Data processing method and device

Abstract

The embodiment of the invention discloses a data processing method and device. An initial query statement and a to-be-evaluated object are obtained, a first feature vector of the to-be-evaluated object is obtained based on an original model, target associated information of the to-be-evaluated object is obtained according to the initial query statement, and the target associated information comprises at least one of an enhanced query statement and summary information; and obtaining an evaluation result representing the correlation between the initial query statement and the to-be-evaluated object through an online model according to the initial query statement, the first feature vector, the target association information, the object title and the object content. Therefore, the understanding ability of the query system can be improved, the user demand can be accurately obtained, and the accuracy of the query result is further improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a data processing method and device. Background Art

[0002] With the development of information technology, efficient access to information has become crucial, especially in the areas of technical documents, contract texts, and knowledge content queries.

[0003] Most existing query systems rely on traditional keyword matching methods or simple retrieval algorithms, which return results by identifying the keywords entered by users and comparing them with stored data. Some systems have introduced natural language processing (NLP) technology to enhance the understanding of query intent, but they still mainly remain at the surface level of keyword matching.

[0004] However, users' queries may contain ambiguous keywords, professional vocabulary, or even spelling errors, making it difficult for existing query systems to accurately capture requirements, resulting in poor relevance of query results. In addition, when dealing with multi-level information structures, the system's understanding and extraction capabilities are limited, affecting retrieval efficiency. Summary of the invention

[0005] In view of this, an object of an embodiment of the present invention is to provide a data processing method and device, which can improve the understanding ability of the query system, accurately obtain user needs, and thus improve the accuracy of the query results.

[0006] In a first aspect, an embodiment of the present invention provides a data processing method, the method comprising:

[0007] Get the initial query statement and the object to be evaluated;

[0008] Acquire a first feature vector of the object to be evaluated based on the original model;

[0009] Acquire target association information of the object to be evaluated according to the initial query statement, wherein the target association information includes at least one of an enhanced query statement and summary information;

[0010] An evaluation result is obtained through an online model according to the initial query statement, the first feature vector, the target association information, the object title and the object content, and the evaluation result is used to characterize the relevance between the initial query statement and the object to be evaluated.

[0011] In some embodiments, the method further comprises:

[0012] Acquire historical data, where the historical data includes historical query statements and historical query results, where the historical query results include at least one candidate object;

[0013] Acquire an enhanced query statement according to the historical query statement and the candidate object by a query enhancement module;

[0014] Generate historical summary information corresponding to each candidate query statement according to the candidate object by a summary generation module, wherein the candidate query statement includes a historical query statement and an enhanced query statement;

[0015] Obtaining a behavior feedback result according to the historical summary information through a relevance evaluation module, wherein the behavior feedback result includes the candidate query statement, the historical summary information, and a label, wherein the label is used to characterize the relevance between the candidate query statement and the corresponding historical summary information;

[0016] The online model is adjusted according to the behavior feedback result.

[0017] In some embodiments, the method further comprises:

[0018] Generate and store the associated information corresponding to the candidate object, wherein the associated information includes the query statement and the corresponding summary information.

[0019] In some embodiments, obtaining the target association information of the object to be evaluated according to the initial query statement is specifically:

[0020] The target association information is acquired from the stored data according to the initial query statement and the object to be evaluated.

[0021] In some embodiments, the online model includes a feature encoding layer, a micro feature extraction layer, a macro feature extraction layer, a first fusion layer, a feedforward neural network layer, a second fusion layer, and an output layer.

[0022] In some embodiments, obtaining the evaluation result according to the initial query statement, the first feature vector, the target association information, the object title and the object content through the online model includes:

[0023] Acquire an input vector through the feature encoding layer, wherein the input vector includes a query vector corresponding to the initial query statement and the enhanced query statement, a title vector corresponding to the object title, a content vector corresponding to the object content, and a summary vector corresponding to the summary information;

[0024] Acquire a micro feature vector according to the query vector and the title vector through the micro feature extraction layer;

[0025] Acquire a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer;

[0026] fusing the microscopic feature vector and the macroscopic feature vector through the first fusion layer to obtain a first fusion vector;

[0027] Obtaining a second feature vector according to the first fusion vector through the feedforward neural network layer;

[0028] fusing the first feature vector and the second feature vector through the second fusion layer to obtain a second fused vector;

[0029] The evaluation result is generated according to the second fusion vector through the output layer.

[0030] In some embodiments, the micro-feature extraction layer includes a conversion sublayer, a kernel pooling sublayer, and a feature extraction sublayer;

[0031] Wherein, obtaining the micro feature vector according to the query vector and the title vector through the micro feature extraction layer includes:

[0032] Obtaining a similarity matrix according to the query vector and the title vector through the conversion sublayer;

[0033] Obtaining a kernel matrix according to the similarity matrix through the kernel pool sublayer;

[0034] The microscopic feature vector is obtained according to the kernel matrix through the feature extraction sublayer.

[0035] In some embodiments, the macro feature extraction layer includes a fusion sublayer, a paragraph-level multi-head attention sublayer, and a first flat sublayer;

[0036] The step of obtaining a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer includes:

[0037] fusing the query vector and the content vector through the fusion sublayer to obtain a third fusion vector;

[0038] Obtain a multidimensional tensor according to the third fusion vector and the summary vector through the paragraph-level multi-head attention sublayer;

[0039] The multidimensional tensor is converted into the macro feature vector by the first flattening sub-layer.

[0040] In some embodiments, the query enhancement module, summary generation module and relevance assessment module are large language models.

[0041] In some embodiments, the method further comprises:

[0042] Generate query results based on the evaluation results and push them to the user.

[0043] In a second aspect, an embodiment of the present invention provides a data processing device, the device comprising:

[0044] A data acquisition unit, used to acquire an initial query statement and an object to be evaluated;

[0045] A first feature vector acquisition unit, configured to acquire a first feature vector of the object to be evaluated based on an original model, wherein the first feature vector is used to characterize the relevance between the object to be evaluated and the initial query statement;

[0046] A target association information acquisition unit, configured to acquire target association information of the object to be evaluated according to the initial query statement, wherein the target association information includes at least one of an enhanced query statement and summary information;

[0047] An evaluation result acquisition unit is used to acquire an evaluation result based on the initial query statement, the first feature vector, the target association information, the object title and the object content through an online model, wherein the evaluation result is used to characterize the correlation between the initial query statement and the object to be evaluated.

[0048] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0049] In a fourth aspect, an embodiment of the present invention provides a computer program product, wherein the computer program product includes a computer program. When the computer program runs on a computer, the computer executes the method described in the first aspect.

[0050] In a fifth aspect, an embodiment of the present invention provides a computer-readable storage medium on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described in the first aspect is implemented.

[0051] The technical solution of the embodiment of the present invention obtains the initial query statement and the object to be evaluated, obtains the first feature vector of the object to be evaluated based on the original model, obtains the target association information of the object to be evaluated according to the initial query statement, and the target association information includes at least one of the enhanced query statement and the summary information, and obtains the evaluation result representing the correlation between the initial query statement and the object to be evaluated according to the initial query statement, the first feature vector, the target association information, the object title and the object content through the online model. In this way, the understanding ability of the query system can be improved, the user needs can be accurately obtained, and the accuracy of the query results can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0053] Figure 1 is a schematic diagram of a data processing system according to an embodiment of the present invention;

[0054] Figure 2 is a schematic diagram of a query system according to an embodiment of the present invention;

[0055] Figure 3 is a schematic diagram of an offline model of an embodiment of the present invention;

[0056] Figure 4 is a flow chart of offline data processing according to an embodiment of the present invention;

[0057] Figure 5 is a schematic diagram of a query statement and a query result according to an embodiment of the present invention;

[0058] Figure 6 is a flow chart of a data processing method according to an embodiment of the present invention;

[0059] Figure 7 is a schematic diagram of an online model of an embodiment of the present invention;

[0060] Figure 8 is a flow chart of obtaining evaluation results according to an embodiment of the present invention;

[0061] Fig. 9 is a flow chart of obtaining a microscopic feature vector according to an embodiment of the present invention;

[0062] Fig.10 is a flow chart of obtaining a macro feature vector according to an embodiment of the present invention;

[0063] Fig.11 is a schematic diagram of a data processing device according to an embodiment of the present invention;

[0064] Fig.12 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The present application is described below based on embodiments, but the present application is not limited to these embodiments. In the detailed description of the present application below, some specific details are described in detail. It is possible for those skilled in the art to fully understand the present application without the description of these details. In order to avoid confusing the essence of the present application, known methods, processes, flows, components and circuits are not described in detail.

[0066] In addition, persons of ordinary skill in the art will appreciate that the drawings provided herein are for illustration purposes and are not necessarily drawn to scale.

[0067] Unless the context clearly requires otherwise, the words "include", "comprising" and similar words throughout the application should be interpreted as including rather than exclusive or exhaustive; that is, the meaning is "including but not limited to".

[0068] In the description of this application, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of this application, unless otherwise specified, the meaning of "plurality" is two or more.

[0069] The solutions described in this specification and in the examples, if they involve the processing of personal information, will be processed on the premise of having a legal basis (such as obtaining the consent of the subject of personal information, or being necessary for the performance of a contract, etc.), and will only be processed within the scope of regulations or agreements. If a user refuses to process personal information other than the necessary information for basic functions, it will not affect the user's use of basic functions.

[0070] Figure 1 Schematic diagram of a data processing system according to an embodiment of the present invention. Figure 1 As shown, the data processing system of the embodiment of the present invention includes a server 1, a network 2 and a client cluster. The client cluster includes clients 3a, ..., 3n. Any client in the client cluster can be connected to the server 1 through the network 2. For example, there is a communication connection between the client 3a and the server 1, and there is a communication connection between the client 3n and the server 1.

[0071] In this embodiment, the client is used to provide an interface for users to interact with the data processing system, which can be implemented through various terminal devices, such as desktop computers, laptops, tablet computers or mobile phones or other data processing terminals. The client is responsible for receiving the query statement input by the user and sending the query statement to the server. In addition, the client is also responsible for receiving and displaying the query results sent by the server 1.

[0072] Server 1 stores objects, and after receiving a query statement sent by a client, obtains query results according to the query statement and pushes them to the client. Server 1 can be a local server or a cloud server, and can be implemented by an independent server or a server cluster composed of multiple servers.

[0073] Furthermore, the server 1 includes a database 11 and a query system 12 .

[0074] The database 11 stores data files available for query, and each data file is referred to as an object in the embodiment of the present invention. The object may be an article, such as a technical document, a contract text, and knowledge content. Each object includes at least an object title and an object content. The object title is the title of the article, and the object content is the content of the article. In some embodiments, the database 11 may include a large-capacity memory, a removable memory, a volatile read-write memory, or a read-only memory (ROM), or any combination thereof. As an example, mass storage may include magnetic disks, optical disks, solid-state drives, etc.; removable storage may include flash drives, floppy disks, optical disks, memory cards, zip disks, magnetic tapes, etc.; volatile read-write storage may include random access memory (Random Access Memory, RAM); RAM may include dynamic RAM (Dynamic RandomAccess Memory, DRAM), double data rate synchronous dynamic RAM (DoubleDate-Rate Synchronous RAM, DDR SDRAM); static RAM (Static Random-Access Memory, SRAM), thyristor RAM (Thyristor-Based Random Access Memory, T-RAM) and zero capacitor RAM (Zero-RAM), etc. As an example, ROM may include mask ROM (Mask Read-Only Memory, MROM), programmable ROM (Programmable Read-Only Memory, PROM), erasable programmable ROM (Programmable Erasable Read-only Memory, PEROM), electrically erasable programmable ROM (Electrically Erasable Programmable read only memory, EEPROM), CD-ROM (CD-ROM), and digital versatile disk ROM, etc. In some embodiments, the database 11 can be implemented on a cloud platform. As an example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, a cross-cloud, a multi-cloud, or other similar, etc., or any combination thereof.

[0075] The query system 12 is used to parse the query statement received from the client, execute the corresponding search algorithm to find the matching object in the database 11, and generate the query result and send it to the client.

[0076] Network 2 can be used for the exchange of information and / or data. Wherein, network 2 can be any type of wired or wireless network, or a combination thereof. In some embodiments, network 2 may include a wired network, a wireless network, a fiber optic network, a telecommunication network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a wide area network (WAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, or a near field communication (NFC) network, etc., or any combination thereof. In some embodiments, network 2 may include one or more network access points. For example, network 2 may include a wired or wireless network access point, such as a base station and / or a network switching node, through which one or more components of a data processing system may be connected to the network to exchange data and / or information.

[0077] Figure 2 Schematic diagram of a query system according to an embodiment of the present invention. Figure 2 As shown, the query system 12 of the embodiment of the present invention includes an original model 121 , an online model 122 and an offline model 123 .

[0078] The offline model 123 is used to execute the offline process, which includes generating behavior feedback results and generating related information. Generating behavior feedback results means generating behavior feedback results based on historical data and adjusting the online model 122. Generating related information means generating related information based on historical data and storing the related information to provide data support for subsequent online processes.

[0079] Figure 3 Schematic diagram of an offline model of an embodiment of the present invention. Figure 3 As shown, the offline model of the embodiment of the present invention includes a query enhancement module L11, a summary generation module L12, a relevance evaluation module L13 and a fine-tuning module L14.

[0080] The query enhancement module L11, the summary generation module L12 and the relevance evaluation module L13 may be implemented by a large language model (LLM).

[0081] Figure 4: is a flowchart of offline data processing according to an embodiment of the present invention. Figure 4 As shown, offline data processing includes the following steps:

[0082] Step S110: Obtain historical data.

[0083] In this embodiment, historical data of querying a database is obtained, the historical data includes historical query statements and historical query results, and the historical query results include at least one candidate object.

[0084] Specifically, historical query statements refer to query statements submitted by users to the server in the past historical period. The historical query statements can be statements in various languages, which can be complete statement descriptions, such as "the basic principle of Bluetooth communication"; or they can be keyword sets, such as "Bluetooth, communication, principle".

[0085] The historical query result is the query result corresponding to the historical query statement. The historical query result includes at least one candidate object. In some embodiments, the historical query result includes the triggered object corresponding to the historical query statement. For example, when a user's historical query statement is "Basic Principles of Bluetooth Communication", 10 related articles are pushed to the user through matching algorithms, and the user clicks to view two of them, then the two clicked articles are candidate objects, that is, there are two candidate objects.

[0086] Perform statistics on historical data to obtain historical query statements and candidate objects in each historical query. Analyze these data to obtain the query statement corresponding to each candidate object. For example, suppose there are three objects A in the database. b1 , A b2 and A b3 The five historical queries are q b1 -q b5 Among them, the query statement and query results are as follows Figure 5 As shown, q b1 The corresponding query result is A b1 ;q b2 The corresponding query result is A b1 , A b2 ;q b3 The corresponding query result is A b2 ;q b4 The corresponding query result is A b2 , A b3 ;q b5 The corresponding query result is A b1 , A b3 .

[0087] By analyzing the above results, we can know that:

[0088] Object A b1 The corresponding query statements include q b1 ,q b2 ,q b5 ;

[0089] Object A b2 The corresponding query statements include q b2 ,q b3 ,q b4 ;

[0090] Object A b3 The corresponding query statements include q b4 ,q b5 .

[0091] Step S120: Obtain an enhanced query statement according to the historical query statement and the candidate object through a query enhancement module.

[0092] In this embodiment, the historical query statement q 0 and a corresponding candidate object A H The query enhancement module L11 is used to input the historical query statement q 0 and the candidate object A H Get the enhanced query statement. Among them, candidate object A H Including object title title and object content content. Among them, the query enhancement module L11 can be implemented by a large language model.

[0093] For example, the information input to the query enhancement module L11 may be as follows:

[0094] “I want to use a search engine to find an article. Below is the title and content of the article.

[0095] Title: "xxx";

[0096] Content: {xxxxxx};

[0097] Please list the most likely popular search terms to find this article.

[0098] Please refer to these popular search terms: {top3 queries};

[0099] Please output x different popular search terms in order."

[0100] The output of the query enhancement module L11 can be shown as follows:

[0101] "xxx, xxx, xxx, xxx, xxx, xxx,...".

[0102] That is to say, firstly, all historical query records are analyzed to obtain the query statement corresponding to each candidate object, and then for each candidate object, the object title and object content of the candidate object, as well as the historical popular query statements of the candidate object are input into the query enhancement module L11, and the query enhancement module L11 enhances the query statement to obtain an enhanced query statement.

[0103] Corresponding to Figure 3 The candidate object input to the query enhancement module L11 is A H The object title is title, the object content is content, and there is one historical query statement input into the query enhancement module L11, specifically q 0 .

[0104] The output of the query enhancement module L11 is a first set Q of enhanced query statements, which includes multiple enhanced query statements q 1 -q k .

[0105] Step S130: Generate historical summary information corresponding to each candidate query statement according to the candidate object through a summary generation module.

[0106] In this embodiment, the historical query statements and the enhanced query statements are used as candidate query statements, and the summary generation module L12 generates historical summary information corresponding to each candidate query statement according to the candidate objects.

[0107] Among them, the summary information is the result after the summary generation module summarizes the object content of the candidate object according to the query statement. Each candidate query statement corresponds to a historical summary information. For the same candidate object, the historical summary information corresponding to different candidate query statements will also be different. For example, assuming that the candidate object is an article on Bluetooth technology, it records the processes of Bluetooth broadcasting, scanning, pairing, binding, etc. For the query statement "the basic principles of Bluetooth", the summary information obtained by the summary generation module L12 is the content related to the basic principles of Bluetooth; for the "Bluetooth pairing process", the summary information obtained by the summary generation module L12 is the content related to Bluetooth pairing.

[0108] As an example, the content input to the summary generation module L12 may be:

[0109] “I want to use a search engine to find an article. Below is the title and content of the article.

[0110] Title: "xx";

[0111] Content: {xxxxxx} The content of the article ends here. The following are the popular search terms to find this article:

[0112] {Basic principles of Bluetooth}, {Bluetooth pairing process}, {xxx}, {xxx}, …;

[0113] Please tell me how many topics this article can be divided into and what each topic discusses. "

[0114] The output of the summary generation module L12 can be:

[0115] “Basic principles of Bluetooth: xxx.

[0116] Bluetooth pairing process: xxx.

[0117] ..."

[0118] Thus, the historical summary information corresponding to each candidate query statement can be obtained.

[0119] Corresponding to Figure 3 In the example, the output of the summary generation module L12 is a second set S of historical summary information, which includes the historical summary information S corresponding to each candidate query statement. 0 -S k .

[0120] Step S140: obtaining a behavior feedback result according to the historical summary information through a relevance evaluation module.

[0121] In this embodiment, the behavior feedback result includes the candidate query statement q i , historical summary information S i and label G i , the tag is used to characterize the relevance between the candidate query statement and the corresponding historical summary information.

[0122] That is to say, the i-th behavior feedback result can be recorded as (q i , S i , G i ). Wherein, Gi is relevant or irrelevant. In some embodiments, G i =Y or N, where Y indicates related and N indicates not related. In some embodiments, G i =0 or 1, 1 means relevant, 0 means irrelevant.

[0123] The relevance evaluation module L13 compares each historical summary information with the object content to obtain a relevance result.

[0124] As an example, the content input to the relevance evaluation module 13 may be:

[0125] "Please help me to evaluate the relevance of the following query to the text. My query is:

[0126] {Basic principles of Bluetooth};

[0127] {Bluetooth pairing process};

[0128] …

[0129] My text is: "xxxxx"

[0130] The correlation results are divided into the following categories: relevant and not relevant.

[0131] Please output in the following format:

[0132] Reason: [one sentence reason];

[0133] Result: [Not relevant / Relevant]"

[0134] The output of the relevance evaluation module 13 is:

[0135] "Bluetooth Basics: An Introduction to…

[0136] Reason: The text introduces xxx and its technical features in detail, which is exactly what the query points to.

[0137] Results: Related

[0138] Bluetooth pairing process: discussed...

[0139] Reason: Although the text mentions the technology used by xxx, the introduction to the specific pairing process is relatively limited.

[0140] Result: Not relevant".

[0141] Furthermore, the behavior feedback result can be expressed as: R i =(q i , S i , G i ).

[0142] Corresponding to Figure 3 In the example, the relevance evaluation module 13 outputs a third set R of behavior feedback results, R including R 0 -R k .

[0143] Step S150: adjusting the online model according to the behavior feedback result.

[0144] In this embodiment, the online model is adjusted according to the behavior feedback result by the fine-tuning module L14. i , S i , G i ), if G i Indicates irrelevant, then as a negative sample, if G iIf the query is related to the object, it is used as a positive sample, and the online model is fine-tuned through the positive and negative samples. By analyzing the above behavior feedback results, the fine-tuning module L14 can more accurately adjust the parameters of the online model to better capture the semantic relationship between the query and the object, thereby improving the relevance and accuracy of the query results.

[0145] It should be noted that, further, in step S150, adjusting the online model according to the behavior feedback result actually means: adjusting the feature coding layer of the online model. That is, after obtaining the behavior feedback result, the prediction result corresponding to each behavior feedback result is obtained through the online model, and the feature coding layer is adjusted to make the prediction result as consistent as possible with the label. In this way, the adjustment of the online model can be completed.

[0146] Specifically, the fine-tuning module L14 is a BGE (Bidirectional Generative Encoder) fine-tuning module, which is a component used to improve the performance of models in natural language processing (NLP) tasks. Specifically, the BGE fine-tuning module uses user behavior feedback results to optimize the online model's understanding and processing capabilities for queries and article summaries.

[0147] Therefore, the online model can be fine-tuned through historical data.

[0148] In some embodiments, the offline model is further used to perform the following steps:

[0149] Step S160: Generate and store the associated information corresponding to the candidate object.

[0150] In this embodiment, for each candidate object, the server obtains the candidate query statement and historical summary information with the label "related", generates the association information corresponding to the candidate object according to the candidate query statement and historical summary information with the label "related", and stores the association information. The association information includes the query statement and the corresponding summary information. The stored association information provides data support for the online process.

[0151] Thus, the offline process can be completed, wherein the offline process can be executed before the query system is deployed to the server, or in an idle period after the query system is deployed to the server.

[0152] Furthermore, after adjusting the online model through the offline process and obtaining the associated information, the user's normal query can be realized, that is, the online process can be executed, wherein the online process is realized by the original model 121 and the online model 122 .

[0153] After obtaining the initial query statement input by the user, the purpose of the online process is to obtain objects related to the initial query statement and push them to the user. Specifically, for each object to be evaluated in the database, the first feature vector of the object to be evaluated is obtained through the original model 121, and then the evaluation result is obtained through the online model 122 according to the initial query statement, the first feature vector, the target association information, the object title and the object content. The evaluation result is used to characterize the relevance between the initial query statement and the object to be evaluated.

[0154] Among them, the original model 121 is the original relevance evaluation model of the query system 12, which is used to evaluate the relevance between the object and the query, which can be implemented through various existing methods, for example, WDL (Wide & Deep Learning), EDCN (Enhanced Deep & Cross Network), DCNV2 (Deep & Cross Network Version 2), AutoInt (Automatic Feature Interaction), XDeepFM (Extreme Deep Factorization Machine), FiBiNETv2 (Feature Importance and Bilinear feature Interaction NETwork version2), etc.

[0155] Figure 6 is a flow chart of a data processing method according to an embodiment of the present invention. Figure 6 As shown, the data processing method of the embodiment of the present invention includes the following steps:

[0156] Step S210: Obtain an initial query statement and an object to be evaluated.

[0157] In this embodiment, the initial query statement is the query statement input by the user for this query, which can be in various languages ​​such as English, Chinese, etc. The initial query statement can be a complete sentence, such as "the basic principle of Bluetooth communication"; or one or more keywords, such as "Bluetooth, communication, principle", etc.

[0158] The object to be evaluated is an object that needs to be evaluated for relevance with the initial query statement, and the object is an article in the database.

[0159] In some embodiments, the object to be evaluated is any object in the database.

[0160] In some embodiments, the object to be evaluated is any one of the partial objects obtained through preliminary screening in the database. The preliminary screening method can be implemented based on various existing methods, such as keyword matching.

[0161] Step S220: Acquire a first feature vector of the object to be evaluated based on the original model.

[0162] In this embodiment, the first feature vector V1 is obtained according to the object to be evaluated and the initial query statement through the original model. The first feature vector V1 is an expression of the correlation between the object to be evaluated and the initial query statement obtained based on the original model.

[0163] Step S230: Acquire target association information of the object to be evaluated according to the initial query statement.

[0164] In this embodiment, as described above, the association information corresponding to each candidate object is generated and stored through step S160, and the association information includes a query statement and corresponding summary information. Thus, target association information is obtained from the stored data according to the initial query statement and the object to be evaluated, and the target association information includes at least one of an enhanced query statement and summary information.

[0165] Specifically, the associated information corresponding to the object to be evaluated is obtained from the stored data, and the query statements and summary information in the associated information are extracted, other query statements that are not repeated with the initial query statements are used as the enhanced query statements, and the summary information corresponding to the initial query statements and the enhanced query statements is determined.

[0166] It should be noted that, when there is not enough historical data, some objects have less associated information, and there may be no initial query statement in the acquired target associated information. In this case, all query statements in the target associated information are used as enhanced query statements, and the summary information corresponding to the enhanced query statements is determined to be required for this step. Alternatively, there may be only initial query statements in the acquired target associated information, but no enhanced query statements. In this case, the enhanced query statement is empty, and the summary information corresponding to the initial query statement is determined to be required for this step. Alternatively, there may be no query statement in the acquired target associated information. In this case, the enhanced query statement can be obtained according to the initial query statement and the object to be evaluated through the query enhancement module, and the summary information corresponding to the initial query statement and the enhanced query statement can be generated through the summary generation module, and the summary information corresponding to the initial query statement and the enhanced query statement can be determined to be required for this step.

[0167] Step S240: obtaining an evaluation result through an online model according to the initial query statement, the first feature vector, the target association information, the object title and the object content.

[0168] In this embodiment, the evaluation result is used to characterize the relevance between the initial query statement and the object to be evaluated.

[0169] Figure 7 Schematic diagram of an online model of an embodiment of the present invention. Figure 7 In the illustrated embodiment, the online model includes a feature encoding layer L21, a micro feature extraction layer L22, a macro feature extraction layer L23, a first fusion layer L24, a feedforward neural network layer L25, a second fusion layer L26, and an output layer L27.

[0170] in, Figure 8 FIG. 1 is a flowchart of obtaining evaluation results according to an embodiment of the present invention. Figure 8 As shown, obtaining the evaluation result through the online model according to the initial query statement, the first feature vector, the target association information, the object title and the object content includes the following steps:

[0171] Step S241, obtaining an input vector through the feature encoding layer.

[0172] In this embodiment, the input vector includes query vectors corresponding to the initial query statement and the enhanced query statement, title vectors corresponding to the object title, content vectors corresponding to the object content, and summary vectors corresponding to the summary information. The feature encoding layer L21 is used to convert text into feature vectors.

[0173] In some embodiments, the feature encoding layer L21 can be implemented by a passage topic-aware encoding layer (PassageTopic-aware Encoding Layer). Passage Topic-aware Encoding Layer is a coding layer used to enhance text understanding in natural language processing (NLP, Neuro-Linguistic Programming). It aims to capture the topic information of a document or passage and incorporate this information into the model's representation of the text, thereby improving the performance of downstream tasks. Passage Topic-aware Encoding Layer introduces topic models (such as latent Dirichlet allocation) or other methods to identify and utilize the topic structure of the text, so that the encoded text representation is richer and can better reflect the true semantics of the text.

[0174] First, input information is obtained, and the input information includes a query statement, an object title and object content of the object to be evaluated, and summary information. The query statement includes an initial query statement and an enhanced query statement, and the enhanced query statement and summary information are determined by the target association information obtained in the above step S230. Then, the input information is converted into an input vector through the feature encoding layer L21.

[0175] Specifically, the object title of the object to be evaluated is converted into a title vector W through the feature encoding layer L21 T , convert the query statement into a query vector W Q , convert the object content of the object to be evaluated into a content vector W C , convert the summary information into a summary vector W S .

[0176] For the object title of the object to be evaluated, convert it into a title vector W T Considering that the object title is relatively short, the embodiment of the present invention retains the embedding vector of each word corresponding to the object title to obtain the title vector W T . Title vector W T =(w T1 , w T2 ,……,w Tn ), where n is the number of words in the object title, and w Ti Represents the embedding vector of the i-th word, i = 1, 2, ..., n.

[0177] For converting the query sentence into a query vector W Q Considering that the query statement is relatively short, the embodiment of the present invention retains the embedding vector of each word corresponding to the query statement to obtain the query vector W Q The query vector W Q =(w Q1 , w Q2 ,……,w Qm ), where m is the number of query statements, w Qi Represents the embedding vector of the i-th query statement, i = 1, 2, ..., m.

[0178] For the object content of the object to be evaluated, convert it into a content vector W C Considering that the object content is relatively long, the embodiment of the present invention aggregates the embedding vector of each word in the object content to obtain W C .

[0179] For converting the summary information into a summary vector W S Considering that the summary information is relatively long, the embodiment of the present invention aggregates the embedding vector of each word in the summary information to obtain W C Among them, WC =(w C1 , w C2 ,……,w Cx ). Among them, w Ci represents the embedding vector of the i-th summary information, i = 1, 2, ..., x. x is the number of summary information.

[0180] Step S242: Obtain a micro feature vector according to the query vector and the title vector through the micro feature extraction layer.

[0181] In this embodiment, the microscopic feature extraction layer L22 extracts the query vector W Q and the title vector Q T Get the microscopic eigenvector W M1 .

[0182] The micro feature extraction layer L22 can be implemented by Micro Wordwise QT SemanticExtractor, which specifically includes a conversion sublayer L221, a kernel pool sublayer L222 and a feature extraction sublayer L223.

[0183] in, Fig. 9 FIG. 1 is a flow chart of obtaining microscopic feature vectors according to an embodiment of the present invention. Fig.11 As shown, obtaining a micro feature vector according to the query vector and the title vector through the micro feature extraction layer includes the following steps:

[0184] Step S2421: Obtain a similarity matrix according to the query vector and the title vector through the conversion sublayer.

[0185] In this embodiment, the conversion sublayer L221 is a Translation Layer, and the query vector W is input to this layer. Q and the title vector W T . In the query vector W Q and the title vector W T After normalization, the conversion sublayer L221 calculates the similarity between each word in the query vector and each word in the title vector, wherein the similarity can be calculated by methods such as dot product and cosine similarity. The similarity results of each word are used to generate a similarity matrix, and each element in the similarity matrix represents the semantic similarity of all word pairs between the query statement and the object title.

[0186] Step S2422: Obtain a kernel matrix according to the similarity matrix through the kernel pool sublayer.

[0187] In this embodiment, the kernel pool sublayer L222 is a Kernel Pooling Layer. The similarity matrix is ​​passed to the kernel pool sublayer L222. The kernel pool sublayer L222 processes the similarity matrix using a Gaussian kernel function to generate multiple kernel matrices. Each kernel matrix corresponds to a different "kernel center", which can be understood as a cluster at different semantic distances. This process can help capture more complex semantic patterns and can model similarities of different scales.

[0188] Step S2423: Obtain the microscopic feature vector according to the kernel matrix through the feature extraction sublayer.

[0189] In this embodiment, the feature extraction sublayer L223 is a soft TF feature extraction layer (Soft-TF Feature Extraction Layer). The feature extraction sublayer L223 is intended to simulate traditional term frequency (TF) statistics, combining the results of each kernel matrix to generate a micro feature vector W M1 .

[0190] Therefore, by combining word-level similarity calculation and high-level semantic pattern recognition, such a model can effectively capture the subtle differences between queries and documents.

[0191] Step S243: Obtain a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer.

[0192] In this embodiment, the macro feature extraction layer L23 extracts the query vector W Q , content vector W C and the summary vector W S Get the macroscopic feature vector W M2 .

[0193] Among them, the macro feature extraction layer L23 can be implemented by the Macro QS Semantic Extractor, which is a neural network module used to process the semantic relationship between query statements and paragraphs, specifically including the fusion sublayer L231, the paragraph-level multi-head attention sublayer L232 and the first flat sublayer L233.

[0194] in, Fig.10 FIG. 1 is a flow chart of obtaining a macro feature vector according to an embodiment of the present invention. Fig.10 As shown, the step of obtaining a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer includes the following steps:

[0195] Step S2431: Fuse the query vector and the content vector through the fusion sublayer to obtain a third fusion vector.

[0196] In this embodiment, the query vector W Q , content vector W C and the summary vector W S Input to the macro feature extraction layer L23. Among them, the query vector W Q , content vector W C Input to the fusion sublayer L231. The fusion sublayer L231 processes the query vector W Q and the content vector W C The fusion is performed to obtain a third fusion vector. The fusion method can be implemented in various ways, such as averaging, splicing, etc.

[0197] Step S2432: Obtain a multidimensional tensor according to the third fusion vector and the summary vector through the paragraph-level multi-head attention sublayer.

[0198] In this embodiment, the paragraph-level multi-head attention sublayer L232 receives the output of the fusion sublayer L231 and the summary vector W S , and obtain a multidimensional tensor according to the third fused vector and the summary vector.

[0199] Among them, the paragraph-level multi-head attention sublayer L232 can be implemented by the Passage-Level MHA layer (multi-head attention mechanism layer). The Passage-Level MHA layer receives the third fusion vector and the summary vector as input, and uses the multi-head attention mechanism (MHA, Multi-Head Attention) to capture the complex semantic relationship between them. Among them, multi-head attention allows the model to focus on different parts of the input at different positions, so as to better understand how the query and the words in the paragraph are related to each other. In this way, the paragraph-level multi-head attention sublayer L232 can learn richer representations that contain not only word-level information, but also sentence structure and context information. In this embodiment, the paragraph-level multi-head attention sublayer L232 outputs a multidimensional tensor, for example, the dimension of the multidimensional tensor can be [batch_size, sequence_length, feature_dimension], where batch_size is the batch size, sequence_length is the sequence length (such as the number of words in a sentence), and feature_dimension is the dimension of each position feature.

[0200] Step S2433: Convert the multidimensional tensor into the macro feature vector through the first flattening sublayer.

[0201] In this embodiment, the multidimensional tensor is passed to the first flattening sublayer L233, wherein the first flattening sublayer L233 may be a Flatten layer. The function of the Flatten layer is to convert the multidimensional tensor into a one-dimensional vector. Specifically, for the three-dimensional tensor from the paragraph-level multi-head attention sublayer L232, the Flatten layer will flatten it into a two-dimensional tensor [batch_size, flattened_features], wherein flattened_features is the product of all dimensions in the original three-dimensional tensor except batch_size, i.e., the macro feature vector of the embodiment of the present invention. This conversion is intended to simplify the input format of the subsequent feed-forward neural network layer (FNN, Feed-Forward Neural Network) so that it can directly process the flattened feature vector.

[0202] Step S244: Fusing the microscopic feature vector and the macroscopic feature vector through the first fusion layer to obtain a first fusion vector.

[0203] In this embodiment, the micro feature vector output by the micro feature extraction layer L22 simulates word frequency statistics, which helps to capture the degree of match between the query and the title or paragraph at the lexical level. The one-dimensional vector output by the macro feature extraction layer L23 represents the complex semantic relationship between the query and the paragraph captured by the multi-head attention mechanism. The micro feature vector and the macro feature vector are fused by the first fusion layer L24 to obtain a first fusion vector. Among them, the fusion method can be concatenation, element-wise addition, element-wise multiplication, weighted average, etc.

[0204] Step S245: Obtain a second feature vector according to the first fusion vector through the feedforward neural network layer.

[0205] In this embodiment, after the fusion is completed, the generated first fused vector will be sent to the feedforward neural network layer L25. Among them, the feedforward neural network layer L25 is FNN (feedforward neural network). FNN is a neural network architecture, which consists of one or more layers of linear transformation (fully connected layer), and each layer is usually followed by a nonlinear activation function (such as ReLU, Sigmoid or Tanh) to introduce nonlinearity. Among them, FNN receives the first fused vector, passes through one or more hidden layers, each hidden layer applies a linear transformation (weight matrix multiplied by the input vector plus the bias term), and then passes the result through the activation function. The second feature vector V2 is then output through the output layer. In this way, FNN can learn how to effectively combine information from the Flatten layer and the soft TF feature extraction layer, thereby improving the model's ability to understand the semantic relationship between queries and documents.

[0206] Step S246: Fusing the first feature vector and the second feature vector through the second fusion layer to obtain a second fusion vector.

[0207] In this embodiment, the first feature vector V1 and the second feature vector V2 are fused through the second fusion layer L26 to obtain a second fused vector.

[0208] The fusion method may be concatenation, element-wise addition, element-wise multiplication, weighted average, or the like.

[0209] Step S247: Generate the evaluation result according to the second fusion vector through the output layer.

[0210] In this embodiment, the evaluation result is generated according to the second fusion vector through the output layer L27.

[0211] The output layer L27 includes a second flat sublayer L271 and an activation function sublayer L272. The second flat sublayer L271 expands the second fusion vector to obtain a third feature vector, and then processes the third feature vector through the activation function sublayer L272 to obtain a correlation prediction result between the query and the object to be evaluated, that is, an evaluation result.

[0212] In some embodiments, the method further comprises:

[0213] Step S250: Generate query results based on the evaluation results and push them to the user.

[0214] In an optional implementation, the evaluation result is relevant or irrelevant, or the evaluation result is 1 or 0 (1 means relevant, 0 means irrelevant). At this time, the evaluation result can be used to obtain the query result, and all objects with relevant evaluation results are extracted as push data, and the push data is sent to the user.

[0215] In an optional implementation, the evaluation result is a relevance score, which is any value between greater than or equal to 0 and less than or equal to 1. At this time, the evaluation result can be used to sort all objects, and push data can be generated according to the sorting, and the push data can be sent to the user.

[0216] Article search is a core component in the question-answering field. Its task is to find the most relevant document from a set of related documents based on a specific query. This process relies on calculating the correlation between the user query and the article, and then retrieving and ranking the documents based on it. However, there are many challenges in the existing correlation calculation. First, the query is usually short and vague, and short queries often indicate that the user's intention is broader, resulting in a large number of recalled documents, which increases the complexity and pressure of sorting. Second, user feedback is sparse and noisy. Click behavior does not necessarily accurately reflect the relevance of the query. There may be cases of misclicks, and the articles that are not clicked are not completely irrelevant. In addition, the title of the article may sometimes seem irrelevant, but its content may be relevant to the query. In actual scenarios, users usually go through multiple behaviors to lock the target article, such as modifying keywords, clicking on documents, etc. Especially in technical documents, users may not be able to fully judge whether the document meets their needs before reading it in depth. Therefore, there are many invalid clicks in user behavior, which leads to users only being able to browse a small number of documents in a limited time and sparse behavior data. However, typical click-through rate estimation methods usually assume that user clicks are valid, or at least mostly valid. Moreover, documents are long and have diverse topics. Knowledge-based documents are usually long and consist of multiple paragraphs. Each document may cover multiple topics. For a query, the entire document may meet the requirement, or only some paragraphs may be relevant.

[0217] With the rapid development of large language models (LLMs), more and more studies have used LLMs to improve the performance of document retrieval. For example, some studies have alleviated the problem of insufficient annotated data by using large models to generate query data. However, these methods still have limitations when dealing with short and ambiguous queries. Other studies have used large models for document ranking and have achieved certain results, but such methods usually assume that the initial retrieved paragraphs are accurate enough, so their applicability to long documents and multi-topic documents is limited. In addition, there are also studies that try to enhance training data through LLMs, showing significant improvements in unsupervised re-ranking and click-through rate prediction, but this method often relies on a large amount of annotated data, while in actual scenarios, user feedback is often sparse and noisy.

[0218] Therefore, in order to solve the above problems, an embodiment of the present invention provides a data processing method, which realizes article query through the cooperation of offline process and online process.

[0219] Among them, the online process is divided into the query enhancement stage, the query-centric paragraph summarization stage, and the relevance evaluation stage. In the query enhancement stage, the original query statements in the historical data and the retrieved single documents are first input into the LLM to generate better search terms (enhanced query statements) that may lead to the document. In the query-centric paragraph summarization stage, the document, the original query statement, and the enhanced query statement are input into the LLM together. The LLM automatically divides the document into topics and generates a summary for each topic. In the relevance evaluation stage, the LLM evaluates the correlation between the enhanced query and the summary. These correlation results are used to fine-tune the feature encoding layer of the online model to improve the effect of subsequent embedding generation.

[0220] In the online process, the fine-tuned feature encoding layer is used to encode the initial query, article title, and summary generated by LLM, and perform semantic extraction at the macro and micro levels respectively. The micro level mainly captures the semantic relationship between the query and the article title at the word level; the macro level uses a multi-head attention mechanism to extract the semantic relationship between the query and the longer summary. Finally, the results of the two semantic extractions are combined with the output of the traditional text search model to calculate the final relevance score between the query and the document.

[0221] Therefore, a query enhancement technology based on a large language model (LLM) is proposed. By inferring and generating more specific and relevant potential search terms, the ambiguity problem of short queries is effectively alleviated, and the accuracy of recall and ranking is improved. By using LLM to model the complex behaviors of user clicks and queries, enhanced training data is generated, so that sparse and noisy data can be better processed and the model's understanding of the user's true intention can be improved. Through a query-centric document segmentation and summary generation method, LLM is used to automatically divide document topics and generate core summaries according to queries, while retaining key information of documents, reducing the complexity of long document processing. In the relevance evaluation stage of LLM-based enhanced queries and summaries, the feature encoding layer is fine-tuned through the generated relevance results to improve the matching effect of long documents and multi-topic articles. At the same time, in the online process, through two-layer semantic extraction, at the micro level, the query is tokenized and the semantic relationship with the article title is extracted at the word level. At the macro level, the semantic relationship between the query and the long summary is extracted using a multi-head attention mechanism, and the semantic matching between the query and the article is fully captured. At the same time, the practicality of the query system is improved by separating the offline and online processes. The entire model is divided into two parts: offline (query enhancement, paragraph summary generation, relevance evaluation) and online (embedding generation and real-time ranking) to ensure the scalability and efficiency of the model in deployment. By connecting multiple semantic extraction results in series and combining them with the results of traditional text search models, a more comprehensive relevance score calculation is achieved.

[0222] The embodiment of the present invention obtains an initial query statement and an object to be evaluated, obtains a first feature vector of the object to be evaluated based on an original model, obtains target association information of the object to be evaluated based on the initial query statement, and the target association information includes at least one of an enhanced query statement and summary information, and obtains an evaluation result representing the correlation between the initial query statement and the object to be evaluated based on the initial query statement, the first feature vector, the target association information, the object title, and the object content through an online model. In this way, the understanding ability of the query system can be improved, user needs can be accurately obtained, and the accuracy of the query results can be improved.

[0223] Fig.11 Schematic diagram of a data processing device according to an embodiment of the present invention. Fig.11As shown, the data processing device of the embodiment of the present invention includes a data acquisition unit 111, a first feature vector acquisition unit 112, a target association information acquisition unit 113 and an evaluation result acquisition unit 114. The data acquisition unit 111 is used to acquire an initial query statement and an object to be evaluated. The first feature vector acquisition unit 112 is used to acquire a first feature vector of the object to be evaluated based on the original model, and the first feature vector is used to characterize the relevance between the object to be evaluated and the initial query statement. The target association information acquisition unit 113 is used to acquire target association information of the object to be evaluated according to the initial query statement, and the target association information includes at least one of an enhanced query statement and summary information. The evaluation result acquisition unit 114 is used to acquire an evaluation result according to the initial query statement, the first feature vector, the target association information, the object title and the object content through an online model, and the evaluation result is used to characterize the relevance between the initial query statement and the object to be evaluated.

[0224] The embodiment of the present invention obtains an initial query statement and an object to be evaluated, obtains a first feature vector of the object to be evaluated based on an original model, obtains target association information of the object to be evaluated based on the initial query statement, and the target association information includes at least one of an enhanced query statement and summary information, and obtains an evaluation result representing the correlation between the initial query statement and the object to be evaluated based on the initial query statement, the first feature vector, the target association information, the object title, and the object content through an online model. In this way, the understanding ability of the query system can be improved, user needs can be accurately obtained, and the accuracy of the query results can be improved.

[0225] Fig.12 Schematic diagram of an electronic device according to an embodiment of the present invention. In this embodiment, the electronic device 6 includes a server, a terminal, etc. Figure 6 As shown, the electronic device 6: includes at least one processor 61; and a memory 62 connected to the at least one processor 61 for communication; and a communication component 63 connected to the scanning device for communication, the communication component 63 receives and sends data under the control of the processor 61; wherein the memory 62 stores instructions that can be executed by at least one processor 61, and the instructions are executed by at least one processor 61 to implement the above-mentioned data processing method.

[0226] Specifically, the electronic device includes: one or more processors 61 and a memory 62, Figure 6 A processor 61 is taken as an example. The processor 61 and the memory 62 may be connected via a bus or other means. Figure 6In the example, the bus connection is used. The memory 62 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 61 executes various functional applications and data processing of the device by running the non-volatile software programs, instructions and modules stored in the memory 62, that is, realizing the above-mentioned data processing method.

[0227] The memory 62 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store a list of options, etc. In addition, the memory 62 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 62 may optionally include a memory remotely arranged relative to the processor 61, and these remote memories may be connected to an external device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0228] One or more modules are stored in the memory 62, and when executed by one or more processors 61, the data processing method in any of the above method embodiments is executed.

[0229] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.

[0230] The embodiment of the present invention obtains an initial query statement and an object to be evaluated, obtains a first feature vector of the object to be evaluated based on an original model, obtains target association information of the object to be evaluated based on the initial query statement, and the target association information includes at least one of an enhanced query statement and summary information, and obtains an evaluation result representing the correlation between the initial query statement and the object to be evaluated based on the initial query statement, the first feature vector, the target association information, the object title, and the object content through an online model. In this way, the understanding ability of the query system can be improved, user needs can be accurately obtained, and the accuracy of the query results can be improved.

[0231] Another embodiment of the present invention relates to a non-volatile storage medium for storing a computer-readable program, wherein the computer-readable program is used for a computer to execute part or all of the above method embodiments.

[0232] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including a number of instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0233] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A data processing method, characterized in that: The method comprises: Get the initial query statement and the object to be evaluated; Acquire a first feature vector of the object to be evaluated based on the original model; Acquire target association information of the object to be evaluated according to the initial query statement, wherein the target association information includes at least one of an enhanced query statement and summary information; An evaluation result is obtained through an online model according to the initial query statement, the first feature vector, the target association information, the object title and the object content, and the evaluation result is used to characterize the relevance between the initial query statement and the object to be evaluated.

2. The method according to claim 1, characterized in that The method further comprises: Acquire historical data, where the historical data includes historical query statements and historical query results, where the historical query results include at least one candidate object; Acquire an enhanced query statement according to the historical query statement and the candidate object by a query enhancement module; Generate historical summary information corresponding to each candidate query statement according to the candidate object by a summary generation module, wherein the candidate query statement includes a historical query statement and an enhanced query statement; Obtaining a behavior feedback result according to the historical summary information through a relevance evaluation module, wherein the behavior feedback result includes the candidate query statement, the historical summary information, and a label, wherein the label is used to characterize the relevance between the candidate query statement and the corresponding historical summary information; The online model is adjusted according to the behavior feedback result.

3. The method according to claim 2, characterized in that The method further comprises: Generate and store the associated information corresponding to the candidate object, wherein the associated information includes the query statement and the corresponding summary information.

4. The method according to claim 3, characterized in that The step of obtaining the target association information of the object to be evaluated according to the initial query statement is specifically: The target association information is acquired from the stored data according to the initial query statement and the object to be evaluated.

5. The method according to claim 1, characterized in that The online model includes a feature encoding layer, a micro feature extraction layer, a macro feature extraction layer, a first fusion layer, a feedforward neural network layer, a second fusion layer, and an output layer.

6. The method according to claim 5, characterized in that The obtaining of the evaluation result according to the initial query statement, the first feature vector, the target association information, the object title and the object content by the online model comprises: Acquire an input vector through the feature encoding layer, wherein the input vector includes a query vector corresponding to the initial query statement and the enhanced query statement, a title vector corresponding to the object title, a content vector corresponding to the object content, and a summary vector corresponding to the summary information; Acquire a micro feature vector according to the query vector and the title vector through the micro feature extraction layer; Acquire a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer; fusing the microscopic feature vector and the macroscopic feature vector through the first fusion layer to obtain a first fusion vector; Obtaining a second feature vector according to the first fusion vector through the feedforward neural network layer; fusing the first feature vector and the second feature vector through the second fusion layer to obtain a second fused vector; The evaluation result is generated according to the second fusion vector through the output layer.

7. The method according to claim 6, characterized in that The microscopic feature extraction layer includes a conversion sublayer, a kernel pool sublayer and a feature extraction sublayer; Wherein, obtaining the micro feature vector according to the query vector and the title vector through the micro feature extraction layer includes: Obtaining a similarity matrix according to the query vector and the title vector through the conversion sublayer; Obtaining a kernel matrix according to the similarity matrix through the kernel pool sublayer; The microscopic feature vector is obtained according to the kernel matrix through the feature extraction sublayer.

8. The method according to claim 6, characterized in that The macro feature extraction layer includes a fusion sublayer, a paragraph-level multi-head attention sublayer and a first flat sublayer; The step of obtaining a macro feature vector according to the query vector, the content vector and the summary vector through the macro feature extraction layer includes: fusing the query vector and the content vector through the fusion sublayer to obtain a third fusion vector; Obtain a multidimensional tensor according to the third fusion vector and the summary vector through the paragraph-level multi-head attention sublayer; The multidimensional tensor is converted into the macro feature vector by the first flattening sub-layer.

9. The method according to claim 2, characterized in that: The query enhancement module, summary generation module and relevance evaluation module are large language models.

10. The method according to claim 1, characterized in that The method further comprises: Generate query results based on the evaluation results and push them to the user.

11. A data processing device, characterized in that: The device comprises: A data acquisition unit, used to acquire an initial query statement and an object to be evaluated; A first feature vector acquisition unit, configured to acquire a first feature vector of the object to be evaluated based on an original model, wherein the first feature vector is used to characterize the relevance between the object to be evaluated and the initial query statement; A target association information acquisition unit, configured to acquire target association information of the object to be evaluated according to the initial query statement, wherein the target association information includes at least one of an enhanced query statement and summary information; An evaluation result acquisition unit is used to acquire an evaluation result based on the initial query statement, the first feature vector, the target association information, the object title and the object content through an online model, wherein the evaluation result is used to characterize the correlation between the initial query statement and the object to be evaluated.

12. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method according to any one of claims 1-10.

13. A computer program product, comprising a computer program, characterized in that: When the computer program is executed on a computer, the computer executes the method according to any one of claims 1 to 10.

14. A computer-readable storage medium storing computer program instructions, characterized in that: The computer program instructions implement the method according to any one of claims 1 to 10 when executed by a processor.

Images