Standardized demand retrieval matching method and system based on user portrait feature fusion

Abstract

The invention belongs to the field of retrieval and matching, and provides a standardized demand retrieval and matching method and system based on user portrait feature fusion, and the method comprises the steps: firstly carrying out character splitting on a standard writing demand, and then carrying out pooling operation and dimension mapping to obtain dimension mapping features; splitting the dimension mapping feature into a plurality of sub-features, and performing multidimensional symmetric cross attention processing on every two sub-features to obtain a cross attention splicing feature; performing multi-time convolution splicing on the cross attention splicing features to realize feature cross fusion to obtain cross fusion features; performing feature mapping on the cross fusion features to obtain three different user portrait index mapping vectors, extracting the standardized compiled user portrait features, and summarizing and concluding the user portrait features to form a feature index set; semantic information contained in a text is learned, and a feature extraction task is converted into a feature index classification task, so that a user portrait is described more accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of retrieval and matching, and specifically relates to a standardized demand retrieval and matching method and system based on user portrait feature fusion. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In the context of the rapid development of information technology today, the demand for standardized document writing is growing, especially in the fields of knowledge management and content creation. However, existing technologies often lack effective automated methods to accurately match and provide relevant knowledge when dealing with user descriptions of standardized knowledge. Traditional solutions rely on static classification systems and simple keyword matching, which are powerless when faced with complex and changing user needs. In addition, due to the lack of in-depth understanding and analysis of user descriptions, these technologies cannot effectively portray user portraits, resulting in low accuracy and efficiency of knowledge matching. Therefore, it is necessary to develop a new technical solution to learn and understand user descriptions through deep learning models, thereby achieving accurate classification and portrait portrayal of user needs, and ultimately achieving the purpose of accurately matching standardized knowledge in the knowledge base. At present, the research on user portrait feature extraction for standardized writing users and matching needs with knowledge based on user portraits is still in its early stages, and there are no mature rules and methods yet. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a standardized demand retrieval and matching method and system based on user portrait feature fusion. The present invention extracts standardized user portrait features based on a deep learning algorithm of feature cross-splicing fusion, summarizes the user portrait features, forms a feature indicator set, learns the semantic information contained in the text, and transforms the feature extraction task into a feature indicator classification task, so that the user portrait can be portrayed more accurately.

[0005] According to some embodiments, a first solution of the present invention provides a standardized demand retrieval and matching method based on user portrait feature fusion, which adopts the following technical solution: The standardized demand retrieval and matching method based on user portrait feature fusion includes: Obtain the standard writing requirements of the standard writer, use the pre-trained feature cross-stitching fusion network model for classification, and obtain three different user portrait indicator mapping vectors; Perform semantic retrieval and matching based on the knowledge in the user portrait index mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

[0006] Furthermore, the standard writing requirements are first split into characters, and then pooling operations and dimension mapping are performed after splitting to obtain dimension mapping features, specifically: Obtain standard writing requirement text data for word embedding and preliminary feature extraction to obtain requirement text features; The text boundary labels in the requirement text features are used to construct a mask matrix, and the requirement text features are decomposed based on the mask matrix; The decomposed features are pooled, and the features after the pooling operation are mapped into a character correlation matrix to obtain dimension mapping features.

[0007] Furthermore, the mask matrix is ​​constructed by using the text boundary labels in the demand text features, specifically: The requirement text features are divided into n+1 text fields according to n text boundary labels; Set a one-dimensional pre-mask matrix. According to the order of the fields, define the elements in the pre-mask matrix with the corresponding order as 1, which is the number of characters in the text field, and set the remaining elements to 0; And so on, finally we get n+1 mask matrices.

[0008] Furthermore, after the dimensional mapping feature is split into multiple sub-features, the sub-features are subjected to multi-dimensional symmetrical cross-attention processing in pairs to obtain a cross-attention splicing feature, specifically: After splitting the dimension mapping feature into multiple sub-features, they are input into multiple multi-dimensional symmetric cross attention modules in pairs; Each multi-dimensional symmetric cross attention module first merges the two inputs to perform multi-head attention operation and batch normalization; Then perform multi-head attention operation on the two inputs and the batch normalized results to obtain two attention features; After concatenating the two attention features and the corresponding input with the batch normalized results, batch normalization and convolution operations are performed, and the features after the two convolution operations are cross-concatenated as the output of each multi-dimensional symmetric cross attention module; After being processed by multiple multi-dimensional symmetric cross-attention modules, the cross-attention splicing features are obtained.

[0009] Furthermore, the cross-attention splicing features are convolved multiple times to achieve feature cross-fusion, and obtain cross-fusion features, specifically: Split the cross-attention concatenation feature into four cross-sub-features; The four cross sub-features are input into four convolution modules for processing respectively, and the outputs of the four convolution modules are spliced ​​in pairs to obtain two spliced ​​outputs and then convolution operations are performed respectively; The features after the two convolution operations are cross-joined and then average-pooled. The features after average-pooling are cross-joined and convolved with the features after the two convolution operations to obtain two cross-convolution features. The two cross-convolution features are combined and inputted to perform two different multi-head attention operations, and finally the cross-fusion features are obtained.

[0010] Furthermore, the cross-fusion features are subjected to feature mapping to obtain three different user feature mapping vectors, specifically: Split the cross-fusion feature into two cross-fusion sub-features and perform convolution operations on each of them; The results of the two convolution operations are then cross-joined, convolved, and flattened; The results of the flattening operation are output using three fully connected layers to obtain three different user feature mapping vectors.

[0011] According to some embodiments, the second solution of the present invention provides a standardized demand retrieval and matching system based on user portrait feature fusion, which adopts the following technical solution: A standardized demand retrieval and matching system based on user portrait feature fusion, including: The user portrait feature mapping module is configured to obtain the standard writing requirements of the standard writer, and use the pre-trained feature cross-stitching fusion network model for classification to obtain three different user portrait indicator mapping vectors; A semantic search and matching module, configured to perform semantic search and matching based on knowledge in a user portrait indicator mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

[0012] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0013] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in the first aspect above.

[0014] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in the first aspect above are implemented.

[0016] According to some embodiments, a fifth aspect of the present invention provides a computer program product or a computer program.

[0017] The present invention provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in the first aspect above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention extracts the features of standardized user portraits based on the deep learning algorithm of the feature cross-splicing fusion network model, summarizes the user portrait features, forms a set of feature indicators, learns the semantic information contained in the text, and transforms the feature extraction task into a feature indicator classification task, so that the user portrait can be portrayed more accurately; cross-splicing, multi-dimensional fusion and other methods are used to extract user portrait features and perform feature fusion, which improves the quality of semantic features, improves the accuracy of the final user portrait indicator classification, and makes the user portrait portrayal more reasonable and more in line with actual application conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 It is a training flow chart of a standardized demand retrieval and matching method based on user portrait feature fusion in an embodiment of the present invention; Figure 2 It is a demand text data set training feature cross-stitching fusion network model in an embodiment of the present invention; Figure 3 is a structural diagram of a character splitting and dimension mapping module in an embodiment of the present invention; Figure 4 is a structural diagram of a multi-dimensional symmetrical cross attention module in an embodiment of the present invention; Figure 5 is a structural diagram of a user portrait feature cross-fusion module in an embodiment of the present invention; Figure 6 It is a structural diagram of the user portrait feature label mapping module in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0024] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0025] Embodiment 1 The present embodiment provides a standardized demand retrieval and matching method based on user portrait feature fusion. The present embodiment uses the method applied to a server as an example for illustration. It is understandable that the method can also be applied to a terminal, and can also be applied to a terminal, a server, and a system, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in this application. In the present embodiment, the method includes the following steps: Obtain the standard writing requirements of the standard writer, use the pre-trained feature cross-stitching fusion network model for classification, and obtain three different user portrait indicator mapping vectors; Perform semantic retrieval and matching based on the knowledge in the user portrait index mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

[0026] In order to accurately divide the user portraits of standard writing users and accurately match the standardized demand knowledge, the standardized demand and knowledge matching method based on cross-stitching and fusion of user portrait features disclosed in the present invention is as follows: Figure 1 As shown, the specific steps include the following steps.

[0027] S1: Obtain the standard writing requirements of standard writers and build a text dataset of standard writing requirements.

[0028] S2: Extract standard writer user portrait feature labels based on the demand text dataset and label the data.

[0029] S3: Build a feature cross-stitching fusion network.

[0030] S4: Use the annotated writing requirement text dataset to train the feature cross-stitching fusion network model. Input into the BERT pre-trained model to obtain features . The feature Input character splitting and dimension mapping module, output feature . Input the multi-dimensional symmetric cross attention module group, and output the feature . The feature Input the user portrait feature cross fusion module and output the feature . Input the user portrait indicator mapping module and output the feature .use Map the user portrait feature labels. Calculating Losses And the optimizer is used to perform gradient optimization on the feature cross-splicing fusion network model.

[0031] S5: Construct user portrait feature label mapping and standardized demand knowledge matching method.

[0032] S6: Deploy the feature cross-stitching fusion network model and test the system using the new standard writer’s requirement text.

[0033] The following is a detailed description of each of the above steps.

[0034] 1. Step S1: Obtain the standard writing requirements of the standard writer and build a standard writing requirements text dataset.

[0035] S1-1: Collect the needs of standard writers for standardization knowledge during the standard drafting process; the above demand information is obtained from multiple channels such as interviews and questionnaires, and the obtained user demand information is stored in text form.

[0036] S1-2: Perform preliminary segmentation on the text information, and replace the punctuation marks between the text with the [SEP] symbol, where [SEP] is used to mark the dividing point between the previous text data and the next text data.

[0037] 2. Step S2: Extract standard writer user portrait feature labels based on the required text dataset and label the data.

[0038] S2-1: The collected standard writing requirements text data set was disassembled and analyzed, and two layers of user portrait feature labels were extracted based on the text content and expert opinions. The first layer is divided into three aspects: industry field, standardization document type and writing purpose. The second layer is an extension of the first layer. In terms of industry field, it is divided into 21 feature labels, including "agriculture, forestry, animal husbandry and fishery", "mining industry", "manufacturing industry", "construction industry", "wholesale and retail industry", "transportation", "warehousing and postal industry", "accommodation and catering industry", "information transmission, software and information technology service industry", "financial industry", "real estate industry", "leasing and business service industry", "scientific research and technical service industry", "water conservancy, environment and public facilities management industry", "resident services, repairs and other service industries", "education", "health and social work", "culture, sports and entertainment industry", "public management", "social security" and "social organization". In terms of standardization document type, it is divided into 7 feature labels, including national standards, industry standards, local standards, enterprise standards, specification standards, procedure standards and guide standards; in terms of writing purpose, it is divided into 5 feature labels, including ensuring quality and safety, promoting efficiency and consistency, supporting technical compatibility and consistency, enhancing international trade and market access, and compliance specifications.

[0039] S2-2: Use 0, 1, and 2 to identify the three aspects of the first level: industry field, standardized document type, and compilation purpose.

[0040] Use 0a, 0b, 0c, 0d, 0e, 0f, 0g, 0h, 0i, 0j, 0k, 0l, 0m, 0n, 0o, 0p, 0q, 0r, 0s, 0t, 0u to identify the industry fields of agriculture, forestry, animal husbandry and fishery, mining, manufacturing, construction, wholesale and retail, transportation, warehousing and postal services, accommodation and catering, information transmission, software and information technology services, finance, real estate, leasing and business services, scientific research and technical services, water conservancy, environment and public facilities management, resident services, repair and other services, education, health and social work, culture, sports and entertainment, public administration, social security and social organizations; 1a, 1b, 1c, 1d, 1e, 1f and 1g are used to identify national standards, industry standards, local standards, enterprise standards, specification standards, procedure standards and guide standards in terms of standardized document types, and 2a, 2b, 2c, 2d and 2e are used to identify the purpose of writing to ensure quality and safety, promote efficiency and consistency, support technical compatibility and consistency, enhance international trade and market access, and compliance standards.

[0041] S2-3: Annotate the data in the standard writing requirement text dataset with reference to expert opinions and user portrait feature labels. Three examples after annotation: ① ,② , ③ .

[0042] S2-4: Divide the standard writing requirement text dataset into training set, validation set and test set in a ratio of 6:2:2.

[0043] 3. Step S3: Build a feature cross-stitching fusion network.

[0044] S3-1: Figure 2 As shown, the feature cross-splicing fusion network model includes a character splitting and dimension mapping module, a multi-dimensional symmetric cross-attention module group, a user portrait feature cross-fusion module, and a user portrait feature label mapping module.

[0045] S3-2: Specifically, the data in the standard writing requirement text dataset Input the BERT pre-trained model, embed the characters and extract the features, and output the required text features .

[0046] S3-3: The standard writing requirement is first split into characters, and then pooling and dimension mapping are performed to obtain dimension mapping features, which are specifically: Obtain standard writing requirement text data for word embedding and preliminary feature extraction to obtain requirement text features; The text boundary labels in the requirement text features are used to construct a mask matrix, and the requirement text features are decomposed based on the mask matrix; The decomposed features are pooled, and the features after the pooling operation are mapped into a character correlation matrix to obtain dimension mapping features.

[0047] The mask matrix is ​​constructed by using the text boundary labels in the demand text features, specifically: The requirement text features are divided into n+1 text fields according to n text boundary labels; Set a one-dimensional pre-mask matrix. According to the order of the fields, define the elements in the pre-mask matrix with the corresponding order as 1, which is the number of characters in the text field, and set the remaining elements to 0; And so on, finally we get n+1 mask matrices.

[0048] Character splitting and dimension mapping modules such as Figure 3 As shown in the figure. In the splitting module, the [SEP] tag in the text data is used to construct a mask matrix. When a text data contains n [SEP]-text boundary tags, the number of mask matrices constructed based on the [SEP] tag is n+1. Use the mask matrix to identify the required text features. After decomposition, it can be decomposed into n+1 requirement text sub-features, recorded as .

[0049] Specifically, the mask matrix is ​​as follows: a pre-mask matrix is ​​set, denoted as is the number of characters. First, In Set the elements of to 0. Use the [SEP] identifier to construct the mask. The mask matrix for the first field Recorded as , where the number of 1s is the number of characters before the first [SEP] identifier, and the number of 0s is the number of other characters. The mask matrix for the second field Recorded as ], where the number of 1s is the number of characters after the first [SEP] identifier and between the second [SEP] identifier, and the subsequent mask construction is the same.

[0050] Will After LSE pooling operation, the pooled output features are obtained The LSE pooling operation can be expressed as the formula: (1); in, is a hyperparameter that can be optimized by deep learning model training; is the total number of points in the feature matrix, It is the first Row, No. Characteristic points of the column; It is the output feature point after LSE pooling operation.

[0051] The pooled output features Perform character relevance matrix mapping, use the character relevance mapping function to perform matrix mapping on the features, and obtain the mapping output features .

[0052] The character relevance mapping function can be expressed as the formula: (2); in, is the input feature matrix, is the learnable parameter matrix, is the learnable bias term, is the output feature matrix. When the input feature matrix The dimension is When , the parameter matrix can be learned The dimension is , ensuring that the output feature matrix is ​​a square matrix. In order to facilitate feature transmission, the output features are mapped Grouping of collections, recorded as dimension mapping features .

[0053] S3-4: After the dimensional mapping feature is split into multiple sub-features, the sub-features are subjected to multi-dimensional symmetrical cross-attention processing in pairs to obtain a cross-attention splicing feature, specifically: After splitting the dimension mapping feature into multiple sub-features, they are input into multiple multi-dimensional symmetric cross attention modules in pairs; Each multi-dimensional symmetric cross attention module first merges the two inputs to perform multi-head attention operation and batch normalization; Then perform multi-head attention operation on the two inputs and the batch normalized results to obtain two attention features; After concatenating the two attention features and the corresponding input with the batch normalized results, batch normalization and convolution operations are performed, and the features after the two convolution operations are cross-concatenated as the output of each multi-dimensional symmetric cross attention module; After being processed by multiple multi-dimensional symmetric cross-attention modules, the cross-attention splicing features are obtained.

[0054] The multi-dimensional symmetric cross attention module group is composed of multiple multi-dimensional symmetric cross attention modules. Figure 4 The multi-dimensional symmetric cross attention module group obtains the output features of the character splitting and dimension mapping modules - dimension mapping features and disassembled to form The dimension mapping sub-feature is recorded as .Will Input multiple multi-dimensional symmetric cross-attention modules.

[0055] The output of the multi-dimensional symmetric cross attention module group is 4 feature matrices, so the number of multi-dimensional symmetric cross attention modules is the same as the number of dimension mapping sub-features Assuming the input is Dimensions map sub-features, the first layer requires Multi-dimensional symmetric cross attention modules are used to output features; the second layer requires Multi-dimensional symmetric cross attention modules are used to output features, and so on until the number of output features is 4. The total number of multi-dimensional symmetric cross attention modules required can be obtained by the formula calculate.

[0056] The multi-dimensional symmetric cross attention module obtains two inputs, which are recorded as input one and input two. and input two .Will and Input the first multi-head attention of the multi-dimensional symmetric cross attention module to obtain the output feature . The feature Input the first batch of normalized modules of the multi-dimensional symmetric cross attention module to obtain the output features .

[0057] Will and Features Input the second multi-head attention of the multi-dimensional symmetric cross attention module to obtain the output attention feature .Will and Features Input the third multi-head attention of the multi-dimensional symmetric cross attention module to obtain the output attention feature .

[0058] Will ,feature and Features Add up the features .Will ,feature and Features Add up the features .

[0059] The characteristics The second batch of normalization modules input into the multi-dimensional symmetric cross attention module outputs the feature . The feature Input the third batch of normalization modules of the multi-dimensional symmetric cross attention module, and the output is the feature .

[0060] The characteristics Input into the first convolution module of the multi-dimensional symmetric cross attention module, and output the feature . The feature Input into the second convolution module of the multi-dimensional symmetric cross attention module, and the output is the feature .

[0061] The characteristics and Features Input the cross-concatenation module of the multi-dimensional symmetric cross attention module, and output the feature The four feature matrices finally output by the above multi-dimensional symmetric cross attention module are recorded as In order to facilitate feature transmission, the feature Grouping, recorded as cross-attention concatenation features .

[0062] S3-5: The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion, and obtain cross-fusion features, specifically: Split the cross-attention concatenation feature into four cross-sub-features; The four cross sub-features are input into four convolution modules for processing respectively, and the outputs of the four convolution modules are spliced ​​in pairs to obtain two spliced ​​outputs and then convolution operations are performed respectively; The features after the two convolution operations are cross-joined and then average-pooled. The features after average-pooling are cross-joined and convolved with the features after the two convolution operations to obtain two cross-convolution features. The two cross-convolution features are combined and inputted to perform two different multi-head attention operations, and finally the cross-fusion features are obtained.

[0063] User portrait feature cross fusion module Figure 5 As shown. Decomposed into 4 cross-features . Input the first convolution module in the user portrait feature cross fusion module, and output the feature . Input the second convolution module in the user portrait feature cross fusion module, and output the feature . Input the third convolution module in the user portrait feature cross fusion module, and output the feature . Input the fourth convolution module in the user portrait feature cross fusion module, and output the feature .

[0064] The characteristics and Input the first cross-joining module in the cross-fusion module of the user portrait feature, and output the feature . and Input the second cross-joining module in the cross-fusion module of the user portrait feature, and output the feature .

[0065] The characteristics Input the fifth convolution module in the user portrait feature cross fusion module, and output the feature . Input the sixth convolution module in the user portrait feature cross fusion module, and output the feature .feature and Input the third cross-joining module in the cross-fusion module of the user portrait feature, and output the feature . Input the average pooling module in the cross-fusion module of the user portrait feature, and output the feature .

[0066] The characteristics and Input the second cross-joining module in the cross-fusion module of the user portrait feature, and output the feature . The feature and Input the second cross-joining module in the cross-fusion module of the user portrait feature, and output the feature . The feature Input the seventh convolution module in the cross-fusion module of the user portrait feature, and output the cross-convolution feature . The feature Input the eighth convolution module in the cross-fusion module of the user portrait feature, and output the cross-convolution feature .

[0067] The characteristics and Input the first multi-head attention module in the user portrait feature cross-fusion module, and output the feature . The feature and Input the second multi-head attention module in the user portrait feature cross-fusion module, and output the feature In order to facilitate feature transmission, the feature and Features Group together and record as features .

[0068] S3-6: The cross-fusion features are subjected to feature mapping to obtain three different user feature mapping vectors, specifically: Split the cross-fusion feature into two cross-fusion sub-features and perform convolution operations on each of them; The results of the two convolution operations are then cross-joined, convolved, and flattened; The results of the flattening operation are output using three fully connected layers to obtain three different user feature mapping vectors.

[0069] User portrait feature label mapping module Figure 6 As shown. Decomposed into 2 cross-fusion sub-features . The feature Input the convolution group 1 in the user portrait feature label mapping module and output the feature . The feature Input the convolution group 2 in the user portrait feature label mapping module, and output the feature .

[0070] The characteristics and Features Input the cross-joining module in the user portrait feature label mapping module and output the feature . Input the convolution group 3 in the user portrait feature label mapping module, and output the feature . Input the flattening operation module in the user portrait feature label mapping module, and output the feature .

[0071] The characteristics Input the fully connected layer group 1 in the user portrait feature label mapping module, and output the user portrait index mapping vector 1 . Input the fully connected layer group 2 in the user portrait feature label mapping module, and output the user portrait index mapping vector 2 . Input the fully connected layer group 3 in the user portrait feature label mapping module, and output the user portrait index mapping vector 3 .

[0072] S3-7: The cross-stitching fusion described in steps S3-4, S3-5 and S3-6 is specifically: The cross-stitching fusion module receives two inputs of the same size, denoted as and . Put the two input and Cross-arrangement can be expressed as (3); This method is used to cross-join and merge features. When the size of the input feature is When the feature size of the output after the cross-stitching fusion module becomes the input feature size .

[0073] 4. Step S4: Use the writing requirement text dataset to obtain a deployable feature cross-stitching fusion network model.

[0074] S4-1: Loss function of feature cross-stitching fusion network model It is obtained by adding the classification loss of user portrait index mapping vector one, the classification loss of user portrait index mapping vector two, and the classification loss of user portrait index mapping vector three. Both loss functions are cross entropy loss functions.

[0075] Among them, the cross entropy loss function is defined as follows: (4); in, is the loss value, Indicates the sample number, represents the number of samples, The embedding representation representing the true value of the label, Indicates the predicted value of user profile label mapping.

[0076] S4-2: During the training and validation process, the SGD optimizer is used to perform gradient optimization on the feature cross-splicing fusion network model.

[0077] 5. Step S5: Construct user portrait feature label mapping and standardized demand knowledge matching method.

[0078] S5-1: Normalize the indicator mapping vector. Input the data into the deployed feature cross-stitching fusion network model to obtain three label classification outputs of the feature cross-stitching fusion network model, namely user portrait indicator mapping vector 1, user portrait indicator mapping vector 2, and user portrait indicator mapping vector 3.

[0079] The three user portrait indicator mapping vectors are normalized and the indicator calculation scores are standardized to the range of [0, 1].

[0080] Use the normalization formula The indicator calculation scores are normalized. is the first one-dimensional feature elements, is the minimum value of all elements in the one-dimensional feature, is the maximum value of all elements in the one-dimensional feature. This step outputs three normalized user portrait index mapping vectors, recorded as , and .

[0081] S5-2: User portrait feature mapping. Read the normalized user portrait indicator mapping vector , and Extract the element value in the mapping vector 0.75 of the element.

[0082] Assumptions The 24 elements in are: [0.47, 0.21, 0.29, 0.56, 0.23, 0.17, 0.64, 0.35, 0.72, 0.09, 0.51, 0.27, 0.89, 0.26, 0.83, 0.41, 0.25, 0.86, 0.28, 0.24, 0.58], where the numbers greater than or equal to 0.75 are 0.89, 0.83, 0.86, and the element indexes corresponding to these three numbers are: 13, 16, 18.

[0083] Use element indexing to extract the industry field that may correspond to this user demand data from the indicator vector ["Agriculture, Forestry, Animal Husbandry and Fishery", "Mining", "Manufacturing", "Construction", "Wholesale and Retail", "Transportation", "Warehousing and Postal Services", "Accommodation and Catering", "Information Transmission, Software and Information Technology Services", "Financial Industry", "Real Estate", "Leasing and Business Services", "Scientific Research and Technical Services", "Water Conservancy, Environment and Public Facilities Management", "Residential Services, Repairs and Other Services", "Education", "Health and Social Work", "Culture, Sports and Entertainment", "Public Administration", "Social Security", "Social Organizations"], that is, the industry fields corresponding to 13, 16, and 18 ["Scientific Research and Technical Services", "Education", "Culture, Sports and Entertainment"]. and User portrait feature mapping method and Similarly, the extracted user portrait features are used to characterize the user portrait.

[0084] S5-3: Matching user demand label features with knowledge. Use the extracted user portrait features to search and match the knowledge in the knowledge base. Use keyword semantic search based on cosine similarity optimization to search and match the content in the knowledge base. Use the TF-IDF vectorizer to map the text in the knowledge base and the extracted user portrait feature label text into vectors, and then use the sum of the cosine similarity algorithm and the Tanimoto coefficient algorithm as the standard calculation method to measure the similarity between the two vectors, calculate the matching score between the two, and sort them according to the matching score to push standard domain knowledge for user portraits. For a t-dimensional vector, the algorithm formula is as follows: (5); in, The text in the knowledge base is mapped to a vector by the TF-IDF vectorizer. The vector formed by mapping the extracted user portrait feature label text through the TF-IDF vectorizer. is the dimension of the vector.

[0085] 6. Step S6: Deploy the feature cross-stitching fusion network model and test the system using the new standard writer's requirement text.

[0086] The above model and method are encapsulated into a standardized demand knowledge matching system and deployed. The system is tested using the demand text of the new standard writer. The new user demand text is input into the standardized demand knowledge matching system according to this method, and the user portrait index mapping vector is output. The user portrait is described by the user portrait feature label mapped by the vector. The obtained user portrait label is used to match and retrieve the knowledge in the knowledge base using keyword retrieval optimized by semantic search.

[0087] The role of the user portrait indicator mapping vector is to calculate the matching degree between the current input demand and the knowledge in the knowledge base. For example, if a text is input, the results of the three indicator mapping vectors are "field: agriculture, forestry, animal husbandry and fishery, standardization document type: national standard, writing purpose: to promote efficiency and consistency".

[0088] A matching algorithm is used to match the knowledge in the knowledge base with the label content, that is, the one with the highest matching similarity is taken as the final matching result.

[0089] Embodiment 2 This embodiment provides a standardized demand retrieval and matching system based on user portrait feature fusion, including: The user portrait feature mapping module is configured to obtain the standard writing requirements of the standard writer, and use the pre-trained feature cross-stitching fusion network model for classification to obtain three different user portrait indicator mapping vectors; A semantic search and matching module, configured to perform semantic search and matching based on knowledge in a user portrait indicator mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

[0090] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0091] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0092] The proposed system can be implemented in other ways. For example, the system embodiment described above is only illustrative, and the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0093] Embodiment 3 This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in the above-mentioned embodiment 1 are implemented.

[0094] Embodiment 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in the first embodiment above are implemented.

[0095] Embodiment 5 This embodiment provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the standardized demand retrieval and matching method based on user portrait feature fusion described in the first embodiment.

[0096] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0097] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0100] 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 a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0101] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A standardized demand retrieval and matching method based on user portrait feature fusion, characterized in that: include: Obtain the standard writing requirements of the standard writer, use the pre-trained feature cross-stitching fusion network model for classification, and obtain three different user portrait indicator mapping vectors; Perform semantic retrieval and matching based on the knowledge in the user portrait index mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

2. The standardized demand retrieval and matching method based on user portrait feature fusion according to claim 1 is characterized in that: The standard writing requirements are first split into characters, and then pooling and dimension mapping are performed to obtain dimension mapping features, which are specifically: Obtain standard writing requirement text data for word embedding and preliminary feature extraction to obtain requirement text features; The text boundary labels in the requirement text features are used to construct a mask matrix, and the requirement text features are decomposed based on the mask matrix; The decomposed features are pooled, and the features after the pooling operation are mapped into a character correlation matrix to obtain dimension mapping features.

3. The standardized demand retrieval and matching method based on user portrait feature fusion according to claim 2 is characterized in that: The mask matrix is ​​constructed by using the text boundary labels in the demand text features, specifically: The requirement text features are divided into n+1 text fields according to n text boundary labels; Set a one-dimensional pre-mask matrix. According to the order of the fields, define the elements in the pre-mask matrix with the corresponding order as 1, which is the number of characters in the text field, and set the remaining elements to 0; And so on, finally we get n+1 mask matrices.

4. The standardized demand retrieval and matching method based on user portrait feature fusion according to claim 1 is characterized in that: After the dimensional mapping feature is split into multiple sub-features, the sub-features are subjected to multi-dimensional symmetrical cross-attention processing in pairs to obtain a cross-attention splicing feature, specifically: After splitting the dimension mapping feature into multiple sub-features, they are input into multiple multi-dimensional symmetric cross attention modules in pairs; Each multi-dimensional symmetric cross attention module first merges the two inputs to perform multi-head attention operation and batch normalization; Then perform multi-head attention operation on the two inputs and the batch normalized results to obtain two attention features; After concatenating the two attention features and the corresponding input with the batch normalized results, batch normalization and convolution operations are performed, and the features after the two convolution operations are cross-concatenated as the output of each multi-dimensional symmetric cross attention module; After being processed by multiple multi-dimensional symmetric cross-attention modules, the cross-attention splicing features are obtained.

5. The standardized demand retrieval and matching method based on user portrait feature fusion according to claim 1 is characterized in that: The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion, and obtain cross-fusion features, specifically: Split the cross-attention concatenation feature into four cross-sub-features; The four cross sub-features are input into four convolution modules for processing respectively, and the outputs of the four convolution modules are spliced ​​in pairs to obtain two spliced ​​outputs and then convolution operations are performed respectively; The features after the two convolution operations are cross-joined and then average-pooled. The features after average-pooling are cross-joined and convolved with the features after the two convolution operations to obtain two cross-convolution features. The two cross-convolution features are combined and inputted to perform two different multi-head attention operations, and finally the cross-fusion features are obtained.

6. The standardized demand retrieval and matching method based on user portrait feature fusion according to claim 1 is characterized in that: The cross-fusion features are subjected to feature mapping to obtain three different user feature mapping vectors, specifically: Split the cross-fusion feature into two cross-fusion sub-features and perform convolution operations on each of them; The results of the two convolution operations are then cross-joined, convolved, and flattened; The results of the flattening operation are output using three fully connected layers to obtain three different user feature mapping vectors.

7. A standardized demand retrieval and matching system based on user portrait feature fusion, characterized in that: include: The user portrait feature mapping module is configured to obtain the standard writing requirements of the standard writer, and use the pre-trained feature cross-stitching fusion network model for classification to obtain three different user portrait indicator mapping vectors; A semantic search and matching module, configured to perform semantic search and matching based on knowledge in a user portrait indicator mapping vector knowledge base; Among them, the pre-trained feature cross-stitching fusion network model is used for classification to obtain three different user portrait indicator mapping vectors, specifically: For standard writing requirements, character splitting is first performed, followed by pooling and dimension mapping to obtain dimension mapping features. After splitting the dimension mapping feature into multiple sub-features, the sub-features are processed with multi-dimensional symmetrical cross attention in pairs to obtain the cross attention splicing feature; The cross-attention splicing features are convolved multiple times to achieve feature cross-fusion and obtain cross-fusion features; Perform feature mapping on the cross-fused features to obtain three different user feature mapping vectors.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in any one of claims 1 to 6 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the steps in the standardized demand retrieval and matching method based on user portrait feature fusion as described in any one of claims 1 to 6 are implemented.

Images