Large language model training method and device and resource retrieval method and device

Abstract

The invention provides a big language model training method and device and a resource retrieval method and device, and relates to the technical field of artificial intelligence, in particular to the field of deep learning and big language models. According to the specific implementation scheme, a base large language model is obtained, the base large language model is used for outputting a semantic identification sequence corresponding to search information based on the search information input by a user, the semantic identification sequence comprises at least one semantic identification, and the semantic identification represents at least one semantic attribute; a first training sample set is obtained, each first training sample in the first training sample set comprises first sample search information, a first sample semantic identifier and sample correlation, and the sample correlation is used for describing the correlation degree between the first sample search information and the first sample semantic identifier; and training the base large language model based on the first training sample set to obtain a target large language model generated after training.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, in particular to the field of deep learning and large language models, and specifically to a large language model training method, a resource retrieval method and a device thereof. Background Art

[0002] With the rapid advancement of Internet technology, users can now easily browse product and advertising information through terminal devices such as smartphones, and conduct precise searches by entering text or uploading images. This efficient retrieval method not only simplifies the information search process, but also helps users quickly locate resources that meet their needs, greatly improving the efficiency of resource acquisition. Summary of the invention

[0003] The present disclosure provides a large language model training method, resource retrieval method, device, equipment and storage medium.

[0004] According to one aspect of the present disclosure, a method for training a large language model is provided, comprising: obtaining a base large language model, wherein the base large language model is used to output a semantic identification sequence corresponding to the search information based on the search information input by a user, the semantic identification sequence includes at least one semantic identification, and the semantic identification represents at least one semantic attribute; obtaining a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identification and a sample correlation, and the sample correlation is used to describe the degree of correlation between the first sample search information and the first sample semantic identification; training the base large language model based on the first training sample set to obtain a target large language model generated after training.

[0005] By marking "sample relevance" in training samples, the model can learn the correlation between search information and semantic identifiers, so that the model can better understand the relationship between search information and semantic identifiers. The target large language model obtained after training can give answers that better meet user needs based on the user's search information, thereby improving the accuracy and relevance of retrieval. This training method can be applied to the training process of large language models in multiple scenarios such as resource recommendation scenarios or search scenarios, and has strong generalization capabilities.

[0006] According to another aspect of the present disclosure, a resource retrieval method is provided, including: obtaining target search information input by a user; processing the target search information based on a trained target large language model to obtain a target semantic identification sequence, wherein the target semantic identification sequence includes at least one target semantic identification, and the target semantic identification represents at least one target semantic attribute; determining a second candidate resource associated with the target semantic identification sequence from the first candidate resource; determining a target resource from the second candidate resource, and displaying the target resource.

[0007] According to one aspect of the present disclosure, a training device for a large language model is provided, comprising: a first acquisition module, used to acquire a base large language model, wherein the base large language model is used to output a semantic identification sequence corresponding to the search information based on the search information input by the user, the semantic identification sequence includes at least one semantic identification, and the semantic identification represents at least one semantic attribute; a second acquisition module, used to acquire a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identification and a sample correlation, and the sample correlation is used to describe the degree of correlation between the first sample search information and the first sample semantic identification; a model training module, used to train the base large language model based on the first training sample set to obtain a target large language model generated after training.

[0008] According to one aspect of the present disclosure, a resource retrieval device is provided, including: an acquisition module, used to acquire target search information input by a user; a processing module, used to process the target search information based on a trained target large language model to obtain a target semantic identification sequence, wherein the target semantic identification sequence includes at least one target semantic identification, and the target semantic identification represents at least one target semantic attribute; a determination module, used to determine a second candidate resource associated with the target semantic identification sequence from a first candidate resource; and a display module, used to determine a target resource from the second candidate resource and display the target resource.

[0009] According to another aspect of the present disclosure, there is provided an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned large language model training method or resource retrieval method.

[0010] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the above-mentioned large language model training method or resource retrieval method.

[0011] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements the above-mentioned large language model training method or resource retrieval method when executed by a processor.

[0012] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0014] Figure 1 It is a schematic diagram of an exemplary implementation of a method for training a large language model according to an exemplary embodiment of the present disclosure.

[0015] Figure 2 It is a schematic diagram of an exemplary implementation of a method for training a large language model according to an exemplary embodiment of the present disclosure.

[0016] Figure 3 It is a schematic diagram of an exemplary implementation of a method for training a large language model according to an exemplary embodiment of the present disclosure.

[0017] Figure 4 It is a schematic diagram of an exemplary implementation of a method for training a large language model according to an exemplary embodiment of the present disclosure.

[0018] Figure 5 It is a schematic diagram of an exemplary implementation of a resource retrieval method according to an exemplary embodiment of the present disclosure.

[0019] Figure 6 It is a schematic diagram of a display interface according to an exemplary embodiment of the present disclosure.

[0020] Figure 7 It is a schematic diagram of an exemplary implementation of a resource retrieval method according to an exemplary embodiment of the present disclosure.

[0021] Figure 8 It is a schematic diagram of a display interface according to an exemplary embodiment of the present disclosure.

[0022] Fig. 9 is a schematic diagram of a large language model training device according to an exemplary embodiment of the present disclosure.

[0023] Fig.10 is a schematic diagram of a resource retrieval device according to an exemplary embodiment of the present disclosure.

[0024] Fig.11 is a schematic diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0026] Deep Learning (DL) is a new research direction in the field of Machine Learning (ML). It is introduced into machine learning to make it closer to its original goal - artificial intelligence. Deep learning is the internal laws and representation levels of learning sample data. The information obtained in the learning process is of great help in the interpretation of data such as text, images and sounds. Its ultimate goal is to enable machines to have analytical learning capabilities like humans and to recognize data such as text, images and sounds. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition that far exceed previous related technologies.

[0027] Artificial Intelligence (AI) is a discipline that studies how computers can simulate certain thought processes and intelligent behaviors of human beings (such as learning, reasoning, thinking, planning, etc.). It includes both hardware-level technologies and software-level technologies. AI hardware technologies generally include computer vision technology, speech recognition technology, natural language processing technology, as well as learning / deep learning, big data processing technology, knowledge graph technology, and other aspects.

[0028] In the technical solution disclosed herein, the acquisition, storage and application of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0029] Figure 1 is a schematic diagram of an exemplary implementation of a large language model training method shown in the present application, such as Figure 1 As shown, the training method of the large language model includes the following steps:

[0030] S101, obtaining a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0031] In some embodiments, the base large language model refers to a large neural network model based on deep learning, such as Generative Pretrained Transformer (GPT), Bidirectional Encoder Representations from Transformers (BERT), etc. These models are pre-trained on a large amount of text data, and can understand the user's search information, and then output a semantic identification sequence related to the search information.

[0032] In the present application, the base large language model is a pre-trained large language model. The base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, and the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0033] For example, if the user's search information is "Please recommend me a pair of women's shoes", after the base large language model processes the user's search information, the semantic identifier sequence output by the base large language model is {1171, 1172, 1175...}, among which "1171", "1172", "1175" and so on are all semantic identifiers, and the semantic identifier represents at least one semantic attribute. For example, "1171" can represent clothing category-shoe category-women's shoes-women's sports shoes; and "1172" can represent clothing category-shoe category-women's shoes-women's high heels; and "1175" can represent clothing category-shoe category-women's shoes-women's casual shoes.

[0034] Among them, taking the above-mentioned "1171" as an example, it can be understood that the first number "1" represents the first level is clothing, the second number "1" represents the second level is shoes, the third number "7" represents the third level is women's shoes, and the fourth number "1" represents the fourth level is women's sports shoes.

[0035] The search information may include text information generated based on the user's input operation, for example, the text information may be obtained based on various types of information such as text, image, audio, etc. input by the user in the input box of the page. This application does not limit this.

[0036] S102, obtaining a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identifier and sample correlation, and the sample correlation is used to describe the correlation degree between the first sample search information and the first sample semantic identifier.

[0037] In the present application, the first training sample set is mainly to enable the target large language model generated after subsequent training to better learn the correlation between the search information and the semantic identifier.

[0038] For example, the first training sample set may include the following first training samples:

[0039] Women's shoes-1171-0.90, indicating that the correlation between the first sample search information "women's shoes" and the first sample semantic identifier "1171" (indicating clothing category-shoe category-women's shoes-women's sports shoes) is 0.90;

[0040] Women's bags-1172-0.10, indicating that the correlation between the first sample search information "women's bags" and the first sample semantic identifier "1172" (indicating clothing-shoes-women's shoes-women's high heels) is 0.10;

[0041] Women's shoes-1175-0.85, indicating that the correlation between the first sample search information "women's shoes" and the first sample semantic identifier "1175" (indicating clothing category-shoe category-women's shoes-women's casual shoes) is 0.85;

[0042] By analogy, the first training sample set includes a large number of first training samples, and each first training sample includes first sample search information, a first sample semantic identifier, and sample relevance.

[0043] S103: Train the base large language model based on the first training sample set to obtain a target large language model generated after training.

[0044] In some embodiments, the base large language model is trained based on the first training sample set obtained above, and the target large language model is obtained after the loss function converges.

[0045] Among them, the target large language model generated after training improves the ability to learn the correlation between search information and semantic identifiers compared to the base large language model.

[0046] The embodiment of the present application proposes a training method for a large language model, including: obtaining a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute; obtaining a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, first sample semantic identifier and sample correlation, and the sample correlation is used to describe the degree of correlation between the first sample search information and the first sample semantic identifier; training the base large language model based on the first training sample set to obtain a target large language model generated after training. The present application marks "sample correlation" in the training samples, so that the model can learn the correlation between the search information and the semantic identifier, so that the model can better understand the relationship between the search information and the semantic identifier. The target large language model obtained after training can give answers that better meet the user's needs based on the user's search information, thereby improving the accuracy and relevance of the retrieval. This training method can be applied to the training process of large language models in multiple scenarios such as resource recommendation scenarios or search scenarios, and has strong generalization capabilities.

[0047] Figure 2 is a schematic diagram of an exemplary implementation of a large language model training method shown in the present application, such as Figure 2 As shown, the training method of the large language model includes the following steps:

[0048] S201, obtaining a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0049] S202, obtaining a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identifier and sample correlation, and the sample correlation is used to describe the correlation degree between the first sample search information and the first sample semantic identifier.

[0050] Regarding the specific implementation of steps S201 to S202, reference may be made to the specific introduction of the relevant parts in the above embodiments, which will not be elaborated here.

[0051] S203: Train the base large language model based on the first training sample set to obtain an intermediate large language model.

[0052] In some embodiments, the base large language model is trained based on the first training sample set obtained above, and an intermediate large language model is obtained after the loss function converges.

[0053] Among them, the intermediate large language model improves the ability to learn the correlation between search information and semantic identifiers compared to the base large language model.

[0054] S204: Acquire a second training sample set, wherein each second training sample in the second training sample set includes second sample search information, a second sample semantic identifier, and a sample preference index.

[0055] It is not difficult to understand that in order to enable users to have a better interactive experience when using the target large language model generated after training to recommend relevant resources to users, it is also necessary to train the intermediate large language model obtained above in combination with the user's preferences.

[0056] For example, the second training sample set may include the following second training samples:

[0057] Women's shoes - 1171 - 90 points (with a full score of 100 points as the standard), indicating that when the second sample search information is "women's shoes", after the women's sports shoes under the second sample semantic identifier "1171" (which means clothing - shoes - women's shoes - women's sports shoes) are recommended to the user, the user's preference for the women's sports shoes under the second sample semantic identifier "1171" is 90 points;

[0058] Women's shoes - 1172 - 60 points (with a full score of 100 points as the standard), indicating that when the second sample search information is "women's shoes", after the women's high-heeled shoes under the second sample semantic identifier "1172" (which means clothing - shoes - women's shoes - women's high-heeled shoes) are recommended to the user, the user's preference for the women's high-heeled shoes under the second sample semantic identifier "1172" is 60 points;

[0059] Women's shoes - 1175 - 80 points (with a full score of 100 points as the standard), which means that when the second sample search information is "women's shoes", after the women's casual shoes under the second sample semantic identifier "1175" (which means clothing - shoes - women's shoes - women's casual shoes) are recommended to the user, the user's preference for the women's casual shoes under the second sample semantic identifier "1175" is 80 points;

[0060] By analogy, the second training sample set includes a large number of second training samples, and each second training sample includes second sample search information, a second sample semantic identifier, and a sample preference index (or can also be understood as a sample preference score).

[0061] S205: Train the intermediate large language model based on the second training sample set to obtain a target large language model generated after training.

[0062] The intermediate large language model is trained based on the second training sample set obtained above, and the target large language model is obtained after the loss function converges.

[0063] Through a two-stage training process (based on relevance and preference index), the embodiment of the present application enables the model to optimize its generation capability on different training data sets and improve its adaptability to practical applications; the second stage of training introduces the preference index of the sample, so that the model can better capture the user's preferences and enhance the user experience.

[0064] Figure 3 is a schematic diagram of an exemplary implementation of a large language model training method shown in the present application, such as Figure 3 As shown, the training method of the large language model includes the following steps:

[0065] S301, determining a sample resource set, and obtaining a sample semantic identifier corresponding to each sample resource in the sample resource set.

[0066] First, determine the sample resource set. After determining the usage scenario, the sample resource set corresponding to the scenario should include as many different types of sample resources as possible. For example, if the application scenario is a product recommendation scenario, the sample resource set may include various types of products.

[0067] In some embodiments, the sample semantic identifier corresponding to each sample resource in the sample resource set may be determined by manual labeling. For example, if a sample resource is a pair of women's sports shoes, the sample resource may be labeled with "1171" (indicating clothing category - shoes category - women's shoes - women's sports shoes) by manual labeling.

[0068] In some embodiments, the sample information corresponding to each sample resource in the sample resource set may be obtained (for example, when the sample resource is a commodity resource, the information of the commodity resource may include the commodity title, commodity promotion text, commodity icon information, commodity description information, etc. of the commodity resource), and then each sample information is processed to obtain the sample semantic identifier corresponding to each sample resource. In this way, by obtaining and processing a variety of information in the sample resource (such as commodity title, promotion text, etc.), the semantic features of the sample resource can be fully captured, so that the sample semantic identifier corresponding to the sample resource finally obtained is more accurate.

[0069] In some embodiments, when obtaining sample information corresponding to each sample resource in a sample resource set, key information can be extracted from the landing details page of any sample resource, and the extracted key information (such as product type, color, size introduction, etc.) is used as the sample information corresponding to the sample resource, and then each sample information is processed to obtain the sample semantic identifier corresponding to each sample resource. In this way, by extracting key information from the landing details page of the sample resource, it can be ensured that the obtained sample information is closer to the actual application scenario, and that the obtained sample information is more focused on content that is highly relevant to user needs and resource characteristics (such as the core attributes, characteristics, usage scenarios, etc. of the product), so that the sample semantic identifier corresponding to the sample resource finally obtained is more accurate.

[0070] In some embodiments, the sample information can be processed based on the following methods to obtain a sample semantic identifier corresponding to the sample resource: encoding the sample information based on an encoder to obtain sample semantic features corresponding to the sample information; processing the sample semantic features based on a clustering algorithm to obtain multiple clusters and a sample semantic identifier corresponding to each cluster, wherein the cluster includes at least one sample semantic feature; and using the sample semantic identifier corresponding to the cluster where the sample resource is located as the sample semantic identifier corresponding to the sample resource.

[0071] Among them, when multiple clusters are obtained by processing the semantic features of the samples based on the clustering algorithm, the clustering can be performed hierarchically for multiple times. For example, the first clustering is clustering on large categories, and then the clusters formed by the first clustering are further clustered for the second level, and so on.

[0072] The clustering algorithm classifies similar sample semantic features into the same cluster, so that the sample features in each cluster have a high degree of similarity. In this way, a representative semantic identifier can be effectively assigned to the sample resources.

[0073] S302: Acquire sample search information related to sample resources.

[0074] Optionally, sample search information related to each sample resource may be obtained from the user history log.

[0075] S303, constructing a first sample pair based on the sample semantic identifier and the sample search information, annotating the sample correlation corresponding to the first sample pair, and constructing a first training sample set based on the first sample pair and the sample correlation.

[0076] The first sample pair includes a positive sample pair and a negative sample pair.

[0077] Specifically, a positive sample pair is constructed based on the sample semantic identifier and the sample search information, and the sample correlation of the positive sample pair is annotated. Among them, the sample semantic identifier and the sample search information in the positive sample pair have a high correlation. For example, the sample search information "women's shoes" and the sample semantic identifier "1171" (indicating clothing category-shoe category-women's shoes-women's sports shoes) have a high correlation, so a high correlation can be set.

[0078] Specifically, a negative sample pair is constructed based on the sample semantic identifier and the sample search information, and the sample correlation of the negative sample pair is annotated. Among them, the sample semantic identifier and the sample search information in the negative sample pair have a low correlation. For example, the sample search information "air conditioner" and the sample semantic identifier "1171" (indicating clothing-shoes-ladies shoes-ladies sports shoes) have a low correlation, so a low correlation can be set.

[0079] S304, constructing a second sample pair based on the sample semantic identifier and the sample search information, annotating and obtaining a sample preference index corresponding to the second sample pair, and constructing a second training sample set based on the second sample pair and the sample preference index.

[0080] A second sample pair is constructed based on the sample semantic identifier and the sample search information, and user interaction data corresponding to the sample resource is obtained, wherein the user interaction data at least includes parameters such as click rate, purchase rate, conversion rate, etc. corresponding to each sample resource.

[0081] Based on the user interaction data, a sample preference index corresponding to the second sample pair is annotated.

[0082] Specifically, taking the example of a sample search information corresponding to multiple sample semantic identifiers, a higher sample preference index is set for the sample semantic identifier corresponding to the sample resource having better user interaction data (for example, higher click-through rate, higher purchase rate and / or higher conversion rate) corresponding to the sample search information; a lower sample preference index is set for the sample semantic identifier corresponding to the sample resource having poor user interaction data (for example, poor click-through rate, poor purchase rate and / or poor conversion rate) corresponding to the sample search information.

[0083] It is not difficult to understand that a sample semantic identifier may include one or more sample resources. When labeling the sample preference index for the sample semantic identifier and sample search information, it is necessary to consider the user interaction data corresponding to all sample resources included in the sample semantic identifier.

[0084] S305, obtaining a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0085] S306: Train the base large language model based on the first training sample set to obtain an intermediate large language model.

[0086] S307: Train the intermediate large language model based on the second training sample set to obtain a target large language model generated after training.

[0087] Regarding the specific implementation of steps S305 to S307, reference may be made to the specific introduction of the relevant parts in the above embodiments, which will not be elaborated here.

[0088] The embodiment of the present application details the construction process of the first training sample set and the second training sample set. The first training sample set includes both positive samples and negative samples, so that the model can learn the correlation between search information and semantic identifiers in the first stage of training, so that the model can better understand the relationship between search information and semantic identifiers. The second stage of training introduces the sample preference index, so that the model can better capture the user's preferences and enhance the user experience.

[0089] Figure 4 is a schematic diagram of an exemplary implementation of a large language model training method shown in the present application, such as Figure 4 As shown, the training method of the large language model includes the following steps:

[0090] A first training sample set is constructed, wherein each first training sample in the first training sample set includes first sample search information, first sample semantic identifier and sample relevance, and the sample relevance is used to describe the degree of relevance between the first sample search information and the first sample semantic identifier. The first training sample set includes both positive samples and negative samples.

[0091] Construct a second training sample set, wherein each second training sample in the second training sample set includes second sample search information, a second sample semantic identifier, and a sample preference index. Taking a second sample search information corresponding to multiple second sample semantic identifiers as an example, a higher sample preference index is set for the second sample semantic identifier corresponding to the sample resource with better user interaction data (for example, higher click-through rate, higher purchase rate, and / or higher conversion rate) corresponding to the second sample search information; and a lower sample preference index is set for the second sample semantic identifier corresponding to the sample resource with poor user interaction data (for example, poor click-through rate, poor purchase rate, and / or poor conversion rate) corresponding to the second sample search information.

[0092] For example, if a second sample search information is "women's shoes", its related second sample semantic identifiers include "1171", "1172" and "1175", etc. After comparing the interaction data of a large number of users, if it is found that the user interaction data corresponding to the sample recommendation resources under the second sample semantic identifier "1171" performs well, a higher sample preference index is set for the second sample pair "women's shoes" + "1171"; if it is found that the user interaction data corresponding to the sample recommendation resources under the second sample semantic identifier "1175" performs moderately, a moderate sample preference index is set for the second sample pair "women's shoes" + "1175"; if it is found that the user interaction data corresponding to the sample recommendation resources under the second sample semantic identifier "1172" performs poorly, a lower sample preference index is set for the second sample pair "women's shoes" + "1172".

[0093] A base large language model is obtained, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0094] The base large language model is trained based on the first training sample set to obtain an intermediate large language model.

[0095] The intermediate large language model is trained based on the second training sample set to obtain a target large language model generated after training.

[0096] Through a two-stage training process (based on relevance and preference index), the embodiment of the present application enables the model to optimize its generation capability on different training data sets and improve its adaptability to practical applications; the second stage of training introduces the preference index of the sample, so that the model can better capture the user's preferences and enhance the user experience.

[0097] Figure 5 is a schematic diagram of an exemplary implementation of a resource retrieval method shown in the present application, such as Figure 5 As shown, the resource retrieval method includes the following steps:

[0098] S501, obtaining target search information input by a user.

[0099] For example, the target search information input by the user may be “recommend me a pair of men's shoes”.

[0100] S502, processing the target search information based on the trained target large language model to obtain a target semantic identifier sequence, wherein the target semantic identifier sequence includes at least one target semantic identifier, and the target semantic identifier represents at least one target semantic attribute.

[0101] The target large language model is generated by training using the training method described in the above embodiment.

[0102] S503: Determine a second candidate resource associated with the target semantic identifier sequence from the first candidate resources.

[0103] It is not difficult to understand that the target large language model obtained by the above training is applied to the upstream of resource retrieval, and its purpose is to output the target semantic identification sequence related to the target search information. In general, the information finally presented to the user on the display interface is generally the relevant information of the target resource finally recommended to the user. Therefore, after determining the target semantic identification sequence related to the target search information upstream, it is also necessary to determine the target resource that needs to be displayed to the user based on the target semantic identification sequence related to the target search information downstream.

[0104] In the present application, after the target large language model outputs the target semantic identification sequence, it is necessary to determine a second candidate resource associated with the target semantic identification sequence from the first candidate resources based on the target semantic identification sequence.

[0105] Exemplarily, if the above-mentioned target large language model outputs a target semantic identification sequence of {1181, 1183, 1189...} based on the target search information "Recommend me a pair of men's shoes", among which "1181", "1183", "1189" and the like are all target semantic identifications, and the target semantic identification represents at least one target semantic attribute. For example, "1181" can represent clothing category-shoe category-men's shoes-men's sports shoes; and "1183" can represent clothing category-shoe category-men's shoes-men's skateboard shoes; and "1189" can represent clothing category-shoe category-men's shoes-men's leather shoes.

[0106] Among them, the first candidate resource is generally the entire candidate resource library. After the target semantic identifier sequence is determined as above, candidate resources related to the target semantic identifiers such as "1181", "1183", "1189", etc. are determined from the first candidate resources as the second candidate resources. It can also be understood that candidate resources related to categories such as men's sports shoes, men's skateboard shoes, and men's leather shoes are determined from the first candidate resources as the second candidate resources.

[0107] S504: Determine a target resource from the second candidate resources and display the target resource.

[0108] In some embodiments, after the second candidate resources are determined as described above, some candidate resources are determined from the second candidate resources as target resources, and the target resources are displayed on the display screen.

[0109] When determining some of the candidate resources from the second candidate resources as target resources, some of the candidate resources can be determined from the second candidate resources as target resources based on the brand, promotion information, user praise, price and other information of each second candidate resource.

[0110] Figure 6 This is a schematic diagram of a display interface shown in this application, such as Figure 6 As shown, the user enters "recommend me a pair of men's shoes" in the input box on the display interface, and the above-determined target resources will be recommended to the user on the display interface. For example, the target resources include: sports shoes 1, sports shoes 2, skateboard shoes 1, skateboard shoes 2, leather shoes 1, etc. The user can click on the target resource to enter the landing details page of the target resource for browsing.

[0111] The embodiment of the present application applies the target large language model obtained by the above training. The target large language model obtained by the above training learns the correlation between search information and semantic identifiers, and can better capture user preferences, so that in resource retrieval scenarios, more suitable target resources can be recommended to users based on the target large language model combined with the downstream related resource screening model, thereby improving user satisfaction.

[0112] Figure 7 is a schematic diagram of an exemplary implementation of a resource retrieval method shown in the present application, such as Figure 7 As shown, the resource retrieval method includes the following steps:

[0113] S701, obtaining target search information input by a user.

[0114] S702, processing the target search information based on the trained target large language model to obtain a target semantic identifier sequence, wherein the target semantic identifier sequence includes at least one target semantic identifier, and the target semantic identifier represents at least one target semantic attribute.

[0115] S703: Determine a second candidate resource associated with the target semantic identifier sequence from the first candidate resources.

[0116] S704: Determine a target resource from the second candidate resources.

[0117] Regarding the specific implementation of steps S701 to S704, reference may be made to the specific introduction of the relevant parts in the above embodiments, which will not be elaborated here.

[0118] S705: Generate prompt information based on the target search information and the target resource.

[0119] In the present application, prompt information may be generated based on the target search information in combination with relevant characteristics of the target resource to help the user better understand the target resource.

[0120] Figure 8 This is a schematic diagram of a display interface shown in this application, such as Figure 8 As shown, the user enters "Recommend me a pair of men's shoes" in the input box on the display interface. The prompt information generated based on the target search information and target resources may be "If you prefer a sporty style, we recommend the following models for you", "If you prefer a fashionable style, we recommend the following models for you", "If you prefer a business style, we recommend the following models for you" and other related prompt information.

[0121] S706, displaying prompt information and displaying target resources.

[0122] After the prompt information is determined, the prompt information and the target resource are displayed on the display interface according to the relationship between the prompt information and the target resource and in combination with a preset display format.

[0123] The embodiment of the present application applies the target large language model obtained by the above training. The target large language model obtained by the above training learns the correlation between search information and semantic identifiers, and can better capture user preferences, so that in resource retrieval scenarios, more suitable target resources can be recommended to users based on the target large language model combined with the downstream related resource screening model. In the embodiment of the present application, prompt information is generated based on the target search information combined with the relevant characteristics of the target resource to help users better understand the target resource and improve user satisfaction.

[0124] Fig. 9 is a schematic diagram of a large language model training device shown in this application, such as Fig. 9 As shown, the large language model training device 900 includes a first acquisition module 901, a second acquisition module 902 and a model training module 903, wherein:

[0125] The first acquisition module 901 is used to acquire a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, and the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute.

[0126] The second acquisition module 902 is used to acquire a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identifier and sample correlation, and the sample correlation is used to describe the degree of correlation between the first sample search information and the first sample semantic identifier.

[0127] The model training module 903 is used to train the base large language model based on the first training sample set to obtain a target large language model generated after training.

[0128] By marking "sample relevance" in training samples, the model can learn the correlation between search information and semantic identifiers, so that the model can better understand the relationship between search information and semantic identifiers. The target large language model obtained after training can give answers that better meet user needs based on the user's search information, thereby improving the accuracy and relevance of retrieval. This training method can be applied to multiple scenarios such as resource recommendation scenarios or search scenarios, and has strong generalization capabilities.

[0129] Furthermore, the model training module 903 is also used to: train the base large language model based on the first training sample set to obtain an intermediate large language model; obtain a second training sample set, wherein each second training sample in the second training sample set includes second sample search information, a second sample semantic identifier and a sample preference index; and train the intermediate large language model based on the second training sample set to obtain a target large language model generated after training.

[0130] Furthermore, the training device 900 for a large language model also includes a sample construction module, which is used to: determine a sample resource set, and obtain a sample semantic identifier corresponding to each sample resource in the sample resource set; obtain sample search information related to the sample resource; construct a first sample pair based on the sample semantic identifier and the sample search information, annotate the sample correlation corresponding to the first sample pair, and construct a first training sample set based on the first sample pair and the sample correlation; and construct a second sample pair based on the sample semantic identifier and the sample search information, annotate the sample preference index corresponding to the second sample pair, and construct a second training sample set based on the second sample pair and the sample preference index.

[0131] Furthermore, the sample construction module is further used to: obtain sample information corresponding to each sample resource in the sample resource set; and process the sample information to obtain a sample semantic identifier corresponding to the sample resource.

[0132] Furthermore, the sample construction module is also used to: for any sample resource, extract key information from the landing details page of the sample resource, and use the extracted key information as the sample information corresponding to the sample resource.

[0133] Furthermore, the sample construction module is also used to: encode the sample information based on the encoder to obtain sample semantic features corresponding to the sample information; process the sample semantic features based on the clustering algorithm to obtain multiple clusters and sample semantic identifiers corresponding to each cluster, wherein the cluster includes at least one sample semantic feature; and use the sample semantic identifier corresponding to the cluster where the sample resource is located as the sample semantic identifier corresponding to the sample resource.

[0134] Furthermore, the sample construction module is also used to: construct positive sample pairs based on sample semantic identification and sample search information, and mark the sample correlation of the positive sample pairs; and construct negative sample pairs based on sample semantic identification and sample search information, and mark the sample correlation of the negative sample pairs.

[0135] Furthermore, the sample construction module is also used to: construct a second sample pair based on the sample semantic identifier and the sample search information; obtain user interaction data corresponding to the sample resource; and annotate the sample preference index corresponding to the second sample pair based on the user interaction data.

[0136] Fig.10 is a schematic diagram of a resource retrieval device shown in this application, such as Fig.10 As shown, the resource search device 1000 includes an acquisition module 1001, a processing module 1002, a determination module 1003 and a display module 1004, wherein:

[0137] The acquisition module 1001 is used to acquire the target search information input by the user;

[0138] The processing module 1002 is used to process the target search information based on the trained target large language model to obtain a target semantic identifier sequence, wherein the target semantic identifier sequence includes at least one target semantic identifier, and the target semantic identifier represents at least one target semantic attribute;

[0139] A determination module 1003 is used to determine a second candidate resource associated with a target semantic identifier sequence from the first candidate resources;

[0140] The display module 1004 is used to determine the target resource from the second candidate resources and display the target resource.

[0141] The device applies the target large language model obtained through the above training. The target large language model obtained through the above training learns the correlation between search information and semantic identifiers, and can better capture user preferences, so that in resource retrieval scenarios, more suitable target resources can be recommended to users based on the target large language model combined with the downstream related resource screening model, thereby improving user satisfaction.

[0142] Furthermore, the display module 1004 is also used to: generate prompt information based on the target search information and the target resource; and display the prompt information.

[0143] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0144] Fig.11A schematic block diagram of an example electronic device 1100 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0145] like Fig.11 As shown, the device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded from a storage unit 1108 into a random access memory (RAM) 1103. In the RAM 1103, various programs and data required for the operation of the device 1100 can also be stored. The computing unit 1101, the ROM 1102, and the RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0146] A number of components in the device 1100 are connected to the I / O interface 1105, including: an input unit 1106, such as a keyboard, a mouse, etc.; an output unit 1107, such as various types of displays, speakers, etc.; a storage unit 1108, such as a disk, an optical disk, etc.; and a communication unit 1109, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1109 allows the device 1100 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0147] The computing unit 1101 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1101 performs the various methods and processes described above, such as a training method or a resource retrieval method for a large language model. For example, in some embodiments, the training method or the resource retrieval method for a large language model may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 1108. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1100 via the ROM 1102 and / or the communication unit 1109. When the computer program is loaded into the RAM 1103 and executed by the computing unit 1101, one or more steps of the training method or the resource retrieval method for the large language model described above may be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured to execute the large language model training method or resource retrieval method in any other appropriate manner (eg, by means of firmware).

[0148] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0149] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0150] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0151] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0152] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0153] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0154] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0155] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a large language model, comprising: Acquire a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information based on the search information input by the user, wherein the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute; Acquire a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identifier, and sample relevance, wherein the sample relevance is used to describe a degree of relevance between the first sample search information and the first sample semantic identifier; The base large language model is trained based on the first training sample set to obtain a target large language model generated after training.

2. The method according to claim 1, wherein: The step of training the base large language model based on the first training sample set to obtain a target large language model generated after training includes: Training the base large language model based on the first training sample set to obtain an intermediate large language model; Acquire a second training sample set, wherein each second training sample in the second training sample set includes second sample search information, a second sample semantic identifier, and a sample preference index; The intermediate large language model is trained based on the second training sample set to obtain a target large language model generated after training.

3. The method according to claim 2, wherein: The method for constructing the first training sample set and the second training sample set includes: Determine a sample resource set, and obtain a sample semantic identifier corresponding to each sample resource in the sample resource set; Acquire sample search information related to the sample resource; constructing a first sample pair based on the sample semantic identifier and the sample search information, annotating and obtaining sample correlations corresponding to the first sample pair, and constructing the first training sample set based on the first sample pair and the sample correlations; and, A second sample pair is constructed based on the sample semantic identifier and the sample search information, a sample preference index corresponding to the second sample pair is annotated, and the second training sample set is constructed based on the second sample pair and the sample preference index.

4. The method according to claim 3, wherein: The obtaining of the sample semantic identifier corresponding to each sample resource in the sample resource set includes: Obtaining sample information corresponding to each sample resource in the sample resource set; The sample information is processed to obtain a sample semantic identifier corresponding to the sample resource.

5. The method according to claim 4, wherein: The acquiring sample information corresponding to each sample resource in the sample resource set includes: For any of the sample resources, key information is extracted from the landing details page of the sample resource, and the extracted key information is used as the sample information corresponding to the sample resource.

6. The method according to claim 5, wherein: The processing of the sample information to obtain a sample semantic identifier corresponding to the sample resource includes: Encoding the sample information based on an encoder to obtain sample semantic features corresponding to the sample information; Processing the sample semantic features based on a clustering algorithm to obtain a plurality of clusters and a sample semantic identifier corresponding to each cluster, wherein the cluster includes at least one of the sample semantic features; The sample semantic identifier corresponding to the cluster where the sample resource is located is used as the sample semantic identifier corresponding to the sample resource.

7. The method according to claim 3, wherein: The first sample pair includes a positive sample pair and a negative sample pair. The step of constructing the first sample pair based on the sample semantic identifier and the sample search information and annotating the sample correlation corresponding to the first sample pair includes: constructing a positive sample pair based on the sample semantic identifier and the sample search information, and annotating sample correlations of the positive sample pair; and, A negative sample pair is constructed based on the sample semantic identifier and the sample search information, and sample correlation of the negative sample pair is annotated.

8. The method according to claim 3, wherein: The step of constructing a second sample pair based on the sample semantic identifier and the sample search information, and annotating to obtain a sample preference index corresponding to the second sample pair, includes: constructing a second sample pair based on the sample semantic identifier and the sample search information; Acquire user interaction data corresponding to the sample resource; Based on the user interaction data, a sample preference index corresponding to the second sample pair is obtained by annotation.

9. A resource retrieval method, comprising: Get the target search information entered by the user; Processing the target search information based on the trained target large language model to obtain a target semantic identifier sequence, wherein the target semantic identifier sequence includes at least one target semantic identifier, and the target semantic identifier represents at least one target semantic attribute; Determine, from the first candidate resources, a second candidate resource associated with the target semantic identifier sequence; A target resource is determined from the second candidate resources, and the target resource is displayed.

10. The method according to claim 9, wherein: The method further comprises: Generate prompt information based on the target search information and the target resource; The prompt information is displayed.

11. A large language model training device, comprising: A first acquisition module is used to acquire a base large language model, wherein the base large language model is used to output a semantic identifier sequence corresponding to the search information input by a user, the semantic identifier sequence includes at least one semantic identifier, and the semantic identifier represents at least one semantic attribute; A second acquisition module is used to acquire a first training sample set, wherein each first training sample in the first training sample set includes first sample search information, a first sample semantic identifier and sample relevance, and the sample relevance is used to describe the degree of relevance between the first sample search information and the first sample semantic identifier; The model training module is used to train the base large language model based on the first training sample set to obtain a target large language model generated after training.

12. The device according to claim 11, wherein The model training module is also used for: Training the base large language model based on the first training sample set to obtain an intermediate large language model; Acquire a second training sample set, wherein each second training sample in the second training sample set includes second sample search information, a second sample semantic identifier, and a sample preference index; The intermediate large language model is trained based on the second training sample set to obtain a target large language model generated after training.

13. The device according to claim 12, wherein: The device also includes a sample construction module, which is used to: Determine a sample resource set, and obtain a sample semantic identifier corresponding to each sample resource in the sample resource set; Acquire sample search information related to the sample resource; constructing a first sample pair based on the sample semantic identifier and the sample search information, annotating and obtaining sample correlations corresponding to the first sample pair, and constructing the first training sample set based on the first sample pair and the sample correlations; and, A second sample pair is constructed based on the sample semantic identifier and the sample search information, a sample preference index corresponding to the second sample pair is annotated, and the second training sample set is constructed based on the second sample pair and the sample preference index.

14. The device according to claim 13, wherein: The sample building module is also used to: Obtaining sample information corresponding to each sample resource in the sample resource set; The sample information is processed to obtain a sample semantic identifier corresponding to the sample resource.

15. The device according to claim 14, wherein: The sample building module is also used to: For any of the sample resources, key information is extracted from the landing details page of the sample resource, and the extracted key information is used as the sample information corresponding to the sample resource.

16. The device according to claim 15, wherein: The sample building module is also used to: Encoding the sample information based on an encoder to obtain sample semantic features corresponding to the sample information; Processing the sample semantic features based on a clustering algorithm to obtain a plurality of clusters and a sample semantic identifier corresponding to each cluster, wherein the cluster includes at least one of the sample semantic features; The sample semantic identifier corresponding to the cluster where the sample resource is located is used as the sample semantic identifier corresponding to the sample resource.

17. The device according to claim 13, wherein: The sample building module is also used to: constructing a positive sample pair based on the sample semantic identifier and the sample search information, and annotating sample correlations of the positive sample pair; and, A negative sample pair is constructed based on the sample semantic identifier and the sample search information, and sample correlation of the negative sample pair is annotated.

18. The device according to claim 13, wherein: The sample building module is also used to: constructing a second sample pair based on the sample semantic identifier and the sample search information; Acquire user interaction data corresponding to the sample resource; Based on the user interaction data, a sample preference index corresponding to the second sample pair is obtained by annotation.

19. A resource retrieval device, comprising: An acquisition module is used to acquire target search information input by a user; A processing module, configured to process the target search information based on a trained target large language model to obtain a target semantic identifier sequence, wherein the target semantic identifier sequence includes at least one target semantic identifier, and the target semantic identifier represents at least one target semantic attribute; A determination module, configured to determine, from the first candidate resources, a second candidate resource associated with the target semantic identifier sequence; A display module is used to determine a target resource from the second candidate resources and display the target resource.

20. The device according to claim 19, wherein The display module is also used for: Generate prompt information based on the target search information and the target resource; The prompt information is displayed.

21. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8 or 9-10.

22. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-8 or 9-10.

23. A computer program product comprising a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-8 or 9-10.

Images