accident multimodal feature is used as the input of the insurance claim review model.

[0138] In this embodiment, the accident text features and the accident multimodal features are input into a preset residual network to obtain a first residual feature. The residual network can effectively solve the gradient vanishing problem in the deep network through its unique structure, thereby improving the efficiency and effect of feature learning. In this network, the input text features and multimodal features will be processed through multiple convolutional layers, which not only extract high-level features, but also retain the original information through residual connections to enhance the representation ability of the model. Through such processing, the first residual feature will be able to better reflect the complex relationship between the accident text and multimodal information.

[0139] At the same time, the accident image features and accident multimodal features are input into another preset residual network to generate a second residual feature. This step ensures the full utilization of image information in the deep network, enabling the model to capture the interaction of important visual features and multimodal information in a high-dimensional feature space. Through the output of these two residual networks, the model can obtain two sets of high-quality residual features. Finally, the first residual feature, the second residual feature, and the accident multimodal feature are combined and used as the input of the insurance claim review model.

[0140] Through the above steps, this combination not only integrates the features of text and images, but also combines the residual features extracted from the residual network, which can prevent the model from forgetting text information. At the same time, the use of residual connection operations can reduce the complex mapping that the model needs to learn, help optimize the model, and avoid the problem of gradient disappearance.

[0141] In the above embodiment, the present application discloses an insurance claim review method based on multimodal feature extraction, which belongs to the field of artificial intelligence technology. The present application obtains the accident description text and the accident scene picture; performs text feature encoding on the accident description text; performs image feature extraction on the accident scene picture; calculates the feature weight value of the accident text feature and the accident image feature based on the preset cross-attention mechanism; uses the feature weight value to perform feature weighted summation on the accident text feature and the accident image feature to obtain the accident multimodal feature; inputs the accident multimodal feature into the pre-trained insurance claim review model, and outputs the accident claim review result. Compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operation and avoid the problem of inconsistent review standards due to differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omission and misjudgment, shortens the cycle of claim review, improves the fairness and transparency of the claim process, and improves the user experience.

[0142] In this embodiment, the insurance claim review method based on multimodal feature extraction runs on an electronic device (e.g. Figure 1 The server shown in the figure) can receive instructions or obtain data through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0143] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned claim application materials information, the above-mentioned claim application materials information can also be stored in a blockchain node.

[0144] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0145] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0146] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0147] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0148] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0149] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of an insurance claim review device based on multimodal feature extraction, and the device embodiment is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0150] like Figure 3 As shown, the insurance claim review device 300 based on multimodal feature extraction described in this embodiment includes:

[0151] The material submission module 301 is used to obtain the claim application materials submitted by the claim applicant, wherein the claim application materials include the accident description text and the accident scene picture;

[0152] The text encoding module 302 is used to perform text feature encoding on the accident description text to obtain accident text features;

[0153] The image processing module 303 is used to extract image features from the accident scene picture to obtain accident image features;

[0154] A weight calculation module 304 is used to calculate feature weight values ​​of accident text features and accident image features based on a preset cross-attention mechanism;

[0155] The feature weighting module 305 is used to perform feature dimension alignment on the accident text features and the accident image features, and use the feature weight value to perform feature weighted summation on the dimensionally aligned accident text features and the accident image features to obtain the accident multimodal features;

[0156] The claim review module 306 is used to input the accident multimodal features into the pre-trained insurance claim review model and output the accident claim review results.

[0157] Furthermore, the text encoding module 302 is specifically used for:

[0158] The accident description text is segmented to obtain accident text segmentation; the accident text segmentation is vectorized to obtain text segmentation vectors; feature extraction and aggregation are performed on the text segmentation vectors to obtain accident text features.

[0159] Furthermore, the image processing module 303 is specifically used for:

[0160] The accident scene picture is input into a preset image feature extraction model, wherein the image feature extraction model is constructed based on a residual network, and the image feature extraction model includes a number of convolutional layers of different specifications, each convolutional layer is connected to a residual block; the accident scene picture is convolved through a number of convolutional layers of different specifications to obtain a number of initial image features; the several initial image features are imported into the corresponding residual blocks to obtain a number of accident image feature maps; the accident image feature maps are used as accident image features, and the accident image features are output.

[0161] Furthermore, the weight calculation module 304 is specifically used for:

[0162] An initial query matrix and an initial key-value pair matrix are constructed based on a cross-attention mechanism; the initial query matrix is ​​updated based on accident text features to obtain a first query matrix, and the initial key-value pair matrix is ​​updated based on accident image features to obtain a first key-value pair matrix; a first feature weight value is calculated based on the first query matrix and the first key-value pair matrix; the initial query matrix is ​​updated based on accident image features to obtain a second query matrix, and the initial key-value pair matrix is ​​updated based on accident text features to obtain a second key-value pair matrix; a second feature weight value is calculated based on the second query matrix and the first key-value pair matrix.

[0163] Furthermore, the first feature weight value is calculated based on the following formula:

[0164]

[0165] Attention(Q T ,K I ,V I ) is the first feature weight value, Q T is the first query matrix, K I is the key matrix in the first key-value pair matrix, V I is the value matrix in the first key-value pair matrix, d k is the dimension of the key matrix, and softmax is a normalization function, which means normalizing the attention scores of the query matrix on different keys;

[0166] The second feature weight value is calculated based on the following formula:

[0167]

[0168] In the formula, Attention(Q I ,K T ,V T ) is the second feature weight value, Q I is the second query matrix, K T is the key matrix in the second key-value pair matrix, V T is the value matrix in the second key-value pair matrix.

[0169] Furthermore, the feature weighting module 305 is specifically used for:

[0170] The accident text features and the accident image features are mapped into the feature space of the same dimension; the first feature weight value is used to perform feature weighting on the dimensionally aligned accident image features to obtain the first weighted feature; the second feature weight value is used to perform feature weighting on the dimensionally aligned accident text features to obtain the second weighted feature; the first weighted feature and the second weighted feature are feature fused to obtain the multimodal feature of the accident.

[0171] Furthermore, the insurance claim review device 300 based on multimodal feature extraction also includes a residual processing module, which is specifically used to:

[0172] The accident text features and the accident multimodal features are input into a preset residual network to obtain a first residual feature; the accident image features and the accident multimodal features are input into a preset residual network to obtain a second residual feature; the feature combination of the first residual feature, the second residual feature and the accident multimodal feature is used as the input of the insurance claim review model.

[0173] In the above embodiment, the present application discloses an insurance claim review device based on multimodal feature extraction, which belongs to the field of artificial intelligence technology. The present application obtains the accident description text and the accident scene picture; performs text feature encoding on the accident description text; performs image feature extraction on the accident scene picture; calculates the feature weight value of the accident text feature and the accident image feature based on the preset cross-attention mechanism; uses the feature weight value to perform feature weighted summation on the accident text feature and the accident image feature to obtain the accident multimodal feature; inputs the accident multimodal feature into the pre-trained insurance claim review model, and outputs the accident claim review result. Compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operation and avoid the problem of inconsistent review standards caused by differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omission and misjudgment, shortens the cycle of claim review, improves the fairness and transparency of the claim process, and improves the user experience.

[0174] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0175] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected and communicated through a system bus. It should be noted that the figure only shows a computer device 4 with a memory 41, a processor 42, and a network interface 43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0176] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0177] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of the insurance claim review method based on multimodal feature extraction, etc. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0178] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run the computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the insurance claim review method based on multimodal feature extraction.

[0179] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0180] In the above embodiment, the present application discloses a computer device, which belongs to the field of artificial intelligence technology. The present application obtains the accident description text and the accident scene picture; performs text feature encoding on the accident description text; performs image feature extraction on the accident scene picture; calculates the feature weight values ​​of the accident text features and the accident image features based on the preset cross-attention mechanism; uses the feature weight values ​​to perform feature weighted summation on the accident text features and the accident image features to obtain the accident multimodal features; inputs the accident multimodal features into the pre-trained insurance claim review model, and outputs the accident claim review results. Compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operation and avoid the problem of inconsistent review standards due to differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omission and misjudgment, shortens the cycle of claim review, improves the fairness and transparency of the claim process, and improves the user experience.

[0181] The present application also provides another embodiment, namely, providing a computer-readable storage medium, which stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to enable the at least one processor to perform the steps of the insurance claim review method based on multimodal feature extraction as described above.

[0182] In the above embodiment, the present application discloses a computer-readable storage medium, which belongs to the field of artificial intelligence technology. The present application obtains the accident description text and the accident scene picture; performs text feature encoding on the accident description text; performs image feature extraction on the accident scene picture; calculates the feature weight values ​​of the accident text features and the accident image features based on the preset cross-attention mechanism; uses the feature weight values ​​to perform feature weighted summation on the accident text features and the accident image features to obtain the accident multimodal features; inputs the accident multimodal features into the pre-trained insurance claim review model, and outputs the accident claim review results. Compared with the traditional manual review method, the automated review process can reduce the subjectivity in human operation and avoid the problem of inconsistent review standards due to differences in experience and ability. At the same time, the comprehensive processing of multimodal features, through comprehensive analysis of claim materials, effectively improves the efficiency and accuracy of claim review, greatly reduces the risk of information omission and misjudgment, shortens the cycle of claim review, improves the fairness and transparency of the claim process, and improves the user experience.

[0183] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0184] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0185] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. An insurance claim review method based on multimodal feature extraction, characterized in that: include: Obtaining claim application materials submitted by the claim applicant, wherein the claim application materials include accident description text and accident scene pictures; Performing text feature encoding on the accident description text to obtain accident text features; Extracting image features from the accident scene picture to obtain accident image features; Calculating feature weight values ​​of the accident text feature and the accident image feature based on a preset cross-attention mechanism; Performing feature dimension alignment on the accident text feature and the accident image feature, and performing feature weighted summation on the dimensionally aligned accident text feature and the accident image feature using the feature weight value to obtain an accident multimodal feature; The multimodal features of the accident are input into a pre-trained insurance claim review model, and the accident claim review result is output.

2. The insurance claim review method based on multimodal feature extraction according to claim 1, characterized in that: The step of encoding the text features of the accident description text to obtain the accident text features specifically includes: Performing word segmentation processing on the accident description text to obtain accident text word segmentation; Vectorizing the accident text segmentation to obtain a text segmentation vector; Feature extraction and aggregation processing are performed on the text segmentation vector to obtain the accident text features.

3. The insurance claim review method based on multimodal feature extraction according to claim 1, characterized in that: The step of extracting image features from the accident scene picture to obtain accident image features specifically includes: Inputting the accident scene picture into a preset image feature extraction model, wherein the image feature extraction model is constructed based on a residual network, and the image feature extraction model includes a plurality of convolutional layers of different specifications, each of which is connected to a residual block; Performing convolution processing on the accident scene image through a plurality of convolution layers of different specifications to obtain a plurality of initial image features; Importing a plurality of the initial image features into corresponding residual blocks to obtain a plurality of accident image feature maps; The accident image feature map is used as the accident image feature, and the accident image feature is output.

4. The insurance claim review method based on multimodal feature extraction according to claim 1, characterized in that: The step of calculating the feature weight values ​​of the accident text features and the accident image features based on the preset cross attention mechanism specifically includes: Construct the initial query matrix and the initial key-value pair matrix based on the cross-attention mechanism; updating the initial query matrix based on the accident text features to obtain a first query matrix, and updating the initial key-value pair matrix based on the accident image features to obtain a first key-value pair matrix; Calculate a first feature weight value based on the first query matrix and the first key-value pair matrix; updating the initial query matrix based on the accident image features to obtain a second query matrix, and updating the initial key-value pair matrix based on the accident text features to obtain a second key-value pair matrix; A second feature weight value is calculated based on the second query matrix and the first key-value pair matrix.

5. The insurance claim review method based on multimodal feature extraction according to claim 4, characterized in that: The first feature weight value is calculated based on the following formula: Attention(Q T ,K I ,V I ) is the first feature weight value, Q T is the first query matrix, K I is the key matrix in the first key-value pair matrix, V I is the value matrix in the first key-value pair matrix, d k is the dimension of the key matrix, and softmax is a normalization function, which means normalizing the attention scores of the query matrix on different keys; The second feature weight value is calculated based on the following formula: In the formula, Attention(Q I ,K T ,V T ) is the second feature weight value, Q I is the second query matrix, K T is the key matrix in the second key-value pair matrix, V T is the value matrix in the second key-value pair matrix.

6. The insurance claim review method based on multimodal feature extraction according to claim 4, characterized in that: The step of aligning the feature dimensions of the accident text features and the accident image features, and performing feature weighted summation on the dimensionally aligned accident text features and the accident image features using the feature weight values ​​to obtain the accident multimodal features specifically includes: Mapping the accident text features and the accident image features into a feature space of the same dimension; Using the first feature weight value to perform feature weighting on the dimensionally aligned accident image features to obtain a first weighted feature; Using the second feature weight value to perform feature weighting on the dimensionally aligned accident text feature to obtain a second weighted feature; The first weighted feature and the second weighted feature are fused to obtain the accident multimodal feature.

7. The insurance claim review method based on multimodal feature extraction according to claim 1, characterized in that: Before the step of inputting the accident multimodal features into the pre-trained insurance claim review model and outputting the accident claim review result, the method further includes: Inputting the accident text feature and the accident multimodal feature into a preset residual network to obtain a first residual feature; Inputting the accident image feature and the accident multimodal feature into a preset residual network to obtain a second residual feature; A feature combination of the first residual feature, the second residual feature and the accident multimodal feature is used as an input of the insurance claim review model.

8. An insurance claim review device based on multimodal feature extraction, characterized in that: include: A material submission module is used to obtain the claim application materials submitted by the claim applicant, wherein the claim application materials include accident description text and accident scene pictures; A text encoding module, used for performing text feature encoding on the accident description text to obtain accident text features; An image processing module is used to extract image features from the accident scene picture to obtain accident image features; A weight calculation module, used for calculating feature weight values ​​of the accident text features and the accident image features based on a preset cross-attention mechanism; A feature weighting module, used for aligning the feature dimensions of the accident text features and the accident image features, and performing feature weighted summation on the dimensionally aligned accident text features and the accident image features using the feature weight values ​​to obtain accident multimodal features; The claim review module is used to input the multimodal features of the accident into a pre-trained insurance claim review model and output the accident claim review result.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, it implements the steps of the insurance claim review method based on multimodal feature extraction as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the insurance claim review method based on multimodal feature extraction as described in any one of claims 1 to 7.

Images