Training method and device of search matching model, computer equipment and storage medium

Abstract

The invention relates to a training method of a search matching model. The method comprises the steps of obtaining a training data set, wherein the training data set comprises an information two-tuple of a target search word and an information object; the target search word in each information two-tuple corresponds to an information inquiry intention, and the information object corresponds to an information providing intention; based on each information two-tuple in the training data set, obtaining a pre-trained twin-tower model; fusing the target search word and the information inquiry intention corresponding to the target search word into a search intention word, and fusing the information object and the information providing intention corresponding to the information object into an information carrier word; performing fine tuning training on the pre-trained double-tower model based on the search intention word and the information carrier word to obtain a trained double-tower model; the trained double-tower model is used for object retrieval. By adopting the method, the accuracy of object retrieval is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer device, storage medium, and computer program product for training a search matching model. Background Technology

[0002] Information retrieval technology aims to help users quickly and accurately obtain the information they need from large-scale datasets. The typical retrieval process involves the user entering search terms, the information retrieval system retrieving preliminary matching results from a document database, and further calculating the degree of match between the search terms and each result to return sorted search results to the user.

[0003] In related technologies, keyword matching or statistical ranking algorithms (such as term frequency-inverse term frequency algorithms) are typically relied upon to determine the degree of matching between search results and search terms. These methods return search results and rank them by analyzing the surface matching degree between query terms and the content of search results.

[0004] However, in reality, there is often a difference between the user's actual query intent and the intent expressed by the search terms. Therefore, the search results obtained by keyword matching based on the search terms may differ significantly from the user's expectations. Consequently, training the model using the search terms and the matching results leads to inaccurate search results output by the model. Summary of the Invention

[0005] Therefore, it is necessary to provide a training method, apparatus, computer device, computer-readable storage medium, and computer program product for a search matching model to address the aforementioned technical problems.

[0006] Firstly, this application provides a method for training a search matching model, the method comprising:

[0007] Obtain a training dataset, which contains information tuples of target search terms and information objects; in each information tuple, the target search term corresponds to an information query intent, and the information object corresponds to an information providing intent;

[0008] Based on each of the information pairs in the training dataset, a pre-trained dual-tower model is obtained;

[0009] The target search term and the information query intent corresponding to the target search term are merged into a search intent term, and the information object and the information providing intent corresponding to the information object are merged into an information carrier term;

[0010] The pre-trained dual-tower model is fine-tuned based on the search intent words and the information carrier words to obtain a trained dual-tower model; the trained dual-tower model is used for object retrieval.

[0011] In one embodiment, obtaining the pre-trained dual-tower model based on each of the information tuples in the training dataset includes:

[0012] The training dataset is input into the dual-tower model, and the matching results of the target search terms and information objects in each information tuple in the training dataset are obtained through the dual-tower model.

[0013] Based on the matching results, the first loss value of the dual-tower model is calculated;

[0014] Based on the first loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the pre-trained dual-tower model is obtained through iterative optimization.

[0015] In one embodiment, the dual-tower model includes a query tower module and a document tower module. Obtaining the matching results between the target search term and the information object in each information tuple of the training dataset through the dual-tower model includes:

[0016] The query tower module analyzes and processes the target search terms and information query intent in each information tuple to obtain a search feature vector.

[0017] The document tower module analyzes and processes the information objects and information provision intentions in each information tuple to obtain information feature vectors.

[0018] Based on the search feature vector and the information feature vector, the matching result between each target search term and the information object is determined.

[0019] In one embodiment, the step of fusing the target search term and the information query intent corresponding to the target search term into a search intent term, and fusing the information object and the information providing intent corresponding to the information object into an information carrier term, includes:

[0020] The target search term and the information query intent corresponding to the target search term are concatenated to obtain the first concatenation result;

[0021] Feature extraction is performed on the first concatenation result to obtain the search intent words;

[0022] The information object and the information provision intent corresponding to the information object are concatenated to obtain a second concatenation result;

[0023] Feature extraction is performed on the second splicing result to obtain the information carrier words.

[0024] In one embodiment, the step of fine-tuning the pre-trained dual-tower model based on the search intent term and the information carrier term to obtain the trained dual-tower model includes:

[0025] A consistency comparison is performed between the search intent word and the information carrier word for each information tuple in the training dataset to obtain the comparison result.

[0026] Based on the comparison results, the second loss value of the dual-tower model is calculated;

[0027] Based on the second loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the trained dual-tower model is obtained through iterative optimization.

[0028] In one embodiment, the method further includes:

[0029] Obtain real-time search terms and matching information objects within the platform;

[0030] The search terms and information objects are matched based on the trained dual-tower model to obtain the matching results of the search terms and information objects;

[0031] The matching results are used to locate the target information object in response to information retrieval requests.

[0032] In one embodiment, obtaining the training dataset includes:

[0033] Obtain the target search terms and the user behavior data associated with those target search terms;

[0034] Based on the first intent extraction model, the target search term, and the user behavior data, the information query intent corresponding to the target search term is extracted.

[0035] Obtain the information object and the target historical search terms associated with the information object;

[0036] Based on the second intent extraction model, the information object, and the target historical search terms, the information providing intent corresponding to the information object is extracted;

[0037] A training dataset is constructed based on the target search terms, the information query intent, the information object, and the information providing intent.

[0038] Secondly, this application also provides a training apparatus for a search matching model, the apparatus comprising:

[0039] The acquisition module is used to acquire a training dataset, which contains information tuples of target search terms and information objects; in each information tuple, the target search term corresponds to an information query intent, and the information object corresponds to an information providing intent.

[0040] The pre-training module is used to obtain a pre-trained dual-tower model based on each of the information tuples in the training dataset.

[0041] The fusion module is used to fuse the target search term and the information query intent corresponding to the target search term into a search intent term, and to fuse the information object and the information providing intent corresponding to the information object into an information carrier term.

[0042] The training module is used to fine-tune the pre-trained dual-tower model based on the search intent words and the information carrier words to obtain a trained dual-tower model; the trained dual-tower model is used for object retrieval.

[0043] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any of the methods described above.

[0044] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements any of the above methods.

[0045] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements any of the above methods.

[0046] The training methods, devices, computer equipment, storage media, and computer program products of the aforementioned search matching model obtain a training dataset. This training dataset not only contains information tuples of target search terms and information objects, but also additionally associates information query intent and information provision intent. This compensates for the shortcomings of short search terms and shallow semantics of information objects in traditional retrieval data, providing a deep semantic association training basis for the dual-tower model and avoiding matching bias caused by insufficient input information. Furthermore, the dual-tower model is first pre-trained based on complete information tuples, allowing the dual-tower model to initially grasp the basic semantic matching logic of the search side and the information provision side. Then, fine-tuning is performed through the fused search intent words and information carrier words. This not only retains the broad-spectrum semantic understanding ability of the pre-trained dual-tower model, but also strengthens the model's ability to capture fine-grained needs through explicit intent fusion, significantly improving the matching accuracy of the dual-tower model, and thus improving the object retrieval accuracy. Attached Figure Description

[0047] Figure 1 This is a diagram illustrating the application environment of an object search method in one embodiment;

[0048] Figure 2 This is a flowchart illustrating an object search method in one embodiment;

[0049] Figure 3 This is a schematic diagram illustrating the training and application of a first intent extraction model in one embodiment;

[0050] Figure 4 This is a schematic diagram illustrating the training and application of the second intent extraction model in one embodiment;

[0051] Figure 5 This is a structural block diagram of an object search device in one embodiment;

[0052] Figure 6 This is a flowchart illustrating the training method of a search matching model in one embodiment;

[0053] Figure 7 This is a flowchart illustrating the two-stage training process of a dual-tower model in one embodiment;

[0054] Figure 8 This is a flowchart illustrating the pre-training process of a dual-tower model in one embodiment;

[0055] Figure 9 This is a flowchart illustrating the steps for determining the matching results between each target search term and the information object in one embodiment;

[0056] Figure 10 This is a flowchart illustrating the steps of constructing the fused search intent terms and information carrier terms in one embodiment;

[0057] Figure 11 This is a flowchart illustrating the two-stage training process for a dual-tower model in one embodiment.

[0058] Figure 12 This is a flowchart illustrating the application method of a trained dual-tower model in one embodiment.

[0059] Figure 13 This is a flowchart illustrating the steps for obtaining the training dataset in one embodiment;

[0060] Figure 14 This is a schematic diagram of the training device for searching and matching models in one embodiment;

[0061] Figure 15 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] The object search method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices that support NFC functionality. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0064] In one embodiment, such as Figure 2 As shown, an object search method is provided. Embodiments of this application apply this method to... Figure 1 The example of server 104 includes the following steps:

[0065] Step 202: Obtain at least one target history selected object associated with the target search term.

[0066] In this embodiment, the target search term is the information that the user inputs or selects on the target application of the terminal. It should be noted that, in addition to text (words, sentences, or instructions), the target search term can also be data of any modality, such as one or more of characters, images, videos, and audio. The target search term can also include any object in the target application, such as published content (e.g., published images, videos, text, comments, etc.), users, activities, topics, live streams, shops, products, group chats, special effects, points of interest (e.g., locations, books, movies, music, podcasts), virtual objects, virtual resources, etc.

[0067] After determining their target search term, users can initiate a search request to the server via their terminal. Upon receiving the search request, the server retrieves the target historical selected objects associated with the target search term. These target historical selected objects refer to one or more objects ultimately selected by the user from the historical search results associated with the target search term (hereinafter referred to as information objects, and historical search results as historical information objects). Historical information objects are information objects retrieved based on a search request initiated at any historical moment; a historical moment refers to any time prior to the current search.

[0068] It should be noted that, similar to the target search term, the information object here can be data of any modality, and can also include any object in the target application.

[0069] Historical information objects associated with the target search term may include at least one of the following:

[0070] Historical information objects obtained based on the target search term; for example, if the target search term determined by the user is the text "food recommendations in a certain place", then the historical information objects are any information objects obtained by the server after any user initiates a search based on "food recommendations in a certain place" at any historical moment.

[0071] Historical information objects obtained from related search terms based on the target search term; for example, if the target search term determined by the user is the text "food recommendations in a certain place", then the historical information objects are any information objects obtained by the server after any user initiates a search based on related search terms such as "food guide in a certain place" or "must-eat list in a certain place" at any historical moment.

[0072] The related search terms for the target search term can be selected from search terms whose similarity to the target search term is greater than a similarity threshold. This similarity can be calculated in any way; for example, when the search term is text, the similarity can be the edit distance or text cosine similarity between the target search term and other search terms; when the search term is an image, the similarity can be the image similarity between the target search term and other search terms.

[0073] After each search, a user may select one or more objects from the search results as their preferred search terms. These selected objects are called the selected objects. Historical selected objects are determined from these historical selected objects; they are the selected objects from the search results obtained when searching based on the target search term or related search terms.

[0074] Here, "selected" means that the user interacts with the object in any preset form, such as: clicking to enter the object's details page, liking, saving, or commenting on the object, or browsing the object for a preset duration, etc. Which specific interaction is considered selected can be set by those skilled in the art according to actual needs.

[0075] The target historical selection objects can be some or all of the historical selection objects. In one embodiment, the target historical selection objects are determined based on the user information corresponding to each historical selection object, and / or the first interaction information corresponding to each historical selection object, wherein:

[0076] When selecting a target historical selected object based on the user information corresponding to each historical selected object, the difference between the user information of the search request corresponding to each historical selected object (hereinafter referred to as the first user information) and the user information corresponding to the target search term (hereinafter referred to as the target user information) can be calculated. That is, the server can obtain the first user information of the user who initiated the search request corresponding to the historical selected object, and obtain the target user information of the user who initiated the search request based on the target search term. The target historical selected object is determined from each historical selected object based on the difference between these two sets of user information. The user information can be any information generated by the user in the target application, such as, but not limited to: interest characteristics, objects viewed by the user in the target application, accounts followed by the user in the target application, etc.

[0077] For example, historical selection objects with relatively small differences between the corresponding first user information and target user information can be selected as target historical selection objects, so that the users who select target historical selection objects are more similar to users who search based on target search terms, making it easier to analyze the search intent of users who search based on target search terms through target historical selection objects.

[0078] When selecting a target historical selected object based on the first interaction information of each historical selected object, the first interaction information includes the number of times each historical selected object was selected during the search and / or the interaction data after being selected. Specifically, when an object is selected by the user once in a search, it is considered that the number of times that object has been selected is incremented by 1. Interaction data refers to any data used to describe the interaction between the user and the historical selected objects, such as: the total time spent browsing historical selected objects, the number of words in comments made when commenting on historical selected objects, whether the user liked, saved, or shared historical selected objects, etc.

[0079] For example, historical selections that have been selected more often or whose interaction data indicates that the user has interacted more frequently with the historical selection can be selected as target historical selections. This allows the target historical selections to represent the objects that most users would choose when searching based on target search terms (or related search terms), which also facilitates the analysis of the user's search intent.

[0080] Step 204: Based on the target search term and the target historical selected objects, extract the information query intent corresponding to the target search term.

[0081] In this embodiment, since the target historical selected objects are information objects that the user has selected in the historical search, the user's information search intent when searching using the target search term can be analyzed based on the target historical selected objects. The information search intent can be obtained by inputting the target search term and each target historical selected object into any pre-trained natural language processing model.

[0082] In one embodiment, the natural language processing model (hereinafter referred to as the first intent extraction model) is trained based on the output of a large language model. The first intent extraction model can be a large language model with fewer parameters, or any other natural language processing model with faster inference speed, to improve the extraction speed of information query intent in applications. The training process of the first intent extraction model includes:

[0083] Input the sample search terms and the selected historical objects of the sample target into the initial first intent extraction model to obtain the initial information query intent;

[0084] Based on the difference between the initial information query intent and the target information query intent, the initial first intent extraction model is trained to obtain a trained first intent extraction model.

[0085] The target information query intent is obtained by inputting the first prompt word into the large language model. The first prompt word is constructed based on the sample search terms and the sample target historical selected objects.

[0086] In this embodiment of the application, combined with Figure 3 As shown, before training the first intent extraction model, the sample search terms and the selected objects from the sample target history are first input into the large language model to obtain the target information query intent output by the large language model, that is... Figure 3 The process is shown in section (a). The sample search term can be any search term obtained from historical search requests. The method for determining the historical target selected object corresponding to the sample search term can be found in the aforementioned embodiment, and will not be repeated here.

[0087] In one example, a first prompt word can be constructed based on sample search terms and selected objects from the sample target history. This first prompt word is then input into a large language model to instruct the model to output a target information query intent that meets the requirements. In addition to the sample search terms and selected objects from the sample target history, the first prompt word may also include instruction text for instructing the large language model. The instruction text can be used to instruct the large language model to generate a target information query intent based on at least one of the following:

[0088] Keywords corresponding to the selected historical objects of the sample target; where keywords can be words contained within the selected historical objects of the sample target, or summary words extracted based on the selected historical objects of the sample target. An example of instruction text indicating the former could be: "Reflect high-frequency elements in the historical selected objects in the output." An example of instruction text indicating the latter could be: "Reflect keywords obtained from summarizing the historical selected objects in the output."

[0089] The instruction large language model generates target information query intent based on keywords of the target historical selected objects in the sample target. This helps the first intent extraction model learn the ability to include keywords of the target historical selected objects in the output information query intent. Subsequently, the matching degree between the information object and the target search term can be calculated by considering whether the keywords of the information object and the target historical selected objects match.

[0090] The application scenarios of the selected historical objects in the sample target; where application scenarios refer to the areas in which the selected historical objects in the sample target can provide assistance to users. For example, when the selected historical objects in the sample target are published content introducing food guides, its application scenarios may be tourism scenarios, dining scenarios, or business banquet scenarios. An example of instructing to generate target information query intent based on application scenarios can be: "When analyzing target information query intent, infer the usage scenario based on the selected historical objects in the target."

[0091] The information provision intent of the target search term refers to the specific information intent that the searched information must possess to satisfy the user's desired information query intent expressed through the target search term. For example, when the target search term is "food recommendations in a certain place," the information provision intent might be restaurant introductions or descriptions of local specialties. An example of instructing users to generate target information query intent based on the information provision intent of the target search term could be: "When analyzing target information query intent, consider what type of content the user wants to see."

[0092] The instruction text can also be used to indicate information that any other large language model needs to refer to when generating the target information query intent, and this application embodiment does not limit this. A first prompt word is constructed for each pair of sample search terms and sample target historical selected objects, and after inputting the first prompt word into the large language model, the target information query intent for each pair of sample search terms and sample target historical selected objects can be obtained.

[0093] Then it can be executed Figure 3 The process shown in section (b) involves training the initial first intent extraction model using the target information query intent. Sample search terms and historical selected objects can be input into the initial first intent extraction model to obtain the initial information query intent output by the model. Any loss function (such as mean squared error, reward model loss, etc.) can be used to calculate the difference between the initial information query intent and the target information query intent corresponding to the sample search terms and historical selected objects, and the parameters of the initial first intent extraction model are adjusted based on this difference. This process is repeated until a predetermined number of training rounds are completed, or until the difference between the initial information query intent output by the initial first intent extraction model and the target information intent is less than a preset threshold, at which point the trained first intent extraction model is obtained.

[0094] In applications, such as Figure 3 As shown in part (c), the server can input the target search term and the target historical selected object into the trained first intent extraction model to extract the information query intent corresponding to the target search term.

[0095] Step 206: Obtain at least one information object and the information providing intent corresponding to the information object. The information providing intent is extracted based on the target historical search term corresponding to the information object. The information object is a historically selected object associated with the target historical search term.

[0096] In this embodiment of the application, the information object is the search result that is initially recalled in the database using the target search term and has at least a certain relevance to the target search term. The server needs to calculate the matching degree between each information object and the target search term in order to perform further operations such as search ranking and search result matching.

[0097] Each information object has a corresponding information provision intent. This intent can be pre-extracted by the server and stored in association with the information object, or it can be extracted in real time after the server initially recalls the information object using target search terms. The information provision intent is extracted based on the information object and its corresponding target historical search terms. The target historical search terms are determined from various historical search terms, which refer to objects that ultimately included the information object when a search was initiated using them.

[0098] The target historical search terms can be some or all of the historical search terms. In one embodiment, the target historical search terms are determined based on the user information corresponding to each historical search term and / or the first interaction information corresponding to each historical selected object, wherein:

[0099] When selecting a target historical search term based on user information corresponding to each historical search term, the difference between the user information of the user who initiated the search request based on each historical search term (hereinafter referred to as second user information) and the user information corresponding to the target search term (hereinafter referred to as target user information) can be calculated. For example, historical selection objects with relatively small differences between the corresponding second user information and target user information can be selected as target historical selection objects. This ensures that users searching based on target historical search terms are more similar to users searching based on the target search term, facilitating analysis of which search term a user inputs before selecting this information object.

[0100] When selecting target historical search terms based on the second interaction information of each historical search term, the second interaction information includes the number of times the information object was selected when searching using historical search terms and / or the interaction data after being selected. Specifically, when an information object is selected once by a user in a search using a certain historical search term, it is considered as the number of times the information object has been selected for that historical search term +1. Interaction data refers to any data used to describe the interaction between the user and the information object when searching using historical search terms, such as: the total time spent browsing the information object, the number of words in comments made when commenting on the information object, whether the user liked, saved, or forwarded the information object, etc.

[0101] For example, historical search terms that users frequently select information objects when searching based on them, or that represent frequent interactions between users and information objects based on their interaction data, can be selected as target historical search terms, so that the target historical search terms can represent the search terms that most users would choose when selecting information objects.

[0102] Based on the information object and target historical search terms, the information provision intent of the information object can be extracted, that is, what specific information the information object can provide to users who view it. The information provision intent can be obtained by inputting the information object and each target historical search term into any pre-trained natural language processing model.

[0103] In one embodiment, the natural language processing model (hereinafter referred to as the second intent extraction model) is trained based on the output of the large language model. The second intent extraction model can be a large language model with fewer parameters, or any other natural language processing model with faster inference speed, to improve the speed of extracting the intent provided by information in the application. The training process of the second intent extraction model includes:

[0104] Input the sample information object and the sample historical search terms into the initial second intent extraction model to obtain the initial information to provide intent;

[0105] Based on the difference between the initial information provision intent and the target information provision intent, the initial second intent extraction model is trained to obtain a trained second intent extraction model.

[0106] The target information provision intent is obtained by inputting the second prompt word into the large language model. The second prompt word is constructed based on the information object and historical related search terms.

[0107] In this embodiment of the application, combined with Figure 4 As shown, before training the second intent extraction model, the sample information object and the sample historical search terms are first input into the large language model to obtain the target information output by the large language model to provide the intent, that is... Figure 4 The process is shown in section (a). The sample information object can be any object in the target application. The method for determining the sample historical search terms corresponding to the sample information object can be found in the method for determining the target historical search terms in the aforementioned embodiments, and will not be repeated here.

[0108] In one example, a second prompt word can be constructed based on the sample information object and the sample historical search terms. This second prompt word is then input into the large language model to instruct the model to output target information that meets the requirements and provides intent. In addition to the sample information object and the sample historical search terms, the second prompt word may also include instruction text for instructing the large language model. The instruction text can be used to instruct the large language model to generate target information and provide intent based on at least one of the following:

[0109] Keywords corresponding to the historical search terms of the sample target; where keywords can be words contained within the historical search terms of the sample target, or summaries extracted based on the historical search terms of the sample target.

[0110] The application scenario of the information object; that is, the area in which the information object can provide assistance to the user. An example of instructing the generation of the intended purpose of the target information in conjunction with the application scenario could be: "When analyzing the intended purpose of the target information, consider the applicable scenario or timing of the information object."

[0111] The information object satisfies the information query intent, that is, what kind of information query intent the user's search terms must have to lead the user to select the information object. An example of instructing the generation of target information provision intent based on the information query intent satisfied by the information object can be: "When analyzing the target information provision intent, consider which specific search needs the information object satisfies."

[0112] It should be noted that the instructions in the second prompt word for the large language model can correspond one-to-one with the instructions in the first prompt word for the large language model. For example, when the first prompt word indicates the intent to query the target information based on the application scenario, the second prompt word can correspondingly indicate the intent to provide the target information based on the application scenario, so that the corresponding parts of the information query intent and the information providing intent can be used when calculating the matching degree later. However, in fact, the instructions in the first prompt word and the instructions in the second prompt word may not have a corresponding relationship, and this application embodiment does not specifically limit this.

[0113] Instruction text can also be used to indicate information that any other large language model should refer to when generating target information to provide intent.

[0114] For each pair of sample information objects and sample target historical search terms, a second prompt word is constructed. After inputting the second prompt word into the large language model, the target information provision intent of each pair of sample information objects and sample target historical search terms can be obtained.

[0115] Then it can be executed Figure 4 The process shown in section (b) involves training the initial second intent extraction model using the target information provision intent. Sample information objects and historical search terms can be input into the initial second intent extraction model to obtain the initial information provision intent output by the model. Any loss function (such as mean squared error, reward model loss, etc.) can be used to calculate the difference between the initial information provision intent and the target information provision intent corresponding to the sample information object and historical search terms. The parameters of the initial second intent extraction model are then adjusted based on this difference. This process is repeated until a predetermined number of training rounds are completed, or until the difference between the initial information provision intent output by the initial second intent extraction model and the target information provision intent is less than a preset threshold, at which point the trained second intent extraction model is obtained.

[0116] In applications, such as Figure 4 As shown in section (c), the server can input each information object in the target application and its corresponding target historical search terms into the trained second intent extraction model to extract the information provision intent corresponding to the information object. Considering that the target historical search terms corresponding to the information object may change with new search requests, the server can also periodically update the information provision intent of each information object.

[0117] Step 208: Based on the target search term, information query intent, each information object and the information providing intent of each information object, determine the degree of matching between each information object and the target search term.

[0118] In this embodiment of the application, an enhanced target search term can be obtained based on the target search term and the information query intent. An enhanced information object can be obtained using the information object and the information providing intent corresponding to the information object. The degree of matching between the enhanced target search term and the enhanced information object is calculated as the degree of matching between the target search term and the information object.

[0119] This application does not limit how the target search terms and information objects are enhanced. For example, operations such as concatenating the target search terms and information query intent, or summing the vectors after converting them into vectors, can be performed to enhance the target search terms. Similarly, operations such as concatenating the information objects and information provision intent, or summing the vectors after converting them into vectors, can also be performed to enhance the information objects.

[0120] The degree of matching can be represented by a similarity score or by labels (e.g., labels may include two: match and non-match). Any text matching algorithm can be used to calculate the degree of matching between the enhanced target search term and the enhanced information object, or the enhanced search term and the enhanced information object can be input into any text similarity calculation model or text matching model to calculate the degree of matching; this application does not limit this approach.

[0121] In one embodiment, the degree of matching is determined by a matching model; the method further includes:

[0122] Identify the target information object selected from among the various information objects, and determine the actual matching degree corresponding to the target information object based on the interaction data after the target information object is selected;

[0123] The matching model is adjusted based on the difference between the degree of matching of the target information object and the actual degree of matching.

[0124] In this embodiment of the application, the degree of matching is calculated by a matching model, which can be any text matching model, such as a large language model.

[0125] The server can perform self-supervised training of the matching model based on user feedback on the calculated matching degree during practical applications. After calculating the matching degree between each information object and the target search term, the server can record the target information object actually selected by the user among each information object, and can further determine the actual matching degree between the target information object and the target search term based on the user's interaction data with the target information object. The server can then calculate the loss of the matching model based on the difference between the actual matching degree and the matching degree of the target information object calculated by the matching model, and adjust the parameters of the matching model based on the loss.

[0126] For example, when the matching degree is represented by labels, if the user's interaction time with the target information object is long enough (e.g., the browsing time of the target information object exceeds a certain threshold), or if the user performs a preset type of interactive operation on the target information object (preset types may include liking, collecting, etc.), then the server can determine that the actual matching degree of the target information object is "match". If the matching model calculates a "mismatch" for the target information object, the server can calculate the loss of the matching model using any loss function applicable to classification tasks, including but not limited to cross-entropy loss, to adjust the matching model and achieve self-supervised training of the matching model.

[0127] The object search method provided in this application, for the target search term input by the user, extracts information query intent (i.e., what query needs the user has when searching with the target search term) based on the target historical selected objects that were associated with the target search term and selected by the user in previous search tasks. For the information object to be matched, it extracts information provision intent (i.e., what query needs the information object can satisfy the user) based on the target historical search terms that the user selected in the search results when searching with it before. The information query intent and information provision intent are then used as supplementary inputs when calculating the matching degree. This application embodiment improves the accuracy of search matching by combining the user's query intent to calculate the matching degree between the target search term and the information object.

[0128] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0129] Based on the same inventive concept, this application also provides an object search apparatus for implementing the object search method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more object search apparatus embodiments provided below can be found in the limitations of the object search method described above, and will not be repeated here.

[0130] In one embodiment, such as Figure 5As shown, an object search device 500 is provided, including: a first acquisition module 502, an extraction module 504, a second acquisition module 506, and a first determination module 508, wherein:

[0131] The first acquisition module 502 is used to acquire at least one target historical selected object associated with the target search term;

[0132] The extraction module 504 is used to extract the information query intent corresponding to the target search term based on the target search term and the target historical selected object;

[0133] The second acquisition module 506 is used to acquire at least one information object and the information providing intent corresponding to the information object. The information providing intent is extracted based on the target historical search term corresponding to the information object. The information object is a historical selected object associated with the target historical search term.

[0134] The first determining module 508 is used to determine the degree of matching between each information object and the target search term based on the target search term, the information query intent, each information object and the information providing intent of each information object.

[0135] In one embodiment, the target historical selected object is determined from a variety of historical selected objects, which are objects selected from the various objects obtained during a search based on the target search term or related search terms of the target search term; and / or,

[0136] The target historical search term is determined from various historical search terms, and the information object is one of the objects selected from the various objects obtained when searching based on the historical search term.

[0137] In one embodiment, the target historical selection object is selected from each of the historical selection objects based on first interaction information, the first interaction information including the number of times each historical selection object was selected during the search and / or the interaction data after being selected; and / or,

[0138] The target historical search terms are selected from each of the historical search terms based on the second interaction information, which includes the number of times the information object was selected during the search and / or the interaction data after it was selected.

[0139] In one embodiment, the information query intent is obtained by inputting the target search term and the target historical selected object into a first intent extraction model;

[0140] The training process of the first intent extraction model includes:

[0141] Input the sample search terms and the selected historical objects of the sample target into the initial first intent extraction model to obtain the initial information query intent;

[0142] Based on the difference between the initial information query intent and the target information query intent, the initial first intent extraction model is trained to obtain a trained first intent extraction model.

[0143] The target information query intent is obtained by inputting a first prompt word into a large language model. The first prompt word is constructed based on the sample search terms and the historical selected objects of the sample target.

[0144] In one embodiment, the first prompt word is further used to instruct the large language model to generate the target information query intent based on at least one of the following:

[0145] The keywords corresponding to the historical selected objects of the sample target;

[0146] The sample targets historically selected the application scenarios of the objects.

[0147] The target search term requires information to provide intent.

[0148] In one embodiment, the information providing intent is obtained by inputting the information object and the target historical search terms corresponding to the information object into a second intent extraction model;

[0149] The training process of the second intent extraction model includes:

[0150] Input the sample information object and the sample historical search terms into the initial second intent extraction model to obtain the initial information to provide intent;

[0151] Based on the difference between the initial information provision intention and the target information provision intention, the initial second intention extraction model is trained to obtain a trained second intention extraction model.

[0152] The target information provision intent is obtained by inputting a second prompt word into a large language model. The second prompt word is constructed based on the information object and the historical related search terms.

[0153] In one embodiment, the second prompt word is further used to instruct the large language model to generate the target information to provide intent based on at least one of the following:

[0154] Keywords corresponding to the historical search terms of the sample target;

[0155] The application scenarios of the information object;

[0156] The information object satisfies the information query intent.

[0157] In one embodiment, the degree of matching between the information object and the target search term is determined by a matching model; the apparatus further includes:

[0158] The second determining module is used to determine the target information object selected among the information objects, and to determine the actual matching degree corresponding to the target information object based on the interaction data after the target information object is selected.

[0159] The adjustment module is used to adjust the matching model based on the difference between the matching degree of the target information object and the actual matching degree.

[0160] Each module in the above-mentioned device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0161] In one embodiment, the matching model can be a dual-tower model built on a large language model decoder. This model learns low-dimensional vector representations of the target search term and the information object through two independent decoders, and then completes the matching task by calculating vector similarity. The advantage of using a dual-tower model is that it can pre-compute and store the vector representation of the information object. When a user initiates a search request, the dual-tower model calculates the vector representation of the target search term in real time, and then retrieves the vector representation of the information object from storage for matching, thus completing the search task and significantly improving the response efficiency of online searches.

[0162] Accordingly, this application is Figure 6 This paper provides a method for training a search matching model, which can be applied to... Figure 1 Taking server 104 as an example, it can be seen that this method can also be applied to terminals, and also to systems composed of terminals and servers. The method includes the following steps 602 to 608, wherein:

[0163] Step 602: Obtain the training dataset.

[0164] The training dataset contains information tuples of target search terms and information objects. Each information tuple contains a target search term corresponding to an information query intent, and an information object corresponding to an information provision intent. These information query intents and information provision intents can be extracted using the first intent extraction model and the second intent extraction model as described in the preceding embodiments.

[0165] Step 604: Based on each information tuple in the training dataset, obtain the pre-trained dual-tower model.

[0166] In implementation, the server obtains the pre-trained dual-tower model based on each information tuple in the training dataset. For example, this pre-training process can be divided into three parts: input encoding, task-driven learning, and iterative parameter optimization.

[0167] In the input encoding stage, the server can first construct prompts for the target search term and the information object respectively. For the target search term side, the prompts could be "Search requirement: Term [XX], Information query intent [User wants to obtain XX information]". For the information object side, the prompts could be "Information content: Object [XX note], Value [Can provide XX information]". The server can then encode the prompts for the target search term and the information object into the input vector of the dual-tower model through word segmentation, embedding mapping, and positional encoding processes.

[0168] In the task-driven learning phase, the server can teach the model to make matching judgments based on weakly supervised tasks. For example, the process of learning using weakly supervised tasks may include: generating weakly supervised labels (such as match / non-match labels) for information tuples based on user behavior (such as whether the information object is clicked after searching for the target search term), and then designing a binary classification matching task, that is, having the dual-tower model calculate the similarity between the target search term and the information object based on the semantic vectors output by the dual-tower modules, and then outputting the match / non-match prediction results through a fully connected layer.

[0169] In the parameter iterative optimization stage, a loss function such as binary cross-entropy can be used to calculate the model's predicted loss value, and the Adam optimizer can be used to control the learning rate of the parameters so that the predicted loss value is used to update the model parameters. After the model parameters have been iterated and optimized for a preset number of rounds, or after the predicted loss value is lower than a certain threshold, a pre-trained dual-tower model can be obtained.

[0170] Step 606: Merge the target search term and the information query intent corresponding to the target search term into a search intent term, and merge the information object and the information providing intent corresponding to the information object into an information carrier term.

[0171] In implementation, after the model pre-training is completed, a preset fusion algorithm can be used to fuse the target search term and the corresponding information query intent into a search intent term, and to fuse the information object and the corresponding information provision intent into an information carrier term. This application embodiment does not limit the preset fusion algorithm, but may include, for example, pre- and post-merging, mean fusion, weighted fusion, etc.

[0172] Step 608: Fine-tune the pre-trained dual-tower model based on the search intent words and information carrier words to obtain the trained dual-tower model.

[0173] The trained dual-tower model is used for object search.

[0174] In practice, after obtaining the search intent words and information carrier words, the terminal further trains the pre-trained dual-tower model based on the search intent words and information carrier words, and obtains the trained dual-tower model by fine-tuning the model parameters.

[0175] In practice, the fused search intent terms are input into the query tower of the pre-trained dual-tower model (i.e., the sub-network used to convert target query terms into vector representations). With the help of the pre-trained semantic extraction capabilities within the tower, a search feature vector containing complete search requirements is generated. At the same time, the information carrier terms are input into the document tower (i.e., the sub-network used to convert information objects into vector representations). With the help of the pre-trained semantic extraction capabilities within the tower, an information provision feature vector is generated.

[0176] Furthermore, the server performs a consistency comparison between the search intent term and the information carrier term (i.e., determines the degree of matching between the search intent term and the information carrier term), and calculates the loss value of the dual-tower model based on the consistency comparison result and the model's preset loss function, in order to fine-tune the dual-tower model training. The loss function can be dynamically adjusted according to the task type; multi-class cross-entropy loss can be used, and if it is necessary to strengthen the learning of difficult examples, focal loss can be introduced to reduce the weight of easy examples, ensuring that the model focuses more on learning high-value samples. Thus, the model parameters of the dual-tower model are fine-tuned using the model loss value (i.e., the second loss value in the following embodiment), and the fine-tuning process of the dual-tower model is determined to have stopped. When the loss value (second loss value) of the dual-tower model meets the preset model training stopping condition, the parameter iterative optimization is completed, and the trained dual-tower model is obtained.

[0177] In a specific embodiment, such as Figure 7 As shown, Figure 7 This demonstrates the complete secondary training process for the dual-tower model. Figure 7 The training process of the dual-tower model is divided into two stages. The first stage is pre-training, in which the dual-tower model can be a Qwen3 LLM Decoder (large language model decoder). Then, the target search term, information query intent, and information object are respectively input into the query tower and document tower of this large language model decoder. Figure 7Taking notes as an example, this section describes how information objects (including note titles, body text, and tags) are processed by a dual-tower model to output a judgment result. This result can be a binary classification matching result, where y represents a match and n represents a mismatch. Based on this matching result and a pre-defined loss function, the dual-tower model is pre-trained. Through multiple iterations to optimize the model parameters, the pre-training process of the dual-tower model is completed.

[0178] The second stage involves enhancing representation modeling. The pre-trained dual-tower model (large language model decoder) is further trained. In the dual-tower model, the left query tower receives the search intent term generated by fusing the target search term and the information intent term, while the right document tower receives the information carrier term generated by fusing the information object and the information provision intent. The left and right towers share parameters. Contrastive learning is used to optimize the representation of the search intent term and the information carrier term, and a consistency comparison result between them is provided. Then, based on this comparison result and a pre-defined loss function, the dual-tower model is trained. Through multiple iterations to optimize the model parameters, the training process of the dual-tower model is completed.

[0179] In this embodiment, a training dataset is obtained. This training dataset not only contains information tuples of target search terms and information objects, but also additionally associates information query intent and information provision intent. This compensates for the shortcomings of short search terms and shallow semantics of information objects in traditional search data, providing a training basis for deep semantic association for the dual-tower model and avoiding matching deviations caused by insufficient input information. Furthermore, the dual-tower model is first pre-trained based on complete information tuples, allowing the dual-tower model to initially grasp the basic semantic matching logic of the search side and the information provision side. Then, fine-tuning is performed through the fused search intent words and information carrier words. This not only retains the broad-spectrum semantic understanding ability of the pre-trained dual-tower model, but also strengthens the model's ability to capture fine-grained needs through explicit intent fusion, greatly improving the matching accuracy of the dual-tower model, and thus improving the object search accuracy.

[0180] In one exemplary embodiment, such as Figure 8 As shown, the specific processing steps of step 604 include:

[0181] Step 801: Input the training dataset into the dual-tower model, and obtain the matching results of the target search term and the information object in each information tuple in the training dataset through the dual-tower model.

[0182] In practice, during the training of the dual-tower model, the server first obtains and loads the training dataset stored locally or synchronized from the server. Then, for each information tuple in the training dataset, it performs structured preprocessing operations according to the input format requirements of the dual-tower model.

[0183] Specifically, the server semantically integrates the target search terms and information query intent in each information tuple to generate structured query-side input text. Simultaneously, it integrates the information object and information provision intent into structured document-side input text. After input preprocessing, the server calls a locally deployed dual-tower model, passing the two types of structured inputs to their respective modules: the query-side input text is sent to the query tower of the dual-tower model, where the built-in Transformer encoder performs semantic parsing and feature extraction, ultimately generating a fixed-dimensional query-side vector; simultaneously, the document-side input text is passed to the document tower, which uses the same encoder structure as the query tower, extracting features of the information object through the same encoding logic to generate a document-side vector matching the dimension of the query-side vector.

[0184] After the query side vector and document side vector are generated, the server uses the model's correlation calculation module to call a preset similarity algorithm to perform numerical correlation analysis on the query side vector and document side vector, and obtains the original numerical values ​​that reflect the degree of semantic similarity between the two.

[0185] Subsequently, the original values ​​are transformed according to preset business rules: if the business rule indicates that a discrete conclusion should be output, the server compares the original value with a preset threshold. If the original value is greater than the preset threshold, the result is judged as "match," or if the original value is less than or equal to the preset threshold, the result is judged as "mismatch." If the business rule indicates that the difference in matching degree should be reflected, the server can normalize the original value to the range of 0 to 1, and generate a matching probability result (expressed as a percentage, for example, if the normalized result is 0.4, then the matching probability result is 40%) or a matching score (which can be expressed as a decimal in the range of 0 to 1, for example, if the normalized result is 0.4, then the matching score is also 0.4).

[0186] Finally, the server stores the matching results corresponding to each information tuple in the local database according to a preset format.

[0187] Step 802: Based on the matching results, calculate the first loss value of the dual-tower model.

[0188] In implementation, after obtaining the matching results for each information tuple in the training dataset, the true label for each information tuple is extracted from the metadata of the training dataset. These labels are pre-annotated based on previously recorded user behavior data. For example, if it is detected that a user has actively clicked on or saved a corresponding information object through a target search term, or stayed on the information object page for a preset threshold (reflecting that the user's needs have been met), the true label for that information tuple is marked as a value representing "match". If the records show that the user did not click on the information object in the tuple, or the core content of the information object has no semantic connection with the target search term, it is marked as a value representing "not match", ensuring that the true label for each tuple can objectively reflect the actual matching situation.

[0189] Subsequently, the server calls a locally preset loss function (such as the binary cross-entropy loss function) based on the current training task type of the dual-tower model. During the calculation, the model's predicted matching probability or matching score for each information tuple is extracted from the matching result data corresponding to each information tuple. Then, the true label and the predicted matching probability or matching score for each information tuple are aligned and calculated one by one according to the loss function formula to obtain the local loss value of the dual-tower model's output for each information tuple.

[0190] An example of how to calculate the local loss value may include: for information tuples with the true label "match", calculating the deviation between the predicted match probability / match score and the preset match probability / preset match score of the "match" label as the local loss value. For binary classification problems, the preset match probability / preset match score of the "match" label can typically be set to 1.

[0191] For information tuples with the true label "not matching", the deviation between the predicted probability / match score and the preset matching probability / preset matching score of the "not matching" label is calculated as the local loss value. For binary classification problems, the preset matching probability / preset matching score of the "not matching" label can usually be set to 0.

[0192] After calculating the local loss values ​​for all information tuples, the summation of all local loss values ​​and the average value are taken to obtain the final first loss value.

[0193] Step 803: Based on the first loss value, adjust the parameters of the dual-tower model through the backpropagation algorithm, and iteratively optimize to obtain the pre-trained dual-tower model.

[0194] In practice, after obtaining the first loss value, the forward computation path of the dual-tower model (from the input layer to the output layer) can be traced back to calculate the degree of influence of each parameter in the model (including the weight matrix, bias term, attention mechanism parameters, etc. of the Transformer encoder in the query tower and document tower) on the loss value, i.e., the parameter gradient.

[0195] The server can generate a corresponding gradient value for each parameter: if the gradient is positive, it means that increasing the parameter will lead to an increase in the loss value, and the parameter needs to be adjusted downward according to the gradient direction; if the gradient is negative, it means that increasing the parameter can reduce the loss value, and the parameter needs to be adjusted upward along the gradient direction, so as to ensure that the parameter adjustment always points in the direction of reducing the loss.

[0196] Subsequently, a locally configured optimizer (such as the Adam optimizer) can be invoked to control the parameter tuning process, thereby improving the tuning accuracy. The optimizer adaptively allocates tuning step sizes to different parameters based on information such as the gradient magnitude and historical tuning records for each parameter.

[0197] For example, for key parameters with large absolute gradient values ​​(i.e., parameters that significantly affect the loss value), the optimizer can assign a larger adjustment range to these parameters to improve the optimization speed of key parameters. For minor parameters with small absolute gradient values, a smaller adjustment step size is used to avoid over-adjustment that could damage the basic performance of the model.

[0198] During the parameter update process, the values ​​of each parameter before and after the adjustment, the gradient change curve, and other data can be recorded in real time in each round of adjustment, so that the optimizer can refer to these data to allocate the adjustment step size later.

[0199] This parameter adjustment process can be iterated multiple times. For example, a validation dataset for the dual-tower model can be created in the same way as the training dataset. After each round of parameter adjustment, the validation dataset can be loaded, and the dual-tower model with the currently adjusted parameters can be used to predict the matching between the target search term and the information object. This calculates the matching accuracy and loss value of the dual-tower model on the validation set, and the results are compared with those from the previous round of parameter adjustment.

[0200] If, after two consecutive rounds of adjustments, the model's accuracy on the validation dataset does not significantly improve, or the first loss value has fallen below a preset threshold (e.g., below 0.1), the model performance can be determined to have stabilized, and the iteration process can be stopped. At this point, the pre-training phase is complete.

[0201] In this embodiment, training data containing target search terms, information objects and their corresponding intentions are input into the dual-tower model, allowing the model to access complete information fused with deep semantics in the initial stage. This avoids semantic learning bias caused by fragmented input information and makes the matching results generated by the model more accurate.

[0202] In one exemplary embodiment, such as Figure 9 As shown, the dual-tower model, as a typical dual-branch deep learning architecture, revolves around a symmetrical design of the query tower module and the document tower module. The two modules maintain consistency in their network structure (some basic parameters can be shared) to ensure semantic space alignment. The query tower module focuses on search-side information processing and outputs a search feature vector; the document tower module, on the other hand, focuses on information-side content, employing the exact same layer structure as the query tower to synchronously process information objects and information intent, ultimately outputting an information feature vector with the same dimension as the search feature vector. Furthermore, the dual-tower model includes a matching computation layer, which receives the feature vectors output by both towers and quantifies semantic relevance using algorithms such as cosine similarity and dot product, thereby generating matching results between the target search term and the information object. Therefore, the specific processing steps in step 801 include:

[0203] Step 901: The query tower module analyzes and processes the target search terms and information query intent in each information tuple to obtain the search feature vector.

[0204] In implementation, the target search terms and their corresponding information query intents in each information tuple are first standardized. Standardization may include, but is not limited to, removing redundant symbols using text cleaning tools and standardizing the format of the Chinese and English texts. Then, semantically concatenating the target search terms and their corresponding information query intents according to a pre-defined structured template (e.g., "search term: [content]; intent: [content]").

[0205] Next, the concatenated target search terms and information query intent (i.e., the concatenated text) are input into the query tower module, which is built on the same Transformer encoder architecture as the document tower. The encoder's process for processing the concatenated text typically includes: extracting basic features through a word embedding layer and assigning different dynamic weights to different words in the concatenated text (e.g., important words should be given higher weights); capturing semantic relationships between words through a self-attention mechanism; and integrating global semantic information through a cross-attention layer.

[0206] Ultimately, the query tower module can compress the high-dimensional encoding result output by the cross-attention layer into a low-dimensional vector of fixed dimensions, i.e., the search feature vector, through the pooling layer.

[0207] Step 902: The document tower module analyzes and processes the information objects and information providing intentions in each information tuple to obtain information feature vectors.

[0208] In implementation, the information objects and their corresponding information provision intentions within each information tuple are first standardized. Standardization may include, but is not limited to: filtering key content from the information objects (such as the title, key paragraphs in the body text, tags, etc.), removing redundant symbols using text cleaning tools, and standardizing the format of the Chinese and English texts. Then, according to a pre-defined structured template (such as "Information Object: [Key Content]; Value: [Content]"), the information objects and their corresponding information provision intentions are semantically concatenated.

[0209] Next, the concatenated information object and the information provision intent (i.e., the concatenated text) are input into the document tower module, which is built on the same encoder architecture as the query tower. Therefore, the encoder's process for processing concatenated text can be referenced from the query tower's process for processing concatenated text, and will not be elaborated here.

[0210] After encoding, the Document Tower module can use pooling layers to compress the high-dimensional encoding result output by the cross-attention layer into a low-dimensional vector of fixed dimensions, i.e., the information feature vector.

[0211] Step 903: Based on the search feature vector and the information feature vector, determine the matching results between each target search term and the information object.

[0212] In implementation, the search feature vector and information feature vector from the same pair of information tuples are retrieved, and the similarity between the search feature vector and information feature vector is calculated using any similarity algorithm (such as cosine similarity or dot product).

[0213] After similarity calculation, the calculated raw values ​​are transformed according to preset business rules: If a discretized matching conclusion needs to be output, the raw similarity value can be compared with a preset threshold. If the raw similarity value is greater than or equal to the threshold, the matching result is determined as "match". If the raw similarity value is less than the threshold, the matching result is determined as "no match". If it is necessary to reflect the difference in matching degree, the raw similarity value can be mapped to the range of 0 to 1.

[0214] Finally, the server can store the matching results and information tuples associated with each information tuple in a way that will be used for subsequent training of the dual-tower model.

[0215] In this embodiment, the query tower module converts the target search term and information query intent into search feature vectors, and the document tower module converts the information object and information provision intent into information feature vectors. Then, based on the similarity calculation of the two types of feature vectors, the matching degree between the target search term and the information object is quantified to obtain the matching result. This avoids the problem of surface semantic interference in traditional text matching and captures deep intent associations, significantly improving the accuracy of the matching result.

[0216] In one exemplary embodiment, such as Figure 10 As shown, after the pre-training of the dual-tower model is completed, in order to further improve the output accuracy of the dual-tower model, a secondary training of the dual-tower model is implemented. In this secondary training process, it is necessary to fuse the target search terms and their corresponding information query intentions, and to fuse the information objects and their corresponding information providing intentions. The specific processing steps of step 606 include:

[0217] Step 1001: Concatenate the target search term with the information query intent corresponding to the target search term to obtain the first concatenation result.

[0218] In implementation, the splicing method can include, but is not limited to, concatenation of phrases before and after the query. In one example, the information query intent can be broken down into multiple semantic units (such as demand type, applicable scenario, preference conditions, etc.). During the splicing process, the terminal uses a semantic deduplication algorithm to identify and filter expressions in the information query intent that are repeated with the search terms. Simultaneously, it constructs a semantic hierarchy using hierarchical delimiters and appends these hierarchical units sequentially after the target search term according to a preset priority, resulting in a structured first splicing result such as "target search term: [semantic unit 1]; [semantic unit 2]...", which facilitates the model's recognition of the first splicing result. After the target search term and information query intent are spliced, the server dynamically adjusts the text length. The final generated first splicing result can be encrypted and stored in the terminal's temporary data buffer.

[0219] Step 1002: Extract features from the first splicing result to obtain the search intent words.

[0220] In implementation, the server can extract features from the first concatenated result through steps including but not limited to word segmentation, attention weight allocation, and semantic integration to obtain search intent words. The final search intent words will undergo length standardization for use in model consistency comparison training. The word segmentation step includes, but is not limited to, decomposing the first concatenated result into lexical units, part-of-speech tagging, and entity recognition. In the attention weight allocation step, higher weights can be assigned to keywords in the target search term and keywords in the information query intent, while lower weights can be assigned to non-core components such as modifying adverbs and conjunctions, so that the model can consider the keywords in the first concatenated result more in subsequent processing.

[0221] Step 1003: The information object and the information providing intent corresponding to the information object are spliced ​​together to obtain the second splicing result.

[0222] In practice, considering that information objects are usually quite long and contain a lot of redundant information, the information objects can be simplified before splicing. For example, key paragraphs and core information of the information objects can be identified through a pre-trained text classification model.

[0223] This allows for the combination of information extracted from the information object and the intent behind that information. For details on the combination process, please refer to the section on combining target search terms and information query intent; it will not be elaborated upon here.

[0224] In an optional example, during the concatenation process of the second concatenation result, the server will also initiate semantic association verification. By calculating the semantic similarity between the main content and the value description, if the similarity is lower than a preset threshold (e.g., 0.4), it is determined that there is a semantic disconnect, and a manual review is automatically prompted to avoid generating logically contradictory concatenated text. After the initial concatenation is completed, the terminal initiates adaptive optimization based on the text length: if the text length exceeds the maximum limit of the model input (e.g., 512 characters), the terminal will adopt a core information priority retention strategy, prioritizing the retention of core elements such as topic words and key value points based on word frequency and semantic weight, and performing gradient compression on decorative content to ensure that the core information can still be completely conveyed after compression; if the text length is lower than the minimum threshold, the terminal will retrieve relevant supplementary information from a preset domain semantic library according to the domain to which the information object belongs, to avoid insufficient semantic information during subsequent feature extraction due to the text being too short.

[0225] Step 1004: Extract features from the second splicing result to obtain the information carrier words.

[0226] In implementation, the same method of feature extraction for the first concatenation result can be used to extract features for the second concatenation result through steps including but not limited to word segmentation, attention weight allocation, and semantic integration.

[0227] In this embodiment, the process achieves bidirectional optimization of the text on the search side and the information side, which not only ensures the integrity and accuracy of the two types of features, but also ensures the alignment of the two in the semantic dimension through symmetrical processing logic. This provides high-quality input for the subsequent matching calculation and loss optimization of the dual-tower model, which can improve the model's learning efficiency and matching accuracy.

[0228] In one exemplary embodiment, such as Figure 11 As shown, after the dual-tower model pre-training is completed, in order to improve the accuracy of the dual-tower model, this embodiment considers further training the dual-tower model, that is, performing consistency comparison between the search intent words and information carrier words generated by the information tuples, and further training the dual-tower model. Specifically, the specific processing procedure of step 606 includes:

[0229] Step 1101: Perform a consistency comparison between the search intent word and the information carrier word for each information tuple in the training dataset to obtain the comparison results.

[0230] In implementation, for each pair of search intent words and information carrier words belonging to the same information tuple, the server inputs the search intent word into the query tower and the information carrier word into the document tower, obtaining the comparison result output by the dual-tower model. According to the output format of the dual-tower model in the aforementioned pre-training process, the comparison result here can be a discrete match / non-match label, or a continuous probability or score between 0 and 1.

[0231] Step 1102: Based on the comparison results, calculate the second loss value of the dual-tower model.

[0232] In implementation, the comparison results are compared with the true labels (match or non-match) of the information tuples. Subsequently, an appropriate loss function is selected according to task requirements, and the second loss value of the dual-tower model is calculated based on the difference between the comparison results and the true labels.

[0233] Optionally, a local loss value can be calculated for the search intent term and information carrier term of each information tuple. Then, a weighted sum and mean calculation can be performed based on the local loss values ​​of each information tuple to obtain the final second loss value. The calculation method of the local loss value can refer to the calculation method of the local loss value when calculating the first loss value in the aforementioned embodiment, and will not be repeated here.

[0234] Step 1103: Based on the second loss value, adjust the parameters of the dual-tower model through the backpropagation algorithm, and iteratively optimize to obtain the trained dual-tower model.

[0235] In implementation, after obtaining the second loss value, the computation path of the dual-tower model is traced backward (from the feature extraction layer of search intent words and information carrier words to the consistency comparison layer) to calculate the gradient impact of each parameter on the second loss value. These parameters may include, but are not limited to, the encoder weights of the query tower and document tower, attention mechanism parameters, and comparison layer coefficients. A positive gradient impact indicates that increasing the parameter value will exacerbate the loss and should be adjusted downward proportionally; a negative gradient impact indicates that increasing the parameter value helps reduce the loss and should be adjusted upward proportionally.

[0236] Subsequently, a locally configured optimizer (such as the Adam optimizer) can be invoked to control the parameter tuning process, thereby improving the tuning accuracy. The optimizer adaptively allocates tuning step sizes to different parameters based on information such as the gradient magnitude and historical tuning records for each parameter.

[0237] For example, for key parameters with large absolute gradient values ​​(i.e., parameters that significantly affect the loss value), the optimizer can assign a larger adjustment range to these parameters to improve the optimization speed of key parameters. For minor parameters with small absolute gradient values, a smaller adjustment step size is used to avoid over-adjustment that could damage the basic performance of the model.

[0238] During the parameter update process, the values ​​of each parameter before and after the adjustment, the gradient change curve, and other data can be recorded in real time in each round of adjustment, so that the optimizer can refer to these data to allocate the adjustment step size later.

[0239] This parameter tuning process can be iterated multiple times. For example, a validation dataset can be created for the second stage of training. After each round of parameter tuning, the validation dataset is loaded, and the dual-tower model with the currently tuned parameters is called to perform matching predictions between the target search term and the information object. This calculates the matching accuracy and loss value of the dual-tower model on the validation set, and the results are compared with those of the previous round of parameter tuning. If, after two consecutive rounds of tuning, the model's accuracy on the validation dataset does not significantly improve, or the first loss value has fallen below a preset threshold (e.g., below 0.1), the model performance can be considered stable, and the iteration process can be stopped. At this point, the pre-training phase is complete.

[0240] In this embodiment, the consistency comparison between the search intent word and the information carrier word in each information tuple can deeply explore the semantic relationship between the two, avoiding the limitations of relying solely on surface text matching. Furthermore, the consistency comparison results are used to further train the dual-tower model, ultimately enabling the trained model to possess stronger semantic understanding and association judgment capabilities, effectively improving the accuracy of matching user needs with information value in real-world search scenarios.

[0241] In one exemplary embodiment, such as Figure 12 As shown, the method also includes:

[0242] Step 1201: Obtain real-time search terms and matching information objects within the platform.

[0243] In practice, after obtaining the trained dual-tower model, it can be applied.

[0244] Specifically, the trained dual-tower model can support real-time search tasks. When the server receives a search request from the terminal, it can retrieve the real-time search terms from the search request and recall information objects (i.e., information objects to be matched) that are at least somewhat related to the real-time search terms from the database. The information objects to be matched can be obtained using any algorithm, such as: term frequency-inverse term frequency algorithm, Boolean matching algorithm, BM25 algorithm, etc.

[0245] Step 1202: Based on the trained dual-tower model, match the search terms and information objects to obtain the matching results.

[0246] The matching results are used to locate the target information object in response to the information search request.

[0247] In implementation, for real-time search terms and information objects to be matched, search intent terms and information carrier terms are generated sequentially according to a preset processing flow: for real-time search terms, the corresponding information query intent is combined in a structured manner to extract search intent terms; for each information object to be matched, its information provision intent is combined to extract information carrier terms. The process of generating search intent terms and information carrier terms has been described in the above embodiments and will not be repeated here.

[0248] Thus, after feature extraction is completed, the locally stored, trained dual-tower model is loaded. The search intent terms are input into the query tower module to generate search feature vectors, and each information carrier term is input into the document tower module to generate corresponding information feature vectors. Subsequently, the dual-tower model calculates the similarity between the search feature vector and each information feature vector, obtains a quantified value using a preset similarity algorithm (such as cosine similarity), and sorts all the information objects to be matched according to the platform's sorting rules (such as similarity values ​​from high to low).

[0249] After sorting, the server integrates the sorted information objects and their corresponding similarity scores into a matching result according to the response format requirements of the search request. This result may include a unique identifier for the information object, a summary of its core content, and a matching score. The matching result can then be displayed on the terminal to respond to the user's information search request.

[0250] In this embodiment, real-time search terms and information objects to be matched are obtained, and then a trained dual-tower model is used for matching. Because the dual-tower model avoids the limitations of traditional search relying on surface-level keyword matching, and can delve deeper into the underlying intent of the search terms and the core value of the information objects, it can significantly improve the accuracy of the matching results.

[0251] In one exemplary embodiment, such as Figure 13 As shown, the specific processing steps of step 602 include:

[0252] Step 1301: Obtain the target search term and the user behavior data associated with the target search term.

[0253] In implementation, a database storing the correspondence between search terms in each search request and the searched information objects is accessed, and search terms that meet the requirements are selected as target search terms. Then, user behavior data associated with these target search terms can be extracted from a database recording user behavior data. This data may include, but is not limited to: user click behavior after triggering a search, i.e., the identifier of the clicked information object, click time, dwell time (duration on a specific information object page), interaction behavior (collection, commenting, and sharing records), and search context behavior (browsing path before the search, subsequent search terms after the search), etc.

[0254] Step 1302: Based on the first intent extraction model, the target search term, and user behavior data, extract the information query intent corresponding to the target search term.

[0255] In implementation, refer to the relevant descriptions in the foregoing embodiments. Based on user behavior data, extract the target historical selected object corresponding to the target search term, and use the first intent extraction model in the foregoing embodiments to extract the information query intent based on the target search term and the target historical selected object.

[0256] Step 1303: Obtain the information object and the target historical search terms associated with the information object.

[0257] Step 1304: Based on the second intent extraction model, the information object, and the target historical search terms, extract the information providing intent corresponding to the information object.

[0258] In implementation, refer to the relevant descriptions in the foregoing embodiments to obtain information objects that are related to the target search terms, determine the target historical search terms corresponding to the information objects, and use the second intent extraction model in the foregoing embodiments to extract the information provision intent based on the information objects and the target historical search terms.

[0259] Step 1305: Construct a training dataset based on the target search terms, information query intent, information object, and information providing intent.

[0260] In implementation, the target search term is paired with one or more information objects that are related to it, forming one or more binary structures consisting of <target search term + information query intent, information object + information provision intent>. Then, a label is added to each binary tuple. The label is generated based on user behavior data. For example, binary tuples with multiple click interactions are labeled as positive samples, and binary tuples with no interaction and low semantic relevance are labeled as negative samples.

[0261] The selected binary structure can then be divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for model parameter learning, the validation set is used for parameter adjustment during the training process, and the test set is used for final performance evaluation.

[0262] In this embodiment, acquiring the target search term and associated user behavior data ensures the integrity of the search-side information. The introduction of user behavior data provides a behavioral basis for subsequent extraction of information query intent, avoiding semantic biases that may result from relying solely on the search term itself. Based on the first intent extraction model, combining the target search term and user behavior data to extract information query intent allows for in-depth exploration of the underlying needs behind user search behavior, transforming vague search terms into clear intent expressions, and making the representation of search-side needs more accurate. Acquiring the information object and historical associated search terms provides rich related information for information-side intent extraction. Historical associated search terms reflect the past association between the information object and user needs, laying the foundation for understanding the value of information. Using the second intent extraction model, combining the information object and historical associated search terms to extract information provision intent clearly defines the user needs that the information object can satisfy, making the information-side value positioning more explicit. Finally, the constructed training dataset integrates the target search term, information query intent, information object, and information provision intent, forming a structurally complete and semantically rich training sample. This ensures both the authenticity and relevance of the data and provides a high-quality training set for the learning of the dual-tower model, effectively improving the efficiency of model training and the final performance.

[0263] In one embodiment, such as Figure 14 As shown, a training device 1400 for a search matching model is provided, the training device 1400 for the search matching model includes:

[0264] The acquisition module 1401 is used to acquire the training dataset, which contains information tuples of target search terms and information objects; in each information tuple, the target search term corresponds to the information query intent, and the information object corresponds to the information providing intent.

[0265] The pre-training module 1402 is used to obtain the pre-trained dual-tower model based on each information tuple in the training dataset.

[0266] The fusion module 1403 is used to merge the target search term and the information query intent corresponding to the target search term into a search intent term, and to merge the information object and the information providing intent corresponding to the information object into an information carrier term;

[0267] Training module 1404 is used to fine-tune the pre-trained dual-tower model based on search intent words and information carrier words to obtain a trained dual-tower model; the trained dual-tower model is used for object retrieval.

[0268] In one embodiment, the pre-training module 1402 is specifically used for:

[0269] The training dataset is input into the dual-tower model, and the matching results of the target search terms and information objects in each information tuple in the training dataset are obtained through the dual-tower model.

[0270] Based on the matching results, calculate the first loss value of the dual-tower model;

[0271] Based on the first loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the pre-trained dual-tower model is obtained through iterative optimization.

[0272] In one embodiment, the dual-tower model includes a query tower module and a document tower module, and the pre-training module 1402 is specifically used for:

[0273] The query tower module analyzes and processes the target search terms and information query intent in each information tuple to obtain the search feature vector;

[0274] The document tower module analyzes and processes the information objects and information provision intentions in each information tuple to obtain information feature vectors.

[0275] Based on the search feature vector and the information feature vector, the matching results between each target search term and the information object are determined.

[0276] In one embodiment, the fusion module 1403 is specifically used for:

[0277] The first concatenation result is obtained by concatenating the target search term with the information query intent corresponding to the target search term.

[0278] Feature extraction is performed on the first concatenation result to obtain the search intent words;

[0279] The information object and the corresponding information provision intent are concatenated to obtain the second concatenation result.

[0280] Feature extraction is performed on the second concatenation result to obtain the information carrier words.

[0281] In one embodiment, the training module 1404 is specifically used for:

[0282] The search intent word and information carrier word of each information tuple in the training dataset are compared for consistency to obtain the comparison results.

[0283] Based on the comparison results, the second loss value of the dual-tower model is calculated;

[0284] Based on the second loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the trained dual-tower model is obtained through iterative optimization.

[0285] In one embodiment, the training device 1400 for the search matching model further includes:

[0286] The acquisition module is used to acquire real-time search terms and information objects to be matched within the platform.

[0287] The matching module is used to match search terms and information objects based on the trained dual-tower model to obtain the matching results of search terms and information objects;

[0288] The matching results are used to locate the target information object in response to information retrieval requests.

[0289] In one embodiment, the acquisition module 1401 is specifically used for:

[0290] Acquire target search terms and related user behavior data;

[0291] Based on the first intent extraction model, target search terms, and user behavior data, the information query intent corresponding to the target search terms is extracted.

[0292] Retrieve the information object and the associated target historical search terms;

[0293] Based on the second intent extraction model, the information object, and the target historical search terms, the information provision intent corresponding to the information object is extracted;

[0294] A training dataset is constructed based on the target search terms, the information query intent, the information object, and the information providing intent.

[0295] Each module in the above-mentioned device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0296] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 15 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an object search method or a training method for a search matching model.

[0297] Those skilled in the art will understand that Figure 15 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0298] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0299] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0300] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0301] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0302] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0303] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0304] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for training a search matching model, characterized in that, The method includes: Obtain a training dataset, which contains information tuples of target search terms and information objects; in each information tuple, the target search term corresponds to an information query intent, and the information object corresponds to an information providing intent; Based on each of the information pairs in the training dataset, a pre-trained dual-tower model is obtained; The target search term and the information query intent corresponding to the target search term are merged into a search intent term, and the information object and the information providing intent corresponding to the information object are merged into an information carrier term; The pre-trained dual-tower model is fine-tuned based on the search intent words and the information carrier words to obtain a trained dual-tower model; the trained dual-tower model is used for object retrieval.

2. The method according to claim 1, characterized in that, The process of obtaining the pre-trained dual-tower model based on each of the information tuples in the training dataset includes: The training dataset is input into the dual-tower model, and the matching results of the target search terms and information objects in each information tuple in the training dataset are obtained through the dual-tower model. Based on the matching results, the first loss value of the dual-tower model is calculated; Based on the first loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the pre-trained dual-tower model is obtained through iterative optimization.

3. The method according to claim 2, characterized in that, The dual-tower model includes a query tower module and a document tower module. The process of obtaining the matching results between the target search term and the information object in each information tuple of the training dataset through the dual-tower model includes: The query tower module analyzes and processes the target search terms and information query intent in each information tuple to obtain a search feature vector. The document tower module analyzes and processes the information objects and information provision intentions in each information tuple to obtain information feature vectors. Based on the search feature vector and the information feature vector, the matching result between each target search term and the information object is determined.

4. The method according to claim 1, characterized in that, The process of fusing the target search term and the corresponding information query intent into a search intent term, and fusing the information object and the corresponding information providing intent into an information carrier term, includes: The target search term and the information query intent corresponding to the target search term are concatenated to obtain the first concatenation result; Feature extraction is performed on the first concatenation result to obtain the search intent words; The information object and the information provision intent corresponding to the information object are concatenated to obtain a second concatenation result; Feature extraction is performed on the second splicing result to obtain the information carrier words.

5. The method according to claim 1, characterized in that, The step of fine-tuning the pre-trained dual-tower model based on the search intent words and the information carrier words to obtain the trained dual-tower model includes: A consistency comparison is performed between the search intent word and the information carrier word for each information tuple in the training dataset to obtain the comparison result. Based on the comparison results, the second loss value of the dual-tower model is calculated; Based on the second loss value, the parameters of the dual-tower model are adjusted through backpropagation algorithm, and the trained dual-tower model is obtained through iterative optimization.

6. The method according to claim 1, characterized in that, The method further includes: Obtain real-time search terms and matching information objects within the platform; The search terms and information objects are matched based on the trained dual-tower model to obtain the matching results of the search terms and information objects; The matching results are used to locate the target information object in response to information retrieval requests.

7. The method according to claim 1, characterized in that, The acquisition of the training dataset includes: Obtain the target search terms and the user behavior data associated with those target search terms; Based on the first intent extraction model, the target search term, and the user behavior data, the information query intent corresponding to the target search term is extracted. Obtain the information object and the target historical search terms associated with the information object; Based on the second intent extraction model, the information object, and the target historical search terms, the information providing intent corresponding to the information object is extracted; A training dataset is constructed based on the target search terms, the information query intent, the information object, and the information providing intent.

8. A training device for a search matching model, characterized in that, The device includes: The acquisition module is used to acquire a training dataset, which contains information tuples of target search terms and information objects; in each information tuple, the target search term corresponds to an information query intent, and the information object corresponds to an information providing intent. The pre-training module is used to obtain a pre-trained dual-tower model based on each of the information tuples in the training dataset. The fusion module is used to fuse the target search term with the information query intent corresponding to the target search term into a search intent term, and to fuse the information object with the information providing intent corresponding to the information object into an information carrier term. The training module is used to fine-tune the pre-trained dual-tower model based on the search intent words and the information carrier words to obtain a trained dual-tower model; the trained dual-tower model is used for object retrieval.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images