Data matching method and device, storage medium and data matching equipment

Abstract

The invention relates to the technical field of deep learning, and discloses a data matching method and device, a storage medium and data matching equipment, and the method comprises the steps: obtaining to-be-matched data and matched target data; keyword extraction is performed on the to-be-matched data and the matched target data through a preset base large model, and the preset base large model is obtained by performing model training according to an association score between a to-be-matched sample keyword and a matched target sample keyword in a current use scene; the association score is determined according to a preset knowledge graph in the current use scene; and matching the keyword extraction result through a preset recommendation model to generate a final matching result. Compared with an existing matching model which is directly used for matching operation in a new scene, the method has the advantages that keyword extraction can be performed by using the preset base large model obtained by performing model training through the association score in the current use scene, and then the keyword extraction result is matched through the preset recommendation model, so that the accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a data matching method, apparatus, storage medium and data matching device. Background Art

[0002] At present, with the development of big model technology, artificial intelligence has increasingly stronger capabilities, especially in intelligent matching (such as matching job information with the resumes of job seekers, etc.).

[0003] However, when the existing recommendation model is used to match the matching data (such as job applicant resumes) with the matching target data (such as job information), due to the large differences in matching data in different fields, when the recommendation model is directly used for matching operations in new scenarios, the feature extraction capability is insufficient, resulting in poor accuracy. Summary of the invention

[0004] The main purpose of this application is to provide a data matching method, apparatus, storage medium and data matching device, aiming to solve the existing technical problem that when the recommendation model is directly used for matching operations in new scenarios, the feature extraction capability is insufficient, resulting in poor accuracy.

[0005] To achieve the above purpose, the present application provides a data matching method, which includes:

[0006] Obtain the data to be matched and the matching target data;

[0007] Keyword extraction is performed on the to-be-matched data and the matching target data through a preset base large model, wherein the preset base large model is obtained through model training according to the association score between the to-be-matched sample keywords and the matching target sample keywords in the current usage scenario, and the association score is determined according to the preset knowledge graph in the current usage scenario;

[0008] The keyword extraction results are matched through the preset recommendation model to generate the final matching results.

[0009] In one embodiment, the step of extracting keywords from the to-be-matched data and the matching target data by using a preset base macro model includes:

[0010] Performing feature extraction on the to-be-matched data through a preset base macro model to obtain a first feature vector, and performing feature extraction on the matching target data through the preset base macro model to obtain a second feature vector;

[0011] The first feature vector and the second feature vector are decoded to obtain a keyword extraction result.

[0012] In one embodiment, after the step of matching the keyword extraction results by using a preset recommendation model to generate a final matching result, the step further includes:

[0013] Acquire new to-be-matched data and new matching target data, and independently input the first feature vector and the second feature vector into the preset base macro model;

[0014] Performing keyword extraction on the new to-be-matched data and the new matching target data based on the first feature vector and the second feature vector by using the preset base macro model;

[0015] The new keyword extraction results are matched through the preset recommendation model to generate new final matching results.

[0016] In one embodiment, the step of matching the keyword extraction results by using a preset recommendation model to generate a final matching result includes:

[0017] Determine a preset processing mode, wherein the preset processing mode is independent processing, splicing processing or difference processing;

[0018] The keyword extraction results are processed according to the preset processing method, and the processed keyword extraction results are matched through the preset recommendation model to generate a final matching result.

[0019] In one embodiment, the step of processing the keyword extraction results according to the preset processing method, matching the processed keyword extraction results through a preset recommendation model, and generating a final matching result includes:

[0020] When the preset processing mode is the independent processing, matching the keyword extraction results by using a preset recommendation model to generate a final matching result;

[0021] When the preset processing mode is the splicing processing, the keyword extraction results are spliced, and the spliced ​​keyword extraction results are matched through a preset recommendation model to generate a final matching result;

[0022] When the preset processing method is the subtraction processing, the keyword extraction results are subtracted item by item, the keyword extraction results after subtracting item by item are spliced, and the spliced ​​keyword extraction results are matched through the preset recommendation model to generate the final matching result.

[0023] In one embodiment, the preset recommendation model includes a single-tower recall recommendation model, a dual-tower recall recommendation model or a convolutional recommendation model, and the convolutional recommendation model is a model that extracts features in a dimensional overall convolution manner.

[0024] In one embodiment, before the steps of obtaining the to-be-matched data and matching the target data, the method further includes:

[0025] Obtaining the sample data to be matched and the matching target sample data in the current usage scenario, and obtaining the preset knowledge graph in the current usage scenario;

[0026] Determine, according to the preset knowledge graph, the sample keywords to be matched corresponding to the sample data to be matched and the matching target sample keywords corresponding to the matching target sample data;

[0027] Determine the association score between the to-be-matched sample keyword and the matching target sample keyword based on the preset knowledge graph;

[0028] The initial base large model is trained through the sample keywords to be matched, the matching target sample keywords and the associated scores to obtain a preset base large model.

[0029] In one embodiment, the step of training the initial base large model through the to-be-matched sample keywords, the matching target sample keywords and the associated scores to obtain the preset base large model includes:

[0030] Filtering the association scores according to a preset screening method to obtain a target association score, and generating a prompt word text based on the target association score;

[0031] The initial base large model is trained through the prompt word text, the sample keywords to be matched and the matching target sample keywords to obtain a preset base large model.

[0032] In one embodiment, the step of screening the relevance scores according to a preset screening method to obtain a target relevance score includes:

[0033] When the preset screening method is the first screening method, screening the association scores based on the first preset threshold value to obtain a target association score;

[0034] When the preset screening method is the second screening method, the relevance scores are sorted, and the relevance scores are screened according to a preset screening quantity to obtain the target relevance scores.

[0035] In one embodiment, before the steps of obtaining the to-be-matched data and matching the target data, the method further includes:

[0036] Acquire the sample data to be matched and the matching target sample data in the current usage scenario, and acquire the standard matching result between the sample data to be matched and the matching target sample data;

[0037] The initial recommendation model is trained by using the to-be-matched sample keywords, the matching target sample keywords and the standard matching results to obtain a preset recommendation model.

[0038] In one embodiment, after the step of matching the keyword extraction results by using a preset recommendation model to generate a final matching result, the step further includes:

[0039] Determine a similarity score of the keyword extraction result, and determine a feedback score corresponding to the final matching result;

[0040] constructing a correlation matrix of the keyword extraction results based on the similarity scores and the feedback scores, and screening the correlation matrix by a second preset threshold;

[0041] The confidence of the keyword extraction result is determined according to the screening result, and the preset knowledge graph is iterated according to the confidence.

[0042] In addition, to achieve the above purpose, the present application also provides a data matching method and device, the device comprising:

[0043] A data acquisition module is used to acquire the data to be matched and the matching target data;

[0044] A keyword extraction module, used to extract keywords from the to-be-matched data and the matching target data through a preset base large model, wherein the preset base large model is obtained through model training based on the association score between the to-be-matched sample keywords and the matching target sample keywords in the current usage scenario, wherein the association score is determined based on a preset knowledge graph in the current usage scenario;

[0045] The result generation module is used to match the keyword extraction results through a preset recommendation model to generate the final matching results.

[0046] In addition, to achieve the above-mentioned purpose, an embodiment of the present application further proposes a storage medium, on which a data matching program is stored, and when the data matching program is executed by a processor, the steps of the data matching method described above are implemented.

[0047] In addition, to achieve the above-mentioned purpose, an embodiment of the present application also proposes a data matching device, which includes: a memory, a processor, and a data matching program stored in the memory and executable on the processor, and when the data matching program is executed by the processor, the steps of the data matching method described above are implemented.

[0048] The embodiments of the present application provide a data matching method, apparatus, storage medium and data matching device, the method comprising: obtaining data to be matched and matching target data; performing keyword extraction on the data to be matched and the matching target data through a preset base large model, the preset base large model is obtained through model training based on the association score between the sample keywords to be matched and the target sample keywords in the current usage scenario, the association score is determined based on the preset knowledge graph in the current usage scenario; matching the keyword extraction results through a preset recommendation model to generate a final matching result.

[0049] Since the present application can determine the correlation score between the sample keywords to be matched and the target sample keywords to be matched according to the preset knowledge graph in the current usage scenario, and then perform model training according to the correlation score to obtain the preset base large model, when the data to be matched and the target data to be matched in the current usage scenario are obtained, the preset base large model can be used to extract keywords from the data to be matched and the target data to be matched, and then the keyword extraction results can be matched through the preset recommendation model to generate the final matching result. Compared with the existing recommendation model, which has insufficient feature extraction capabilities and poor accuracy when it is directly used for matching operations in new scenarios, the present application can first use the preset base large model obtained by model training through the correlation score in the current usage scenario to extract keywords, and then match the keyword extraction results through the preset recommendation model, so that when the recommendation model is directly used for matching operations in new scenarios, keyword extraction can be first performed through the preset base large model, thereby improving accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0052] Figure 1 A schematic diagram of the data matching device structure of the hardware operating environment involved in the embodiment of the present application;

[0053] Figure 2 This is a flow chart of the first embodiment of the data matching method of this application;

[0054] Figure 3 This is a flow chart of the second embodiment of the data matching method of this application;

[0055] Figure 4 This is a schematic diagram of the structure of a preset matching model in the second embodiment of the data matching method of this application;

[0056] Figure 5 This is a schematic diagram of the structure of the large base model in the second embodiment of the data matching method of this application;

[0057] Figure 6 This is a flow chart of the third embodiment of the data matching method of the present application;

[0058] Figure 7 This is a schematic diagram of the structure of a preset matching model in the third embodiment of the data matching method of this application;

[0059] Figure 8 This is a flow chart of the fourth embodiment of the data matching method of the present application;

[0060] Fig. 9 This is a schematic diagram of the structure of a preset matching model in the fourth embodiment of the data matching method of this application;

[0061] Fig.10 This is a structural block diagram of the first embodiment of the data matching method device of this application.

[0062] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0063] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0064] Reference Figure 1 , Figure 1 This is a schematic diagram of the data matching device structure of the hardware operating environment involved in the embodiment of the present application.

[0065] like Figure 1As shown, the data matching device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The wired interface of the user interface 1003 may be a USB interface in this application. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable memory (Non-volatile Memory, NVM), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0066] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the data matching device, and may include more or less components than those shown in the figure, or combine certain components, or arrange the components differently.

[0067] like Figure 1 As shown, the memory 1005 identified as a computer storage medium may include an operating system, a network communication module, a user interface module, and a data matching program.

[0068] exist Figure 1 In the data matching device shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the user device; the data matching device calls the data matching program stored in the memory 1005 through the processor 1001, and executes the data matching method provided in the embodiment of the present application.

[0069] It should be noted that, at present, with the development of big model technology, artificial intelligence has increasingly stronger capabilities, especially in intelligent matching (such as matching job information with the resumes of job seekers, etc.).

[0070] However, when the existing recommendation model is used to match the matching data (such as job applicant resumes) with the matching target data (such as job information), due to the large differences in matching data in different fields, when the recommendation model is directly used for matching operations in new scenarios, the feature extraction capability is insufficient, resulting in poor accuracy.

[0071] Therefore, in order to solve the above defects, this embodiment provides a data matching method, which can determine the correlation score between the sample keywords to be matched and the target sample keywords to be matched according to the preset knowledge graph in the current usage scenario, and then perform model training according to the correlation score to obtain a preset base large model. When the data to be matched and the target data to be matched in the current usage scenario are obtained, the preset base large model can be used to extract keywords from the data to be matched and the target data to be matched, and then the keyword extraction results can be matched through the preset recommendation model to generate the final matching result. Compared with the existing recommendation model, when it is directly used for matching operations in new scenarios, the feature extraction capability is insufficient and the accuracy is poor. In this embodiment, the preset base large model obtained by model training through the correlation score in the current usage scenario can be used to extract keywords, and then the keyword extraction results can be matched through the preset recommendation model, so that when the recommendation model is directly used for matching operations in new scenarios, the preset base large model can be used to extract keywords first, thereby improving the accuracy.

[0072] For ease of understanding, the following combination Figures 2 to 9 The data matching method provided in the embodiment of the present application is introduced in detail.

[0073] Reference Figure 2 , Figure 2 The flowchart of the first embodiment of the data matching method of the present application is shown in FIG. 1 , and the first embodiment of the data matching method of the present application is proposed. Figure 2 As shown, in this embodiment, the specific method includes:

[0074] Step S10: Acquire the data to be matched and the matching target data.

[0075] It is understandable that the method of this embodiment can be applied on a data matching device, wherein the above-mentioned data matching device can be any device used to perform matching tasks, such as a computer, etc., and this embodiment does not limit this. The execution subject of the embodiment of this method can adopt the above-mentioned data matching device (referred to as the device), and this device is used as an example to illustrate this embodiment and the following embodiments.

[0076] It is also understandable that the above-mentioned matching task can be a matching task in any scenario. For example, the matching task in the person-job matching scenario can be a task of matching each job seeker's resume with the job information; the matching task in the question-answering scenario can be a task of matching each document with the question, etc. This embodiment can be described using the person-job matching scenario, but is not limited thereto.

[0077] It should be understood that the above-mentioned to-be-matched data may refer to data that needs to be evaluated or compared in the matching task, such as numerous job seekers' resumes, etc. The above-mentioned matching target data may refer to data used as a comparison benchmark in the matching task, such as job information of recruitment positions, etc.

[0078] In actual use, the above-mentioned device can obtain the information of the positions currently to be recruited as the above-mentioned matching target data, and can obtain the resume of each job seeker as the above-mentioned data to be matched.

[0079] Step S20: extract keywords from the data to be matched and the matching target data through a preset base large model, the preset base large model is obtained by model training based on the association score between the sample keywords to be matched and the target sample keywords in the current usage scenario, and the association score is determined based on the preset knowledge graph in the current usage scenario.

[0080] It should be noted that the above-mentioned preset base model can be a model with a keyword extraction function. The above-mentioned keywords can be words that represent key data in the data to be matched and the matching target data. For example, in the person-job matching task, the keywords of the matching target data can be: software engineer or hardware engineer, etc., and the keywords of the data to be matched can be: education, major and skills, etc. This embodiment does not limit this.

[0081] In actual use, the above-mentioned device can input the data to be matched and the matching target data into the preset base large model, and the preset base large model can extract keywords from the data to be matched and the matching target data to obtain keyword extraction results.

[0082] Step S30: Match the keyword extraction results through a preset recommendation model to generate a final matching result.

[0083] It should also be noted that the above-mentioned preset recommendation model can be the model used to perform the matching task in this embodiment, and can be any model that can implement the matching operation, such as a single-tower recall model or a dual-tower recall model, etc. This embodiment does not limit this.

[0084] Therefore, when in use, the preset base large model can be used to perform keyword extraction on the data to be matched and the matching target data respectively, to obtain the keyword extraction results corresponding to the data to be matched and the keyword extraction results corresponding to the matching target data, and then the preset recommendation model can be used to match the keyword extraction results corresponding to the data to be matched with the keyword extraction results corresponding to the matching target data, so as to generate the final matching result.

[0085] In order to solve the defect that the preset recommendation model has insufficient feature extraction capability when used for matching operations in new scenarios, this embodiment can use a preset base model to extract keywords before the preset recommendation model, which can also be understood as performing feature extraction in the previous step. In order to enable the preset recommendation model to accurately extract keywords, the preset matching model in this embodiment is obtained by model training based on the association score between the sample keywords to be matched and the matching target sample keywords in the current usage scenario, and the association score is determined based on the preset knowledge graph in the current usage scenario.

[0086] It is understandable that the current usage scenario may be a scenario where a matching task is currently required, i.e., a person-job matching scenario in this embodiment. The preset knowledge graph may store a graph of the relationship between the sample keywords to be matched and the matching target keywords in the current usage scenario, and the relationship is the association score.

[0087] It is also understandable that the above-mentioned to-be-matched sample keywords may be keywords of the to-be-matched data in the preset knowledge graph that are consistent with the current usage scenario, and the above-mentioned matching target sample keywords may be keywords of the matching target data in the preset knowledge graph that are consistent with the current usage scenario. The above-mentioned association score may be a score used to characterize the degree of association between the to-be-matched sample keywords and the matching target sample keywords.

[0088] Furthermore, in a specific implementation, the above-mentioned device may first determine the correlation score between the sample keywords to be matched and the target sample keywords to be matched according to the preset knowledge graph in the current usage scenario, and then perform model training according to the correlation score to obtain a preset base large model. When the data to be matched and the target data to be matched in the current usage scenario are obtained, the preset base large model may be used to extract keywords from the data to be matched and the target data to be matched, and then the keyword extraction results may be matched through the preset recommendation model to generate the final matching result. Compared with the existing recommendation model, which has insufficient feature extraction capability and poor accuracy when directly used for matching operations in new scenarios, this embodiment may first use the preset base large model obtained by model training through the correlation score in the current usage scenario to extract keywords, and then match the keyword extraction results through the preset recommendation model, so that when the recommendation model is directly used for matching operations in new scenarios, keyword extraction may be first performed through the preset base large model, thereby improving accuracy.

[0089] Furthermore, in order to obtain a preset matching model that matches the current usage scenario, in this embodiment, before the above step S10, the following is further included:

[0090] Step S01: Obtain sample data to be matched and matching target sample data in the current usage scenario, and obtain a preset knowledge graph in the current usage scenario.

[0091] It should be understood that the above-mentioned sample data to be matched can be data that needs to be evaluated or compared during training in the current usage scenario, such as job applicant resumes, etc. The above-mentioned matching target sample data can be data used as a comparison benchmark in the current usage scenario, such as job information, etc.

[0092] Step S02: Determine the sample keywords to be matched corresponding to the sample data to be matched and the matching target sample keywords corresponding to the matching target sample data according to the preset knowledge graph.

[0093] It should also be understood that in this embodiment, different preset knowledge graphs are set according to different current usage scenarios, and each preset knowledge graph may store the sample keywords to be matched corresponding to the sample data to be matched in the corresponding current usage scenario, and store the matching target sample keywords corresponding to the matching target sample data in the corresponding current usage scenario.

[0094] Step S03: Determine the association score between the to-be-matched sample keyword and the matching target sample keyword based on the preset knowledge graph.

[0095] It should be noted that the above-mentioned preset knowledge graph may also store the correlation scores between each to-be-matched sample keyword and each matching target sample keyword.

[0096] In actual use, during the training process, the above device can obtain the sample data to be matched and the matching target sample data in the current usage scenario, and obtain the preset knowledge graph in the current usage scenario, and then query the preset knowledge graph to obtain the keywords involved in the sample data to be matched as the above sample keywords to be matched, and obtain the keywords involved in the matching target sample data as the above matching target sample keywords. And query the preset knowledge graph to obtain the association score between each sample keyword to be matched and each target sample keyword.

[0097] Step S04: The initial base large model is trained by using the to-be-matched sample keywords, the matching target sample keywords and the associated scores to obtain a preset base large model.

[0098] After obtaining the to-be-matched sample keywords, the matching target sample keywords and the associated scores, the initial base large model can be trained based on the to-be-matched sample keywords, the matching target sample keywords and the associated scores to obtain the above-mentioned preset base large model.

[0099] It should be emphasized that in order to enable the above-mentioned preset recommendation model to have the ability to match the keyword extraction results, after training the initial base large model, the pre-trained recommendation model (i.e., the subsequent initial recommendation model) can be trained with the sample keywords to be matched and the matching target sample keywords to obtain the above-mentioned preset recommendation model.

[0100] This embodiment can determine the correlation score between the sample keywords to be matched and the target sample keywords to be matched according to the preset knowledge graph in the current usage scenario, and then perform model training according to the correlation score to obtain a preset base large model. When the data to be matched and the target data to be matched in the current usage scenario are obtained, the preset base large model can be used to extract keywords from the data to be matched and the target data to be matched, and then the keyword extraction results can be matched through the preset recommendation model to generate the final matching result. Compared with the existing recommendation model, when it is directly used for matching operations in new scenarios, the feature extraction capability is insufficient and the accuracy is poor. This embodiment can first use the preset base large model obtained by model training through the correlation score in the current usage scenario to extract keywords, and then match the keyword extraction results through the preset recommendation model, so that when the recommendation model is directly used for matching operations in new scenarios, keyword extraction can be performed first through the preset base large model, thereby improving accuracy.

[0101] Reference Figure 3 , Figure 3 This is a flow chart of the second embodiment of the data matching method of the present application. Based on the above-mentioned first embodiment, the second embodiment of the data matching method of the present application is proposed.

[0102] like Figure 3 As shown, in order to train the initial base model and the initial recommendation model respectively, as Figure 3 As shown, in this embodiment, the above step S04 includes:

[0103] Step S041: screening the relevance scores according to a preset screening method to obtain target relevance scores, and generating prompt word text based on the target relevance scores;

[0104] Step S042: The initial base large model is trained by using the prompt word text, the sample keywords to be matched and the matching target sample keywords to obtain a preset base large model.

[0105] It should be noted that the above prompt word text may be a text used to input the intention of the base large model. The above preset screening method may be a method for screening the associated scores. Since not all associated scores need to be trained, this embodiment may screen the associated scores based on the preset screening method, thereby improving the training effect.

[0106] For ease of understanding, refer to Figure 4 as well as Figure 5 To explain, Figure 4 This is a schematic diagram of the structure of the preset base large model and the preset recommended model in the second embodiment of the data matching method of this application. Figure 5 This is a schematic diagram of the structure of a preset base large model in the second embodiment of the data matching method of this application;

[0107] like Figure 4 and Figure 5 As shown, in this embodiment, during training, the above sample data to be matched can be recorded as data A (i.e. Figure 4 and Figure 5 Data A), the matching target sample data is recorded as data B (i.e. Figure 4 and Figure 5 After the above device obtains the above sample data to be matched and the matching target sample data, it can be input into the base large model (i.e. Figure 4 The input layer of the large model of the base Figure 5 Token input area).

[0108] At the same time, you can also filter the data from the preset knowledge graph (i.e. Figure 4 The obtained association score is used as the target association score, and then the prompt word text is generated according to the target association score and input into the initial base model for keyword extraction (i.e. Figure 4 Then, the extraction result can be compared with the keywords of the sample to be matched and the keywords of the target sample, and the model training can be performed according to the comparison result to obtain the above-mentioned preset base large model;

[0109] Specifically, when the initial base large model performs keyword extraction, it can be implemented by prompting the initial base large model, that is, generating a prompt word text for keyword extraction and inputting it into the initial base large model, and the initial base large model then performs keyword extraction based on the prompt word text (i.e. Figure 5 Prompt is generated, that is Figure 5 prompt based on knowledge graph).

[0110] It should be emphasized that when the base large model generation performance is insufficient or when there is insufficient data, the multi-head attention mechanism can also be used for keyword extraction.

[0111] Furthermore, in this embodiment, two preset screening methods may be provided, which are respectively recorded as a first screening method and a second screening method. Specifically, the step of screening the relevance scores according to the preset screening methods to obtain the target relevance scores includes:

[0112] Step S0411: When the preset screening method is the first screening method, the association scores are screened based on a first preset threshold to obtain a target association score.

[0113] It is understandable that the first preset threshold may be a score threshold for screening associated scores, which may be set according to actual conditions, and this embodiment does not impose any limitation on this.

[0114] When it is determined that the current preset screening method is the first screening method, during training, the association score between each sample keyword to be matched and each matching target keyword can be obtained according to the preset knowledge graph in the current usage scenario, and the association score higher than the above-mentioned first preset threshold can be screened out, and the obtained result is used as the above-mentioned target association score, and then when the prompt word text is generated based on the target association score, it is input into the initial base large model for keyword extraction.

[0115] Exemplarily, if the correlation score x between keyword a1 in the to-be-matched sample keywords and keyword b1 in the matching target sample keywords is higher than the first preset threshold, then the correlation score x can be used as the target correlation score, and then a prompt word text can be generated according to the target correlation score. The generated prompt word text can be "the correlation score between keyword a1 and keyword b1 is x, and keyword extraction is performed based on the above information". The prompt word text is then input into the initial base large model, and the initial base large model extracts keywords for the to-be-matched sample data and the matching target sample data according to the prompt word text, and obtains the keywords corresponding to the to-be-matched sample data (i.e. Figure 4 The key information A) and the keywords corresponding to the matching target sample data (i.e. Figure 4 Key information B).

[0116] It should be emphasized that the above-mentioned association scores higher than the first preset threshold can be generated in the form of a list, and of course can also be generated in other forms, which is not limited in this embodiment.

[0117] Step S0412: when the preset screening method is the second screening method, the relevance scores are sorted, and the relevance scores are screened according to a preset screening quantity to obtain the target relevance scores.

[0118] It should be understood that the above-mentioned preset screening number can be the number used to screen the associated scores, which can be set according to actual conditions. This embodiment does not limit this. This embodiment uses k for illustration.

[0119] After determining that the current preset screening method is the second screening method, during training, the association score between each sample keyword to be matched and each matching target keyword can be obtained according to the preset knowledge graph in the current usage scenario, and each association score is sorted from large to small, and then k association scores are selected in sequence starting from the highest association score as the above-mentioned target association score, and then the prompt word text is generated based on the target association score and input into the initial base large model for keyword extraction.

[0120] For example, if the correlation score x between keyword a1 in the to-be-matched sample keywords and keyword b1 in the matching target sample keywords is ranked in the top k, then the correlation score x can be used as the target correlation score, and then a prompt word text can be generated according to the target correlation score. The generated prompt word text can be "the correlation score between keyword a1 and keyword b1 is x, and keyword extraction is performed based on the above information". The prompt word text is then input into the initial base large model, and the initial base large model extracts keywords for the to-be-matched sample data and the matching target sample data according to the prompt word text, and obtains the keywords corresponding to the to-be-matched sample data (i.e. Figure 4 The key information A) and the keywords corresponding to the matching target sample data (i.e. Figure 4 Key information B).

[0121] Furthermore, if Figure 5 As shown, in this embodiment, the initial base model may include: an input layer (i.e. Figure 5 Token input area), encoding layer, decoding layer and key information layer;

[0122] The encoding layer is connected to the input layer and the decoding layer respectively, and the key information layer is also connected to the decoding layer.

[0123] In actual use, when the input layer receives the sample data to be matched, the matching target sample data and the prompt word text, it can be transmitted to the encoding layer for encoding and feature extraction, and then the generated feature vector can be output to the decoding layer for decoding, and then the decoding result can be transmitted to the key information layer for output, so as to obtain the corresponding keywords.

[0124] Similarly, when the preset base large model is in use, the matching data and the matching target data can be encoded and feature extracted through the encoding layer, and then decoded and output to obtain the keyword extraction result. The specific process is:

[0125] The step of extracting keywords from the to-be-matched data and the matching target data by using the preset base macro model includes:

[0126] Step S21: extracting features of the to-be-matched data through a preset base macro model to obtain a first feature vector, and extracting features of the matching target data through the preset base macro model to obtain a second feature vector;

[0127] Step S22: Decode the first feature vector and the second feature vector to obtain a keyword extraction result.

[0128] It should be noted that the above-mentioned first feature vector can be a feature vector after the input layer of the preset base large model encodes the data to be matched and extracts features, and the above-mentioned second feature vector can be a feature vector after the input layer of the preset base large model encodes the matching target data and extracts features.

[0129] After obtaining the first feature vector and the second feature vector, the first feature vector may be decoded to obtain the to-be-matched keyword of the to-be-matched data, and the second feature vector may be decoded to obtain the matching target keyword of the matching target data.

[0130] As another implementation, continue as Figure 4 and Figure 5 As shown in the figure, when the amount of data in the preset knowledge graph is insufficient, it can also be obtained from external databases (i.e. Figure 4 In the external database, search, etc. to retrieve relevant data and obtain supplementary information, that is, Figure 5 The external database query) obtains additional correlation scores to supplement and then generates the prompt word text (i.e. Figure 5 in the supplementary data) for keyword extraction (i.e. Figure 4 Supplement key information C);

[0131] For example, when there are overlapping keywords, prompt text may be generated. For example, the generated prompt text may be "The correlation scores between the supplementary keywords a2 and a3 and the keyword b2 are high, please refer to the above information to extract keywords".

[0132] After obtaining the target association score, the above-mentioned prompt word text can be generated according to the target association score, and the initial base large model is trained through the sample keywords to be matched, the matching target sample keywords and the prompt word text to obtain the preset base large model.

[0133] Furthermore, in order to enable the preset recommendation model to have the function of matching keywords, in this embodiment, before the above step S10, it also includes:

[0134] Step S05: obtaining the sample data to be matched and the matching target sample data in the current usage scenario, and obtaining the standard matching result between the sample data to be matched and the matching target sample data;

[0135] Step S06: training the initial recommendation model through the to-be-matched sample keywords, the matching target sample keywords and the standard matching results to obtain a preset recommendation model.

[0136] It should be noted that after the initial base model is trained, the initial recommendation model can be trained. After the initial base model is used to extract keywords, the sample keyword extraction results can be input into the initial recommendation model (i.e. Figure 4 In the recommendation model algorithm), the above-mentioned device can also obtain the standard matching result between the sample data to be matched and the matching target sample data. The standard matching result can be a result used to characterize whether the sample data to be matched and the matching target sample data match each other in the current usage scenario. It can be obtained in advance through human feedback or by annotating data.

[0137] After obtaining the standard matching results, the sample keywords can be input as training samples into the initial recommendation model for matching training to obtain sample matching results, and the standard matching results are used as sample labels. The initial recommendation model is trained based on the standard matching results and sample matching results to obtain the preset recommendation model.

[0138] It should also be emphasized that during training, the above sample keyword extraction results (i.e. Figure 5 The key features of data A and key features of data B are saved and input into the initial base model separately as feature input to ensure that the original information invariants exist in the subsequent process.

[0139] Similarly, when extracting keywords from the preset large base model, the features obtained during this extraction can also be saved and used as separate input for the next extraction. The specific process is as follows:

[0140] After the above step S30, the method further includes:

[0141] Step S40: acquiring new to-be-matched data and new matching target data, and independently inputting the first feature vector and the second feature vector into the preset base macro model;

[0142] Step S50: extracting keywords from the new to-be-matched data and the new matching target data based on the first feature vector and the second feature vector by using the preset base macro model;

[0143] Step S60: Match the new keyword extraction result through the preset recommendation model to generate a new final matching result.

[0144] It should be noted that the new data to be matched may be the data to be matched in the next matching, and the new matching target data may be the matching target data in the next matching.

[0145] In actual use, when the next match is performed, when the new data to be matched and the new matching target data are input into the encoding layer of the preset base large model, the first feature vector (i.e. Figure 5 The key features of data A) and the second eigenvector (i.e. Figure 5 The data B and key features in the data (key features) can also be independently input into the encoding layer for encoding and feature extraction, and then decoded through the decoding layer, and the decoding results are transmitted to the key information layer for output, so as to obtain the keyword extraction results of the data to be matched and the keyword extraction results of the new matching target data, and finally the new keyword extraction results are matched through the preset recommendation model to generate a new final matching result for the next matching.

[0146] Furthermore, considering that the corresponding association score may not be stored in the preset knowledge graph in the current scenario when extracting the first keyword in the first training, when generating the prompt word text, the form without the association score can be directly generated, for example, "What are the core keywords in this text?" is directly generated. Therefore, the base large model can extract keywords without the association score during the first training.

[0147] Furthermore, when using the above-mentioned preset matching model, the data to be matched and the matching target data can be first input into the base large model for keyword extraction, and then the obtained keyword extraction results are transmitted to the recommendation model for matching, thereby generating the final matching results.

[0148] Reference Figure 6 , Figure 6 This is a flow chart of the third embodiment of the data matching method of the present application. Based on the above embodiments, the third embodiment of the data matching method of the present application is proposed.

[0149] Considering that different devices have different computing power levels, in this embodiment, Figure 6 As shown, step S30 includes:

[0150] Step S31: determining a preset processing mode, wherein the preset processing mode is independent processing, splicing processing or difference processing;

[0151] Step S32: Process the keyword extraction results according to the preset processing method, and match the processed keyword extraction results through the preset recommendation model to generate the final matching results.

[0152] It should be noted that the above-mentioned preset processing method can be a method for processing the extracted keywords. In this embodiment, any one of the above-mentioned independent processing, splicing processing or difference processing can be adopted. And because the sequence will be adjusted to the form of a matrix after processing, in this embodiment, the result after processing by the preset processing method is recorded as a feature matrix, that is, the keyword extraction result after the above processing.

[0153] Furthermore, the above step S32 includes:

[0154] Step S321: When the preset processing mode is the independent processing, the keyword extraction results are matched by a preset recommendation model to generate a final matching result.

[0155] It should also be noted that, since the output of the preset base model is generally expressed in the form of a token sequence when performing keyword extraction in this embodiment, and the length of each token sequence is related to the embedding layer design of the preset base model, the length of the token sequence can be referred to as the above-mentioned element output length in this embodiment, denoted as L m Provide explanation.

[0156] When the above-mentioned preset processing method is independent processing, if the length of the keyword of the extracted data to be matched (that is, the above-mentioned key information A) is recorded as n a , the length of the keyword (i.e. the key information B) extracted to match the target data is recorded as n b , then the size of the feature matrix corresponding to the to-be-matched data generated after independent processing of the keywords can be [n a , L m ], the size of the feature matrix corresponding to the generated matching target data can be [n b , L m ]. However, since the element output length of the preset base large model output is L m Therefore, it can be understood that when the preprocessing method is independent processing, the keyword extraction results are not processed, and the keyword extraction results are directly matched through the preset recommendation model to obtain the final matching results.

[0157] Step S322: when the preset processing mode is the splicing processing, the keyword extraction results are spliced, and the spliced ​​keyword extraction results are matched through a preset recommendation model to generate a final matching result.

[0158] When the above-mentioned preset processing method is splicing processing, the keywords can be spliced, and the size of the feature matrix corresponding to the to-be-matched data and the size of the feature matrix corresponding to the matching target data generated after the processing can both be [n a +n b , L m ]. Then, the concatenated keyword extraction results are matched through the preset recommendation model to obtain the final matching results.

[0159] Step S323: When the preset processing method is the subtraction processing, the keyword extraction results are subtracted item by item, the keyword extraction results after the subtraction are spliced, and the spliced ​​keyword extraction results are matched through the preset recommendation model to generate a final matching result.

[0160] When the above-mentioned preset processing method is difference processing, the keywords are subjected to difference processing item by item, and the keywords subjected to difference processing item by item are concatenated with the keywords before the preset processing, and the size of the feature matrix corresponding to the to-be-matched data generated after the processing and the size of the feature matrix corresponding to the matching target data can both be [n a +1,n b +1, L m ]. Then, the concatenated keyword extraction results are matched through the preset recommendation model to obtain the final matching results.

[0161] Similarly, when the initial matching model is trained, the to-be-matched sample keywords of the matching sample data and the matching target sample keywords of the matching target sample data will also be processed according to the above preset processing method. Figure 7 , Figure 7 Schematic diagram of the structure of the preset matching model in the third embodiment of the data matching method of this application, such as Figure 7 As shown, after the initial base large model performs keyword extraction, the obtained sample keyword extraction results (i.e. Figure 7 The key information A, key information B and supplementary key information C) are processed (i.e. Figure 7 In the preliminary feature processing), the corresponding sample feature matrix is ​​obtained (i.e. Figure 7 Then, the obtained sample feature matrix is ​​input into the initial recommendation model (i.e. Figure 7 Recommended model algorithm) to obtain sample matching results (i.e. Figure 7 Generate recommendation results in .

[0162] Furthermore, in this embodiment, a variety of recommendation models may be set, specifically, the preset recommendation model includes a single-tower recall recommendation model, a dual-tower recall recommendation model or a convolutional recommendation model, and the convolutional recommendation model is a model that extracts features in a dimensional overall convolution manner.

[0163] It is understandable that the preset recommendation model in this embodiment may adopt the above-mentioned single-tower recall recommendation model, dual-tower recall recommendation model or convolution recommendation model.

[0164] For the single-tower recall model, it can be adapted to the above preset processing method as concatenation processing, specifically, the sequence of the obtained sample feature matrix can be directly subjected to a fully connected network to obtain the sample matching result;

[0165] For the dual-tower recall model, it can be adapted to the above-mentioned preset processing mode as independent processing. Specifically, the sample feature matrix corresponding to the obtained sample data to be matched is input into a feature network based on full connection, and a recommended feature vector is output. The sample feature matrix corresponding to the obtained matching target sample data is input into another feature network based on full connection, and another recommended feature vector is output. Then, the two recommended feature vectors are input into the final recommendation recall model to obtain the sample matching result.

[0166] For the convolution recommendation model, it can be adapted to the above-mentioned preset processing method as difference processing. Specifically, it can be any model that achieves matching through convolution. In order to increase the specificity of utilizing embedding and thus improve accuracy, this embodiment improves the traditional convolution method and adopts a feature extraction method according to the overall convolution method of dimensions. Specifically, when matching the keyword extraction results, the feature matrix of the keyword extraction results is subjected to feature extraction according to the overall convolution method of dimensions, and matching is performed according to the feature extraction results to obtain the final matching result.

[0167] When extracting features in a dimensional overall convolution manner, compared to the traditional convolution kernel convolution operation, this embodiment can perform convolution on each element in the feature matrix, with the position of the current element as the center, and all layer elements in the local area within the preset range are convolved with the weight parameters in the corresponding convolution kernel. For ease of understanding, the sample feature matrix is ​​used for explanation, as follows:

[0168] First, the weight parameters of the traditional convolution kernel are recorded as That is, the parameter matrix of the convolution kernel is defined as a two-dimensional array of (S+1)×(S+1), where S is the size of the convolution kernel minus 1, and c is i，j Represents the weight parameter of the convolution kernel in the i-th row and j-th column;

[0169] Since the size of the obtained sample feature matrix can be [na +1,n b +1, L m ], and then in this embodiment, the sample feature matrix after the convolution operation is recorded as C, and the definition Among them C m，n，p represents the element of the nth row, mth column, and pth layer in the sample feature matrix C after the convolution operation. Q represents the sample feature matrix at the time of input. The sample feature matrix at the time of input is a three-dimensional matrix, so Q x，y，z It can be expressed as the element of the sample feature matrix at the xth position in the first dimension (i.e., the above row), the yth position in the second dimension (i.e., the above column), and the zth position in the third dimension (i.e., the above layer) at the input, Q x，y， : It can be represented as the sample feature matrix at the xth position in the first dimension, the yth position in the second dimension, and all elements in the third dimension (i.e., all layers) at the time of input;

[0170] and then to It can be understood that with (m, n) as the center, in the local area of ​​(S+1)×(S+1), all the elements in the third dimension of the local area are convolved with the convolution kernel (i.e. the above *), and (S+1)×(S+1) is the above preset range, which can be consistent with the size of the convolution kernel. Through the above operation, it is possible to extract features by overall convolution of dimensions.

[0171] It should be emphasized that the convolution method in this embodiment can be normal convolution, hole convolution, etc., with an indefinite step size.

[0172] Reference Figure 8 , Figure 8 This is a flow chart of the fourth embodiment of the data matching method of the present application. Based on the above embodiments, the fourth embodiment of the data matching method of the present application is proposed.

[0173] In order to update the preset knowledge graph in the current usage scenario, such as Figure 8 In the present embodiment, after the step of matching the keyword extraction results by the preset recommendation model to generate the final matching results, the following step is further included:

[0174] Step S70: Determine the similarity score of the keyword extraction result, and determine the feedback score corresponding to the final matching result.

[0175] It should be noted that the similarity score may be a score of the similarity between the keyword extraction result of the to-be-matched data and the keyword extraction result of the matching target data. The feedback score may be a score that is manually fed back based on the final matching result.

[0176] Reference Fig. 9 , Fig. 9 Schematic diagram of the structure of the preset matching model in the fourth embodiment of the data matching method of this application, such as Fig. 9 As shown in the figure, in actual use, after the recommendation model matches the keyword extraction results to generate the final matching results, the cosine similarity method can be used to calculate the similarity between the keyword extraction results of each to-be-matched data and the keyword extraction results of each matching target data, and the similarity score between the two token vectors can be obtained, which is recorded as At the same time, the final matching result can also be pushed to the user, and the user can evaluate the final matching result, which may include but is not limited to: unmatched, matched, and highly matched. Different grades can be assigned different scores, and then the corresponding score can be queried according to the grade evaluated by the user as the above feedback score (i.e. Fig. 9 The manual feedback result), denoted as s n .

[0177] Step S80: constructing a correlation matrix of the keyword extraction results based on the similarity scores and the feedback scores, and screening the correlation matrix by a second preset threshold.

[0178] Getting a similarity score and feedback ratings n After that, the similarity score can be The product of each keyword extraction result in the to-be-matched data and the keyword extraction result in the matching target data is obtained by multiplying the result with the feedback score, and then a matrix is ​​formed based on each product as the above correlation matrix.

[0179] It should be emphasized that in scoring similarity With feedback ratings n When performing multiplication, corresponding weights may be set to adjust the multiplication result, and the specific weight setting is not limited in this embodiment.

[0180] After obtaining the above-mentioned correlation matrix, the correlation matrix can be filtered by the second preset threshold. Specifically, each product in the correlation matrix higher than the second preset threshold is recorded as an existence relationship as a screening result. The second preset threshold can be set according to the actual situation, and this embodiment does not limit this.

[0181] Step S90: Determine the confidence of the keyword extraction result according to the screening result, and iterate the preset knowledge graph according to the confidence.

[0182] After obtaining the screening results, the product of each keyword extraction result of the to-be-matched data in the screening results and each keyword extraction result of the matching target data is formatted and converted into a confidence format, and the conversion result is used as the above confidence. Then, the confidence can be used as the association score corresponding to the keyword extraction result to iterate the preset knowledge graph (i.e. Figure 5 knowledge graph interaction).

[0183] After iterating the preset knowledge graph, a new association score can be obtained based on the new preset knowledge graph query after iteration, and a new prompt word text can be generated again based on the new association score, for example, "the association score between keyword a1 and keyword b1 is x, and keyword extraction is performed based on the above information." Input it into the preset base large model, and the preset base large model can extract keywords again based on the new prompt word text, and process the new keyword extraction results according to the above preset processing method (i.e. Fig. 9 After the preliminary feature processing in the previous step, the preset recommendation model matches the new keyword extraction results again to generate a new final matching result (i.e. Fig. 9 Self-reinforcement learning question and answer via knowledge graph relationships);

[0184] As another implementation method, the preset base large model can also be fine-tuned through the iterated preset knowledge graph, that is, a new association score can be obtained based on the iterated new preset knowledge graph query, and a new prompt word text can be generated again based on the new association score, for example, "the association score between keyword a1 and keyword b1 is x, and keyword extraction is performed based on the above information." Input into the preset base large model, the preset base large model can fine-tune itself based on the new prompt word text, so that the subsequent generated matching results are more accurate (i.e. Fig. 9 In the process, the knowledge graph relationship is used to score keywords, generate prompt words, and fine-tune the large model).

[0185] At the same time, when generating the prompt word text, keywords can also be specified manually to adjust the final matching results.

[0186] Furthermore, in the first matching during training, since no knowledge graph is stored, after the initial recommendation model performs a matching operation, the above-mentioned method can be used to determine the sample similarity score of the sample keyword extraction result, and determine the sample feedback score corresponding to the sample matching result, and construct a correlation matrix of the sample keyword extraction result based on the sample similarity score and the sample feedback score, and filter the sample correlation matrix through the above-mentioned second preset threshold, and form the screening results obtained by filtering into a preset knowledge graph in the current usage scenario, and construct the knowledge graph (i.e. Fig. 9 (Construction of knowledge graph in ).

[0187] In the non-first matching during training, since some association scores have been stored, the process is similar to the above-mentioned iterative process and will not be described in detail in this embodiment.

[0188] In addition, an embodiment of the present application further proposes a storage medium, on which a data matching program is stored. When the data matching program is executed by a processor, the data matching method described above is implemented.

[0189] In addition, refer to Fig.10 , Fig.10 This is a structural block diagram of the first embodiment of the data matching method and device of this application; Fig.10 As shown, the embodiment of the present application also proposes a data matching method device, the device comprising:

[0190] The data acquisition module 101 is used to acquire the data to be matched and the matching target data;

[0191] A keyword extraction module 102 is used to extract keywords from the to-be-matched data and the matching target data through a preset base large model, wherein the preset base large model is obtained by model training based on the association score between the to-be-matched sample keywords and the matching target sample keywords in the current usage scenario, and the association score is determined based on a preset knowledge graph in the current usage scenario;

[0192] The result generation module 103 is used to match the keyword extraction results through a preset recommendation model to generate a final matching result.

[0193] This embodiment can determine the correlation score between the sample keywords to be matched and the target sample keywords to be matched according to the preset knowledge graph in the current usage scenario, and then perform model training according to the correlation score to obtain a preset base large model. When the data to be matched and the target data to be matched in the current usage scenario are obtained, the preset base large model can be used to extract keywords from the data to be matched and the target data to be matched, and then the keyword extraction results can be matched through the preset recommendation model to generate the final matching result. Compared with the existing recommendation model, when it is directly used for matching operations in new scenarios, the feature extraction capability is insufficient and the accuracy is poor. This embodiment can first use the preset base large model obtained by model training through the correlation score in the current usage scenario to extract keywords, and then match the keyword extraction results through the preset recommendation model, so that when the recommendation model is directly used for matching operations in new scenarios, keyword extraction can be performed first through the preset base large model, thereby improving accuracy.

[0194] As an implementation mode, the data acquisition module 101 is also used to obtain the sample data to be matched and the matching target sample data in the current usage scenario, and obtain the preset knowledge graph in the current usage scenario; determine the sample keywords to be matched corresponding to the sample data to be matched and the matching target sample keywords corresponding to the matching target sample data according to the preset knowledge graph; determine the association score between the sample keywords to be matched and the matching target sample keywords based on the preset knowledge graph; train the initial base large model through the sample keywords to be matched, the matching target sample keywords and the association score to obtain the preset base large model.

[0195] Based on the above-mentioned first embodiment of the data matching method and device of the present application, a second embodiment of the data matching method and device of the present application is proposed.

[0196] In this embodiment, the data acquisition module 101 is also used to filter the association scores according to a preset filtering method to obtain target association scores, and generate prompt word text based on the target association scores; the initial base large model is trained through the prompt word text, the sample keywords to be matched and the matching target sample keywords to obtain a preset base large model.

[0197] As an implementation mode, the data acquisition module 101 is also used to, when the preset screening method is the first screening method, filter the association scores based on a first preset threshold to obtain a target association score; when the preset screening method is the second screening method, sort the association scores, and filter the association scores according to a preset screening quantity to obtain the target association score.

[0198] As an implementation mode, the keyword extraction module 102 is also used to perform feature extraction on the data to be matched through a preset base large model to obtain a first feature vector, and to perform feature extraction on the matching target data through the preset base large model to obtain a second feature vector; and to decode the first feature vector and the second feature vector to obtain a keyword extraction result.

[0199] As an implementation mode, the data acquisition module 101 is also used to obtain the sample data to be matched and the matching target sample data in the current usage scenario, and obtain the standard matching results between the sample data to be matched and the matching target sample data; the initial recommendation model is trained through the keywords of the sample to be matched, the keywords of the matching target sample and the standard matching results to obtain a preset recommendation model.

[0200] As an implementation mode, the result generation module 103 is also used to obtain new data to be matched and new matching target data, and independently input the first feature vector and the second feature vector into the preset base large model; perform keyword extraction on the new data to be matched and the new matching target data based on the first feature vector and the second feature vector through the preset base large model; match the new keyword extraction results through the preset recommendation model to generate a new final matching result.

[0201] Based on the above-mentioned embodiments of the data matching method and device of the present application, a third embodiment of the data matching method and device of the present application is proposed.

[0202] In this embodiment, the result generation module 103 is also used to determine a preset processing method, which is independent processing, splicing processing or difference processing; the keyword extraction results are processed according to the preset processing method, and the processed keyword extraction results are matched through a preset recommendation model to generate a final matching result.

[0203] As an implementation mode, the result generation module 103 is also used to match the keyword extraction results through a preset recommendation model to generate a final matching result when the preset processing mode is the independent processing; to splice the keyword extraction results when the preset processing mode is the splicing processing, and to match the spliced ​​keyword extraction results through a preset recommendation model to generate a final matching result; and to differentiate the keyword extraction results item by item when the preset processing mode is the difference processing, to splice the keyword extraction results after the difference item by item, and to match the spliced ​​keyword extraction results through a preset recommendation model to generate a final matching result.

[0204] As an implementation mode, the initial matching model includes a single-tower recall recommendation model, a dual-tower recall recommendation model or a convolutional recommendation model, and the convolutional recommendation model is a model that extracts features in a dimensional overall convolution manner.

[0205] Based on the above-mentioned embodiments of the data matching method and device of the present application, a fourth embodiment of the data matching method and device of the present application is proposed.

[0206] In this embodiment, the result generation module 103 is also used to determine the similarity score of the keyword extraction result and determine the feedback score corresponding to the final matching result; construct a correlation matrix of the keyword extraction result based on the similarity score and the feedback score, and filter the correlation matrix through a second preset threshold; determine the confidence of the keyword extraction result according to the filtering result, and iterate the preset knowledge graph according to the confidence.

[0207] Other embodiments or specific implementations of the data matching method device described in this application can refer to the above-mentioned method embodiments and will not be repeated here.

[0208] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0209] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0210] 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 a read-only memory image (ROM) / random access memory (RAM), a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0211] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A data matching method, characterized in that: The method comprises: Obtain the data to be matched and the matching target data; Keyword extraction is performed on the to-be-matched data and the matching target data through a preset base large model, wherein the preset base large model is obtained through model training according to the association score between the to-be-matched sample keywords and the matching target sample keywords in the current usage scenario, and the association score is determined according to the preset knowledge graph in the current usage scenario; The keyword extraction results are matched through the preset recommendation model to generate the final matching results.

2. The method according to claim 1, characterized in that The step of extracting keywords from the to-be-matched data and the matching target data by using the preset base macro model includes: Performing feature extraction on the to-be-matched data through a preset base macro model to obtain a first feature vector, and performing feature extraction on the matching target data through the preset base macro model to obtain a second feature vector; The first feature vector and the second feature vector are decoded to obtain a keyword extraction result.

3. The method according to claim 2, characterized in that After the step of matching the keyword extraction results by using the preset recommendation model to generate the final matching results, the method further includes: Acquire new to-be-matched data and new matching target data, and independently input the first feature vector and the second feature vector into the preset base macro model; Perform keyword extraction on the new to-be-matched data and the new matching target data based on the first feature vector and the second feature vector by using the preset base macro model; The new keyword extraction results are matched through the preset recommendation model to generate new final matching results.

4. The method according to claim 1, characterized in that The step of matching the keyword extraction results by using a preset recommendation model to generate a final matching result includes: Determine a preset processing mode, wherein the preset processing mode is independent processing, splicing processing or difference processing; The keyword extraction results are processed according to the preset processing method, and the processed keyword extraction results are matched through the preset recommendation model to generate a final matching result.

5. The method according to claim 4, characterized in that The step of processing the keyword extraction results according to the preset processing method, and matching the processed keyword extraction results through a preset recommendation model to generate a final matching result includes: When the preset processing mode is the independent processing, matching the keyword extraction results by using a preset recommendation model to generate a final matching result; When the preset processing mode is the splicing processing, the keyword extraction results are spliced, and the spliced ​​keyword extraction results are matched through a preset recommendation model to generate a final matching result; When the preset processing method is the subtraction processing, the keyword extraction results are subtracted item by item, the keyword extraction results after subtracting item by item are spliced, and the spliced ​​keyword extraction results are matched through the preset recommendation model to generate the final matching result.

6. The method according to claim 1, characterized in that The preset recommendation model includes a single-tower recall recommendation model, a dual-tower recall recommendation model or a convolutional recommendation model, and the convolutional recommendation model is a model that extracts features in a dimensional overall convolution manner.

7. The method according to claim 1, characterized in that Before the steps of obtaining the data to be matched and matching the target data, the method further includes: Obtaining the sample data to be matched and the matching target sample data in the current usage scenario, and obtaining the preset knowledge graph in the current usage scenario; Determine, according to the preset knowledge graph, the sample keywords to be matched corresponding to the sample data to be matched and the matching target sample keywords corresponding to the matching target sample data; Determine the association score between the to-be-matched sample keyword and the matching target sample keyword based on the preset knowledge graph; The initial base large model is trained through the sample keywords to be matched, the matching target sample keywords and the associated scores to obtain a preset base large model.

8. The method according to claim 7, characterized in that The step of training the initial base large model by using the to-be-matched sample keywords, the matching target sample keywords and the associated scores to obtain the preset base large model includes: Filtering the relevance scores according to a preset filtering method to obtain target relevance scores, and generating prompt word text based on the target relevance scores; The initial base large model is trained through the prompt word text, the sample keywords to be matched and the matching target sample keywords to obtain a preset base large model.

9. The method according to claim 8, characterized in that The step of screening the association scores according to a preset screening method to obtain a target association score includes: When the preset screening method is the first screening method, screening the association scores based on the first preset threshold value to obtain a target association score; When the preset screening method is the second screening method, the relevance scores are sorted, and the relevance scores are screened according to a preset screening quantity to obtain the target relevance scores.

10. The method according to claim 1, characterized in that Before the steps of obtaining the data to be matched and matching the target data, the method further includes: Acquire the sample data to be matched and the matching target sample data in the current usage scenario, and acquire the standard matching result between the sample data to be matched and the matching target sample data; The initial recommendation model is trained by using the to-be-matched sample keywords, the matching target sample keywords and the standard matching results to obtain a preset recommendation model.

11. The method according to any one of claims 1 to 10, characterized in that After the step of matching the keyword extraction results by using the preset recommendation model to generate the final matching results, the method further includes: Determine a similarity score of the keyword extraction result, and determine a feedback score corresponding to the final matching result; constructing a correlation matrix of the keyword extraction results based on the similarity scores and the feedback scores, and screening the correlation matrix by a second preset threshold; The confidence of the keyword extraction result is determined according to the screening result, and the preset knowledge graph is iterated according to the confidence.

12. A data matching device, characterized in that: The device comprises: A data acquisition module is used to acquire the data to be matched and the matching target data; A keyword extraction module, used to extract keywords from the to-be-matched data and the matching target data through a preset base large model, wherein the preset base large model is obtained through model training based on the association score between the to-be-matched sample keywords and the matching target sample keywords in the current usage scenario, wherein the association score is determined based on a preset knowledge graph in the current usage scenario; The result generation module is used to match the keyword extraction results through a preset recommendation model to generate the final matching results.

13. A storage medium, characterized in that: The storage medium stores a data matching program, and when the data matching program is executed by the processor, the steps of the data matching method according to any one of claims 1 to 11 are implemented.

14. A data matching device, characterized in that: The data matching device comprises: a memory, a processor, and a data matching program stored in the memory and executable on the processor, wherein the data matching program implements the steps of the data matching method according to any one of claims 1 to 11 when executed by the processor.

Images