Insurance recommendation method and device, equipment and storage medium

Abstract

The invention belongs to the field of artificial intelligence, and relates to an insurance recommendation method, which comprises the following steps: acquiring insurance text information of a to-be-recommended insurance, and extracting to-be-recommended insurance features in the insurance text information; extracting target text information and target image information on the target page information; performing feature extraction on the target image information to obtain target image features; segmenting the target text information to obtain a plurality of target clauses, respectively extracting features of each target clause and fusing the features to obtain target text features; fusing the target text feature and the target image feature to obtain a target fusion feature, and calculating the similarity between the target fusion feature and the to-be-recommended insurance feature to obtain a recommendation value; and performing insurance pushing on the user according to the recommendation value. The invention further provides an insurance recommendation device and equipment and a storage medium. In addition, the invention also relates to a block chain technology, and user browsing information and the like can be stored in a block chain. According to the invention, the insurance pushing accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and finance, and in particular to an insurance recommendation method. Background Art

[0002] With the rapid development of big data technology, information overload has become a major problem facing modern society. As an effective means to deal with information overload, the importance of recommendation systems is becoming increasingly prominent.

[0003] Traditional recommendation systems mainly rely on two types of methods: collaborative filtering and content-based recommendation. However, the collaborative filtering method recommends similar content by analyzing the user's historical behavior, which has the problem of cold start and is not effective for recommending new users or new content.

[0004] The content-based recommendation method mainly recommends content based on text information. In certain scenarios, for example, when recommending the content of a plan PPT, the PPT content often contains a large amount of non-text information and less text information, resulting in content-based recommendations ignoring non-text information and only recommending less text, resulting in a lower recommendation accuracy. Summary of the invention

[0005] The purpose of the embodiments of the present application is to propose an insurance recommendation method, device, equipment and storage medium, the main purpose of which is to improve the accuracy of insurance recommendations.

[0006] In order to solve the above technical problems, the embodiment of the present application provides an insurance recommendation method, which adopts the following technical solution:

[0007] Obtaining insurance text information of the insurance to be recommended, and extracting features of the insurance to be recommended from the insurance text information of the insurance to be recommended using a first text feature extraction model;

[0008] Obtaining the page dwell time of the user in the PPT document, determining the target PPT page information according to the page dwell time, and extracting the target text information and the target image information from the target PPT page information;

[0009] Using a pre-trained image feature extraction model to extract features of the target image information, to obtain target image features;

[0010] Divide the target text information into a plurality of target clauses, extract the features of each target clause using a second text feature extraction model to obtain a plurality of target clause features, and perform weighted fusion on all the target clause features to obtain a target text feature;

[0011] The target text feature and the target image feature are fused to obtain a target fused feature, and a recommendation value of the insurance to be recommended is calculated according to the target fused feature and the feature of the insurance to be recommended;

[0012] If the recommendation value is greater than a preset recommendation threshold, the insurance to be recommended is pushed to the user as the target recommended insurance.

[0013] Furthermore, the extracting features of the insurance to be recommended from the insurance text information of the insurance to be recommended by using the first text feature extraction model includes:

[0014] Vectorizing the insurance text information to obtain an insurance text information vector;

[0015] Performing a convolution operation on the insurance text information vector using multiple convolution kernels of the first text feature extraction model to obtain multiple local initial vectors of the insurance text information vector;

[0016] Using the maximum pooling layer of the first text feature extraction model to screen the multiple local initial vectors to obtain multiple local key vectors;

[0017] The fully connected layer of the first text feature extraction model is used to perform nonlinear transformation and combination operations on the multiple local key vectors to obtain the features of the insurance to be recommended.

[0018] Furthermore, the step of obtaining the page dwell time of the user in the PPT document and determining the target PPT page information according to the page dwell time includes:

[0019] Monitor the switching events in the PPT document;

[0020] Obtaining the timestamp of the switching event, and calculating the page dwell time of the user on each PPT page according to the timestamp to obtain an initial page dwell time set;

[0021] According to a preset residence time threshold, abnormal data of the initial page residence time set is removed to obtain a target page residence time set;

[0022] The PPT page with the longest dwell time is selected from the target page dwell time set as the target PPT page information.

[0023] Furthermore, the use of a pre-trained image feature extraction model to extract features from the target image information to obtain target image features includes:

[0024] Inputting the target image information into the pre-trained image feature extraction model;

[0025] Performing multi-scale feature extraction on the target image information to obtain image features of multiple different scales;

[0026] The multiple image features of different scales are fused to obtain the target image features.

[0027] Furthermore, the second text feature extraction model is used to extract the features of each target clause to obtain a plurality of target clause features, including:

[0028] Performing a preprocessing operation on each of the target clauses to obtain a standard target clause;

[0029] Processing the standard target clause using the second text feature extraction model to obtain a positive order target clause vector and a reverse order target clause vector;

[0030] The forward target clause vector and the reverse target clause vector are subjected to feature fusion to obtain a plurality of target clause features.

[0031] Furthermore, the weighted fusion of all the target clause features to obtain the target text features includes:

[0032] Calculating the attention weight between each of the target clause features and the features of the insurance to be recommended;

[0033] According to the attention weight, each of the target clause features is weightedly summed to obtain the target text feature.

[0034] Furthermore, the calculating of the attention weight between each target clause feature and the feature of the insurance to be recommended includes:

[0035] Performing dot product and difference operations on each of the target clause features and the features of the insurance to be recommended to obtain a dot product vector and a difference vector;

[0036] Concatenate the dot product vector, the difference vector, the target clause feature, and the feature of the insurance to be recommended to obtain a concatenated vector;

[0037] The attention weight between each target clause feature and the feature of the insurance to be recommended is calculated based on the concatenated vector and the preset activation function.

[0038] In order to solve the above technical problems, the embodiment of the present application further provides an insurance recommendation device, which adopts the following technical solution:

[0039] A text feature extraction device, used to obtain insurance text information of the insurance to be recommended, and extract features of the insurance to be recommended from the insurance text information of the insurance to be recommended using a first text feature extraction model;

[0040] A page information acquisition device, used to acquire the page dwell time of the user in the PPT document, determine the target PPT page information according to the page dwell time, and extract the target text information and the target image information from the target PPT page information;

[0041] An image feature extraction device is used to extract features from the target image information using a pre-trained image feature extraction model to obtain target image features;

[0042] A text feature fusion device, used for dividing the target text information into a plurality of target clauses, extracting the features of each target clause using a second text feature extraction model to obtain a plurality of target clause features, and weightedly fusing all the target clause features to obtain a target text feature;

[0043] A target feature fusion device, used for fusing the target text feature and the target image feature to obtain a target fusion feature, and calculating a recommendation value of the insurance to be recommended based on the target fusion feature and the feature of the insurance to be recommended;

[0044] The target insurance pushing device is used to push the insurance to be recommended as the target recommended insurance to the user if the recommendation value is greater than a preset recommendation threshold.

[0045] In order to solve the above technical problems, an embodiment of the present application also provides a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the insurance recommendation method as described above.

[0046] In order to solve the above technical problems, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, the insurance recommendation method as described above is implemented.

[0047] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0048] By accurately extracting key features from insurance texts, the accuracy of subsequent insurance recommendations can be improved; and through automated feature extraction, manual intervention can be avoided, the error rate can be reduced, and the accuracy and efficiency of subsequent recommendations can be improved;

[0049] Accurately locate key content through user browsing behavior and use it for subsequent analysis, which can improve the accuracy of insurance recommendations;

[0050] The target image information is subjected to feature extraction using a pre-trained image feature extraction model to obtain target image features, which can automatically extract the features of the target image information, avoid misidentification caused by manual recognition, and improve the efficiency of feature extraction; and the trained image feature extraction model is used to extract image features of multiple scales and fuse the obtained target image features, which can improve the accuracy and efficiency of subsequent insurance matching;

[0051] By splitting the target text information into multiple target clauses, the problem of missing semantics in long texts is solved, and the accuracy of text analysis is improved. Analysis based on target clauses can capture the local structure and information in the text. This fine-grained processing enables the model to more accurately extract the features of each clause, which improves the accuracy of subsequent insurance recommendations.

[0052] By fusing the target text features and the target image features to obtain the target fusion features, it is possible to utilize information from different modalities, thereby solving the problem of low accuracy caused by only using text information for recommendation in the prior art, and improving the accuracy of insurance recommendations; and according to the target fusion features and the features of the insurance to be recommended, the recommendation value of the insurance to be recommended is calculated, and the recommendation value can be calculated by combining rich features, thereby improving the accuracy of insurance recommendations;

[0053] By calculating the recommendation value, accurate matching is achieved and invalid recommendations are reduced, thereby improving the accuracy and efficiency of insurance recommendations and increasing user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] 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.

[0055] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0056] Figure 2 A flowchart of an embodiment of the insurance recommendation method according to the present application;

[0057] Figure 3 is a structural schematic diagram of an embodiment of an insurance recommendation device according to the present application;

[0058] Figure 4 It is a structural schematic diagram of an embodiment of a device according to the present application. DETAILED DESCRIPTION

[0059] 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.

[0060] 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.

[0061] 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.

[0062] 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.

[0063] 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.

[0064] 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.

[0065] 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 .

[0066] It should be noted that the insurance recommendation method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the insurance recommendation device is generally arranged in the server / terminal device.

[0067] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to the implementation requirements.

[0068] Continue to refer Figure 2 , showing a flow chart of an embodiment of an insurance recommendation method according to the present application.

[0069] In this embodiment, all insurance text information of the insurances to be recommended is obtained from the preset insurance category library, and the insurance features to be recommended of each insurance text information are extracted using the pre-trained text feature extraction model;

[0070] Obtain the PPT page information corresponding to the longest stay time of a user on a certain PPT page when browsing a PPT document, and extract target text information and target image information from the PPT page information; perform feature extraction operations on the target image information using a pre-trained image extraction model to obtain target image features; split the target text information into several target sub-clauses, and use the pre-trained text feature extraction model to extract the features of each target sub-clause to obtain several target sub-clause features; calculate the weights between the target sub-clause features and the insurance features to be recommended through an attention mechanism, and sum all the target sub-clause features according to the weighted sum method based on the weights to obtain target text features; fuse the target text features and the target image features through TFN to obtain target fusion features; calculate the similarity between the target fusion features and the insurance features to be recommended to obtain a recommendation value, and when the recommendation value is greater than a preset threshold, push the insurance to be recommended as the target insurance to the user. The insurance recommendation method includes the following steps:

[0071] S201. Obtain the insurance text information of the insurance to be recommended, and extract the insurance features to be recommended from the insurance text information of the insurance to be recommended using a first text feature extraction model.

[0072] In this embodiment, the above-mentioned insurance to be recommended refers to one of the insurance types in the preset insurance type library; the above-mentioned first text feature extraction model refers to a pre-trained convolutional neural network (CNN), and the convolutional neural network consists of an input layer, several convolutional layers, a max pooling layer, a fully connected layer, and an output layer.

[0073] In this embodiment, obtain all the insurance types from the preset insurance type library, and use each insurance type as the insurance to be recommended. Taking one insurance type as the insurance to be recommended as an example, obtain the insurance text information of the insurance to be recommended, and perform preprocessing on the insurance text information. The preprocessing operations include removing irrelevant characters (such as punctuation marks, etc.), removing stop words (such as "le", "de", etc.), and word segmentation operations to obtain the preprocessed insurance text information;

[0074] Input the preprocessed insurance text information into the first text feature extraction model, use the first text feature extraction model to extract the text features of the insurance text information, and use the text features as the insurance features to be recommended.

[0075] In one embodiment, the step of extracting the insurance features to be recommended from the insurance text information of the insurance to be recommended using the first text feature extraction model includes:

[0076] Vectorize the insurance text information to obtain an insurance text information vector;

[0077] Performing a convolution operation on the insurance text information vector using multiple convolution kernels of the first text feature extraction model to obtain multiple local initial vectors of the insurance text information vector;

[0078] Using the maximum pooling layer of the first text feature extraction model to screen the multiple local initial vectors to obtain multiple local key vectors;

[0079] The fully connected layer of the first text feature extraction model is used to perform nonlinear transformation and combination operations on multiple local key vectors to obtain the features of the insurance to be recommended.

[0080] In this embodiment, it is described above that the first text feature extraction model is composed of an input layer, several convolutional layers, a maximum pooling layer, a fully connected layer and an output layer; first, an embedding technology (such as embedding technology) is used to vectorize the insurance text information to obtain an insurance text information vector, and the insurance text information vector is input into the first text feature extraction model through the input layer, and a plurality of convolutional layers are used to perform convolution operations on the insurance text information vector to extract a plurality of local initial feature vectors, such as "insurance cost", "insurance coverage", etc.; the plurality of local initial feature vectors are processed through the maximum pooling layer to reduce the dimensions of the plurality of local initial vectors and remove redundant information, and then the most important features are screened out to obtain a plurality of local key vectors; the plurality of key vectors are nonlinearly transformed and combined through the fully connected layer to obtain the insurance features to be recommended, and the insurance features to be recommended are output through the output layer.

[0081] In this embodiment, through the combination of multiple convolutional layers and maximum pooling layers, the key features of insurance text information can be automatically captured, thereby improving the accuracy of subsequent recommendations for insurance.

[0082] In an implementable example A, there are 10 types of insurance in the preset insurance type library. The insurance text information corresponding to each insurance type is extracted respectively, and the text features of each insurance text information are extracted through a pre-trained text feature extraction model, and each text feature is used as a feature to be recommended.

[0083] In another embodiment, before extracting the features of the insurance to be recommended from the insurance text information of the insurance to be recommended using the first text feature extraction model, the method further includes:

[0084] Obtain an insurance text dataset, and divide the insurance text dataset into a training set and a test set according to a preset ratio;

[0085] Constructing an initial text feature extraction model, and using the training set to train the initial text feature extraction model to obtain an intermediate text feature extraction model;

[0086] Testing the intermediate text feature extraction model using a test set to obtain a loss value;

[0087] When the loss value is less than or equal to a preset loss threshold, obtaining the first text feature extraction model;

[0088] When the loss value is greater than the preset loss threshold, the initial text feature extraction model is continuously trained using the training set until the loss value is less than or equal to the preset loss threshold, and the training is completed to obtain the first text feature extraction model.

[0089] In this embodiment, a large number of insurance text data sets are obtained, and the large number of insurance text data sets are split into training sets and test sets according to a preset ratio; the initial text feature extraction model is trained by the training set to obtain an intermediate text feature extraction model; the intermediate text feature extraction model is tested by using the test set to obtain a loss value, and it is determined whether the loss value is less than or equal to a preset loss threshold. When the loss value is less than or equal to the preset loss threshold, the training is completed to obtain a first text feature extraction model. When the loss value is greater than the preset loss threshold, the convolutional neural network model is continuously trained by using the training set until the loss value is less than or equal to the preset loss threshold, the training is terminated to obtain the first text feature extraction model.

[0090] In this embodiment, by accurately extracting key features from the insurance text, the accuracy of subsequent insurance recommendations can be improved; and through automated feature extraction, human intervention is avoided, the error rate is reduced, and the accuracy and efficiency of subsequent recommendations are improved.

[0091] S202, obtaining the page dwell time of the user in the PPT document, determining target PPT page information according to the page dwell time, and extracting target text information and target image information from the target PPT page information.

[0092] In this embodiment, the above-mentioned PPT document includes multiple PPT pages, each PPT page includes text information and image information related to insurance, the text information refers to text information related to insurance, such as PPT text information, PPT title and other contents, and the image information refers to insurance-related charts, etc.

[0093] In this embodiment, when a user browses a PPT document through a terminal, the user's dwell time on each PPT page in the PPT document is obtained, the target PPT page information is determined based on the dwell time, and the target text information and target image information are extracted from the PPT page information, wherein the terminal includes a mobile terminal (mobile phone, iPad, etc.) and a fixed terminal (computer, etc.).

[0094] In one embodiment, the step of obtaining the page dwell time of the user in the PPT document and determining the target PPT page information according to the page dwell time includes:

[0095] Monitor the switching events in the PPT document;

[0096] Obtaining the timestamp of the switching event, and calculating the page dwell time of the user on each PPT page according to the timestamp to obtain an initial page dwell time set;

[0097] According to a preset residence time threshold, abnormal data of the initial page residence time set is removed to obtain a target page residence time set;

[0098] The PPT page with the longest dwell time is selected from the target page dwell time set as the target PPT page information.

[0099] In this embodiment, a monitoring script (such as JavaScript, Python, etc.) is embedded in the PPT document in advance, and the switching events in the PPT document are monitored by the monitoring script. The switching event refers to an event triggered by the user switching to the next page or the previous page. The timestamp of each switching event is recorded, and the user's stay time on each page is calculated based on the timestamp. All the stay times are collected to obtain the initial page stay time set;

[0100] In order to prevent abnormal data generated by situations such as users leaving the page but not closing it, a residence time threshold is designed in advance. The residence time threshold refers to an interval range. The data in the initial page residence time set that is not within the residence time threshold is removed as abnormal data to obtain the target page residence time set. The residence time of each page in the target page residence time set is sorted, and the PPT page corresponding to the longest page residence time is selected as the target PPT page information.

[0101] In this embodiment, by determining the target PPT page information based on the user's stay time on the PPT page, it is possible to accurately understand what insurance content the user is interested in, and by analyzing the PPT page information that the user is interested in, the accuracy of subsequent insurance recommendations can be improved.

[0102] Continuing with the above-mentioned feasible example A, the dwell time of the user on each PPT page when browsing the PPT document is monitored, all the dwell time is filtered using a preset time threshold, the dwell time with the longest duration is selected from the filtered dwell time set, and the PPT page corresponding to the longest dwell time is used as the target PPT page, and the target text information and target image information are obtained from the PPT page.

[0103] In this embodiment, key content is accurately located through the user's browsing behavior, and subsequent analysis is based on the key content, which can improve the accuracy of insurance recommendations.

[0104] S203: Using a pre-trained image feature extraction model to extract features of the target image information to obtain target image features.

[0105] In this embodiment, the pre-trained image feature extraction model refers to a ConvNeXt model, which is a variant of a convolutional neural network (CNN), and consists of an input layer, a deep separable convolution layer (Depthwise Separable Convolution Layer), a layer normalization layer (Layer Normalization Layer), a GELU activation function layer (Gaussian Error Linear Unit Activation Function Layer) and an output layer, wherein the deep separable convolution layer uses a large convolution kernel (such as a 7*7 convolution kernel), and each convolution block uses a residual connection.

[0106] In this embodiment, a preprocessing operation is performed on the target image information, and the preprocessed target image information is input into a pre-trained image feature extraction model, and the features of the pre-processed target image information are extracted by the pre-trained image feature extraction model to obtain the target image features.

[0107] The above-mentioned preprocessing operation includes image denoising, grayscale processing, image enhancement and normalization processing on the target image information to obtain the preprocessed target image information; wherein, the image denoising method is used to reduce the noise of the target image information and improve the image quality; the grayscale processing method is used to process the target image information, which can reduce the calculation amount of subsequent processing and improve the efficiency and accuracy of subsequent recognition steps; the image enhancement method is used to improve the contrast and clarity of the target image information, so that the information in the target image information is more prominent, which is convenient for the subsequent feature extraction step; the normalization processing method can convert the target image information into a numerical range that can be recognized by the pre-trained image feature extraction model, avoiding the problem of subsequent unrecognition.

[0108] Specifically, the target image information is denoised by using Gaussian filtering, each pixel in the target image information is scanned, and the value of the central pixel of the template is replaced by the weighted average grayscale value of the pixels in the area determined by the template to obtain the denoised target image information; each pixel of the color target image information is converted from the RGB color space to the grayscale space to obtain the target image information after grayscale processing; the pixel distribution of the target image information is adjusted by using the histogram equalization function to obtain the target image information after image enhancement; the target image information is normalized by using the image normalization method to complete the preprocessing steps.

[0109] In one embodiment, before extracting features from the target image information using a pre-trained image feature extraction model to obtain target image features, the method further includes:

[0110] Obtain an image data set, and divide the image data set into a training set and a test set according to a preset ratio;

[0111] Constructing an initial image feature extraction model, and training the initial image feature extraction model using the training set to obtain an intermediate image feature extraction model;

[0112] Using the test set to test the intermediate image feature extraction model to obtain a loss value;

[0113] When the loss value is less than or equal to a preset loss threshold, a trained image feature extraction model is obtained.

[0114] In this embodiment, a large amount of image data is obtained and compiled into an image data set, and the image data set is split according to a preset ratio to obtain a training set and a test set; an initial image feature extraction model is constructed, the initial image feature extraction model is trained by the training set to obtain an intermediate image feature extraction model, the intermediate image feature extraction model is tested by the test set to obtain a loss value, and it is determined whether the loss value is less than or equal to a preset loss threshold. When the loss value is less than or equal to the preset loss threshold, the training is completed to obtain a trained image feature extraction model. When the loss value is greater than the preset loss threshold, the initial image feature extraction model is continued to be trained using the training set until the loss value is less than or equal to the preset loss threshold, the training is terminated to obtain a trained image feature extraction model.

[0115] In one embodiment, the step of extracting features from the target image information using a pre-trained image feature extraction model to obtain target image features includes:

[0116] Inputting the target image information into the pre-trained image feature extraction model;

[0117] Performing multi-scale feature extraction on the target image information to obtain image features of multiple different scales;

[0118] The multiple image features of different scales are fused to obtain the target image features.

[0119] In this embodiment, the pre-trained image feature extraction model mentioned above includes an input layer, a deep separable convolution layer, a layer normalization layer, a GELU activation function layer, a fully connected layer and an output layer. The target image information is input into the model through the input layer, and the target image information is convolved using the deep separable convolution layer. The convolution operation includes: extracting features of the target image information through a larger convolution kernel and a smaller number of channels. After the feature extraction is completed, the layer normalization layer is entered for normalization. In the next stage, a smaller convolution kernel and a larger number of channels than the previous stage are used to extract features of the target image information to obtain features of multiple scales. Finally, the features of multiple scales are fused through the fully connected layer to obtain the target image features, and the target image features are output through the output layer. The layer normalization layer normalizes the output of each convolution layer, and the RELU activation function layer performs a nonlinear transformation on the normalized activation value.

[0120] In this embodiment, a pre-trained image feature extraction model is used to extract multi-scale features of the target image, which can capture features of different scales and levels and improve the accuracy of subsequent insurance recommendation matching.

[0121] Continuing with the above-mentioned feasible example A, the target image information obtained from the PPT page is input into the pre-trained model, and the target image information is subjected to multiple convolution operations through the model to finally obtain image features of multiple scales, and the image features of multiple scales are fused to finally obtain the target image features.

[0122] In this embodiment, a pre-trained image feature extraction model is used to perform feature extraction on the target image information to obtain target image features, which can automatically extract the features of the target image information, avoid misidentification caused by manual recognition, and improve the efficiency of feature extraction; and the trained image feature extraction model is used to extract image features of multiple scales and fuse the obtained target image features, which can improve the accuracy and efficiency of subsequent insurance matching.

[0123] S204, dividing the target text information into a plurality of target clauses, extracting the features of each target clause using a second text feature extraction model, obtaining a plurality of target clause features, and fusing all the target clause features to obtain target text features.

[0124] In this embodiment, the second text feature extraction model refers to a Bi-LSTM (bidirectional long short-term memory network) model. The Bi-LSTM model combines the bidirectional model and the gating mechanism of LSTM, and is composed of two independent LSTM networks. When the Bi-LSTM model processes data, it will be input into the two LSTM networks in forward and reverse order to perform feature extraction, and output two output vectors. The two data vectors are concatenated to finally obtain the target vector.

[0125] Since the model recognition has a maximum character limit, if the target text information is directly encoded as a single sentence, information will be lost; therefore, the target text information needs to be segmented to avoid semantic loss and improve the accuracy of target text analysis.

[0126] In this embodiment, it is mentioned in step S202 that the target text information contains multiple text information, so the text information is first concatenated to obtain a complete text string, and a punctuation segmentation method is used to segment the text string according to preset punctuation marks such as periods, question marks, exclamation marks, etc. to obtain several target clauses. For each target clause, feature extraction is performed using the second text feature extraction model to obtain several target clause features, and the weight corresponding to each target clause feature is calculated, and the final target text feature is obtained by summing the weights.

[0127] In one embodiment, the feature of each target clause is extracted using the second text feature extraction model to obtain a plurality of target clause features, including:

[0128] Performing a preprocessing operation on each of the target clauses to obtain a standard target clause;

[0129] Processing the standard target clause using the second text feature extraction model to obtain a positive order target clause vector and a reverse order target clause vector;

[0130] The forward target clause vector and the reverse target clause vector are subjected to feature fusion to obtain a plurality of target clause features.

[0131] In this embodiment, a preprocessing operation is performed on each of the target clauses to obtain a standard target clause, wherein the preprocessing operation includes but is not limited to word segmentation processing and vectorization processing (through word embedding technology, etc.), and each standard target clause is input into the second text feature extraction model, and each standard target clause is processed by a forward LSTM layer and a reverse LSTM layer respectively, and a forward target clause vector and a reverse target clause vector are output, and the forward target clause vector and the reverse target clause vector are finally obtained to obtain a plurality of target clause features.

[0132] In this embodiment, the model processes each target clause in forward and reverse order, which can capture information in different directions within the clause, and then fuses the forward target clause vector and the reverse target clause vector through the feature fusion step to obtain several target clause features, forming a richer feature representation, which improves the accuracy of subsequent insurance recommendation matching.

[0133] In one embodiment, all the target clause features are merged, and the target text features include:

[0134] Calculating the attention weight between each of the target clause features and the features of the insurance to be recommended;

[0135] According to the attention weight, each of the target clause features is weightedly summed to obtain the target text feature.

[0136] In this embodiment, the attention weight between each target clause feature and the feature of the insurance to be recommended is calculated by the Attention Unit, so that each target clause feature has a corresponding attention weight. According to the attention weight corresponding to each clause feature, all target clause features are weighted and summed to obtain the target text feature.

[0137] In this embodiment, by calculating the attention weight, the model can identify that the target clause features are more relevant to the features of the insurance to be recommended, avoiding equal weight treatment of all target clause features and improving the accuracy of subsequent insurance to be recommended.

[0138] In another embodiment, the calculating the attention weight between each of the target clause features and the features of the insurance to be recommended includes:

[0139] Performing dot product and difference operations on each of the target clause features and the features of the insurance to be recommended to obtain a dot product vector and a difference vector;

[0140] Concatenate the dot product vector, the difference vector, the target clause feature, and the feature of the insurance to be recommended to obtain a concatenated vector;

[0141] The attention weight between each target clause feature and the feature of the insurance to be recommended is calculated based on the concatenated vector and the preset activation function.

[0142] In this embodiment, taking a target clause as an example, the target clause feature is u, and the feature of the insurance to be recommended is v. The target clause feature is subjected to dot product and difference operations with the feature of the insurance to be recommended to obtain a dot product vector u*v and a difference vector uv. u, v, u*v, and uv are concatenated to obtain a concatenated vector l[u, v, u*v, uv]. Based on the concatenated vector and the ReLU activation function, the attention weight between the target clause feature and the feature of the insurance to be recommended is calculated.

[0143] The following is a calculation formula for the attention weight between each target clause feature and the insurance feature to be recommended:

[0144] The output vector formula of the fully connected layer is:

[0145] z=ω*l+b

[0146] Among them, l refers to the concatenation vector [u, v, u*v, uv], z refers to the output vector of the fully connected layer, ω refers to the weight matrix of the fully connected layer, and b refers to the bias vector of the fully connected layer.

[0147] The formula for the activation function output is:

[0148] a=σ(z)

[0149] Here, a refers to the output of the activation function, σ refers to the activation function, and z refers to the output vector of the fully connected layer.

[0150] The output a of the activation function is output to a linear classification layer, and the processing formula of the linear classification layer is:

[0151] output_weights=W′*a+b′

[0152] Among them, output_weights is the attention weight, W′ is the weight matrix of the linear classification layer, and b′ is the bias vector of the linear classification layer.

[0153] In this embodiment, through dot product and difference operations, the correlation and difference between the target clause features and the features of the insurance to be recommended can be captured; by concatenating all the vectors, a rich feature vector can be obtained, thereby improving the accuracy of subsequent weight calculations; by calculating the attention weights, the model can more accurately identify which target clause features are most relevant to the features of the insurance to be recommended, thereby improving the accuracy of subsequent insurance recommendations.

[0154] Continuing with the above-mentioned feasible example A, the text content, text title, etc. obtained from the PPT page are all concatenated together to obtain a text string, and then they are split according to predefined punctuation marks to obtain several target clauses, and the features of each target clause are extracted to obtain several target clause features. The attention weight between each target clause feature and the feature of the insurance to be recommended is calculated through the attention mechanism, and according to the attention weight corresponding to each target clause, all target clause features are fused to obtain the target text feature.

[0155] In this embodiment, by splitting the target text information into multiple target clauses, the problem of missing semantics in long texts is solved and the accuracy of text analysis is improved. Analysis based on target clauses can capture the local structure and information in the text. This fine-grained processing enables the model to more accurately extract the features of each clause, thereby improving the accuracy of subsequent insurance recommendations.

[0156] S205: Fusing the target text feature and the target image feature to obtain a target fused feature, and calculating a recommendation value of the insurance to be recommended based on the target fused feature and the feature of the insurance to be recommended.

[0157] In this embodiment, the target text features and the target image features are fused through a tensor fusion network (TFN) to obtain a target fusion feature, and the similarity between the target fusion feature and the feature of the insurance to be recommended is calculated through a preset similarity calculation method to obtain the recommendation value of the insurance to be recommended. The tensor fusion network is composed of a modal embedding subnetwork, a tensor fusion layer, and a fully connected layer; the tensor fusion network is a multimodal fusion method specifically used to process data from different modalities (such as images and texts) and fuse them into a unified feature representation.

[0158] In this embodiment, the fusion method for fusing the target text feature and the target image feature is:

[0159] The target text features and the target image features are respectively input into the modal embedding subnetwork of TFN, and the target text features and the target image features are further processed and transformed by the modal embedding subnetwork to generate richer modal embedding vectors; the modal embedding vectors of the target text features and the target image vector are input into the tensor fusion layer, and the modal embedding vectors of the target text features and the target image vector are subjected to outer product operation by the tensor fusion layer to obtain the target fusion tensor; the target fusion features are extracted according to the target fusion tensor.

[0160] In one embodiment, the step of calculating the recommendation value of the insurance to be recommended according to the target fusion feature and the feature of the insurance to be recommended includes:

[0161] Calculate the similarity value between the target fusion feature and the insurance feature to be recommended by using a preset similarity calculation method;

[0162] The similarity value is used as the recommendation value of the insurance to be recommended.

[0163] In this embodiment, the preset similarity calculation method includes but is not limited to cosine similarity (CosineSimilarity), Euclidean Distance (Euclidean Distance), etc. The similarity calculation method is used to calculate the similarity of the target fusion feature and the feature of the insurance to be recommended to obtain a similarity value, and the similarity value is used as the recommendation value of the insurance to be recommended.

[0164] Continuing with the above-mentioned feasible example A, the target image features and the target text features are tensor fused to obtain the target fused features. The similarity between the target fused features and the features of the insurance to be recommended is calculated by the cosine similarity calculation method to obtain the similarity value and use the similarity value as the recommendation value of the corresponding insurance to be recommended.

[0165] In this embodiment, by fusing the target text features and the target image features, a target fusion feature is obtained, which can utilize information from different modalities, thereby solving the problem of low accuracy caused by the existing recommendation using only text information, and improving the accuracy of insurance recommendations; and based on the target fusion feature and the features of the insurance to be recommended, the recommendation value of the insurance to be recommended is calculated, and the recommendation value can be calculated in combination with rich features, thereby improving the accuracy of insurance recommendations.

[0166] S206: If the recommendation value is greater than a preset recommendation threshold, the insurance to be recommended is pushed to the user as a target recommended insurance.

[0167] In this embodiment, two push solutions are designed, one of which is:

[0168] A preset recommendation threshold is set in advance, and the recommendation value is compared with the preset recommendation threshold. If the recommendation value is greater than the preset recommendation threshold, the insurance to be recommended is pushed to the user as the target recommended insurance; if the recommendation value is less than or equal to the preset recommendation threshold, it is determined that the insurance to be recommended is not needed by the user and is not recommended.

[0169] Second:

[0170] A preset recommendation threshold is set in advance, the recommendation value is compared with the preset recommendation threshold, and the recommended insurances corresponding to the recommended values ​​less than or equal to the preset recommendation threshold are excluded, and all the recommended insurances greater than the preset recommendation threshold are sorted, and the highest ranked recommended insurance is selected as the target recommended insurance and pushed to the user.

[0171] In this embodiment, by calculating the recommendation value, accurate matching is achieved, and invalid recommendations are reduced, thereby improving the accuracy and efficiency of insurance recommendations and improving user satisfaction.

[0172] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned user browsing information, the above-mentioned user browsing information, etc. can also be stored in a node of a blockchain.

[0173] 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.

[0174] 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.

[0175] 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.

[0176] 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).

[0177] 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.

[0178] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of an insurance recommendation device, and the device embodiment is Figure 2 Corresponding to the method embodiment shown, the apparatus can be specifically applied to various computer devices.

[0179] like Figure 3 As shown, the insurance recommendation device 300 described in this embodiment includes: a text feature extraction device 301, a page information acquisition device 302, an image feature extraction device 303, a text feature fusion device 304, a target feature fusion device 305, and a target insurance push device 306. Among them:

[0180] The text feature extraction device 301 is used to obtain the insurance text information of the insurance to be recommended, and extract the features of the insurance to be recommended from the insurance text information of the insurance to be recommended using the first text feature extraction model;

[0181] In one embodiment, the extracting the features of the insurance to be recommended from the insurance text information of the insurance to be recommended by using the first text feature extraction model includes:

[0182] Vectorizing the insurance text information to obtain an insurance text information vector;

[0183] Performing a convolution operation on the insurance text information vector using multiple convolution kernels of the first text feature extraction model to obtain multiple local initial vectors of the insurance text information vector;

[0184] Using the maximum pooling layer of the first text feature extraction model to screen the multiple local initial vectors to obtain multiple local key vectors;

[0185] The fully connected layer of the first text feature extraction model is used to perform nonlinear transformation and combination operations on multiple local key vectors to obtain the features of the insurance to be recommended.

[0186] The page information acquisition means 302 is used to acquire the page dwell time of the user in the PPT document, determine the target PPT page information according to the page dwell time, and extract the target text information and the target image information from the target PPT page information;

[0187] In one embodiment, the step of obtaining the page dwell time of the user in the PPT document and determining the target PPT page information according to the page dwell time includes:

[0188] Monitor the switching events in the PPT document;

[0189] Obtaining the timestamp of the switching event, and calculating the page dwell time of the user on each PPT page according to the timestamp to obtain an initial page dwell time set;

[0190] According to a preset residence time threshold, abnormal data of the initial page residence time set is removed to obtain a target page residence time set;

[0191] The PPT page with the longest dwell time is selected from the target page dwell time set as the target PPT page information.

[0192] The image feature extraction device 303 is used to extract features from the target image information using a pre-trained image feature extraction model to obtain target image features;

[0193] In one embodiment, the step of extracting features from the target image information using a pre-trained image feature extraction model to obtain target image features includes:

[0194] Inputting the target image information into the pre-trained image feature extraction model;

[0195] Performing multi-scale feature extraction on the target image information to obtain image features of multiple different scales;

[0196] The multiple image features of different scales are fused to obtain the target image features.

[0197] The text feature fusion device 304 is used to divide the target text information into a plurality of target clauses, extract the features of each target clause using the second text feature extraction model to obtain a plurality of target clause features, and perform weighted fusion on all the target clause features to obtain the target text feature;

[0198] In one embodiment, the feature of each target clause is extracted using the second text feature extraction model to obtain a plurality of target clause features, including:

[0199] Performing a preprocessing operation on each of the target clauses to obtain a standard target clause;

[0200] Processing the standard target clause using the second text feature extraction model to obtain a positive order target clause vector and a reverse order target clause vector;

[0201] The forward target clause vector and the reverse target clause vector are subjected to feature fusion to obtain a plurality of target clause features.

[0202] In one embodiment, all the target clause features are merged, and the target text features include:

[0203] Calculating the attention weight between each of the target clause features and the features of the insurance to be recommended;

[0204] According to the attention weight, each of the target clause features is weightedly summed to obtain the target text feature.

[0205] In one embodiment, the calculating of the attention weight between each target clause feature and the feature of the insurance to be recommended includes:

[0206] Performing dot product and difference operations on each of the target clause features and the features of the insurance to be recommended to obtain a dot product vector and a difference vector;

[0207] Concatenate the dot product vector, the difference vector, the target clause feature, and the feature of the insurance to be recommended to obtain a concatenated vector;

[0208] The attention weight between each target clause feature and the feature of the insurance to be recommended is calculated based on the concatenated vector and the preset activation function.

[0209] The target feature fusion device 305 is used to fuse the target text feature and the target image feature to obtain a target fusion feature, and calculate the recommendation value of the insurance to be recommended according to the target fusion feature and the feature of the insurance to be recommended;

[0210] The target insurance pushing device 306 is used to push the insurance to be recommended as the target recommended insurance to the user if the recommendation value is greater than a preset recommendation threshold.

[0211] In this embodiment, by accurately extracting key features from insurance texts, the accuracy of subsequent insurance recommendations can be improved; and by automated feature extraction, manual intervention is avoided, the error rate is reduced, and the accuracy and efficiency of subsequent recommendations are improved;

[0212] Accurately locate key content through user browsing behavior and use it for subsequent analysis, which can improve the accuracy of insurance recommendations;

[0213] The target image information is subjected to feature extraction using a pre-trained image feature extraction model to obtain target image features, which can automatically extract the features of the target image information, avoid misidentification caused by manual recognition, and improve the efficiency of feature extraction; and the trained image feature extraction model is used to extract image features of multiple scales and fuse the obtained target image features, which can improve the accuracy and efficiency of subsequent insurance matching;

[0214] By splitting the target text information into multiple target clauses, the problem of missing semantics in long texts is solved, and the accuracy of text analysis is improved. Analysis based on target clauses can capture the local structure and information in the text. This fine-grained processing enables the model to more accurately extract the features of each clause, which improves the accuracy of subsequent insurance recommendations.

[0215] By fusing the target text features and the target image features to obtain the target fusion features, it is possible to utilize information from different modalities, thereby solving the problem of low accuracy caused by only using text information for recommendation in the prior art, and improving the accuracy of insurance recommendations; and according to the target fusion features and the features of the insurance to be recommended, the recommendation value of the insurance to be recommended is calculated, and the recommendation value can be calculated by combining rich features, thereby improving the accuracy of insurance recommendations;

[0216] By calculating the recommendation value, accurate matching is achieved and invalid recommendations are reduced, thereby improving the accuracy and efficiency of insurance recommendations and increasing user satisfaction.

[0217] In order to solve the above technical problems, the present application embodiment also provides a device (computer device). Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0218] 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.

[0219] 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.

[0220] 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 equipped on the computer device 4, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. 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 recommendation method, 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.

[0221] 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 computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the insurance recommendation method.

[0222] 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.

[0223] In this embodiment, by accurately extracting key features from insurance texts, the accuracy of subsequent insurance recommendations can be improved; and by automated feature extraction, manual intervention is avoided, the error rate is reduced, and the accuracy and efficiency of subsequent recommendations are improved;

[0224] Accurately locate key content through user browsing behavior and use it for subsequent analysis, which can improve the accuracy of insurance recommendations;

[0225] The target image information is subjected to feature extraction using a pre-trained image feature extraction model to obtain target image features, which can automatically extract the features of the target image information, avoid misidentification caused by manual recognition, and improve the efficiency of feature extraction; and the trained image feature extraction model is used to extract image features of multiple scales and fuse the obtained target image features, which can improve the accuracy and efficiency of subsequent insurance matching;

[0226] By splitting the target text information into multiple target clauses, the problem of missing semantics in long texts is solved, and the accuracy of text analysis is improved. Analysis based on target clauses can capture the local structure and information in the text. This fine-grained processing enables the model to more accurately extract the features of each clause, which improves the accuracy of subsequent insurance recommendations.

[0227] By fusing the target text features and the target image features to obtain the target fusion features, it is possible to utilize information from different modalities, thereby solving the problem of low accuracy caused by only using text information for recommendation in the prior art, and improving the accuracy of insurance recommendations; and according to the target fusion features and the features of the insurance to be recommended, the recommendation value of the insurance to be recommended is calculated, and the recommendation value can be calculated by combining rich features, thereby improving the accuracy of insurance recommendations;

[0228] By calculating the recommendation value, accurate matching is achieved and invalid recommendations are reduced, thereby improving the accuracy and efficiency of insurance recommendations and increasing user satisfaction.

[0229] The present application also provides another embodiment, namely, providing a storage medium (computer-readable storage medium), wherein the computer-readable storage medium 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 recommendation method as described above.

[0230] During the implementation process, the computer-readable storage medium of the present application can improve the accuracy of subsequent insurance recommendations by accurately extracting key features from insurance texts; and through automated feature extraction, manual intervention is avoided, the error rate is reduced, and the accuracy and efficiency of subsequent recommendations are improved;

[0231] Accurately locate key content through user browsing behavior and use it for subsequent analysis, which can improve the accuracy of insurance recommendations;

[0232] The target image information is subjected to feature extraction using a pre-trained image feature extraction model to obtain target image features, which can automatically extract the features of the target image information, avoid misidentification caused by manual recognition, and improve the efficiency of feature extraction; and the trained image feature extraction model is used to extract image features of multiple scales and fuse the obtained target image features, which can improve the accuracy and efficiency of subsequent insurance matching;

[0233] By splitting the target text information into multiple target clauses, the problem of missing semantics in long texts is solved, and the accuracy of text analysis is improved. Analysis based on target clauses can capture the local structure and information in the text. This fine-grained processing enables the model to more accurately extract the features of each clause, which improves the accuracy of subsequent insurance recommendations.

[0234] By fusing the target text features and the target image features to obtain the target fusion features, it is possible to utilize information from different modalities, thereby solving the problem of low accuracy caused by only using text information for recommendation in the prior art, and improving the accuracy of insurance recommendations; and according to the target fusion features and the features of the insurance to be recommended, the recommendation value of the insurance to be recommended is calculated, and the recommendation value can be calculated by combining rich features, thereby improving the accuracy of insurance recommendations;

[0235] By calculating the recommendation value, accurate matching is achieved and invalid recommendations are reduced, thereby improving the accuracy and efficiency of insurance recommendations and increasing user satisfaction.

[0236] The non-Company software tools or components appearing in the embodiments of this application are merely examples and do not represent actual use.

[0237] 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.

[0238] 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 recommendation method, characterized in that: The steps include: Obtaining insurance text information of the insurance to be recommended, and extracting features of the insurance to be recommended from the insurance text information of the insurance to be recommended using a first text feature extraction model; Obtaining the page dwell time of the user in the PPT document, determining the target PPT page information according to the page dwell time, and extracting the target text information and the target image information from the target PPT page information; Using a pre-trained image feature extraction model to extract features of the target image information, to obtain target image features; Divide the target text information into a plurality of target clauses, extract the features of each target clause using a second text feature extraction model to obtain a plurality of target clause features, and perform weighted fusion on all the target clause features to obtain a target text feature; The target text feature and the target image feature are fused to obtain a target fused feature, and a recommendation value of the insurance to be recommended is calculated according to the target fused feature and the feature of the insurance to be recommended; If the recommendation value is greater than a preset recommendation threshold, the insurance to be recommended is pushed to the user as the target recommended insurance.

2. The insurance recommendation method according to claim 1, characterized in that: The extracting features of the insurance to be recommended from the insurance text information of the insurance to be recommended by using the first text feature extraction model includes: Vectorizing the insurance text information to obtain an insurance text information vector; Performing a convolution operation on the insurance text information vector using multiple convolution kernels of the first text feature extraction model to obtain multiple local initial vectors of the insurance text information vector; Using the maximum pooling layer of the first text feature extraction model to screen the multiple local initial vectors to obtain multiple local key vectors; The fully connected layer of the first text feature extraction model is used to perform nonlinear transformation and combination operations on the multiple local key vectors to obtain the features of the insurance to be recommended.

3. The insurance recommendation method according to claim 1, characterized in that: The step of obtaining the page dwell time of the user in the PPT document and determining the target PPT page information according to the page dwell time includes: Monitor the switching events in the PPT document; Obtaining the timestamp of the switching event, and calculating the page dwell time of the user on each PPT page according to the timestamp to obtain an initial page dwell time set; According to a preset residence time threshold, abnormal data of the initial page residence time set is removed to obtain a target page residence time set; The PPT page with the longest dwell time is selected from the target page dwell time set as the target PPT page information.

4. The insurance recommendation method according to claim 1, characterized in that: The step of extracting features from the target image information using a pre-trained image feature extraction model to obtain target image features includes: Inputting the target image information into the pre-trained image feature extraction model; Performing multi-scale feature extraction on the target image information to obtain image features of multiple different scales; The multiple image features of different scales are fused to obtain the target image features.

5. The insurance recommendation method according to claim 1, characterized in that: The second text feature extraction model is used to extract the features of each target clause to obtain a plurality of target clause features, including: Performing a preprocessing operation on each of the target clauses to obtain a standard target clause; Processing the standard target clause using the second text feature extraction model to obtain a positive order target clause vector and a reverse order target clause vector; The forward target clause vector and the reverse target clause vector are subjected to feature fusion to obtain a plurality of target clause features.

6. The insurance recommendation method according to claim 1, characterized in that: The weighted fusion of all the target clause features to obtain the target text features includes: Calculating the attention weight between each of the target clause features and the features of the insurance to be recommended; According to the attention weight, each of the target clause features is weightedly summed to obtain the target text feature.

7. The insurance recommendation method according to claim 6, characterized in that: The calculating the attention weight between each target clause feature and the feature of the insurance to be recommended includes: Performing dot product and difference operations on each of the target clause features and the features of the insurance to be recommended to obtain a dot product vector and a difference vector; Concatenate the dot product vector, the difference vector, the target clause feature, and the feature of the insurance to be recommended to obtain a concatenated vector; The attention weight between each target clause feature and the feature of the insurance to be recommended is calculated based on the concatenated vector and the preset activation function.

8. An insurance recommendation device, characterized in that: The device comprises: A text feature extraction device, used to obtain insurance text information of the insurance to be recommended, and extract features of the insurance to be recommended from the insurance text information of the insurance to be recommended using a first text feature extraction model; A page information acquisition device, used to acquire the page dwell time of the user in the PPT document, determine the target PPT page information according to the page dwell time, and extract the target text information and the target image information from the target PPT page information; An image feature extraction device is used to extract features from the target image information using a pre-trained image feature extraction model to obtain target image features; A text feature fusion device, used for dividing the target text information into a plurality of target clauses, extracting the features of each target clause using a second text feature extraction model to obtain a plurality of target clause features, and weightedly fusing all the target clause features to obtain a target text feature; A target feature fusion device, used for fusing the target text feature and the target image feature to obtain a target fusion feature, and calculating a recommendation value of the insurance to be recommended based on the target fusion feature and the feature of the insurance to be recommended; The target insurance pushing device is used to push the insurance to be recommended as the target recommended insurance to the user if the recommendation value is greater than a preset recommendation threshold.

9. A computer device, characterized in that: The computer device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the insurance recommendation method as described in any one of claims 1 to 7.

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

Images