Recommended model training method

Abstract

Embodiments of the present specification provide a recommendation model training method, wherein the recommendation model training method comprises: constructing an initial recommendation model corresponding to a target business based on a pre-trained language model; reading search party information in historical business information corresponding to the target business, and determining recommended party information having a business association relationship with the search party information; constructing a sample group containing positive samples and negative samples based on the search party information and the recommended party information; processing the positive samples and the negative samples contained in the sample group by using the initial recommendation model, obtaining a positive score sequence corresponding to the positive samples, and a negative score sequence corresponding to the negative samples; calculating a loss value based on the positive score sequence and the negative score sequence, and adjusting the initial recommendation model based on the loss value until a target recommendation model satisfying a training stop condition is obtained. Subsequently, the recommendation task can be performed on the basis of the recommendation of the search party information, thereby improving the accuracy of the recommendation and improving the user satisfaction.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of computer technology, and in particular to a method for training a recommendation model. Background Technology

[0002] With the rapid development of the internet, information is emerging endlessly on various internet platforms, requiring users to use search engines to retrieve the content they need from this vast amount of information. Search engines can perform searches based on user input, match information that is highly relevant to the user's input, and provide feedback, thus offering information recommendation services to users.

[0003] In current technologies, search engines rely solely on the text content entered by the user. This results in a lack of semantic generalization within specific vertical domains and an inability to adequately consider other user information. Consequently, the recommended content is of poor quality and has low accuracy. Therefore, a more effective method is urgently needed to address these issues. Summary of the Invention

[0004] In view of the above, embodiments of this specification provide a method for training a recommendation model. One or more embodiments of this specification also relate to a recommendation model training apparatus, a recommendation model application method, a recommendation model application apparatus, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a method for training a recommendation model is provided, comprising:

[0006] Construct an initial recommendation model corresponding to the target business based on a pre-trained language model;

[0007] Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0008] Based on the searcher information and the recommended party information, a sample group containing positive and negative samples is constructed;

[0009] The initial recommendation model is used to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0010] The loss value is calculated based on the positive score sequence and the negative score sequence, and the parameters of the initial recommendation model are tuned based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0011] According to a second aspect of the embodiments of this specification, a recommendation model training apparatus is provided, comprising:

[0012] The model building module is configured to build an initial recommendation model corresponding to the target business based on a pre-trained language model;

[0013] The determination module is configured to read searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0014] The sample construction module is configured to construct a sample group containing positive and negative samples based on the searcher information and the recommended party information;

[0015] The processing module is configured to use the initial recommendation model to process the positive samples and the negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0016] The training module is configured to calculate a loss value based on the positive score sequence and the negative score sequence, and to tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0017] According to a third aspect of the embodiments of this specification, an information search method is provided, comprising:

[0018] Obtain the searcher information entered by the searcher on the information recommendation platform related to the target business;

[0019] Determine the recommended party information contained in the information recommendation platform;

[0020] The recommended party information and the search party information are input into the target recommendation model for processing to obtain the target recommended party information;

[0021] Search feedback information is generated based on the target recommendation information and then fed back to the searcher through the information recommendation platform.

[0022] According to a fourth aspect of the embodiments of this specification, an information search device is provided, comprising:

[0023] The acquisition module is configured to acquire the searcher information entered by the searcher on the information recommendation platform associated with the target business;

[0024] The determination module is configured to determine the recommended party information contained in the information recommendation platform;

[0025] The processing module is configured to input the recommended party information and the search party information into the target recommendation model for processing to obtain the target recommended party information;

[0026] The feedback module is configured to generate search feedback information based on the target recommended party information and send it back to the search party through the information recommendation platform.

[0027] According to a fifth aspect of the embodiments of this specification, a model training method is provided, applied to a cloud-side device, comprising:

[0028] In response to the model training request submitted by the terminal device, an initial recommendation model corresponding to the target service is constructed based on the pre-trained language model;

[0029] Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0030] Based on the searcher information and the recommended party information, a sample group containing positive and negative samples is constructed;

[0031] The initial recommendation model is used to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0032] The loss value is calculated based on the positive score sequence and the negative score sequence, and the parameters of the initial recommendation model are tuned based on the loss value until the target recommendation model that meets the training stopping condition is obtained.

[0033] The target recommendation model is fed back to the edge device.

[0034] According to a sixth aspect of the embodiments of this specification, a model training apparatus is provided, applied to a cloud-side device, comprising:

[0035] The build module is configured to respond to model training requests submitted by the edge device by building an initial recommendation model corresponding to the target business based on a pre-trained language model;

[0036] The determination module is configured to read searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0037] The construction module is configured to construct a sample group containing positive and negative samples based on the searcher information and the recommended party information;

[0038] The processing module is configured to use the initial recommendation model to process the positive samples and the negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0039] The training module is configured to calculate a loss value based on the positive score sequence and the negative score sequence, and to tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0040] The feedback module is configured to feed back the target recommendation model to the edge device.

[0041] According to a fifth aspect of the embodiments of this specification, a method for applying a recommendation model to a computing node is provided, comprising:

[0042] Upon receiving an information search request for the information to be searched, the information to be searched and the information to be recommended corresponding to the information to be searched are input into the target pre-trained language model to obtain the search vector corresponding to the information to be searched and the recommendation vector corresponding to the information to be recommended.

[0043] The vector to be searched is input into the search processing model to obtain the search intent vector;

[0044] Based on the recommended prediction model, the search intent vector and the recommendation vector are processed, and the feedback information corresponding to the information search request is determined according to the processing result.

[0045] The computing node deploys a target pre-trained language model, a search-side processing model, and a recommended-side prediction model, which are obtained by splitting the target recommendation model in the above method.

[0046] According to a sixth aspect of the embodiments of this specification, a recommendation model application apparatus is provided, applied to a computing node, comprising:

[0047] The receiving module is configured to, upon receiving an information search request submitted for the information to be searched, input the information to be searched and the information to be recommended corresponding to the information to be searched into the target pre-trained language model to obtain the search vector corresponding to the information to be searched and the recommendation vector corresponding to the information to be recommended.

[0048] The input module is configured to input the vector to be searched into the search processing model to obtain the intent vector to be searched.

[0049] The processing module is configured to process the search intent vector and the recommendation vector based on the recommended party prediction model, and determine the feedback information corresponding to the information search request based on the processing result; wherein, the computing node deploys a target pre-trained language model, a search party processing model, and a recommended party prediction model, which are obtained by splitting the target recommendation model in the above method.

[0050] According to a ninth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0051] Memory and processor;

[0052] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the above-described recommended model training method.

[0053] According to a tenth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described recommended model training method.

[0054] According to an eleventh aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described recommended model training method.

[0055] One embodiment of this specification constructs an initial recommendation model corresponding to a target business based on a pre-trained language model. It reads searcher information from historical business information corresponding to the target business and identifies recommended party information with business relevance to the searcher information. Based on the searcher and recommended party information, it constructs a sample group containing positive and negative samples. The initial recommendation model processes the positive and negative samples in the sample group to obtain positive score sequences for positive samples and negative score sequences for negative samples. It calculates loss values ​​based on the positive and negative score sequences and tunes the initial recommendation model based on the loss values ​​until a target recommendation model that meets the training stopping condition is obtained. The initial recommendation model corresponding to the target business is constructed based on the pre-trained language model, and then the sample group is constructed based on the historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporation of searcher information, thereby improving recommendation accuracy and user satisfaction. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of model training for a recommended model training method provided in one embodiment of this specification;

[0057] Figure 2 This is a flowchart illustrating a recommended model training method provided in one embodiment of this specification;

[0058] Figure 3 This is a model structure diagram in a recommended model training method provided in one embodiment of this specification;

[0059] Figure 4 This is a flowchart illustrating the processing steps of a recommendation model training method provided in one embodiment of this specification.

[0060] Figure 5 This is a flowchart illustrating a recommendation model training method provided in one embodiment of this specification.

[0061] Figure 6 This is a schematic diagram of the structure of a recommended model training device provided in one embodiment of this specification;

[0062] Figure 7 This is a flowchart illustrating an information search method provided in one embodiment of this specification;

[0063] Figure 8 This is a flowchart illustrating a recommendation model training method provided in one embodiment of this specification.

[0064] Figure 9 This is a schematic diagram of the structure of an information search device provided in one embodiment of this specification;

[0065] Figure 10 This is a flowchart illustrating a model training method provided in one embodiment of this specification;

[0066] Figure 11 This is a schematic diagram of the structure of a model training device provided in one embodiment of this specification;

[0067] Figure 12 This is a flowchart illustrating a recommendation model application method provided in one embodiment of this specification;

[0068] Figure 13 This is a schematic diagram of the structure of a recommended model application device provided in one embodiment of this specification;

[0069] Figure 14 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0070] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0071] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0072] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0073] Furthermore, 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, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0074] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0075] PTM: Pre-trained Model.

[0076] BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, is a pre-trained language model.

[0077] GPT: generative pre-trained transformer, a type of pre-trained language model.

[0078] Fine-tuning: In this context, it refers to fine-tuning the model, such as training some model parameters or training the model with a small amount of data and a low learning rate.

[0079] es: ElasticSearch, an open-source full-text search engine.

[0080] With the rapid development of the internet, information is emerging endlessly on various internet platforms, requiring users to use search engines to retrieve the content they need from this vast amount of information. In business opportunity searches, users' search intentions are diverse and personalized, potentially including searching for titles, materials, company names, project names, categories, etc.

[0081] Current business opportunity search retrieval technologies only focus on text-based search retrieval (such as Elasticsearch; based on publicly available information, some products also use simple vector retrieval, such as BERT which directly extracts vectors for retrieval). This leads to problems such as a lack of semantic generalization for vertical domains and an inability to effectively balance user intent and personalized needs. Consequently, the recommended content quality is poor, the recommendation accuracy is low, and user satisfaction is low.

[0082] Based on this, one embodiment of this specification provides a recommendation model training method that incorporates user attributes, intents, and other information when performing information searches based on user input, thereby improving the accuracy of information searches. Figure 1 This is a schematic diagram of a recommended model training method provided in one embodiment of this specification, as shown below. Figure 1 As shown, in the procurement and sales scenario, a pre-trained language model with information prediction capabilities is trained based on corpus information within the procurement and sales domain. A shared projection layer, dimensionality reduction layer, intent network, and feature network are determined. Based on the concatenation of the shared projection layer, dimensionality reduction layer, intent network, and feature network of the pre-trained language model, an initial recommendation model corresponding to the target business is constructed. Business logs corresponding to the target business are obtained. Based on the historical business information recorded in the business logs, the search party information (as the seller) is determined, and the recommended party information with business relevance to the search party information is also determined. A sample group containing positive and negative samples is constructed based on the search party information and the recommended party information. The positive and negative samples contained in the sample group are input into the initial recommendation model for processing. The positive and negative samples are vectorized through the pre-trained model, and then the output of the pre-trained model is subjected to dimensionality reduction and projection processing through the shared projection layer and dimensionality reduction layer; finally, the positive and negative samples are scored through the intent network and feature network to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample. Based on the preset loss function, the loss value is calculated for the positive score sequence and the negative score sequence, and the parameters of the initial recommendation model are tuned based on the loss value until the target recommendation model that meets the training stopping condition is obtained.

[0083] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample set is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporated searcher information, thereby improving recommendation accuracy and user satisfaction.

[0084] Furthermore, the recommendation model training method provided in this specification can also be applied to the insurance field, specifically for searching for available insurance for policyholders and for searching for policyholders for insurance service providers. That is, it recommends matching insurance products to policyholders and recommends potential policyholders to insurance service providers. The goal is to train a target recommendation model for matching and recommending between policyholders and insurance service providers.

[0085] The recommendation model training method provided in this manual can also be applied to the field of education, including the mutual selection or matching between tutors and students. This involves recommending students with high professional relevance who meet the student's search criteria to tutors, and recommending tutors with high professional relevance who meet the tutor's search criteria to students. The training aims to develop a target recommendation model for matching and recommending between tutors and students.

[0086] The recommendation model training method provided in this manual can also be applied to the logistics field, for mutual selection or matching between transportation task publishers and transportation task receivers. It recommends transportation task receivers with high professional relevance to the publishers and who meet the search criteria for transportation task receivers, and vice versa. The training is for a target recommendation model used for matching and recommending between transportation task publishers and transportation task receivers.

[0087] This specification provides a method for training a recommendation model, and also relates to a recommendation model training device, a recommendation model application method, a recommendation model application device, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0088] See Figure 2 , Figure 2 A flowchart of a recommended model training method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0089] Step S202: Construct an initial recommendation model corresponding to the target business based on the pre-trained language model.

[0090] Specifically, a pre-trained language model refers to a language model that has been trained on large-scale data in advance for a target domain, enabling the model to recognize the language of that domain. Target domains include, but are not limited to, finance, procurement and sales, education, insurance, transportation, and other fields. Correspondingly, the target business is the business within the target domain. In the case of the procurement and sales domain, the target business is the procurement and sales business; in the case of the transportation domain, the target business is the transportation task posting and receiving business. The initial recommendation model is a model with the ability to recommend information related to the target business.

[0091] Based on this, the business domain to which the target business belongs is defined as the target domain. Domain business information within the target domain is acquired, and the pre-trained model is trained based on this information. Dimensionality reduction and projection network layers, as well as intent and feature networks, are constructed corresponding to the target business. An initial recommendation model corresponding to the target business is built based on the pre-trained language model, dimensionality reduction and projection network layers, intent and feature networks. The initial recommendation model can contain multiple network layers; the intent network can be a network built based on a multilayer perceptron.

[0092] Furthermore, in existing technologies, when constructing recommendation models for performing recommendation tasks, they are typically trained based on a single existing model. However, this training method results in models with weak recommendation capabilities and low prediction accuracy. Therefore, when constructing the initial recommendation model, at least one network layer can be added to the pre-trained language model, along with networks capable of intent prediction and feature extraction. The specific implementation is as follows:

[0093] In response to the model building request associated with the target business, a dimensionality reduction network layer and a projection network layer, as well as an intent network and a feature network, are constructed; the dimensionality reduction network layer, the projection network layer, the intent network, the feature network, and the pre-trained language model are then concatenated according to preset concatenation rules to obtain an initial recommendation model.

[0094] Specifically, the model building request can be a computer instruction submitted by business personnel targeting the target business, used to build the initial recommendation model; the dimensionality reduction network layer is used to reduce the dimensionality of the feature vectors, thereby converting high-dimensional vectors into low-dimensional vectors for easier subsequent processing; the projection network layer is used to perform linear projection on the feature vectors to achieve embedding; the intent network can be built based on support vector machines and is used to determine the intent vectors; the feature network can be a feature gate network and is used to determine the feature vectors; the splicing rules can be set according to actual needs, and the splicing rules represent the connection relationship between the pre-trained language model, the dimensionality reduction network layer, the projection network layer, and the intent network and feature network.

[0095] Based on this, the system receives model building requests submitted by business personnel targeting the target business. These requests may include information such as network name and network type. Based on these requests, dimensionality reduction and projection network layers, as well as intent and feature networks, are constructed. Preset splicing rules are used to splice the various network layers and networks. Based on these rules, the system determines the positional information of the dimensionality reduction and projection network layers, as well as the intent and feature networks, relative to the pre-trained language model. Then, based on this positional information, the dimensionality reduction layer, projection network layer, intent network, feature network, and pre-trained language model are spliced ​​according to the preset splicing rules to obtain an initial recommendation model.

[0096] For example, after training a pre-trained language model using transportation domain corpora based on transportation business, a dimensionality reduction network layer is constructed to reduce the dimensionality of vectors, and a shared projection layer is constructed to project the vectors. An intent network is built based on a multilayer perceptron, and a feature network is built based on a gate network. The positional information of the dimensionality reduction network layer, the shared projection layer, the intent network, and the feature network is determined, and then these positional information are concatenated to obtain an initial recommendation model composed of the pre-trained language model, the dimensionality reduction network layer, the shared projection layer, the intent network, and the feature network. This model is then used to perform transportation recommendation tasks.

[0097] In summary, the dimensionality reduction network layer, the projection network layer, the intent network, the feature network, and the pre-trained language model are concatenated according to preset concatenation rules to obtain the initial recommendation model. This enriches the model structure of the initial recommendation model, improves its predictive ability, and achieves accurate prediction.

[0098] The training of the pre-trained language model includes: acquiring domain historical information associated with the target business; preprocessing the domain historical information and determining training sample pairs based on the processing results; inputting the training samples in the training sample pairs into the initial pre-trained language model to obtain model prediction information; calculating the pre-training loss value based on the model prediction information and the sample labels contained in the training sample pairs, and tuning the parameters of the initial pre-trained language model based on the pre-training loss value until a pre-trained language model that meets the pre-training training stopping condition is obtained.

[0099] Specifically, domain history information refers to the business corpus existing in the domain corresponding to the target business, including but not limited to proper nouns, domain keywords, business participant information, material information, etc. in the domain corresponding to the target business; preprocessing includes but is not limited to format adjustment, deduplication, information shuffling, random dynamic pruning, etc. of domain history information; training samples can be samples obtained by masking the preprocessed domain history information; the corresponding sample labels are the domain history information before masking.

[0100] Based on this, the target business within the target domain is identified, and the domain historical information associated with the target business is obtained. The domain historical information undergoes preprocessing such as format adjustment, deduplication, information fragmentation, and random dynamic pruning. Initial samples are determined based on the processing results, and these initial samples are masked to obtain training sample pairs. The training samples from the training sample pairs are input into the initial pre-trained language model to obtain model prediction information. The pre-training loss value is calculated based on the model prediction information and the sample labels contained in the training sample pairs. The initial pre-trained language model is then tuned based on the pre-training loss value to obtain an updated initial pre-trained language model. The next training sample pair is selected, and the training samples from the next training sample pair are input into the updated initial pre-trained model, and so on, until a pre-trained language model that meets the pre-training stopping condition is obtained.

[0101] In summary, by preprocessing and masking the historical information of the target business domain to obtain training samples and sample labels, and then training the initial pre-trained language model based on the training samples and sample labels, a pre-trained language model with information prediction capabilities is obtained, which facilitates the subsequent construction of the initial recommendation model and improves the prediction ability of the initial recommendation model.

[0102] Step S204: Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information.

[0103] Specifically, after constructing the initial recommendation model corresponding to the target business based on the pre-trained language model, the searcher information can be read from the historical business information corresponding to the target business, and the recommended party information with business relevance to the searcher information can be determined. Here, historical business information refers to the business information generated during the provision of business services by the target business. Historical business information can be used to record the behavioral information of all parties involved in the service when the target business provides services. That is to say, if the target business provides services to the searcher and the recommended party, then the information generated by the searcher and the recommended party when using the services provided by the target business is the target business information. The target business information includes, but is not limited to, the searcher behavior information and / or searcher attribute information of the searcher, and the recommended party behavior information and / or recommended party attribute information of the recommended party.

[0104] The searcher information includes the searcher's behavioral information and / or attribute information, while the recommended party information includes the recommended party's behavioral information and / or attribute information. Searcher information includes, but is not limited to, the searcher's address information, business content information, material information, browsing behavior information, historical cooperation information, and information on following and being followed. Correspondingly, recommended party information includes, but is not limited to, the recommended party's address information, business content information, material information, historical cooperation information, and browsing information. For example, in a transportation business, when the searcher is the transportation task publisher, the recommended party is the transportation task recipient. The searcher information includes, but is not limited to, the transportation task publisher's historical task publication information, transportation material information, transportation route information, and information on following and saving the transportation task recipient. The transportation task recipient information includes, but is not limited to, the transportation task recipient's transport route information, transportation material information, address information, and information on following and cooperating with the transportation task publisher. Business relationships refer to the cooperative relationship, cooperative tendency relationship, or behavioral relationship between the searcher and the recommended party.

[0105] Based on this, after constructing the initial recommendation model corresponding to the target business based on the pre-trained language model, the business log corresponding to the target business is read, historical business information is read from the business log corresponding to the target business, and the search party and the recommended party with business relationship with the search party information are determined from the historical business information. The search party information corresponding to the search party and the recommended party information corresponding to the recommended party are read from the historical business information.

[0106] Furthermore, in existing technologies, when acquiring information for constructing training samples, training samples are typically constructed based on the function selection and corresponding information of the model to be trained. For example, in a product recommendation model, training samples are usually constructed by selecting product information that the recommended user has interacted with, without considering other user information. This method of selecting sample information has certain limitations, as the information selection is not comprehensive enough. Therefore, in this embodiment, when selecting searcher information and recommended user information, as much information as possible can be collected based on business logs. The specific implementation is as follows:

[0107] Historical business information is obtained from the business logs corresponding to the target business, and searcher information containing searcher behavior information and / or searcher attribute information is read from the historical business information; recommended party behavior information and / or recommended party attribute information are determined from the historical business information as recommended party information that has a business relationship with the searcher information.

[0108] Specifically, business logs are used to record user information generated by all users when providing services for the target business. In a procurement and sales scenario, users include suppliers and purchasers; in an insurance scenario, users include insurance companies and policyholders. Searcher behavior information refers to the searcher's search behavior, browsing behavior, cooperation behavior, etc., related to the target business. Searcher attribute information refers to the searcher's address, name, business scope, materials, etc. Correspondingly, recommended party information refers to the recommended party's search behavior, browsing behavior, cooperation behavior, etc., related to the target business; recommended party attribute information refers to the recommended party's address, name, business scope, materials, etc.

[0109] Based on this, the business logs corresponding to the target business are obtained, historical business information within a preset time range is retrieved from the business logs, and the searcher business information corresponding to the searcher is determined from the historical business information. The searcher behavior information and / or searcher attribute information are read from the searcher business information, and the searcher information is composed of the searcher behavior information and / or searcher attribute information. The recommended party business information corresponding to the recommended party is determined from the historical business information, and the recommended party behavior information and / or recommended party attribute information are read from the recommended party business information. The recommended party behavior information and / or recommended party attribute information are used as the recommended party information that has a business relationship with the searcher information.

[0110] Continuing with the previous example, when the target business is transportation, the search party can be the transportation task publisher, and the recommended party is the transportation task recipient. The behavioral and attribute information of both the transportation task publisher and recipient is stored in the transportation business log. By reading the transportation business log, one can obtain publisher information such as the address, origin, transportation equipment, transportation materials, and browsing behavior of the transportation task publisher, as well as recipient information such as the address, number of transportation task executions, transportation materials, and transportation equipment of the transportation task recipient.

[0111] In summary, by obtaining searcher information (containing searcher behavior information and / or searcher attribute information) and recommender information (containing recommender behavior information and / or recommended party attribute information) from the business logs corresponding to the target business, comprehensive information acquisition can be achieved, thereby improving the richness of the samples when constructing training samples in the future.

[0112] Step S206: Construct a sample group containing positive and negative samples based on the searcher information and the recommended party information.

[0113] Specifically, after reading the searcher information from the historical business information corresponding to the target business and determining the recommended party information that has a business relationship with the searcher information, a sample group containing positive and negative samples can be constructed based on the searcher information and the recommended party information.

[0114] Positive samples refer to sample information constructed based on the searcher's information and the recommended party's information when identifying searchers and recommended parties with existing or potential cooperative relationships. Conversely, negative samples refer to sample information constructed based on the searcher's information and the recommended party's information when identifying searchers and recommended parties with no or potential cooperative relationships.

[0115] Based on this, the searcher information is read from the historical business information corresponding to the target business. After identifying the recommended party information that has a business relationship with the searcher information, positive sample information is extracted from the recommended party information based on the searcher information. Positive samples are then constructed based on the searcher information and the positive sample information. Negative sample information is determined by performing a first-stage sampling on the recommended party information. Negative samples are then constructed based on the negative sample information and the searcher information. Finally, a sample group is constructed based on the positive and negative samples.

[0116] Furthermore, in existing technologies, when constructing samples for model training, positive and negative samples are typically constructed separately. For example, a question and its corresponding correct answer are used as positive samples, and a question and its incorrect answer are used as negative samples, so that model training is based on both positive and negative samples. However, this method of sample construction and model training cannot combine positive and negative samples to tune the model parameters during a single training session. Therefore, training sample construction can be achieved by constructing sample groups containing both positive and negative samples, as detailed below:

[0117] Based on the searcher information, a first recommended party information is selected from the recommended party information, and positive intent information is extracted from the first recommended party information as positive sample information; a first-stage sampling is performed on the recommended party information, and negative sample information is determined according to the sampling results; a positive sample is composed of the searcher information and the positive sample information, and a negative sample is composed of the searcher information and the negative sample information; a sample group containing the positive sample and the negative sample is constructed.

[0118] Specifically, the first recommended party information refers to the information used to construct a positive sample. Correspondingly, the positive intent information in the first recommended party information refers to the recommended party information that has a cooperative tendency or has already reached a cooperative agreement with the search party. Whether there is a cooperative tendency or a cooperative agreement can be determined by the search party's actions towards the recommended party, such as clicking and browsing for a certain period of time, or completing a transaction between the search party and the recommended party. The first-stage sampling refers to sampling from the recommended party information, selecting the recommended party information as negative sampling information. The first-stage sampling can be random sampling from the recommended party information or sampling after determining the search party information. In this case, the recommended party information that is not related to the search party information is selected from the recommended party information. For example, if search party A is identified, recommended party B that does not have a cooperative relationship or intention with search party A is selected during sampling, and the information corresponding to recommended party B is used as the sampling result. Negative sample information is the attribute, behavior, and other information of the recommended party in the recommended party information. Negative samples are recommended party information that has not reached a cooperation agreement with the search party or does not have a cooperative tendency. Whether there is a cooperative tendency or a cooperation agreement can be determined by the search party's actions towards the recommended party, such as: clicking and immediately exiting, not clicking, etc.

[0119] Based on this, the system selects the first recommended party information that has a correlation with the searcher information from the recommended party information, and extracts the publicly available information of the recommended party from the first recommended party information as positive intent information. This positive intent information is then used as positive sample information. Public information refers to the information used to initiate the cooperation and recommend the information to the searcher; for example, if the recommended party is a transportation task recipient, then the transportation information (transportation capacity information, transportation price information, etc.) of the transportation task recipient is considered public information. A first-stage sampling is performed on the recommended party information, and negative sample information is determined based on the sampling results. Positive samples are composed of searcher information and positive sample information, and negative samples are composed of searcher information and negative sample information, thus constructing sample groups containing both positive and negative samples.

[0120] In summary, by constructing a sample set containing both positive and negative samples, and then using the sample set as input during model training, the initial recommendation model can process both positive and negative samples simultaneously, thereby improving the model's training efficiency.

[0121] Furthermore, in the first stage of sampling, incorrect answers to the question are usually directly used as negative samples. This sampling method is relatively simple. This embodiment provides two sampling methods, which are implemented as follows:

[0122] Based on the searcher information, a second recommended party information is selected from the recommended party information, and negative intent information is extracted from the second recommended party information as negative sample information, or random sampling is performed in the recommended party information, and negative sample information is determined according to the sampling results.

[0123] Specifically, the second recommender information refers to the information used to construct negative samples. Correspondingly, the negative intent information in the second recommender information is the recommender information that does not have a cooperative tendency with the search party or has not reached a cooperative agreement. Random sampling means randomly selecting recommenders from the recommender information and determining negative sample information based on the selected recommender information.

[0124] Based on this, a second recommended party information that is not related to the searcher information is selected from the recommended party information, and the public information of the recommended party is extracted from the second recommended party information as negative intent information, and the negative intent information is used as negative sample information; or random sampling is directly performed in the recommended party information, and negative sample information is determined based on the selected recommended party information.

[0125] In summary, negative sample information can be determined by random sampling or sampling negative samples, thereby improving the diversity of negative sample information determination.

[0126] Furthermore, in the first stage of sampling, incorrect answers to the question are usually directly used as negative samples, and the first stage sampling is performed using a single sampling method. This embodiment provides a method that combines two sampling methods to implement the first stage sampling, as detailed below:

[0127] Based on the searcher information and a preset ratio, the recommended information is randomly sampled to obtain third recommended information; a fourth recommended information is selected from the recommended information according to the preset ratio and sample type; negative intent information is extracted from the third recommended information as first negative sample information, and negative intent information is extracted from the fourth recommended information as second negative sample information; the first negative sample information and the second negative sample information are used as the negative sample information.

[0128] Specifically, the preset ratio refers to the sampling ratio pre-set according to actual needs. The sampling ratio represents the ratio between the number of samples obtained through random sampling and the number of samples obtained through the first stage of sampling. The sample type can be determined by the recommended party information. If the recommended party information includes operation information such as not being clicked by the search party or exiting immediately after being clicked by the search party, it indicates that the sample type is a negative sample. The fourth recommended party information refers to the information used to construct the negative sample, and the third recommended party is a recommended party that has not reached a cooperation agreement with the search party or has no tendency to cooperate with the search party. Correspondingly, the second negative sample is the actual negative sample. Since the first negative sample is obtained through random sampling, the first negative sample can be either an actual positive sample or an actual negative sample. When the first negative sample is an actual positive sample, it is defaulted to a negative sample.

[0129] Based on this, a preset ratio is established between random sampling and the first-stage sampling. Random sampling is performed on the recommended party information based on the searcher information and the preset ratio to obtain third recommended party information. Then, fourth recommended party information matching the sample type is selected from the recommended party information according to the preset ratio and sample type. Negative intent information is extracted from the third recommended party information as first negative sample information, and negative intent information is extracted from the fourth recommended party information as second negative sample information. The first and second negative sample information constitute the negative sample information obtained through sampling.

[0130] Following the previous example, we determine the party issuing the transportation task, all recipients of the transportation tasks included in the transportation business, and the sampling ratio of 3:1. We then conduct random sampling and first-stage sampling at a 3:1 ratio among the transportation business recipients. The sample information obtained through these two sampling methods is used as negative sample information to construct a negative sample.

[0131] In summary, the first negative sample information and the second negative sample information are determined based on a preset ratio, and the negative sample information is composed of the first negative sample information and the second negative sample information. By sampling in proportion using two sampling methods, the diversity of negative sample information is improved.

[0132] Step S208: Use the initial recommendation model to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0133] Specifically, after constructing a sample group containing positive and negative samples based on the searcher's information and the recommended party's information, the initial recommendation model can be used to process the positive and negative samples in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0134] Here, the positive score sequence refers to the score sequence obtained by scoring positive samples based on the initial recommendation model. A positive sample can contain at least one positive subsample, and the score of each positive subsample constitutes the positive score sequence of the positive sample. Positive subsamples can be sample information determined based on information about recommended parties with a cooperative relationship or intention to cooperate with the searcher. For example, if the searcher is a transportation task publisher, then recommended parties with a cooperative relationship or intention to cooperate with the transportation task publisher can be selected as positive subsamples. The positive score sequence represents the probability that the recommended party corresponding to a positive subsample and the searcher information corresponding to the positive sample will cooperate. Correspondingly, the negative score sequence... A column refers to a scoring sequence obtained by scoring negative samples based on the initial recommendation model. A negative sample may contain at least one negative subsample, and the score of each negative subsample constitutes the negative score sequence of the negative sample. Positive subsamples can be sample information determined based on the information of the recommended party that does not have a cooperative relationship or intention to cooperate with the search party. For example, if the search party is a transportation task publisher, then the recommended party that does not have a cooperative relationship or intention to cooperate with the transportation task publisher, or whose cooperative intention is unclear, can be selected as the target of negative subsamples. The negative score sequence represents the probability that the recommended party corresponding to the negative subsample and the search party information corresponding to the negative sample will cooperate.

[0135] Based on this, after constructing a sample group containing positive and negative samples based on the searcher's information and the recommended party's information, the positive samples in the sample group are processed using the initial recommendation model to obtain the positive score sequence corresponding to the positive samples. At the same time, the negative samples in the sample group are processed using the initial recommendation model to obtain the negative score sequence corresponding to the negative samples.

[0136] Furthermore, in existing technologies, during model training, when scoring positive and negative samples, since positive and negative samples are input into the model as two separate samples for processing, scoring is also performed separately for each sample. However, this approach only processes one sample per training cycle, resulting in low processing efficiency. Therefore, processing efficiency can be improved by processing positive and negative samples simultaneously during a single model training cycle, as detailed below:

[0137] Using the pre-trained language model included in the initial recommendation model, semantic recognition is performed on the searcher information, positive sample information, and negative sample information contained in the sample group to obtain the searcher semantic vector corresponding to the searcher information, the positive sample vector corresponding to the positive sample information, and the negative sample vector corresponding to the negative sample information. The network layer included in the initial recommendation model is then used to process the searcher semantic vector, the positive sample vector, and the negative sample vector to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample.

[0138] Specifically, the search party semantic vector refers to the search party information in vector form output by the pre-trained language model after the search party information is input into the pre-trained language model for processing; the positive sample vector refers to the positive sample information in vector form output by the pre-trained language model after the positive sample information is input into the pre-trained language model for processing; the negative sample vector refers to the negative sample information in vector form output by the pre-trained language model after the negative sample information is input into the pre-trained language model for processing; the network layer is used to realize the reprocessing of search direction vectors, positive sample training and negative sample vectors, realize the linear projection and dimensionality reduction of vectors, and realize feature fusion.

[0139] Based on this, the searcher information, positive sample information, and negative sample information contained in the sample group are determined. These information are then input into a pre-trained language model. The pre-trained language model included in the initial recommendation model is used to perform semantic recognition on the searcher information, positive sample information, and negative sample information, respectively. The information is then vectorized to obtain the searcher semantic vector corresponding to the searcher information, the positive sample vector corresponding to the positive sample information, and the negative sample vector corresponding to the negative sample information. The network layers included in the initial recommendation model are determined. These network layers are used to process the searcher semantic vector, positive sample vector, and negative sample vector. The searcher semantic vector is integrated into the positive sample vector to obtain the positive score sequence corresponding to the positive sample, and the searcher semantic vector is integrated into the negative sample vector to obtain the negative score sequence corresponding to the negative sample.

[0140] In summary, the model structure based on the initial recommendation model enables the simultaneous processing of positive and negative samples during a single model training iteration, thereby improving the training efficiency of the initial recommendation model.

[0141] Furthermore, in existing technologies, recommendation models are typically built based on existing models, using a single model for construction. Even if methods exist for building recommendation models based on two or more models, they simply use the output of one model as the input of the next, failing to achieve model fusion and collaboration. Therefore, model concatenation can be implemented based on model functionality to achieve collaborative training, as detailed below:

[0142] The search party semantic vector, the positive sample vector, and the negative sample vector are projected and dimensionality reduced using the first network layer in the initial recommendation model to obtain the search party embedding vector corresponding to the search party semantic vector, the positive sample embedding vector corresponding to the positive sample vector, and the negative sample embedding vector corresponding to the negative sample vector. The search party embedding vector and the positive sample embedding vector are then input into the second network layer in the initial recommendation model for scoring to obtain the positive score sequence corresponding to the positive sample. The search party embedding vector and the negative sample embedding vector are then input into the second network layer in the initial recommendation model for scoring to obtain the negative score sequence corresponding to the negative sample.

[0143] Specifically, the first network layer may include a shared projection network layer and a dimensionality reduction network layer, used to implement vector projection and dimensionality reduction respectively; the second network layer includes an intent layer and a feature layer, used to score positive and negative samples.

[0144] Based on this, the first network layer in the initial recommendation model is determined. This first network layer is used to project and reduce the dimensionality of the search semantic vector to obtain the search embedding vector corresponding to the search semantic vector. Similarly, the first network layer is used to project and reduce the dimensionality of positive sample vectors to obtain the positive sample embedding vectors corresponding to the positive sample vectors. Likewise, the first network layer is used to project and reduce the dimensionality of negative sample vectors to obtain the negative sample embedding vectors corresponding to the negative sample vectors. The second network layer in the initial recommendation model is then determined. The search embedding vector and the positive sample embedding vector are input into the second network layer for scoring to obtain the positive score sequence corresponding to the positive samples. Finally, the search embedding vector and the negative sample embedding vector are input into the second network layer in the initial recommendation model for scoring to obtain the negative score sequence corresponding to the negative samples.

[0145] In summary, the network layer consists of a first network layer and a second network layer. The first network layer is used to project and reduce the dimensions of the search semantic vector, positive sample vector, and negative sample vector, respectively. The second network layer further realizes the fusion between the search embedding vector and the negative sample embedding vector, as well as between the search embedding vector and the positive sample embedding vector, so as to score the positive and negative samples based on the vector fusion.

[0146] Furthermore, in existing technologies, when processing positive and negative samples using models, processing is typically done separately based on models with specific capabilities. However, this approach limits model training to only positive and negative samples, failing to incorporate other information. Therefore, by planning and designing the model's architecture, it is possible to incorporate other information besides positive and negative samples during model training. The specific implementation is as follows:

[0147] The initial recommendation model includes a second network layer; the intent layer in the second network layer is used to identify the intent of the searcher embedding vector, and the searcher intent vector is determined based on the identification result; the searcher intent vector and the positive sample embedding vector are input to the feature layer in the second network layer for scoring to obtain the positive score sequence corresponding to the positive sample; the searcher intent vector and the negative sample embedding vector are input to the feature layer in the second network layer for scoring to obtain the negative score sequence corresponding to the negative sample.

[0148] Specifically, the intent layer is used to recognize the intent of the input information; the feature layer is used to combine with the intent layer to achieve feature fusion and scoring of positive and negative samples.

[0149] Based on this, the second network layer included in the initial recommendation model is determined, along with the intent layer and feature layer contained within it. The intent layer within the second network layer is used to perform intent recognition on the search party embedding vector containing user information and user intent, and the search party intent vector is determined based on the recognition results. The search party intent vector and positive sample embedding vector are input into the feature layer within the second network layer for scoring; the search party intent vector is incorporated into the positive sample embedding vector to obtain a positive score sequence corresponding to positive samples. Similarly, the search party intent vector and negative sample embedding vector are input into the feature layer within the second network layer for scoring; the search party intent vector is incorporated into the negative sample embedding vector to obtain a negative score sequence corresponding to negative samples.

[0150] Following the previous example, Figure 3 As shown, Figure 3 In the diagram (a), the network corresponding to the intent layer is represented. The intent layer can be a network with a fully connected layer built on a multilayer perceptron and is used to realize intent recognition. Figure 3 In the diagram (b), the network corresponding to the feature layer is shown. The feature layer can be a network built based on a gate network, used to achieve intent fusion and to score positive and negative samples.

[0151] In summary, by performing intent recognition on the searcher's embedding vector through the intent layer, the searcher's intent vector is obtained. This searcher's intent vector is then fused with the positive and negative sample embedding vectors, respectively. This allows the searcher's information to be incorporated into the initial recommendation model trained based on positive and negative samples, thereby improving the prediction ability and information fusion ability of the trained target recommendation model.

[0152] Step S210: Calculate the loss value based on the positive score sequence and the negative score sequence, and tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0153] Specifically, after processing the positive and negative samples in the sample group using the initial recommendation model to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample, the loss value can be calculated based on the positive and negative score sequences. The parameters of the initial recommendation model can then be tuned based on the loss value until a target recommendation model that meets the training stopping condition is obtained. The loss value can be a function value calculated based on a preset loss function, the positive score sequence, and the negative score sequence. The training stopping condition can be that the number of training iterations reaches a preset threshold, the preset training time is reached, or the model is tested on a test set to determine that it has reached a preset prediction accuracy.

[0154] Based on this, after processing the positive and negative samples in the sample group using the initial recommendation model to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample, a loss function for model training is determined. The positive and negative score sequences are input into the loss function to calculate the loss value corresponding to the sample group, and the initial recommendation model is updated by tuning the parameters based on the loss value. Then, the next sample group is obtained, and the positive and negative samples in the next sample group are input into the updated initial recommendation model, and so on, until the target recommendation model that meets the training stopping condition is obtained.

[0155] In practical applications, since the pre-trained language model in the initial recommendation model is based on domain corpora, during the overall training of the initial recommendation model, some layers of the pre-trained language model can be fine-tuned based on the loss values ​​obtained by inputting sample groups into the initial recommendation model. This includes fine-tuning the dimensionality reduction and projection network layers included in the initial recommendation model. This achieves the goal of parameter tuning of the initial recommendation model, enabling a comprehensive adjustment of the initial recommendation model. After subsequent training, the initial recommendation model will be more aligned with the business scenario and provide better recommendation services.

[0156] Furthermore, in existing technologies, when positive or negative samples are input into the model for processing, they are usually processed sequentially. Only after one positive or negative sample is processed will the next positive or negative sample be input for processing. However, this processing method is inefficient and cannot achieve simultaneous processing and comparison of positive and negative samples. Therefore, in this embodiment, the positive and negative samples contained in the sample group can be input into the initial recommendation model for processing simultaneously, and there is at least one negative sample in a sample group. The specific implementation is as follows:

[0157] When there are at least two sample groups, and each sample group contains at least one negative sample, the step of calculating the loss value based on the positive score sequence and the negative score sequence includes: determining at least one negative sample contained in each of the at least two sample groups, and determining the negative score sequence corresponding to the at least one negative sample; selecting target negative score sequences corresponding to each of the at least two sample groups from the negative score sequences corresponding to each of the at least two sample groups; calculating the initial loss value corresponding to each of the at least two sample groups based on the positive score sequence and the target negative score sequence corresponding to each of the at least two sample groups; and summing the initial loss values ​​corresponding to each of the at least two sample groups to obtain the loss value.

[0158] Specifically, when training the initial recommendation model, it can be used to process one sample group or to process multiple sample groups simultaneously. The target negative score sequence refers to the highest score sequence among the negative score sequences corresponding to each sample group. The initial loss value refers to the loss value corresponding to each sample group when at least two sample groups are processed simultaneously using the initial recommendation model. The loss value corresponding to the model input consisting of at least two sample groups can be calculated using the following formula (1).

[0159] Loss=sum(max(0,margin-(posscore-max(negscore))) (1)

[0160] Where, posscore represents the sequence of positive scores for positive samples in a sample group; negscore represents the sequence of negative scores for negative samples in a sample group; and the parameters in the margin loss function are constants that can be set as needed.

[0161] Based on this, given at least two sample groups, and each sample group contains at least one negative sample, the at least one negative sample in each of the at least two sample groups is determined. The negative score sequence corresponding to the at least one negative sample in each sample group is determined. Among the negative score sequences corresponding to each sample group, the sequence with the largest negative score is selected as the target negative score sequence for each of the at least two sample groups. Based on the positive score sequences and target negative score sequences corresponding to the at least two sample groups, the initial loss value for each sample group is calculated. The initial loss values ​​corresponding to the at least two sample groups are summed to obtain the total loss value.

[0162] Following the previous example, if there are three sample groups input into the initial recommendation model at the same time, and each sample group contains one positive sample and four negative samples, calculate the scores of the positive and negative samples in each sample group based on the initial recommendation model, and calculate the loss value based on the above formula (1).

[0163] In summary, by processing at least two sample groups simultaneously based on the initial recommendation model, with each sample group containing at least one negative sample, the simultaneous processing and comparison of positive and negative samples can be achieved, thereby improving the training efficiency of the initial recommendation model and obtaining better training results.

[0164] The following is in conjunction with the appendix Figure 4 Taking the recommendation model training method provided in this specification as an example in the application of recommendation model training, the recommendation model training method will be further explained. Among them, Figure 4 The flowchart of a recommendation model training method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0165] Step S402: In response to the model building request related to procurement and sales business, construct the dimensionality reduction network layer and the projection network layer, as well as the intent network and the feature network.

[0166] Step S404: The dimensionality reduction network layer, projection network layer, intent network, and feature network are concatenated with the pre-trained language model according to the preset concatenation rules to obtain the initial recommendation model.

[0167] Step S406: Obtain historical business information from the business log corresponding to the procurement and sales business, and read the product supplier information containing supplier behavior information and supplier attribute information from the historical business information.

[0168] Step S408: Determine the product demander behavior information and product demander attribute information from the historical business information, as product demander information that has a business relationship with the product supplier information.

[0169] Step S410: Based on the product supplier information, select the first product demander information from the product demander information, and extract positive intent information from the first product demander information as positive sample information.

[0170] Step S412: Randomly sample the product demander information based on the product supplier information and the preset sampling ratio to obtain the second product demander information.

[0171] Step S414: Select the third-party product demand information from the product demand information according to the preset sampling ratio and sample type to complete the first stage of sampling.

[0172] Step S416: Extract negative intent information from the second commodity demand information as the first negative sample information, and extract negative intent information from the third commodity demand information as the second negative sample information.

[0173] Step S418: Use the first negative sample information and the second negative sample information as negative sample information.

[0174] Step S420: A positive sample is formed by combining product supplier information and positive sample information, and a negative sample is formed by combining product supplier information and negative sample information, and a sample group containing positive and negative samples is constructed.

[0175] Step S422: Use the pre-trained language model included in the initial recommendation model to perform semantic recognition on the product supplier information, positive sample information and negative sample information contained in the sample group, respectively, to obtain the supplier semantic vector corresponding to the product supplier information, the positive sample vector corresponding to the positive sample information, and the negative sample vector corresponding to the negative sample information.

[0176] Step S424: Using the first network layer contained in the network layer of the initial recommendation model, project and reduce the supplier semantic vector, positive sample vector and negative sample vector respectively to obtain the supplier embedding vector corresponding to the supplier semantic vector, the positive sample embedding vector corresponding to the positive sample vector, and the negative sample embedding vector corresponding to the negative sample vector.

[0177] Step S426: Determine the second network layer contained in the network layer of the initial recommendation model, use the intent layer contained in the second network layer to perform intent recognition on the supplier embedding vector, and determine the supplier intent vector based on the recognition result.

[0178] Step S428: Input the supplier intent vector and the positive sample embedding vector into the feature layer contained in the second network layer for scoring, and obtain the positive score sequence corresponding to the positive sample.

[0179] Step S430: Input the supplier intent vector and the negative sample embedding vector into the feature layer contained in the second network layer for scoring, and obtain the negative score sequence corresponding to the negative sample.

[0180] Step S432: Calculate the loss value based on the positive score sequence and the negative score sequence, and tune the parameters of the initial recommendation model based on the loss value until the target recommendation model that meets the training stopping condition is obtained.

[0181] In practical applications, such as Figure 5As shown, the initial recommendation model can be composed of a pre-trained language model, a dimensionality reduction layer, a linear projection layer, a multilayer perceptron, and a gate network. When training the initial recommendation model in the procurement and sales field, supplier information is obtained, positive samples and negative samples are determined, and the input sample group of the initial recommendation model is composed of supplier information, positive samples and negative samples. Supplier information can correspond to information such as historical region, historical material, query, discrete features, etc. Positive samples and negative samples can correspond to information such as material, region, category, etc. Supplier information, positive samples and negative samples are input into the pre-trained language model for embedding processing, and then the processing results are subjected to dimensionality reduction and linear projection. The supplier information is processed using a feature-based multilayer perceptron to obtain user vector and intent vector; the feature vectors corresponding to positive samples and negative samples are processed, and the user vector and intent vector are fused to obtain potential intent, as well as the scores of positive samples and negative samples. The loss value is calculated based on the preset loss function (the above formula (1)), and then the parameters of the initial recommendation model are tuned based on the loss value, and so on until the target recommendation model that meets the training stopping condition is obtained.

[0182] This embodiment provides a personalized search retrieval method based on user intent recognition. The target recommendation model adopts a two-stage training approach. The first stage uses vertical domain corpora for self-supervised training, enabling the pre-trained language model to adapt to the vertical domain corpora and better learn and generalize to specific domain knowledge and professional terminology. The second stage uses user behavior information from the procurement and sales application to train some layers of the pre-trained model and a new network, achieving the fusion of semantic and personalized information. A recommendation model combining personalization and search intent is designed to meet search business needs.

[0183] In existing technologies, personalized information is typically represented by directly modeling a user's historical behavioral information. However, each user's personalized needs are different. Therefore, a gate network for user personalization information is established to determine the degree of personalization to be incorporated. Furthermore, since user searches have varying intents—in specific scenarios, they might be searching for information such as merchants, categories, titles, materials, or descriptions—this embodiment categorizes different user information into different search intents, establishing an intent network that directly models search intents. This achieves personalized search while effectively capturing user intent. Therefore, it enables accurate personalized search retrieval without explicitly building an intent understanding service, thereby improving the matching accuracy of search services.

[0184] In summary, an initial recommendation model corresponding to procurement and sales operations is constructed based on a pre-trained language model. Then, a sample set is built based on historical information from the procurement and sales operations to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as product supplier information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporation of product supplier information, thereby improving recommendation accuracy and user satisfaction.

[0185] Corresponding to the above method embodiments, this specification also provides recommended model training apparatus embodiments. Figure 6 A schematic diagram of a recommended model training apparatus provided in one embodiment of this specification is shown. Figure 6 As shown, the device includes:

[0186] The model building module 602 is configured to build an initial recommendation model corresponding to the target business based on a pre-trained language model.

[0187] The determination module 604 is configured to read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0188] The sample construction module 606 is configured to construct a sample group containing positive and negative samples based on the searcher information and the recommended party information;

[0189] The processing module 608 is configured to use the initial recommendation model to process the positive samples and the negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0190] The training module 610 is configured to calculate a loss value based on the positive score sequence and the negative score sequence, and to tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0191] In an optional embodiment, the sample construction module 606 is further configured to:

[0192] Based on the searcher information, a first recommended party information is selected from the recommended party information, and positive intent information is extracted from the first recommended party information as positive sample information.

[0193] The first phase of sampling is performed on the recommended party information, and negative sample information is determined based on the sampling results;

[0194] Positive samples are composed of the searcher information and the positive sample information, and negative samples are composed of the searcher information and the negative sample information.

[0195] Construct a sample group containing the positive sample and the negative sample.

[0196] In an optional embodiment, the sample construction module 606 is further configured to:

[0197] Based on the searcher information, a second recommended party is selected from the recommended party information, and negative intent information is extracted from the second recommended party information as negative sample information, or

[0198] Random sampling is performed on the recommended party information, and negative sample information is determined based on the sampling results.

[0199] In an optional embodiment, the sample construction module 606 is further configured to:

[0200] Based on the searcher information and a preset ratio, the recommended party information is randomly sampled to obtain third recommended party information;

[0201] According to the preset ratio and sample type, select the fourth recommended party information from the recommended party information;

[0202] Negative intent information is extracted from the third recommended party information as the first negative sample information, and negative intent information is extracted from the fourth recommended party information as the second negative sample information.

[0203] The first negative sample information and the second negative sample information are used as the negative sample information.

[0204] In an optional embodiment, the model building module 602 is further configured to:

[0205] In response to the model building request associated with the target business, a dimensionality reduction network layer and a projection network layer, as well as an intent network and a feature network, are constructed.

[0206] The dimensionality reduction network layer, the projection network layer, the intent network, the feature network, and the pre-trained language model are concatenated according to a preset concatenation rule to obtain an initial recommendation model.

[0207] In an optional embodiment, the determining module 604 is further configured to:

[0208] Obtain historical business information from the business logs corresponding to the target business, and read searcher information containing searcher behavior information and / or searcher attribute information from the historical business information;

[0209] The recommended party's behavioral information and / or attribute information are determined from the historical business information and used as recommended party information that has a business relationship with the search party information.

[0210] In an optional embodiment, the sample processing module 608 is further configured to:

[0211] The search semantic vector, the positive sample vector, and the negative sample vector are projected and dimensionality reduced using the first network layer in the network layer of the initial recommendation model to obtain the search embedding vector corresponding to the search semantic vector, the positive sample embedding vector corresponding to the positive sample vector, and the negative sample embedding vector corresponding to the negative sample vector.

[0212] The search embedding vector and the positive sample embedding vector are input into the second network layer contained in the network layer of the initial recommendation model for scoring, so as to obtain the positive score sequence corresponding to the positive sample;

[0213] The search embedding vector and the negative sample embedding vector are input into the second network layer contained in the network layer of the initial recommendation model for scoring, thereby obtaining the negative score sequence corresponding to the negative sample.

[0214] In an optional embodiment, the processing module 608 is further configured to:

[0215] Determine the second network layer included in the network layers of the initial recommendation model;

[0216] The intent layer included in the second network layer is used to perform intent recognition on the search party embedding vector, and the search party intent vector is determined based on the recognition result;

[0217] The search intent vector and the positive sample embedding vector are input into the feature layer contained in the second network layer for scoring, and the positive score sequence corresponding to the positive sample is obtained.

[0218] The search intent vector and the negative sample embedding vector are input into the feature layer contained in the second network layer for scoring, thereby obtaining the negative score sequence corresponding to the negative sample.

[0219] In an optional embodiment, the training module 610 is further configured to:

[0220] Identify at least one negative sample in each of at least two sample groups, and determine the negative score sequence corresponding to at least one negative sample;

[0221] Select the target negative score sequence corresponding to at least two sample groups from the negative score sequences corresponding to at least two sample groups respectively.

[0222] Calculate the initial loss value for each of the at least two sample groups based on the positive score sequence and the target negative score sequence.

[0223] The initial loss values ​​corresponding to at least two sample groups are summed to obtain the loss value.

[0224] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample set is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporated searcher information, thereby improving recommendation accuracy and user satisfaction.

[0225] The above is an illustrative scheme of a recommendation model training device according to this embodiment. It should be noted that the technical solution of this recommendation model training device and the technical solution of the recommendation model training method described above belong to the same concept. For details not described in detail in the technical solution of the recommendation model training device, please refer to the description of the technical solution of the recommendation model training method described above.

[0226] See Figure 7 , Figure 7 A flowchart of an information search method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0227] Step S702: Obtain the searcher information entered by the searcher on the information recommendation platform related to the target business;

[0228] Step S704: Determine the recommended party information contained in the information recommendation platform;

[0229] Step S706: Input the recommended party information and the search party information into the target recommendation model for processing to obtain the target recommended party information;

[0230] Step S708: Generate search feedback information based on the target recommended party information, and feed it back to the search party through the information recommendation platform.

[0231] In practical applications, such as Figure 8As shown, after receiving the search information input by the searcher, the query field and user behavior information contained in the search information are processed based on the independent first part (independent part 1); the business opportunity release information, which serves as the recommended party information, is processed based on the independent third part (independent part 3). At this time, the independent second part (independent part 2) processes the search information and the business opportunity release information, which serves as the recommended party information, to obtain the business opportunity vector corresponding to the business opportunity release information, as well as the user vector and user intent corresponding to the query field and user behavior information. The target recommendation model (model) corresponds independently to the pre-trained model in the independent second part. The business opportunity vector is stored in a vector library. When the user vector and intent vector are retrieved from the vector library based on the vector retrieval engine, the multi-path vectors in the vector library are subjected to multi-path mixing operation. The vectors are sorted according to the matching degree between the business opportunity vector and the user vector and intent vector, and the target vector is selected and fed back to the searcher.

[0232] Furthermore, information search can be a service provided by a service provider. Users with information search needs (searchers) can purchase and use the information search service provided by the service provider. The server implements information search through a target recommendation model. After the service provider sends feedback information to the user through the information recommendation platform, the user can determine the search results based on the feedback information. If the user is dissatisfied with or does not agree with the search results, they can provide feedback on their satisfaction to the recommendation platform, which will then inform the service provider to further fine-tune the target recommendation model, thereby optimizing the target recommendation model.

[0233] In summary, by inputting the information of the recommended party and the searcher into the target recommendation model for processing, the target recommended party information is obtained, thereby improving the accuracy of identifying the target recommended party information and enhancing the information search capabilities of the recommendation platform. This provides users with a better search experience.

[0234] Corresponding to the above method embodiments, this specification also provides embodiments of an information search device. Figure 9 A schematic diagram of the structure of an information search device according to one embodiment of this specification is shown. Figure 9 As shown, the device includes:

[0235] The acquisition module 902 is configured to acquire the searcher information entered by the searcher in the information recommendation platform associated with the target business;

[0236] The determination module 904 is configured to determine the recommended party information contained in the information recommendation platform;

[0237] Processing module 906 is configured to input the recommended party information and the search party information into the target recommendation model for processing to obtain target recommended party information;

[0238] Feedback module 908 is configured to generate search feedback information based on the target recommended party information and feed it back to the search party through the information recommendation platform.

[0239] In summary, by inputting the information of the recommended party and the searcher into the target recommendation model for processing, the target recommended party information is obtained, thereby improving the accuracy of identifying the target recommended party information and enhancing the information search capabilities of the recommendation platform. This provides users with a better search experience.

[0240] The above is an illustrative scheme of an information search device according to this embodiment. It should be noted that the technical solution of this information search device and the technical solution of the information search method described above belong to the same concept. For details not described in detail in the technical solution of the information search device, please refer to the description of the technical solution of the information search method described above.

[0241] See Figure 10 , Figure 10 A flowchart of a model training method according to an embodiment of this specification is shown. This model training method is applied to a cloud-side device and specifically includes the following steps.

[0242] Step S1002: In response to the model training request submitted by the terminal device, construct the initial recommendation model corresponding to the target service based on the pre-trained language model;

[0243] Step S1004: Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0244] Step S1006: Construct a sample group containing positive and negative samples based on the searcher information and the recommended party information;

[0245] Step S1008: Use the initial recommendation model to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0246] Step S1010: Calculate the loss value based on the positive score sequence and the negative score sequence, and tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained;

[0247] Step S1012: Feed back the target recommendation model to the edge device.

[0248] In practical applications, cloud-side devices can provide model training services, receiving model training requests from edge devices and training the models provided by the edge devices. In response to the model training requests from the edge devices, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Historical business information corresponding to the target task carried in the model training request is determined. Searcher information is read from the historical business information, and recommended party information with business relevance to the searcher information is identified based on this historical business information. A sample group containing positive and negative samples is constructed based on the searcher and recommended party information. The initial recommendation model is used to process the positive and negative samples in the sample group, obtaining positive score sequences for positive samples and negative score sequences for negative samples. Based on a preset loss function, loss values ​​are calculated for the positive and negative score sequences, and the initial recommendation model is tuned based on the loss values. This process continues until a target recommendation model that meets the training stopping condition is obtained, and the target recommendation model is then fed back to the edge device.

[0249] Furthermore, the edge devices can be operated by users who require model training. Users provide samples for model training, along with their model requirements. These samples are then given to the cloud-based device for training. Once the cloud-based device has trained the target recommendation model, it sends the model back to the user, who then applies it on the edge device. The cloud-based device provides model building and training services to meet the needs of different users.

[0250] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample group is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporation of searcher information, thereby improving recommendation accuracy and user satisfaction. Training the initial recommendation model on the cloud-side device for the edge device reduces the resource consumption of the edge device.

[0251] Corresponding to the above method embodiments, this specification also provides embodiments of a model training device. Figure 11 A schematic diagram of a model training apparatus provided in one embodiment of this specification is shown. Figure 11 As shown, this device is applied to cloud-side equipment and includes:

[0252] Module 1102 is configured to build an initial recommendation model corresponding to the target service based on a pre-trained language model in response to a model training request submitted by the end device.

[0253] The determination module 1104 is configured to read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information;

[0254] The construction module 1106 is configured to construct a sample group containing positive and negative samples based on the searcher information and the recommended party information;

[0255] The processing module 1108 is configured to use the initial recommendation model to process the positive samples and the negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples.

[0256] The training module 1110 is configured to calculate a loss value based on the positive score sequence and the negative score sequence, and to tune the parameters of the initial recommendation model based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

[0257] Feedback module 1112 is configured to feed back the target recommendation model to the edge device.

[0258] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample group is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. In subsequent recommendation tasks, recommendations can also be made based on the incorporation of searcher information, thereby improving recommendation accuracy and user satisfaction. Training the initial recommendation model on the cloud-side device for the edge device reduces the resource consumption of the edge device.

[0259] The above is an illustrative scheme of a model training device according to this embodiment. It should be noted that the technical solution of this model training device and the technical solution of the model training method described above belong to the same concept. For details not described in detail in the technical solution of the model training device, please refer to the description of the technical solution of the model training method described above.

[0260] See Figure 12 , Figure 12 A flowchart is shown of a recommendation model application method provided according to an embodiment of this specification. The recommendation model application method is applied to a computing node and specifically includes the following steps.

[0261] Step S1202: Upon receiving an information search request submitted for the information to be searched, the information to be searched and the information to be recommended corresponding to the information to be searched are input into the target pre-trained language model to obtain the search vector corresponding to the information to be searched and the recommendation vector corresponding to the information to be recommended.

[0262] Step S1204: Input the vector to be searched into the search processing model to obtain the intent vector to be searched;

[0263] Step S1206: Process the search intent vector and the recommendation vector based on the recommended party prediction model, and determine the feedback information corresponding to the information search request based on the processing result; wherein, the computing node deploys the target pre-trained language model, the search party processing model and the recommended party prediction model, which are obtained by splitting the target recommendation model in the above method.

[0264] Specifically, the search party processing model is used to generate search direction variables based on the output of the target pre-trained language model; the recommended party prediction model is used to generate recommended direction variables based on the output of the target pre-trained language model; the information to be searched can be information provided by the search party for searching for recommended parties; correspondingly, the information to be recommended is the recommended party information associated with the target business of the search party; the search vector is the vector representation of the information to be searched; the recommended vector is the vector representation of the information to be recommended.

[0265] Based on this, the target recommendation model is decomposed according to its functional layers, consisting of a target pre-trained language model, a search-side processing model, and a recommended-side prediction model. These three models are deployed separately on computing nodes, achieving decoupling and allowing them to jointly provide services. Upon receiving an information search request for the information to be searched, the information to be searched and its corresponding recommended information are input into the target pre-trained language model to obtain the search vector and the recommended vector. The search vector is then input into the search-side processing model to obtain the search intent vector. The recommended-side prediction model processes the search intent vector and the recommended vector, fusing them into a recommended vector. The feedback information corresponding to the information search request is determined based on the processing result.

[0266] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample group is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. The target recommendation model is decomposed according to its functional layers, consisting of a target pre-trained language model, a searcher processing model, and a recommended user prediction model, which are deployed separately to computing nodes, achieving decoupling of these three components. In subsequent recommendation tasks, recommendations are made based on the incorporated searcher information, thereby improving recommendation accuracy and user satisfaction.

[0267] Corresponding to the above method embodiments, this specification also provides recommended model application device embodiments. Figure 13 A schematic diagram of a recommended model application device provided in one embodiment of this specification is shown. Figure 13 As shown, this device is applied to a computing node and includes:

[0268] The receiving module 1302 is configured to, upon receiving an information search request submitted for the information to be searched, input the information to be searched and the information to be recommended corresponding to the information to be searched into the target pre-trained language model to obtain the search vector corresponding to the information to be searched and the recommendation vector corresponding to the information to be recommended.

[0269] Input module 1304 is configured to input the vector to be searched into the search processing model to obtain the intent vector to be searched.

[0270] The processing module 1306 is configured to process the search intent vector and the recommendation vector based on the recommendation prediction model, and determine the feedback information corresponding to the information search request based on the processing result; wherein, the computing node deploys a target pre-trained language model, a search processing model, and a recommendation prediction model, which are obtained by splitting the target recommendation model in the above method.

[0271] In summary, an initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. Then, a sample group is built based on historical information of the target business to complete the training of the initial recommendation model. This allows the initial recommendation model to incorporate more information, such as searcher information, during training, improving the model's generalization ability. The target recommendation model is decomposed according to its functional layers, consisting of a target pre-trained language model, a searcher processing model, and a recommended user prediction model, which are deployed separately to computing nodes, achieving decoupling of these three components. In subsequent recommendation tasks, recommendations are made based on the incorporated searcher information, thereby improving recommendation accuracy and user satisfaction.

[0272] The above is an illustrative scheme of a recommendation model application device according to this embodiment. It should be noted that the technical solution of this recommendation model application device and the technical solution of the recommendation model application method described above belong to the same concept. For details not described in detail in the technical solution of the recommendation model application device, please refer to the description of the technical solution of the recommendation model application method described above.

[0273] Figure 14 A structural block diagram of a computing device 1400 according to one embodiment of this specification is shown. The components of the computing device 1400 include, but are not limited to, a memory 1410 and a processor 1420. The processor 1420 is connected to the memory 1410 via a bus 1430, and a database 1450 is used to store data.

[0274] The computing device 1400 also includes an access device 1440, which enables the computing device 1400 to communicate via one or more networks 1460. Examples of such networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 1440 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0275] In one embodiment of this specification, the above-described components of the computing device 1400 and Figure 14 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 14 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0276] The computing device 1400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 1400 can also be a mobile or stationary server.

[0277] The processor 1420 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above method.

[0278] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the above method belong to the same concept, and all details not described in detail in the technical solution of the computing device can be referred to the description of the technical solution of the above method.

[0279] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described method.

[0280] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the method described above belong to the same concept, and all details not described in detail in the technical solution of the storage medium can be referred to the description of the technical solution of the method described above.

[0281] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described method.

[0282] The above is an illustrative example of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the method described above belong to the same concept. Details not described in detail in the technical solution of the computer program can be found in the description of the technical solution of the method described above.

[0283] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0284] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0285] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0286] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0287] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for training a recommendation model, comprising: An initial recommendation model corresponding to the target business is constructed based on a pre-trained language model. The initial recommendation model includes a first network layer and a second network layer. The first network layer is used to perform dimensionality reduction and projection processing on the output of the pre-trained language model, and the second network layer is used to score positive and negative samples. Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information; Based on the searcher information and the recommended party information, a sample group containing positive and negative samples is constructed; The initial recommendation model is used to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples. The loss value is calculated based on the positive score sequence and the negative score sequence, and the parameters of the initial recommendation model are tuned based on the loss value until a target recommendation model that meets the training stopping condition is obtained.

2. The method according to claim 1, wherein constructing a sample group containing positive and negative samples based on the searcher information and the recommended party information comprises: Based on the searcher information, a first recommended party information is selected from the recommended party information, and positive intent information is extracted from the first recommended party information as positive sample information. The first phase of sampling is performed on the recommended party information, and negative sample information is determined based on the sampling results; Positive samples are composed of the searcher information and the positive sample information, and negative samples are composed of the searcher information and the negative sample information. Construct a sample group containing the positive sample and the negative sample.

3. The method according to claim 2, wherein performing a first-stage sampling in the recommended party information and determining negative sample information based on the sampling results includes: Based on the searcher information, a second recommended party is selected from the recommended party information, and negative intent information is extracted from the second recommended party information as negative sample information, or Random sampling is performed on the recommended party information, and negative sample information is determined based on the sampling results.

4. The method according to claim 2, wherein performing a first-stage sampling in the recommended party information and determining negative sample information based on the sampling results includes: Based on the searcher information and a preset ratio, the recommended party information is randomly sampled to obtain third recommended party information; According to the preset ratio and sample type, select the fourth recommended party information from the recommended party information; Negative intent information is extracted from the third recommended party information as the first negative sample information, and negative intent information is extracted from the fourth recommended party information as the second negative sample information. The first negative sample information and the second negative sample information are used as the negative sample information.

5. The method according to claim 1, wherein constructing the initial recommendation model corresponding to the target business based on the pre-trained language model comprises: In response to the model building request associated with the target business, a dimensionality reduction network layer and a projection network layer, as well as an intent network and a feature network, are constructed. The dimensionality reduction network layer, the projection network layer, the intent network, the feature network, and the pre-trained language model are concatenated according to a preset concatenation rule to obtain an initial recommendation model.

6. The method according to claim 1, wherein processing the positive samples and negative samples contained in the sample group using the initial recommendation model to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples comprises: Using the pre-trained language model included in the initial recommendation model, semantic recognition is performed on the search party information, positive sample information and negative sample information contained in the sample group to obtain the search party semantic vector corresponding to the search party information, the positive sample vector corresponding to the positive sample information and the negative sample vector corresponding to the negative sample information. The search semantic vector, the positive sample vector, and the negative sample vector are processed using the network layers contained in the initial recommendation model to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample.

7. The method according to claim 6, wherein processing the search semantic vector, the positive sample vector, and the negative sample vector using the network layers included in the initial recommendation model to obtain the positive score sequence corresponding to the positive sample and the negative score sequence corresponding to the negative sample includes: The search semantic vector, the positive sample vector, and the negative sample vector are projected and dimensionality reduced using the first network layer in the network layer of the initial recommendation model to obtain the search embedding vector corresponding to the search semantic vector, the positive sample embedding vector corresponding to the positive sample vector, and the negative sample embedding vector corresponding to the negative sample vector. The search embedding vector and the positive sample embedding vector are input into the second network layer contained in the network layer of the initial recommendation model for scoring, so as to obtain the positive score sequence corresponding to the positive sample; The search embedding vector and the negative sample embedding vector are input into the second network layer contained in the network layer of the initial recommendation model for scoring, thereby obtaining the negative score sequence corresponding to the negative sample.

8. The method according to claim 7, wherein inputting the search embedding vector and the positive sample embedding vector into a second network layer included in the network layer of the initial recommendation model for scoring to obtain a positive score sequence corresponding to the positive sample; and inputting the search embedding vector and the negative sample embedding vector into a second network layer included in the network layer of the initial recommendation model for scoring to obtain a negative score sequence corresponding to the negative sample, comprises: Determine the second network layer included in the network layers of the initial recommendation model; The intent layer included in the second network layer is used to perform intent recognition on the search party embedding vector, and the search party intent vector is determined based on the recognition result; The search intent vector and the positive sample embedding vector are input into the feature layer contained in the second network layer for scoring, and the positive score sequence corresponding to the positive sample is obtained. The search intent vector and the negative sample embedding vector are input into the feature layer contained in the second network layer for scoring, thereby obtaining the negative score sequence corresponding to the negative sample.

9. The method according to claim 1, wherein when the sample group consists of at least two samples and the sample group contains at least one negative sample, the step of calculating the loss value based on the positive score sequence and the negative score sequence comprises: Identify at least one negative sample in each of at least two sample groups, and determine the negative score sequence corresponding to at least one negative sample; Select the target negative score sequence corresponding to at least two sample groups from the negative score sequences corresponding to at least two sample groups respectively. Based on the positive score sequences and target negative score sequences corresponding to at least two sample groups, calculate the initial loss values ​​corresponding to at least two sample groups respectively; The initial loss values ​​corresponding to at least two sample groups are summed to obtain the loss value.

10. A method for applying a recommendation model to a computing node, comprising: Upon receiving an information search request for the information to be searched, the information to be searched and the information to be recommended corresponding to the information to be searched are input into the target pre-trained language model to obtain the search vector corresponding to the information to be searched and the recommendation vector corresponding to the information to be recommended. The vector to be searched is input into the search processing model to obtain the search intent vector. Based on the recommended prediction model, the search intent vector and the recommendation vector are processed, and the feedback information corresponding to the information search request is determined according to the processing result. The computing node deploys a target pre-trained language model, a search-side processing model, and a recommended-side prediction model, which are obtained by splitting the target recommendation model in any one of claims 1-9.

11. An information retrieval method, comprising: Obtain the searcher information entered by the searcher on the information recommendation platform related to the target business; Determine the recommended party information contained in the information recommendation platform; The recommended party information and the search party information are input into the target recommendation model according to any one of claims 1-10 for processing to obtain the target recommended party information; Search feedback information is generated based on the target recommendation information and then fed back to the searcher through the information recommendation platform.

12. A model training method applied to cloud-side devices, comprising: In response to a model training request submitted by an edge device, an initial recommendation model corresponding to the target service is constructed based on a pre-trained language model. The initial recommendation model includes a first network layer and a second network layer. The first network layer is used to perform dimensionality reduction and projection processing on the output of the pre-trained language model, and the second network layer is used to score positive and negative samples. Read the searcher information from the historical business information corresponding to the target business, and determine the recommended party information that has a business relationship with the searcher information; Based on the searcher information and the recommended party information, a sample group containing positive and negative samples is constructed; The initial recommendation model is used to process the positive samples and negative samples contained in the sample group to obtain the positive score sequence corresponding to the positive samples and the negative score sequence corresponding to the negative samples. The loss value is calculated based on the positive score sequence and the negative score sequence, and the parameters of the initial recommendation model are tuned based on the loss value until the target recommendation model that meets the training stopping condition is obtained. The target recommendation model is fed back to the edge device.

13. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 12.

14. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 12.

Images