Data matching method, device and equipment

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a data matching method, device and equipment, and the method comprises the steps: obtaining to-be-matched data and matched target data; extracting to-be-matched features according to the to-be-matched data, and extracting matched target features according to the matched target data; determining a currently used preset cue word template by using the to-be-matched features and the matched target features, the preset cue word template being obtained by training based on different types of to-be-matched sample features and corresponding matched target sample features, the preset cue word template being composed of trainable variables; and matching the to-be-matched features with the matching target features according to a preset cue word template to obtain a matching result. Compared with the prior art, matching can be performed through the preset prompt word template which is more adaptive to the to-be-matched feature and the matching target feature, so that feature omission can be reduced, and the matching accuracy is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a data matching method, device and equipment. Background Art

[0002] At present, with the development of big model technology, artificial intelligence has increasingly stronger capabilities, especially in the field of intelligent matching (for example, when matching people with jobs, it is necessary to match jobs with job seekers' resumes, etc.).

[0003] In the existing matching process, some feature classifications (such as work experience, salary, etc.) are generally defined in advance, and then the matching data (such as job applicant resumes) and matching target data (such as positions) are extracted and matched according to these feature classifications. However, since the matching data and matching target data are generally in text form, and the text expression methods are diverse, matching by manually defining feature classifications may not cover all features, resulting in some features being omitted during matching, which leads to poor subsequent matching accuracy. Summary of the invention

[0004] The main purpose of the present application is to provide a data matching method, device and equipment, aiming to solve the technical problem that the existing matching method through artificially defined feature classification is prone to missing features, resulting in poor subsequent matching 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] Extracting to-be-matched features according to the to-be-matched data, and extracting matching target features according to the matching target data;

[0008] Determine a currently used preset prompt word template by using the to-be-matched feature and the matching target feature, wherein the preset prompt word template is obtained by training based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables;

[0009] The to-be-matched feature is matched with the matching target feature according to the preset prompt word template to obtain a matching result.

[0010] In one embodiment, the step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result includes:

[0011] Inputting and constructing the to-be-matched feature, the matching target feature, and the preset prompt word template according to a preset construction format to obtain an input construction result;

[0012] The to-be-matched feature is matched with the matching target feature based on the input construction result to obtain a matching result.

[0013] In one embodiment, the preset construction format is a construction format in which the to-be-matched feature, the matching target feature, and the preset prompt word template are spliced ​​in any order.

[0014] In one embodiment, the input structure is established as follows:

[0015] Matrix concatenation is performed in the order of the to-be-matched feature, the matching target feature, and the preset prompt word template;

[0016] Matrix concatenation is performed in the order of the to-be-matched feature, the preset prompt word template, and the matching target feature;

[0017] Matrix concatenation is performed in the order of the matching target feature, the feature to be matched, and the preset prompt word template;

[0018] Perform matrix concatenation in the order of the matching target feature, the preset prompt word template, and the feature to be matched;

[0019] Matrix concatenation is performed in the order of the preset prompt word template, the feature to be matched, and the matching target feature;

[0020] Or, matrix series connection is performed in the order of the preset prompt word template, the matching target feature and the feature to be matched.

[0021] In one embodiment, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a job,

[0022] The input construction result is: the result of the concatenation of the features of the job seeker's resume, the preset prompt word template, and the feature matrix of the job description;

[0023] The input construction result is: the characteristics of the job seeker's resume, the characteristics of the job description, and the result of the concatenation of the preset prompt word template matrix;

[0024] The input construction result is: the characteristics of the job description, the characteristics of the job seeker's resume and the result of the concatenation of the preset prompt word template matrix;

[0025] The input construction result is: the result of the concatenation of the characteristics of the job description, the preset prompt word template and the characteristic matrix of the job seeker's resume;

[0026] The input construction result is: the result of the concatenation of the preset prompt word template, the features of the job seeker's resume, and the feature matrix of the job description;

[0027] Alternatively, the input construction result is: a result of concatenating the preset prompt word template, the features of the job description, and the feature matrix of the job seeker's resume.

[0028] In one embodiment, before the steps of obtaining the to-be-matched data and matching the target data, the method further includes:

[0029] Obtain an initial prompt word template containing trainable variables;

[0030] Obtaining a current iteration parameter based on the initial prompt word template, and determining a corresponding gradient parameter based on a preset loss function, the current iteration parameter, the to-be-matched sample feature, and the matching target sample feature;

[0031] Determine a matrix estimate according to the gradient parameter, and obtain a next iteration model parameter according to the matrix estimate and the current iteration model parameter;

[0032] The trainable variables are updated by the next iteration model parameters to obtain a preset prompt word template.

[0033] In one embodiment, the step of determining a matrix estimate according to the gradient parameter, and obtaining a next iteration model parameter according to the matrix estimate and the current iteration model parameter comprises:

[0034] Determine a first-order matrix estimate and a second-order matrix estimate according to the gradient parameters;

[0035] The next iteration model parameters are obtained based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix and the second-order estimation matrix.

[0036] In one embodiment, the step of obtaining the next iterative model parameters based on the preset learning rate, the preset constant, the current iterative model parameters, the first-order estimation matrix, and the second-order estimation matrix includes:

[0037] Obtaining the next iteration model parameters through a preset parameter update formula based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix, and the second-order estimation matrix;

[0038] The preset parameter update formula is:

[0039]

[0040] Wherein, η is the preset learning rate, is the first-order estimation matrix, is the second-order estimation matrix, ∈ is the preset constant, θ k are the model parameters of the current iteration.

[0041] In one embodiment, after the step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result, the step further includes:

[0042] When the amount of data of the matching target data is higher than a preset threshold, obtaining a hidden layer prediction score corresponding to the matching result;

[0043] The matching target data are sorted based on the hidden layer prediction scores to obtain a sorting result.

[0044] In one embodiment, the step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result includes:

[0045] Matching the to-be-matched feature and the matching target feature according to the preset prompt word template to obtain a prediction probability, and obtaining a matching result according to the prediction probability;

[0046] When the amount of data of the matching target data is higher than a preset threshold, the step of obtaining the hidden layer prediction score corresponding to the matching result includes:

[0047] When the data volume of the matching target data is higher than a preset threshold, a hidden layer prediction score corresponding to the matching result is determined according to the prediction probability.

[0048] In one embodiment, the step of extracting the to-be-matched features according to the to-be-matched data and extracting the matching target features according to the matching target data includes:

[0049] Performing format detection on the to-be-matched data and the matching target data respectively;

[0050] When the format detection results are all the target preset formats, extracting the to-be-matched features according to the to-be-matched data, and extracting the matching target features according to the matching target data;

[0051] When there is a format that is not in the preset format of the target in the format detection result, format conversion is performed on the object to be converted, and features to be matched are extracted based on the data to be matched after the format conversion, and matching target features are extracted based on the matching target data after the format conversion, and the object to be converted is the data to be matched that is not in the preset format of the target and / or the matching target data that is not in the preset format of the target.

[0052] In one embodiment, when there is a format that is not the target preset format in the format detection result, the step of performing format conversion on the object to be converted includes:

[0053] When the first preset format exists in the format detection result, text extraction is performed on the object to be converted according to a preset database, and format conversion is performed on the object to be converted according to the extraction result;

[0054] When the second preset format exists in the format detection result, the object to be converted is subjected to category analysis, a corresponding format conversion method is determined according to the category analysis result, and the format conversion is performed on the object to be converted according to the corresponding format conversion method.

[0055] In one embodiment, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a job position, the target preset format is a txt format and the first preset format is a PDF format;

[0056] The step of extracting text from the object to be converted according to the preset database and converting the format of the object to be converted according to the extraction result includes:

[0057] When the object to be converted is a single column, extracting text from the object to be converted according to a preset database, and converting the format of the object to be converted according to the extraction result;

[0058] When the object to be converted is a double column, text is extracted from each single column of the object to be converted according to a preset database, the extraction results are spliced, and the format of the object to be converted is converted according to the splicing result.

[0059] In one embodiment, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a job, the target preset format is a txt format and the second preset format is a word format;

[0060] The step of performing category analysis on the object to be converted, determining a corresponding format conversion method according to the category analysis result, and performing format conversion on the object to be converted according to the corresponding format conversion method includes:

[0061] Performing category analysis on the object to be converted, and acquiring a corresponding preset conversion module according to the category analysis result, wherein the preset conversion module has a corresponding format conversion method;

[0062] The format of the object to be converted is converted by each of the preset conversion modules according to the corresponding format conversion method.

[0063] In addition, to achieve the above purpose, the present application also provides a data matching device, which includes:

[0064] A data acquisition module is used to acquire the data to be matched and the matching target data;

[0065] A feature extraction module, used to extract to-be-matched features according to the to-be-matched data, and to extract matching target features according to the matching target data;

[0066] A template determination module, used to determine a currently used preset prompt word template using the to-be-matched feature and the matching target feature, wherein the preset prompt word template is obtained by training based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables;

[0067] The data matching module is used to match the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result.

[0068] 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.

[0069] The embodiments of the present application provide a data matching method, apparatus and device, the method comprising: obtaining data to be matched and matching target data; extracting features to be matched according to the data to be matched, and extracting matching target features according to the matching target data; determining a currently used preset prompt word template using the features to be matched and the matching target features, the preset prompt word template being trained based on different types of features of samples to be matched and corresponding features of matching target samples, the preset prompt word template being composed of trainable variables; matching the features to be matched with the matching target features according to the preset prompt word template to obtain a matching result.

[0070] Since the present application can first obtain each preset prompt word template through training of different types of to-be-matched sample features and corresponding matching target sample features, the preset prompt word template can be composed of trainable variables. When in use, the to-be-matched data features are first extracted to obtain the to-be-matched features, and the matching target data features are extracted to obtain the matching target features, and then the corresponding preset prompt word template is determined according to the to-be-matched features and the matching target features, and finally the to-be-matched features and the matching target features are matched according to the preset prompt word template to obtain the matching result. Compared with the existing matching method of artificially defining feature classification, the present application can first obtain the preset prompt word template between different types of to-be-matched sample features and matching target sample features, and then match them through the preset prompt word template that is more suitable between the to-be-matched features and the matching target features, thereby reducing feature omissions and improving the accuracy of matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] 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.

[0072] 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.

[0073] Figure 1 A schematic diagram of the data matching device structure of the hardware operating environment involved in the embodiment of the present application;

[0074] Figure 2 This is a flow chart of the first embodiment of the data matching method of this application;

[0075] Figure 3 This is a flowchart of format conversion in the first embodiment of the data matching method of this application;

[0076] Figure 4 A schematic diagram of the matching process in the first embodiment of the data matching method of this application;

[0077] Figure 5 This is a flow chart of the second embodiment of the data matching method of this application;

[0078] Figure 6 This is a flow chart of the third embodiment of the data matching method of the present application;

[0079] Figure 7 This is a structural block diagram of the first embodiment of the data matching device of this application.

[0080] 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

[0081] 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.

[0082] 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.

[0083] like Figure 1 As 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 can be connected to a display screen (Display), and the optional user interface 1003 can include a standard wired interface and a wireless interface. The wired interface of the user interface 1003 can be a USB interface in this application. The network interface 1004 can optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 can 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 can also be a storage device independent of the aforementioned processor 1001.

[0084] 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.

[0085] 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.

[0086] 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 steps of the data matching method provided in the embodiment of the present application.

[0087] It should be noted that, at present, with the development of big model technology, artificial intelligence has increasingly stronger capabilities, especially in the field of intelligent matching (for example, when matching people with jobs, it is necessary to match jobs with job seekers' resumes, etc.).

[0088] In the existing matching process, some feature classifications (such as work experience, salary, etc.) are generally defined in advance, and then the matching data (such as job applicant resumes) and matching target data (such as positions) are extracted and matched according to these feature classifications. However, since the matching data and matching target data are generally in text form, and the text expression methods are diverse, matching by manually defining feature classifications may not cover all features, resulting in some features being omitted during matching, which leads to poor subsequent matching accuracy.

[0089] Therefore, in order to solve the above defects, this embodiment provides a data matching method, which can first obtain various preset prompt word templates through training of different types of sample features to be matched and corresponding matching target sample features. The preset prompt word template can be composed of trainable variables. When used, the features of the data to be matched are first extracted to obtain the features to be matched, and the features of the matching target data are extracted to obtain the matching target features. Then, the corresponding preset prompt word template is determined according to the features to be matched and the matching target features. Finally, the features to be matched and the matching target features are matched according to the preset prompt word template to obtain the matching results. Compared with the existing matching method using artificially defined feature classification, this embodiment can first obtain preset prompt word templates between different types of sample features to be matched and matching target sample features, and then match them through preset prompt word templates that are more suitable between the features to be matched and the matching target features, thereby reducing feature omissions and improving the accuracy of matching.

[0090] For ease of understanding, the following combination Figures 2 to 6 The data matching method provided in the embodiment of the present application is introduced in detail.

[0091] 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:

[0092] Step S10: Acquire the data to be matched and the matching target data.

[0093] It is understandable that the method of this embodiment can be applied to a data matching device, wherein the above-mentioned data matching device can be any device that implements the matching function, 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 (hereinafter referred to as the device), and this device is used as an example to illustrate this embodiment and the following embodiments.

[0094] It should be understood that the above-mentioned data to be matched may be data that needs to be evaluated or compared in the matching task. The above-mentioned matching target data may be data used as a comparison benchmark in the matching task. The scenario used in this embodiment may be any scenario that needs to be matched, but for the convenience of subsequent description, this embodiment uses the scenario of person-job matching for description, and then in the person-job matching, the data to be matched may be the resumes of many job seekers, and the matching target data may be the job description of the recruitment position in the person-job matching.

[0095] It should also be understood that, generally in job recruitment, the recruiter will describe the job being recruited, and the description can be used as the above-mentioned matching target data. At the same time, the recruiter will also receive resumes submitted by job seekers, and these resumes can be used as the above-mentioned target data to be matched.

[0096] Step S20: extracting to-be-matched features according to the to-be-matched data, and extracting matching target features according to the matching target data.

[0097] It should be noted that, considering that the data to be matched and the matching target data may contain features that are useless for matching, thereby affecting the subsequent matching effect, in this embodiment, feature extraction can be performed on the data to be matched and the matching target data respectively, and features required for subsequent matching, such as name, education, professional courses, project experience, and internship experience, can be extracted, and this embodiment does not limit this. The matching features corresponding to the data to be matched and the matching target features corresponding to the matching target data are obtained.

[0098] It should also be noted that the above-mentioned matching target data, such as job descriptions, are generally concise and organized texts, so feature extraction is not required, and this embodiment does not impose any restrictions on this.

[0099] Step S30: Determine the currently used preset prompt word template using the to-be-matched features and the matching target features, wherein the preset prompt word template is trained based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables.

[0100] It should be noted that the preset prompt word template can be a virtual text description template set when matching, and the actual embodiment of the preset prompt word template can be in the form of a discrete matrix, that is, it can be a template formed by the series connection of discrete prompt words (prompt) and continuous prompt words (i.e., prompt embedding). The preset prompt word template can be composed of trainable variables, and the trainable variables can be in the form of vectors. The matrix is ​​formed by these vectors, so the preset prompt word template is in the form of a matrix. By pre-training different types of sample features to be matched and corresponding matching target sample features, the trainable vectors therein are continuously trained, so that the trainable vectors in the trained preset prompt word template carry the matching relationship between the sample features to be matched and the matching target sample features, so that when matching is performed later, the model can better understand the relationship between the features to be matched and the matching target sample features, and perform more comprehensive feature extraction and matching operations, thereby reducing the omission of features. Since the preset prompt word template needs to be obtained through training, the training process can be the process of updating the above-mentioned trainable vectors.

[0101] It should be emphasized that the number of trainable variables in the above preset prompt word template can be set according to actual conditions, and this embodiment does not impose any restrictions on this.

[0102] It should also be noted that in this embodiment, the above-mentioned device may be provided with a preset matching model for performing matching tasks, and the preset matching model may be any model trained to obtain a preset prompt word template, such as a pre-trained language model (Pre-trained Language Models, PLMs), etc., which is not limited in this embodiment. For example, the Chinese-BERT-WWM model or the Chinese-BERT-WWM-EXT model may be used.

[0103] It is understandable that the above-mentioned to-be-matched sample features may be features obtained by extracting features of the to-be-matched sample data, and the above-mentioned matching target sample features may be features obtained by extracting features of the matching target sample data. The above-mentioned to-be-matched sample data may be to-be-matched data for training, and the above-mentioned matching target sample data may be matching target data for training. Before use, an untrained preset matching model may be recorded as an initial matching model, and an initial prompt word template may be provided in the initial matching model, and the initial prompt word template may be composed of untrained trainable variables.

[0104] Different types of to-be-matched sample data and matching target sample data may be obtained first, and then the initial prompt word template, the current type of to-be-matched sample data and the matching target sample data may be input together into the initial matching model. At the same time, the actual matching results corresponding to the to-be-matched sample data and the matching target sample data may also be input. The actual matching results may be obtained through manual annotation, and of course may also be obtained through other methods, which are not limited in this embodiment.

[0105] After inputting into the initial matching model, the initial matching model can first extract features from the sample data to be matched, obtain features of the sample data to be matched, and extract features from the matching target sample data, obtain features of the matching target sample. Then, the initial prompt word template is trained through the features of the sample data to be matched, the features of the matching target sample, and the actual matching results, and the discrete matrix (i.e., the training variables) therein is continuously adjusted. When the training is completed, the preset matching model can be obtained, and the preset matching model can contain preset prompt word templates corresponding to various types of sample data to be matched and the data to be matched.

[0106] In actual use, the data to be matched and the matching target data can be input into the preset matching model. After the preset matching model performs feature extraction to obtain the features to be matched and the matching target features, the adapted preset prompt word template can be selected from each preset prompt word template according to the features to be matched and the matching target features.

[0107] Step S40: matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result.

[0108] After obtaining the corresponding preset prompt word template, the to-be-matched feature and the matching target feature can be matched according to the preset prompt word template, thereby obtaining a "match" or "mismatch" matching result. Therefore, in this embodiment, since the matching between the to-be-matched feature and the matching target feature can be completed by the preset prompt word template with trainable variables, the preset prompt word template can carry the matching relationship between the to-be-matched sample feature of the current type and the matching target sample feature, so that when matching, the model can better understand the relationship between the to-be-matched feature and the matching target sample feature, reduce feature omissions, and improve the accuracy of matching.

[0109] Furthermore, considering that the data to be matched and the matching target data may exist in various formats, such as a resume in Portable Document Format (PDF), a resume in Word format, or a resume in Text File (txt) format, different file processing methods may be inconsistent. Therefore, in order to facilitate subsequent processing, it is necessary to perform format conversion in advance to convert the data to be matched and the matching target data into a unified format. The specific process is that the above step S20 includes:

[0110] Step S21: performing format detection on the to-be-matched data and the matching target data respectively.

[0111] Step S22: when all format detection results are target preset formats, extracting to-be-matched features according to the to-be-matched data, and extracting matching target features according to the matching target data.

[0112] Reference Figure 3 , Figure 3 This is a flow chart of format conversion in the first embodiment of the data matching method of this application, such as Figure 3 As shown, it should be understood that after obtaining the to-be-matched data and the matching target data, the above-mentioned device may first perform format detection on the to-be-matched data and the matching target data (ie Figure 3 Specifically, the file name suffixes of the data to be matched and the target data to be matched are determined respectively, and the format of the data to be matched and the format of the target data to be matched are determined according to the file name suffixes, and each determined format is used as the above-mentioned format detection result.

[0113] For example, when it is detected that the file name suffix of the to-be-matched data or the matching target data is ".txt", it can be determined that the corresponding format is txt format (i.e. Figure 3 TXT); when the file name suffix of the to-be-matched data or the matching target data is detected as ".docx", it can be determined that the corresponding format is word format (i.e. Figure 3 in Word); when the file name suffix of the to-be-matched data or the matching target data is detected to be ".pdf", it can be determined that the corresponding format is PDF format (i.e. Figure 3 (PDF).

[0114] Since the subsequent model generally supports files in txt format, in this embodiment, all files in non-txt format can be converted (i.e. Figure 3 txt format (i.e. Figure 3TXT). Then, when detecting the format, the txt format can be set as the above-mentioned target preset format. Then, after determining the detection results of each format, it can be determined whether the format detection results of the data to be matched and the format detection results of the matching target data are both the target preset format.

[0115] When it is determined that the format detection results of the data to be matched and the format detection results of the matching target data are both in txt format, it can be explained that the data in this format can be processed subsequently, so there is no need to perform format conversion, that is, directly extract the features to be matched from the data to be matched, and extract the matching target features from the matching target data.

[0116] Step S23: When there is a format that is not in the preset format of the target in the format detection result, format conversion is performed on the object to be converted, and features to be matched are extracted based on the data to be matched after the format conversion, and matching target features are extracted based on the matching target data after the format conversion, and the object to be converted is the data to be matched that is not in the preset format of the target and / or the matching target data that is not in the preset format of the target.

[0117] It should be noted that when the format detection results of the to-be-matched data and the format detection results of the matching target data are not in the txt format, such as the PDF format or the word format, it can be said that the format cannot be processed subsequently, so format conversion is required. It can also be understood that all the to-be-matched data and the matching target data that are not in the txt format are converted into the txt format.

[0118] It should also be noted that the above-mentioned object to be converted can be an object whose format detection result is not in the target preset format, that is, the data to be matched that is not in the txt format and / or the matching target data that is not in the txt format. When performing format conversion, an object that is not in the txt format can be determined as the above-mentioned object to be converted, and the format conversion is performed on the object to be converted, and then the matching features are extracted from the subsequently obtained data to be matched after the format conversion, and the matching target features are extracted from the matching target data after the format conversion.

[0119] It should be emphasized that since the matching target data (such as job descriptions) are generally txt files, they do not need to go through the above-mentioned format conversion process. On the contrary, the data to be matched (such as job applicant resumes) are generally in PDF or word format files, and are often uniquely formatted. There are also pictures, tables and other content in them. Therefore, the above-mentioned format conversion will be used more for the data to be matched.

[0120] Furthermore, considering that different formats may have different conversion methods, in this embodiment, when there is a format detection result that is not the target preset format, the step of performing format conversion on the object to be converted includes:

[0121] Step S231: when the first preset format exists in the format detection result, text extraction is performed on the object to be converted according to a preset database, and format conversion is performed on the object to be converted according to the extraction result.

[0122] It should be noted that the first preset format may be a PDF format. The preset database may be a database that extracts data in PDF format and converts it into a txt format. In actual use, when it is determined that the data to be matched and / or the matching target data are in PDF format, the device may extract the data in the data to be matched and / or the matching target data in PDF format through the preset database, and convert it into a text corresponding to the txt format, thereby obtaining the data to be matched after format conversion and / or the matching target data after format conversion.

[0123] Furthermore, considering that there are usually single-column and double-column situations in PDF format, for example, PDF resumes have single-column and double-column situations, the above-mentioned device can first detect whether it is a single column or a double column when converting the format of the to-be-matched data and / or the matching target data in PDF format; if it is a single column, it can be extracted and converted through the preset database in the above manner; if it is a double column, each single column can be extracted and converted through the preset database in the above manner, and then the conversion results can be spliced.

[0124] Specifically, when the data to be matched is a job seeker's resume and the target data to be matched is a job description of a job, the target preset format is a txt format and the first preset format is a PDF format;

[0125] The step of extracting text from the object to be converted according to the preset database and converting the format of the object to be converted according to the extraction result includes:

[0126] When the object to be converted is a single column, extracting text from the object to be converted according to a preset database, and converting the format of the object to be converted according to the extraction result;

[0127] When the object to be converted is a double column, text is extracted from each single column of the object to be converted according to a preset database, the extraction results are spliced, and the format of the object to be converted is converted according to the splicing result.

[0128] In actual use, specifically, when the data to be matched is a resume of a job seeker and the target data to be matched is a job description of a job position, when the format is determined to be PDF format, it can be determined first whether the data to be matched and / or the target data to be matched is single-column or double-column. When it is single-column, the data in the data to be matched and / or the target data to be matched in PDF format can be extracted through a preset database and converted into text corresponding to the txt format to complete the format conversion;

[0129] When it is a double column, the data to be matched and / or the data in the matching target data in PDF format can be extracted from each single column through the preset database and converted into corresponding text in txt format. Then the extraction results of each single column are spliced ​​to complete the format conversion.

[0130] Step S232: When the second preset format exists in the format detection result, category analysis is performed on the object to be converted, a corresponding format conversion method is determined according to the category analysis result, and format conversion is performed on the object to be converted according to the corresponding format conversion method.

[0131] It is understandable that the second preset format can be a word format. Considering that a file in word format generally contains content such as pictures and tables in addition to text, different content has different format conversion methods. Therefore, in actual use, after determining that the data to be matched and / or the matching target data are in word format, the above-mentioned device can first perform category analysis on the content in the data to be matched and / or the matching target data to determine which categories exist therein, and obtain the corresponding format conversion method according to the category analysis result, and perform format conversion on the content of the corresponding category according to each format conversion method, thereby completing the format conversion of the data to be matched and / or the matching target data.

[0132] Furthermore, in order to realize the format conversion of data in word format, in this embodiment, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a recruitment position, the target preset format is txt format and the second preset format is word format;

[0133] The step of performing category analysis on the object to be converted, determining a corresponding format conversion method according to the category analysis result, and performing format conversion on the object to be converted according to the corresponding format conversion method includes:

[0134] Performing category analysis on the object to be converted, and acquiring a corresponding preset conversion module according to the category analysis result, wherein the preset conversion module has a corresponding format conversion method;

[0135] The format of the object to be converted is converted by each of the preset conversion modules according to the corresponding format conversion method.

[0136] It should be understood that the preset conversion module can be a plug-in module for converting the format of the corresponding category of content, which can be set according to actual conditions, and this embodiment does not limit this. The preset conversion module can include a format conversion method for converting the format of the category of content.

[0137] In actual use, specifically, when the data to be matched is a job applicant's resume, and the matching target data is a job description of a recruitment position, after being determined to be in Word format, a category analysis may be performed on the data to be matched and / or the matching target data to determine the categories contained therein as the above-mentioned category analysis results, and then the corresponding preset conversion module may be called according to the contained categories, and the format conversion of the content of the corresponding category may be performed according to the format conversion method contained in the preset conversion module.

[0138] Similarly, during training, the sample data to be matched and the matching target sample data can also undergo the above-mentioned format detection and format conversion operations and then the initial prompt word template can be trained to obtain the corresponding preset prompt word template. Since this process is consistent with the above-mentioned actual use process, this embodiment will not elaborate on it.

[0139] At the same time, since the previous format conversion may cause word order disorder during the conversion process, the feature extraction performed in this embodiment can prevent the disordered word order from affecting subsequent matching, thereby further improving the accuracy.

[0140] Furthermore, considering that there may be multiple matching target data during matching, for example, there may be multiple job descriptions, and if "match" or "no match" is directly output, it is not convenient for the user to check which job applicant is more suitable for. Therefore, in this embodiment, after the above step S40, the following is further included:

[0141] Step S50: when the data quantity of the matching target data is higher than a preset threshold, obtaining a hidden layer prediction score corresponding to the matching result;

[0142] Step S60: sorting the matching target data based on the hidden layer prediction scores to obtain a sorting result.

[0143] It should be understood that the preset threshold may be a threshold for determining whether there are multiple matching target data. In this embodiment, the preset threshold may be set to 1. The hidden layer prediction score may be a score for predicting the matching data and the matching target data in the hidden layer of the preset matching model.

[0144] In actual use, after obtaining the matching result, the above device can determine whether the number of data of the matching target data currently to be matched is greater than 1;

[0145] If it is not higher than 1, it means that there is only one job description that needs to be matched, that is, there is one job, so there is no need to perform subsequent sorting, and the matching results can be directly output;

[0146] When it is higher than 1, it means that there are more than one job descriptions that need to be matched, that is, there are at least two jobs, so refer to Figure 4 , Figure 4 This is a schematic diagram of the matching process in the first embodiment of the data matching method of this application, such as Figure 4 As shown, the above device can be matched from a preset matching model (i.e. Figure 4 The hidden layer prediction scores of each matching result are obtained from the hidden layer of the PLMs matching model (i.e. Figure 4 The matching scores in the hidden layer are, for example, 0.635, 0.325, 0.856, ..., 0.413, 0.246, and 0.656), and then the matching target data (i.e., each position) is sorted based on the hidden layer prediction score, either in ascending order or in descending order, so that the sorting results of each position (i.e., Figure 4 Topk sorting, such as Position 1, Position 2, and Position 3).

[0147] Furthermore, in order to obtain the hidden layer prediction score, in this embodiment, the step S40 includes:

[0148] Step S41: matching the to-be-matched feature and the matching target feature according to the preset prompt word template to obtain a prediction probability, and obtaining a matching result according to the prediction probability;

[0149] Accordingly, the above step S50 includes:

[0150] Step S51: when the data volume of the matching target data is higher than a preset threshold, determining a hidden layer prediction score corresponding to the matching result according to the prediction probability.

[0151] It should be noted that the above prediction probability may be the probability obtained by the hidden layer when the preset matching model matches the to-be-matched feature and the matching target feature according to the preset prompt word template. Therefore, this embodiment may directly use the prediction probability as the above hidden layer prediction score.

[0152] When matching, the preset matching model can directly obtain the above prediction probability in its hidden layer, and select the result with the highest probability as the above matching result according to the prediction probability, and can also use the prediction probability as the corresponding hidden layer prediction score.

[0153] In this embodiment, each preset prompt word template can be obtained by first training different types of to-be-matched sample features and corresponding matching target sample features. The preset prompt word template can be composed of trainable variables. When used, the to-be-matched data features are first extracted to obtain the to-be-matched features, and the matching target data features are extracted to obtain the matching target features. Then, the corresponding preset prompt word template is determined according to the to-be-matched features and the matching target features. Finally, the to-be-matched features and the matching target features are matched according to the preset prompt word template to obtain the matching results. Compared with the existing matching method of artificially defining feature classification, this embodiment can first obtain preset prompt word templates between different types of to-be-matched sample features and matching target sample features, and then match them through preset prompt word templates that are more suitable between the to-be-matched features and the matching target features, thereby reducing feature omissions and improving the accuracy of matching.

[0154] Reference Figure 5 , Figure 5 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.

[0155] like Figure 5 As shown, further, in order to input the preset prompt word template into the preset matching model, in this embodiment, the above step S40 includes:

[0156] Step S411: inputting and constructing the to-be-matched feature, the matching target feature, and the preset prompt word template according to a preset construction format to obtain an input construction result;

[0157] Step S412: matching the to-be-matched feature with the matching target feature based on the input construction result to obtain a matching result.

[0158] It is understandable that the above-mentioned preset construction format can be an input format supported by the preset matching model. For ease of understanding, in this embodiment, the set of to-be-matched features of all to-be-matched data is denoted as J, that is, all job seekers' resumes are denoted as J, and all to-be-matched features in each to-be-matched data are denoted as j m , where m is the number of data to be matched. For example, if there are m resumes of job seekers, the matching features of the first resume to the mth resume can be recorded as j1, j2, ..., j m , and then J = {j1, j2, ..., j m}; where each feature j to be matched is mThe essence of may be a vector, and the vectors of the to-be-matched features of the to-be-matched data constitute a matrix and are input into the above-mentioned preset matching model. That is, after the preset matching model obtains the to-be-matched data in the form of text, it can convert each to-be-matched data into a machine-recognizable vector form through operations such as format conversion and feature extraction (that is, the above-mentioned j m ), and then all the matching features j1, j2, ..., j m Construct the above matrix J.

[0159] Similarly, if the set of matching target features of all matching target data is recorded as C, that is, all job descriptions are recorded as C, and all matching target features in each matching target data are recorded as C n , where n is the number of matching target data. For example, if there are n positions, the matching target features of the first position to the matching target features of the nth position can be recorded as C1, C2, ..., C n , and then C={C1,C2,...,C n}. Among them, each matching target feature C of the matching target data n The essence of may be a vector, and the vectors of the matching target features of the matching target data constitute a matrix and are input into the above-mentioned preset matching model. That is, after the preset matching model obtains the matching target data in the form of text, it can convert each matching target data into a machine-recognizable vector form (that is, the above-mentioned C n ), and then all the matching target features C1, C2, ..., C n Construct the above matrix C.

[0160] Then, the to-be-matched features, the matching target features and the preset prompt word template may be input and constructed according to a preset construction format. When inputting and constructing, the preset construction format is a construction format in which the to-be-matched features, the matching target features and the preset prompt word template are spliced ​​in any order.

[0161] That is, since there are three inputs (the feature to be matched, the matching target feature, and the preset prompt word template), there are six construction formats in total. In this embodiment, the preset construction formats include but are not limited to the following:

[0162] The first one can be in the format of "features to be matched + matching target features + preset prompt word template";

[0163] The second type can be in the format of "feature to be matched + preset prompt word template + matching target feature";

[0164] The third type can be in the format of "matching target features + features to be matched + preset prompt word template";

[0165] The fourth type can be in the format of "matching target features + preset prompt word template + features to be matched";

[0166] The fifth type may be in the format of "preset prompt word template + features to be matched + matching target features";

[0167] The sixth type may be in the format of "preset prompt word template + matching target features + features to be matched".

[0168] Specifically, if Input is the input construction result, J is the feature to be matched, and Prompt temple is the preset prompt word template, and C is the matching target feature;

[0169] The above six preset construction formats can correspond to the following six input construction results, namely the first one corresponds to Input = J + C + Prompt temple ; The second type corresponds to Input = J + Prompt temple +C; the third type corresponds to Input=C+J+Prompt temple ; The fourth type corresponds to Input=C+Prompt temple +J; The fifth type corresponds to Input = Prompt temple +J+C; the sixth type corresponds to Input=Prompt temple +C+J.

[0170] It should be emphasized that since all three are in the form of matrices, when performing input construction, the essence of its construction can be to connect the matrices in series, that is, when establishing the input structure, the input structure can be established in the following manner: connect the matrices in series in the order of the features to be matched, the matching target features, and the preset prompt word template; connect the matrices in series in the order of the features to be matched, the preset prompt word template, and the matching target features; connect the matrices in series in the order of the matching target features, the features to be matched, and the preset prompt word template; connect the matrices in series in the order of the matching target features, the preset prompt word template, and the features to be matched; connect the matrices in series in the order of the preset prompt word template, the features to be matched, and the matching target features; or, connect the matrices in series in the order of the preset prompt word template, the matching target features, and the features to be matched.

[0171] The first type corresponds to matrix series connection in the order of the features to be matched, the matching target features and the preset prompt word template; the second type corresponds to matrix series connection in the order of the features to be matched, the preset prompt word template and the matching target features; the third type corresponds to matrix series connection in the order of the matching target features, the features to be matched and the preset prompt word template; the fourth type corresponds to matrix series connection in the order of the matching target features, the preset prompt word template and the features to be matched; the fifth type corresponds to matrix series connection in the order of the preset prompt word template, the features to be matched and the matching target features; the sixth type corresponds to matrix series connection in the order of the preset prompt word template, the matching target features and the features to be matched.

[0172] Similarly, if the data to be matched is a job applicant's resume and the matching target data is a job description of a recruitment position, the input construction result corresponding to the first type is: the result of the concatenation of the characteristics of the job applicant's resume, the preset prompt word template, and the characteristic matrix of the job description; the input construction result corresponding to the second type is: the result of the concatenation of the characteristics of the job applicant's resume, the characteristics of the job description, and the preset prompt word template matrix; the input construction result corresponding to the third type is: the result of the concatenation of the characteristics of the job description, the characteristics of the job applicant's resume, and the preset prompt word template matrix; the input construction result corresponding to the fourth type is: the result of the concatenation of the characteristics of the job description, the preset prompt word template, and the characteristic matrix of the job applicant's resume; the input construction result corresponding to the fifth type is: the result of the concatenation of the preset prompt word template, the characteristics of the job applicant's resume, and the characteristic matrix of the job description; the input construction result corresponding to the sixth type is: the result of the concatenation of the preset prompt word template, the characteristics of the job description, and the characteristic matrix of the job applicant's resume.

[0173] For the convenience of subsequent description, this embodiment uses the format of "feature to be matched + preset prompt word template + matching target feature" as the above preset construction format for description, but is not limited to this.

[0174] Specifically, if the input construction result is recorded as Input, and the preset prompt word template is recorded as Prompt temple Since the preset prompt word template is actually a discrete matrix, Prompt temple ={x1,x2,…,x k}, where x k It can represent the kth trainable vector in the preset prompt word template. All trainable vectors form a matrix, that is, the above trainable variables are obtained. Then the above input construction result can be Input = J + Prompt temple+C, that is, the result of matrix concatenation in the order of the feature to be matched, the preset prompt word template, and the matching target feature.

[0175] Then the input structure result can be input into the above-mentioned preset matching model (i.e. Figure 4 After the resume text and job description text are constructed and input, the prompt word is [resume]{x1,x2,…,x k}[position]?), the preset matching model can complete the matching operation of the to-be-matched data and the matching target data according to the input construction result to obtain the matching result.

[0176] Exemplarily, if the above preset prompt word template can be "Is it suitable for the position:", after input construction, the input construction result Input=J+"Is it suitable for the position:"+C can be obtained, or it can be understood that the input construction result is: Is XX job seeker's resume suitable for position: XX position.

[0177] It should be emphasized that "Is it suitable for the position:" will be represented as a trainable variable, that is, as a number of trainable vectors to form a matrix. That is, during the training process, the trainable variables in the initial prompt word template can be continuously updated through back propagation to obtain the above-mentioned preset prompt word template.

[0178] As another example, if this embodiment uses the format of "features to be matched + matching target features + preset prompt word template" as the above-mentioned preset construction format, then the above-mentioned input construction result can be Input=J+C+"whether suitable", where "whether suitable" is used as the above-mentioned preset prompt word template, or it can be understood that the input construction result is: whether the resume of XX job seeker is suitable for XX position.

[0179] In this embodiment, each preset prompt word template can be first obtained by training different types of to-be-matched sample features and corresponding matching target sample features. The preset prompt word template can be composed of trainable variables. When used, the to-be-matched data features are first extracted to obtain the to-be-matched features, and the matching target data features are extracted to obtain the matching target features. Then, the corresponding preset prompt word template is determined according to the to-be-matched features and the matching target features. Finally, the to-be-matched features and the matching target features are matched according to the preset prompt word template to obtain the matching results. Compared with the existing matching method of artificially defining feature classification, this embodiment can first obtain different types of preset prompt word templates between the to-be-matched sample features and the matching target sample features, and then match them through the preset prompt word template that is more suitable between the to-be-matched features and the matching target features, thereby reducing feature omissions and improving the accuracy of matching.

[0180] At the same time, when using traditional matching algorithms, the features contained in resumes are diverse, and the requirements in job descriptions are also complex and changeable. In order to cope with these different features and requirements, the algorithm needs to build different network structures for matching, which not only increases the difficulty of design, but also makes the network structure extremely complex. This complexity not only increases the computational burden of the algorithm, but also increases the risk of error, and also limits the maintainability and scalability of the algorithm. However, since this embodiment directly uses a pre-trained language model as a preset matching model, the job applicant's resume and job description can be directly input at the same time, so the structure is simple, the computational burden is low, and the maintainability and scalability can also be improved.

[0181] 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.

[0182] In order to obtain the above preset prompt word template, in this embodiment, Figure 6 As shown, before the above step S10, it also includes:

[0183] Step S01: obtaining an initial prompt word template containing trainable variables;

[0184] It should be noted that the initial prompt word template may be an untrained prompt word template, which may include trainable variables before training, that is, the aforementioned several trainable vectors.

[0185] Step S02: obtaining a current iteration parameter based on the initial prompt word template, and determining a corresponding gradient parameter based on a preset loss function, the current iteration parameter, the to-be-matched sample feature, and the matching target sample feature.

[0186] It is understandable that the above current iteration parameters may be system parameters before updating in the initial matching model, and for the convenience of subsequent description, may be recorded as θ k The above gradient parameters can be the parameters corresponding to the gradient of the loss function during training, which can be recorded as where θ k+1 are the updated system parameters in the initial matching model.

[0187] Then, when updating, the specific update formula can be:

[0188]

[0189] However, since this updating method does not take into account the historical information of the gradient, another updating method is used in this embodiment. Specifically, the above-mentioned preset loss function can be the loss function in the initial matching model. Since this embodiment is used in the matching scenario, the cross entropy loss is used as the above-mentioned preset loss function in this embodiment, which can be used to measure the difference between two probability distributions, and the loss is calculated by measuring the probability distribution between the predicted result and the true result.

[0190] If the preset loss function is recorded as Loss, then the preset loss function in this embodiment can be

[0191]

[0192] where y i represents the true label distribution, that is, the actual matching result mentioned above, It represents the label distribution predicted by the model, that is, the prediction result during training, and n represents the total number of categories of matching results.

[0193] In this embodiment, since there are only "match" and "mismatch", n=2, that is, the output results have only two categories of 0 and 1, so the above preset loss function can also be written as:

[0194]

[0195] Among them, y represents the true label distribution, that is, the actual matching result mentioned above, Represents the label distribution predicted by the model, that is, the prediction result during training.

[0196] In actual use, the current iteration parameter θ can be used first k The matching prediction is performed on the features of the sample to be matched and the features of the target sample to obtain the prediction result, and then the prediction result is compared with the actual matching result, the value of the preset loss function is calculated, and then the loss function is calculated relative to θ through the back propagation algorithm. k The gradient of

[0197] Step S03: Determine a matrix estimate according to the gradient parameters, and obtain next iterative model parameters according to the matrix estimate and the current iterative model parameters.

[0198] It should be noted that the above matrix estimation may be an estimation of the gradient matrix during iteration. Specifically, it may include first-order matrix estimation and second-order matrix estimation. Therefore, in this embodiment, the above step S032 includes:

[0199] Step S031: determining a first-order matrix estimation and a second-order matrix estimation according to the gradient parameters;

[0200] Step S032: Obtain next iterative model parameters based on a preset learning rate, a preset constant, the current iterative model parameters, the first-order estimation matrix, and the second-order estimation matrix.

[0201] In getting the gradient parameters Then, based on the gradient parameter Determine the first-order matrix estimate and the second-order matrix estimate, where the first-order matrix estimate can be the weighted average of the exponential of the gradient parameter, which can be written as The second-order matrix estimate can be the exponentially weighted average of the square of the gradient parameter, which can be written as

[0202]

[0203] It should also be noted that the above-mentioned preset learning rate can be set according to actual conditions, and this embodiment does not limit this. The above-mentioned next iteration model parameter can be the updated system parameter in the initial matching model, that is, the above-mentioned θ k+1 .

[0204] Furthermore, the above step S032 includes:

[0205] Obtaining the next iteration model parameters through a preset parameter update formula based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix, and the second-order estimation matrix;

[0206] The preset parameter update formula is:

[0207]

[0208] Wherein, η is the preset learning rate, is the first-order estimation matrix, is the second-order estimation matrix, ∈ is the preset constant, θ k are the model parameters of the current iteration.

[0209] It is understandable that after obtaining the first-order estimation matrix and the second-order estimation matrix, the next iteration model parameters can be obtained by the following preset parameter update formula. The specific preset parameter update formula is:

[0210]

[0211] Where η is the preset learning rate, is relative to θ k The first-order estimation matrix of is relative to θ k The second-order estimation matrix of .

[0212] It should also be emphasized that the present embodiment adds a preset constant to the above-mentioned preset parameter update formula, namely the above-mentioned ∈. The preset constant ∈ can be any constant greater than 0 and less than 1. The purpose of the preset constant can be to prevent the denominator from being 0, so its value can be very small. It can be set according to the actual situation, and this embodiment will not elaborate on this.

[0213] Step S04: updating the trainable variables through the next iteration model parameters to obtain a preset prompt word template.

[0214] After obtaining the next iteration model parameter θ k+1 After that, the trainable variables in the initial prompt word template can be updated, that is, the initial matching model is trained to obtain the above-mentioned preset matching model, that is, the above-mentioned preset prompt word template can be obtained.

[0215] 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 steps of the data matching method described above are implemented.

[0216] In addition, refer to Figure 7 , Figure 7 This is a structural block diagram of the first embodiment of the data matching device of this application; Figure 7 As shown, the embodiment of the present application also proposes a data matching device, which includes:

[0217] The data acquisition module 701 is used to acquire the data to be matched and the matching target data;

[0218] A feature extraction module 702 is used to extract to-be-matched features according to the to-be-matched data, and to extract matching target features according to the matching target data;

[0219] The template determination module 703 is used to determine the currently used preset prompt word template by using the to-be-matched feature and the matching target feature, wherein the preset prompt word template is obtained by training based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables;

[0220] The data matching module 704 is used to match the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result.

[0221] In this embodiment, each preset prompt word template can be obtained by first training different types of to-be-matched sample features and corresponding matching target sample features. The preset prompt word template can be composed of trainable variables. When used, the to-be-matched data features are first extracted to obtain the to-be-matched features, and the matching target data features are extracted to obtain the matching target features. Then, the corresponding preset prompt word template is determined according to the to-be-matched features and the matching target features. Finally, the to-be-matched features and the matching target features are matched according to the preset prompt word template to obtain the matching results. Compared with the existing matching method of artificially defining feature classification, this embodiment can first obtain preset prompt word templates between different types of to-be-matched sample features and matching target sample features, and then match them through preset prompt word templates that are more suitable between the to-be-matched features and the matching target features, thereby reducing feature omissions and improving the accuracy of matching.

[0222] As an implementation mode, the feature extraction module 702 is also used to perform format detection on the data to be matched and the matching target data respectively; when the results of each format detection are all in the target preset format, extract the features to be matched according to the data to be matched, and extract the matching target features according to the matching target data; when there is a format that is not in the target preset format in the format detection results, perform format conversion on the object to be converted, and extract the features to be matched according to the data to be matched after the format conversion, and extract the matching target features according to the matching target data after the format conversion, the object to be converted is the data to be matched that is not in the target preset format and / or the matching target data that is not in the target preset format.

[0223] As an implementation mode, the feature extraction module 702 is also used to, when a first preset format exists in the format detection result, perform text extraction on the object to be converted according to a preset database, and perform format conversion on the object to be converted according to the extraction result; when a second preset format exists in the format detection result, perform category analysis on the object to be converted, determine a corresponding format conversion method according to the category analysis result, and perform format conversion on the object to be converted according to the corresponding format conversion method.

[0224] As an implementation mode, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a recruitment position, the target preset format is a txt format and the first preset format is a PDF format;

[0225] The feature extraction module 702 is also used to extract text from the object to be converted according to a preset database when the object to be converted is a single column, and perform format conversion on the object to be converted according to the extraction result; when the object to be converted is a double column, extract text from each single column of the object to be converted according to the preset database, splice each extraction result, and perform format conversion on the object to be converted according to the splicing result.

[0226] As an implementation mode, when the data to be matched is a resume of a job seeker and the matching target data is a job description of a recruitment position, the target preset format is a txt format and the second preset format is a word format;

[0227] The feature extraction module 702 is also used to perform category analysis on the object to be converted, and obtain a corresponding preset conversion module based on the category analysis result, wherein the preset conversion module has a corresponding format conversion method; and perform format conversion on the object to be converted according to the corresponding format conversion method through each preset conversion module.

[0228] As an implementation mode, the data matching module 704 is also used to obtain the hidden layer prediction score corresponding to the matching result when the data quantity of the matching target data is higher than a preset threshold; sort the matching target data based on the hidden layer prediction score to obtain the sorting result.

[0229] As an implementation mode, the data matching module 704 is also used to match the to-be-matched features and the matching target features according to the preset prompt word template, obtain a prediction probability, and obtain a matching result based on the prediction probability; when the data volume of the matching target data is higher than a preset threshold, determine the hidden layer prediction score corresponding to the matching result based on the prediction probability.

[0230] Based on the above-mentioned first embodiment of the data matching device of the present application, a second embodiment of the data matching device of the present application is proposed.

[0231] In this embodiment, the data matching module 704 is also used to input the to-be-matched feature, the matching target feature and the preset prompt word template according to a preset construction format to obtain an input construction result; and match the to-be-matched feature with the matching target feature based on the input construction result to obtain a matching result.

[0232] As an implementation manner, the preset construction format is a construction format in which the to-be-matched feature, the matching target feature and the preset prompt word template are spliced ​​in any order.

[0233] As an implementation mode, the input structure is established in the following manner: matrix series connection is performed in the order of the features to be matched, the matching target features and the preset prompt word template; matrix series connection is performed in the order of the features to be matched, the preset prompt word template and the matching target features; matrix series connection is performed in the order of the matching target features, the features to be matched and the preset prompt word template; matrix series connection is performed in the order of the matching target features, the preset prompt word template and the features to be matched; matrix series connection is performed in the order of the preset prompt word template, the features to be matched and the matching target features; or, matrix series connection is performed in the order of the preset prompt word template, the matching target features and the features to be matched.

[0234] As an implementation mode, when the data to be matched is a job applicant's resume and the matching target data is a job description of a recruitment position, the input construction result is: the result of the concatenation of the features of the job applicant's resume, the preset prompt word template, and the feature matrix of the job description; the input construction result is: the result of the concatenation of the features of the job applicant's resume, the features of the job description, and the preset prompt word template matrix; the input construction result is: the result of the concatenation of the features of the job description, the features of the job applicant's resume, and the preset prompt word template matrix; the input construction result is: the result of the concatenation of the features of the job description, the preset prompt word template, and the feature matrix of the job applicant's resume; the input construction result is: the result of the concatenation of the preset prompt word template, the features of the job applicant's resume, and the feature matrix of the job description; or, the input construction result is: the result of the concatenation of the preset prompt word template, the features of the job description, and the feature matrix of the job applicant's resume.

[0235] Based on the above-mentioned embodiments of the data matching device of the present application, a third embodiment of the data matching device of the present application is proposed.

[0236] In this embodiment, the data acquisition module 701 is also used to obtain an initial prompt word template containing trainable variables; obtain current iteration parameters based on the initial prompt word template, and determine corresponding gradient parameters based on a preset loss function, the current iteration parameters, the sample features to be matched, and the matching target sample features; determine matrix estimates based on the gradient parameters, and obtain next iteration model parameters based on the matrix estimates and the current iteration model parameters; update the trainable variables through the next iteration model parameters to obtain a preset prompt word template.

[0237] In one embodiment, the data acquisition module 701 is also used to determine the first-order matrix estimation and the second-order matrix estimation according to the gradient parameters; and obtain the next iterative model parameters based on the preset learning rate, the preset constant, the current iterative model parameters, the first-order estimation matrix and the second-order estimation matrix.

[0238] In one embodiment, the data acquisition module 701 is further used to obtain the next iteration model parameters based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix and the second-order estimation matrix through a preset parameter update formula; the preset parameter update formula is: Wherein, η is the preset learning rate, is the first-order estimation matrix, is the second-order estimation matrix, ∈ is the preset constant, θ k are the model parameters of the current iteration.

[0239] Other embodiments or specific implementations of the data matching device described in this application can refer to the above-mentioned method embodiments and will not be described in detail here.

[0240] 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.

[0241] 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.

[0242] 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 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.

[0243] 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; Extracting to-be-matched features according to the to-be-matched data, and extracting matching target features according to the matching target data; Determine a currently used preset prompt word template by using the to-be-matched feature and the matching target feature, wherein the preset prompt word template is obtained by training based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables; The to-be-matched feature is matched with the matching target feature according to the preset prompt word template to obtain a matching result.

2. The method according to claim 1, characterized in that The step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result includes: Inputting and constructing the to-be-matched feature, the matching target feature, and the preset prompt word template according to a preset construction format to obtain an input construction result; The to-be-matched feature is matched with the matching target feature based on the input construction result to obtain a matching result.

3. The method according to claim 2, characterized in that The preset construction format is a construction format in which the to-be-matched feature, the matching target feature and the preset prompt word template are spliced ​​in any order.

4. The method according to claim 3, characterized in that Create the input structure as follows: Matrix concatenation is performed in the order of the to-be-matched feature, the matching target feature, and the preset prompt word template; Matrix concatenation is performed in the order of the to-be-matched feature, the preset prompt word template, and the matching target feature; Matrix concatenation is performed in the order of the matching target feature, the feature to be matched, and the preset prompt word template; Matrix concatenation is performed in the order of the matching target feature, the preset prompt word template, and the feature to be matched; Matrix concatenation is performed in the order of the preset prompt word template, the to-be-matched feature, and the matching target feature; Or, matrix series connection is performed in the order of the preset prompt word template, the matching target feature and the feature to be matched.

5. The method according to claim 3, characterized in that When the data to be matched is a job seeker's resume and the matching target data is a job description of a job, The input construction result is: the result of the concatenation of the features of the job seeker's resume, the preset prompt word template, and the feature matrix of the job description; The input construction result is: the characteristics of the job seeker's resume, the characteristics of the job description, and the result of the concatenation of the preset prompt word template matrix; The input construction result is: the characteristics of the job description, the characteristics of the job seeker's resume and the result of the concatenation of the preset prompt word template matrix; The input construction result is: the result of the concatenation of the characteristics of the job description, the preset prompt word template and the characteristic matrix of the job seeker's resume; The input construction result is: the result of the concatenation of the preset prompt word template, the features of the job seeker's resume, and the feature matrix of the job description; Alternatively, the input construction result is: a result of concatenating the preset prompt word template, the features of the job description, and the feature matrix of the job seeker's resume.

6. 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: Obtain an initial prompt word template containing trainable variables; Obtaining a current iteration parameter based on the initial prompt word template, and determining a corresponding gradient parameter based on a preset loss function, the current iteration parameter, the to-be-matched sample feature, and the matching target sample feature; Determine a matrix estimate according to the gradient parameter, and obtain a next iteration model parameter according to the matrix estimate and the current iteration model parameter; The trainable variables are updated by the next iteration model parameters to obtain a preset prompt word template.

7. The method according to claim 6, characterized in that The step of determining a matrix estimate according to the gradient parameter, and obtaining a next iterative model parameter according to the matrix estimate and the current iterative model parameter comprises: Determine a first-order matrix estimate and a second-order matrix estimate according to the gradient parameters; The next iteration model parameters are obtained based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix and the second-order estimation matrix.

8. The method according to claim 7, characterized in that The step of obtaining the next iterative model parameters based on the preset learning rate, the preset constant, the current iterative model parameters, the first-order estimation matrix and the second-order estimation matrix comprises: Obtaining the next iteration model parameters through a preset parameter update formula based on a preset learning rate, a preset constant, the current iteration model parameters, the first-order estimation matrix, and the second-order estimation matrix; The preset parameter update formula is: Wherein, η is the preset learning rate, is the first-order estimation matrix, is the second-order estimation matrix, ∈ is the preset constant, θ k are the model parameters of the current iteration.

9. The method according to claim 1, characterized in that After the step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result, the method further includes: When the amount of data of the matching target data is higher than a preset threshold, obtaining a hidden layer prediction score corresponding to the matching result; The matching target data are sorted based on the hidden layer prediction scores to obtain a sorting result.

10. The method according to claim 9, characterized in that The step of matching the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result includes: Matching the to-be-matched feature and the matching target feature according to the preset prompt word template to obtain a prediction probability, and obtaining a matching result according to the prediction probability; When the amount of data of the matching target data is higher than a preset threshold, the step of obtaining the hidden layer prediction score corresponding to the matching result includes: When the data volume of the matching target data is higher than a preset threshold, a hidden layer prediction score corresponding to the matching result is determined according to the prediction probability.

11. The method according to claim 1, characterized in that The step of extracting the to-be-matched features according to the to-be-matched data and extracting the matching target features according to the matching target data comprises: Performing format detection on the to-be-matched data and the matching target data respectively; When the format detection results are all the target preset formats, extracting the to-be-matched features according to the to-be-matched data, and extracting the matching target features according to the matching target data; When there is a format that is not in the preset format of the target in the format detection result, format conversion is performed on the object to be converted, and features to be matched are extracted based on the data to be matched after the format conversion, and matching target features are extracted based on the matching target data after the format conversion, and the object to be converted is the data to be matched that is not in the preset format of the target and / or the matching target data that is not in the preset format of the target.

12. The method according to claim 11, characterized in that When there is a format that is not the target preset format in the format detection result, the step of performing format conversion on the object to be converted comprises: When the first preset format exists in the format detection result, text extraction is performed on the object to be converted according to a preset database, and format conversion is performed on the object to be converted according to the extraction result; When the second preset format exists in the format detection result, the object to be converted is subjected to category analysis, a corresponding format conversion method is determined according to the category analysis result, and the format conversion is performed on the object to be converted according to the corresponding format conversion method.

13. The method according to claim 12, characterized in that When the data to be matched is a job seeker's resume and the target data to be matched is a job description of a job, the target preset format is a txt format and the first preset format is a PDF format; The step of extracting text from the object to be converted according to the preset database and converting the format of the object to be converted according to the extraction result includes: When the object to be converted is a single column, extracting text from the object to be converted according to a preset database, and converting the format of the object to be converted according to the extraction result; When the object to be converted is a double column, text is extracted from each single column of the object to be converted according to a preset database, the extraction results are spliced, and the format of the object to be converted is converted according to the splicing result.

14. The method according to claim 12, characterized in that When the data to be matched is a job seeker's resume and the target data to be matched is a job description of a job, the target preset format is a txt format and the second preset format is a word format; The step of performing category analysis on the object to be converted, determining a corresponding format conversion method according to the category analysis result, and performing format conversion on the object to be converted according to the corresponding format conversion method includes: Performing category analysis on the object to be converted, and acquiring a corresponding preset conversion module according to the category analysis result, wherein the preset conversion module has a corresponding format conversion method; The format of the object to be converted is converted by each of the preset conversion modules according to the corresponding format conversion method.

15. 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 feature extraction module, used to extract to-be-matched features according to the to-be-matched data, and to extract matching target features according to the matching target data; A template determination module, used to determine a currently used preset prompt word template using the to-be-matched feature and the matching target feature, wherein the preset prompt word template is obtained by training based on different types of to-be-matched sample features and corresponding matching target sample features, and the preset prompt word template is composed of trainable variables; The data matching module is used to match the to-be-matched feature with the matching target feature according to the preset prompt word template to obtain a matching result.

16. 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 14 when executed by the processor.

Images